{"doc_id": "02567fd428a675ca91a0c6786f47f3e35881bcbd", "text": ["document", ":", "Deep", "Label", "Distribution", "Learning", "With", "Label", "Ambiguity", "Convolutional", "Neural", "Networks", "(", "ConvNets", ")", "have", "achieved", "excellent", "recognition", "performance", "in", "various", "visual", "recognition", "tasks", ".", "A", "large", "labeled", "training", "set", "is", "one", "of", "the", "most", "important", "factors", "for", "its", "success", ".", "However", ",", "it", "is", "difficult", "to", "collect", "sufficient", "training", "images", "with", "precise", "labels", "in", "some", "domains", "such", "as", "apparent", "age", "estimation", ",", "head", "pose", "estimation", ",", "multi", "-", "label", "classification", "and", "semantic", "segmentation", ".", "Fortunately", ",", "there", "is", "ambiguous", "information", "among", "labels", ",", "which", "makes", "these", "tasks", "different", "from", "traditional", "classification", ".", "Based", "on", "this", "observation", ",", "we", "convert", "the", "label", "of", "each", "image", "into", "a", "discrete", "label", "distribution", ",", "and", "learn", "the", "label", "distribution", "by", "minimizing", "a", "Kullback", "-", "Leibler", "divergence", "between", "the", "predicted", "and", "ground", "-", "truth", "label", "distributions", "using", "deep", "ConvNets", ".", "The", "proposed", "DLDL", "(", "Deep", "Label", "Distribution", "Learning", ")", "method", "effectively", "utilizes", "the", "label", "ambiguity", "in", "both", "feature", "learning", "and", "classifier", "learning", ",", "which", "help", "prevent", "the", "network", "from", "over", "-", "fitting", "even", "when", "the", "training", "set", "is", "small", ".", "Experimental", "results", "show", "that", "the", "proposed", "approach", "produces", "significantly", "better", "results", "than", "state", "-", "of", "-", "the", "-", "art", "methods", "for", "age", "estimation", "and", "head", "pose", "estimation", ".", "At", "the", "same", "time", ",", "it", "also", "improves", "recognition", "performance", "for", "multi", "-", "label", "classification", "and", "semantic", "segmentation", "tasks", ".", "Label", "distribution", ",", "deep", "learning", ",", "age", "estimation", ",", "head", "pose", "estimation", ",", "semantic", "segmentation", ".", "section", ":", "Introduction", "Convolutional", "Neural", "Networks", "(", "ConvNets", ")", "have", "achieved", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "various", "visual", "recognition", "tasks", "such", "as", "image", "classification", ",", "object", "detection", "and", "semantic", "segmentation", ".", "The", "availability", "of", "a", "huge", "set", "of", "training", "images", "is", "one", "of", "the", "most", "important", "factors", "for", "their", "success", ".", "However", ",", "it", "is", "difficult", "to", "collect", "sufficient", "training", "images", "with", "unambiguous", "labels", "in", "domains", "such", "as", "age", "estimation", ",", "head", "pose", "estimation", ",", "multi", "-", "label", "classification", "and", "semantic", "segmentation", ".", "Therefore", ",", "exploiting", "deep", "learning", "methods", "with", "limited", "samples", "and", "ambiguous", "labels", "has", "become", "an", "attractive", "yet", "challenging", "topic", ".", "Why", "is", "it", "difficult", "to", "collect", "a", "large", "and", "accurately", "labeled", "training", "set", "?", "Firstly", ",", "it", "is", "difficult", "(", "even", "for", "domain", "experts", ")", "to", "provide", "exact", "labels", "to", "some", "tasks", ".", "For", "example", ",", "the", "pixels", "close", "to", "object", "boundaries", "are", "very", "difficult", "to", "label", "for", "annotators", "in", "semantic", "segmentation", ".", "In", "addition", ",", "pixel", "labeling", "is", "a", "time", "-", "consuming", "task", "that", "may", "limit", "the", "amount", "of", "training", "samples", ".", "Another", "example", "is", "that", "people", "’s", "apparent", "age", "and", "head", "pose", "is", "difficult", "to", "describe", "with", "an", "accurate", "number", ".", "Secondly", ",", "it", "is", "very", "hard", "to", "gather", "complete", "and", "sufficient", "data", ".", "For", "example", ",", "it", "is", "difficult", "to", "build", "an", "age", "dataset", "covering", "people", "from", "1", "to", "85", "years", "old", ",", "and", "ensure", "that", "every", "age", "in", "this", "range", "has", "enough", "associated", "images", ".", "Similar", "difficulties", "arise", "in", "head", "pose", "estimation", ",", "where", "head", "poses", "are", "usually", "collected", "at", "a", "small", "set", "of", "angles", "with", "a", "10", "or", "15", "increment", ".", "Thus", ",", "the", "publicly", "available", "age", ",", "head", "pose", "and", "semantic", "segmentation", "datasets", "are", "small", "scale", "compared", "to", "those", "in", "image", "classification", "tasks", ".", "These", "aforementioned", "small", "datasets", "have", "a", "common", "characteristic", ",", "i.e.", ",", "label", "ambiguity", ",", "which", "refers", "to", "the", "uncertainty", "among", "the", "ground", "-", "truth", "labels", ".", "On", "one", "hand", ",", "label", "ambiguity", "is", "unavoidable", "in", "some", "applications", ".", "We", "usually", "predict", "another", "person", "’s", "age", "in", "a", "way", "like", "“", "around", "25", "”", ",", "which", "indicates", "using", "not", "only", "25", ",", "but", "also", "neighboring", "ages", "to", "describe", "the", "face", ".", "And", ",", "different", "people", "may", "have", "different", "guesses", "towards", "the", "same", "face", ".", "Similar", "situations", "also", "hold", "for", "other", "types", "of", "tasks", ".", "The", "labels", "of", "pixels", "at", "object", "boundaries", "are", "difficult", "to", "annotate", "because", "of", "the", "inherent", "ambiguity", "of", "these", "pixels", "in", "semantic", "segmentation", ".", "On", "the", "other", "hand", ",", "label", "ambiguity", "can", "also", "happen", "if", "we", "are", "not", "confident", "in", "the", "labels", "we", "provide", "for", "an", "image", ".", "In", "the", "multi", "-", "label", "classification", "task", ",", "some", "objects", "are", "clearly", "visible", "but", "difficult", "to", "recognize", ".", "This", "type", "of", "objects", "are", "annotated", "as", "Difficult", "in", "the", "PASCAL", "Visual", "Object", "Classes", "(", "VOC", ")", "classification", "challenge", ",", "e.g.", ",", "the", "chair", "in", "the", "third", "image", "of", "the", "first", "row", "in", "Fig", ".", "[", "reference", "]", ".", "[", "Age", "estimation", "]", "[", "Head", "pose", "estimation", "]", "[", "Multi", "-", "label", "classification", "]", "[", "Semantic", "segmentation", "]", "There", "are", "two", "main", "types", "of", "labeling", "methods", ":", "single", "-", "label", "recognition", "(", "SLR", ")", "and", "multi", "-", "label", "recognition", "(", "MLR", ")", ".", "SLR", "assumes", "one", "image", "or", "pixel", "has", "one", "label", "and", "MLR", "assumes", "that", "one", "image", "or", "pixel", "may", "be", "assigned", "multiple", "labels", ".", "Both", "SLR", "and", "MLR", "aim", "to", "answer", "the", "question", "of", "which", "labels", "can", "be", "used", "to", "describe", "an", "image", "or", "pixel", ",", "but", "they", "can", "not", "describe", "the", "label", "ambiguity", "associated", "with", "it", ".", "Label", "ambiguity", "will", "help", "improve", "recognition", "performance", "if", "it", "can", "be", "reasonably", "exploited", ".", "In", "order", "to", "utilize", "label", "correlation", "(", "which", "may", "be", "considered", "as", "a", "consequence", "of", "label", "ambiguity", "in", "some", "applications", ")", ",", "Geng", "et", "al", ".", "proposed", "a", "label", "distribution", "learning", "(", "LDL", ")", "approach", "for", "age", "estimation", "and", "head", "pose", "estimation", ".", "Recently", ",", "some", "improvements", "of", "LDL", "have", "been", "proposed", ".", "Xing", "et", "al", ".", "proposed", "two", "algorithms", "named", "LDLogitBoost", "and", "AOSO", "-", "LDLogitBoost", "to", "learn", "general", "models", "to", "relax", "the", "maximum", "entropy", "model", "in", "traditional", "LDL", "methods", ".", "Furthermore", ",", "He", "et", "al", ".", "generated", "age", "label", "distributions", "through", "weighted", "linear", "combination", "of", "the", "input", "image", "’s", "label", "and", "its", "context", "-", "neighboring", "samples", ".", "However", ",", "these", "methods", "are", "suboptimal", "because", "they", "only", "utilize", "the", "correlation", "of", "neighboring", "labels", "in", "classifier", "learning", ",", "but", "not", "in", "learning", "the", "visual", "representations", ".", "Deep", "ConvNets", "have", "natural", "advantages", "in", "feature", "learning", ".", "Existing", "ConvNet", "frameworks", "can", "be", "viewed", "as", "classification", "and", "regression", "models", "based", "on", "different", "optimization", "objective", "functions", ".", "In", "many", "cases", ",", "the", "softmax", "loss", "and", "loss", "are", "used", "in", "deep", "ConvNet", "models", "for", "classification", "and", "regression", "problems", ",", "respectively", ".", "The", "softmax", "loss", "maximizes", "the", "estimated", "probability", "of", "the", "ground", "-", "truth", "class", "without", "considering", "other", "classes", ",", "and", "the", "loss", "minimizes", "the", "squared", "difference", "between", "the", "estimated", "values", "of", "the", "network", "and", "the", "ground", "-", "truth", ".", "These", "methods", "have", "achieved", "satisfactory", "performance", "in", "some", "domains", "such", "as", "image", "classification", ",", "human", "pose", "estimation", "and", "object", "detection", ".", "However", ",", "existing", "deep", "learning", "methods", "can", "not", "utilize", "the", "label", "ambiguity", "information", ".", "Moreover", ",", "a", "well", "-", "known", "fact", "is", "that", "learning", "a", "good", "ConvNet", "requires", "a", "lot", "of", "images", ".", "In", "order", "to", "solve", "the", "issues", "mentioned", "above", ",", "we", "convert", "both", "traditional", "SLR", "and", "MLR", "problems", "to", "label", "distribution", "learning", "problems", ".", "Every", "instance", "is", "assigned", "a", "discrete", "label", "distribution", "according", "to", "its", "ground", "-", "truth", ".", "The", "label", "distribution", "can", "naturally", "describe", "the", "ambiguous", "information", "among", "all", "possible", "labels", ".", "Through", "deep", "label", "distribution", "learning", ",", "the", "training", "instances", "associated", "with", "each", "class", "label", "is", "significantly", "increased", "without", "actually", "increase", "the", "number", "of", "the", "total", "training", "examples", ".", "Fig", ".", "[", "reference", "]", "intuitively", "shows", "four", "examples", "of", "label", "distribution", "for", "different", "recognition", "tasks", ".", "Then", ",", "we", "utilize", "a", "deep", "ConvNet", "to", "learn", "the", "label", "distribution", "in", "both", "feature", "learning", "and", "classifier", "learning", ".", "Since", "we", "learn", "label", "distribution", "with", "deep", "ConvNets", ",", "we", "call", "our", "method", "DLDL", ":", "Deep", "Label", "Distribution", "Learning", ".", "The", "benefits", "of", "DLDL", "are", "summarized", "as", "follows", ":", "DLDL", "is", "an", "end", "-", "to", "-", "end", "learning", "framework", "which", "utilizes", "the", "label", "ambiguity", "in", "both", "feature", "learning", "and", "classifier", "learning", ";", "DLDL", "not", "only", "achieves", "more", "robust", "performance", "than", "existing", "classification", "and", "regression", "methods", ",", "but", "also", "effectively", "relaxes", "the", "requirement", "for", "large", "amount", "of", "training", "images", ",", "e.g.", ",", "a", "training", "face", "image", "with", "ground", "-", "truth", "label", "25", "is", "also", "useful", "for", "predicting", "faces", "at", "age", "24", "or", "26", ";", "DLDL", "(", "only", "single", "model", "without", "ensemble", ")", "achieves", "better", "performance", "than", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "age", "and", "head", "pose", "estimation", "tasks", ".", "DLDL", "also", "improves", "the", "performance", "for", "multi", "-", "label", "classification", "and", "semantic", "segmentation", ".", "The", "rest", "of", "this", "paper", "is", "organized", "as", "follows", ".", "We", "first", "review", "the", "related", "work", "in", "Section", "[", "reference", "]", ".", "Then", ",", "Section", "[", "reference", "]", "proposes", "the", "DLDL", "framework", ",", "including", "the", "DLDL", "problem", "definition", ",", "DLDL", "theory", ",", "label", "distribution", "construction", "and", "training", "details", ".", "After", "that", ",", "the", "experiments", "are", "reported", "in", "Section", "[", "reference", "]", ".", "Finally", ",", "Section", "[", "reference", "]", "presents", "discussions", "and", "the", "conclusion", "is", "given", "in", "Section", "[", "reference", "]", ".", "section", ":", "Related", "Work", "In", "the", "past", "two", "decades", ",", "many", "efforts", "have", "been", "devoted", "to", "visual", "recognition", ",", "including", "at", "least", "image", "classification", ",", "object", "detection", ",", "semantic", "segmentation", ",", "and", "facial", "attribute", "(", "apparent", "age", "and", "head", "pose", ")", "estimation", ".", "These", "works", "can", "be", "divided", "into", "two", "streams", ".", "Earlier", "research", "was", "mainly", "based", "on", "hand", "-", "crafted", "features", ",", "while", "more", "recent", "ones", "are", "usually", "deep", "learning", "methods", ".", "In", "this", "section", ",", "we", "briefly", "review", "these", "related", "approaches", ".", "Methods", "based", "on", "hand", "-", "crafted", "features", "usually", "include", "two", "stages", ".", "The", "first", "stage", "is", "feature", "extraction", ".", "The", "second", "stage", "learns", "models", "for", "recognition", ",", "detection", "or", "estimation", "using", "these", "features", ".", "SVM", ",", "random", "forest", "and", "neural", "networks", "have", "commonly", "been", "used", "during", "the", "learning", "stage", ".", "In", "addition", ",", "Geng", "et", "al", ".", "proposed", "the", "label", "distribution", "learning", "approach", "to", "utilize", "the", "correlation", "among", "adjacent", "labels", ",", "which", "further", "improved", "performance", "on", "age", "estimation", "and", "head", "pose", "estimation", ".", "Although", "important", "progresses", "have", "been", "made", "with", "these", "features", ",", "the", "hand", "-", "crafted", "features", "render", "them", "suboptimal", "for", "particular", "tasks", "such", "as", "age", "or", "head", "pose", "estimation", ".", "More", "recently", ",", "learning", "feature", "representation", "has", "shown", "great", "advantages", ".", "For", "example", ",", "Lu", "et", "al", ".", "tried", "to", "learn", "cost", "-", "sensitive", "local", "binary", "features", "for", "age", "estimation", ".", "Deep", "learning", "has", "substantially", "improved", "upon", "the", "state", "-", "of", "-", "the", "-", "art", "in", "image", "classification", ",", "object", "detection", ",", "semantic", "segmentation", "and", "many", "other", "vision", "tasks", ".", "In", "many", "cases", ",", "the", "softmax", "loss", "is", "used", "in", "deep", "models", "for", "classification", ".", "Besides", "classification", ",", "deep", "ConvNets", "have", "also", "been", "trained", "for", "regression", "tasks", "such", "as", "head", "pose", "estimation", "and", "facial", "landmark", "detection", ".", "In", "regression", "problems", ",", "the", "training", "procedure", "usually", "optimizes", "a", "squared", "loss", "function", ".", "Satisfactory", "performance", "has", "also", "been", "obtained", "by", "using", "Tukey", "’s", "biweight", "function", "in", "human", "pose", "estimation", ".", "In", "terms", "of", "model", "architecture", ",", "deep", "ConvNet", "models", "which", "use", "deeper", "architecture", "and", "smaller", "convolution", "filters", "(", "e.g.", ",", "VGG", "-", "Nets", "and", "VGG", "-", "Face", ")", "are", "very", "powerful", ".", "Nevertheless", ",", "these", "deep", "learning", "methods", "do", "not", "make", "use", "of", "the", "presence", "of", "label", "ambiguity", "in", "the", "training", "set", ",", "and", "usually", "require", "a", "large", "amount", "of", "training", "data", ".", "A", "latest", "approach", ",", "in", "Inception", "-", "v3", ",", "is", "based", "on", "label", "smoothing", "(", "LS", ")", ".", "Instead", "of", "only", "using", "the", "ground", "-", "truth", "label", ",", "they", "utilize", "a", "mixture", "of", "the", "ground", "-", "truth", "label", "and", "a", "uniform", "distribution", "to", "regularize", "the", "classifier", ".", "However", ",", "LS", "is", "limited", "to", "the", "uniform", "distribution", "among", "labels", "rather", "than", "mining", "labels", "’", "ambiguous", "information", ".", "We", "believe", "that", "label", "ambiguity", "is", "too", "important", "to", "ignore", ".", "If", "we", "make", "good", "use", "of", "the", "ambiguity", ",", "we", "expect", "the", "required", "number", "of", "training", "images", "for", "some", "tasks", "could", "be", "effectively", "reduced", ".", "In", "this", "paper", ",", "we", "focus", "on", "how", "to", "exploit", "the", "label", "ambiguity", "in", "deep", "ConvNets", ".", "Age", "and", "head", "pose", "estimation", "from", "still", "face", "images", "are", "suitable", "applications", "of", "the", "proposed", "research", ".", "In", "addition", ",", "we", "also", "extend", "our", "works", "to", "multi", "-", "label", "classification", "and", "semantic", "segmentation", ".", "section", ":", "The", "Proposed", "DLDL", "Approach", "In", "this", "section", ",", "we", "firstly", "give", "the", "definition", "of", "the", "DLDL", "problem", ".", "Then", ",", "we", "present", "the", "DLDL", "theory", ".", "Next", ",", "we", "propose", "the", "construction", "methods", "of", "label", "distribution", "for", "different", "recognition", "tasks", ".", "Finally", ",", "we", "briefly", "introduce", "the", "DLDL", "architecture", "and", "training", "details", ".", "subsection", ":", "The", "deep", "label", "distribution", "learning", "problem", "Given", "an", "input", "image", ",", "we", "are", "interested", "in", "estimating", "a", "category", "output", "(", "e.g.", ",", "age", "or", "head", "pose", "angles", ")", ".", "For", "two", "input", "images", "and", "with", "ground", "-", "truth", "labels", "and", ",", "and", "are", "supposed", "to", "be", "similar", "to", "each", "other", "if", "the", "correlation", "of", "and", "is", "strong", ",", "and", "vice", "versa", ".", "For", "example", ",", "the", "correlation", "between", "faces", "aged", "32", "and", "33", "should", "be", "stronger", "than", "that", "between", "faces", "aged", "32", "and", "64", ",", "in", "terms", "of", "facial", "details", "that", "reflect", "the", "age", "(", "e.g.", ",", "skin", "smoothness", ")", ".", "In", "other", "words", ",", "we", "expect", "high", "correlation", "among", "input", "images", "with", "similar", "outputs", ".", "The", "label", "distribution", "learning", "approach", "exploited", "such", "correlations", "in", "the", "machine", "learning", "phase", ",", "but", "used", "features", "that", "are", "extracted", "ignoring", "these", "correlations", ".", "The", "proposed", "DLDL", "approach", ",", "however", ",", "is", "an", "end", "-", "to", "-", "end", "deep", "learning", "method", "which", "utilizes", "such", "correlation", "information", "in", "both", "feature", "learning", "and", "classifier", "learning", ".", "We", "will", "also", "extend", "DLDL", "to", "handle", "other", "types", "of", "label", "ambiguity", "beyond", "correlation", ".", "To", "fulfill", "this", "goal", ",", "instead", "of", "outputting", "a", "single", "value", "for", "an", "input", ",", "DLDL", "quantizes", "the", "range", "of", "possible", "values", "into", "several", "labels", ".", "For", "example", ",", "in", "age", "estimation", ",", "it", "is", "reasonable", "to", "assume", "that", ",", "and", "it", "is", "a", "common", "practice", "to", "estimate", "integer", "values", "for", "ages", ".", "Thus", ",", "we", "can", "define", "the", "set", "as", "the", "ordered", "label", "set", "for", "age", "estimation", ".", "The", "task", "of", "DLDL", "is", "then", "to", "predict", "a", "label", "distribution", ",", "where", "is", "the", "estimated", "probability", "that", "should", "be", "predicted", "to", "be", "years", "old", ".", "By", "estimating", "an", "entire", "label", "distribution", ",", "the", "deep", "learning", "machine", "is", "forced", "to", "take", "care", "of", "the", "ambiguity", "among", "labels", ".", "Specifically", ",", "the", "input", "space", "of", "our", "framework", "is", ",", "where", ",", "and", "are", "the", "height", ",", "width", ",", "and", "number", "of", "channels", "of", "the", "input", "image", ",", "respectively", ".", "DLDL", "predicts", "a", "label", "distribution", "vector", ",", "where", "is", "the", "label", "set", "defined", "for", "a", "specific", "task", "(", "e.g.", ",", "the", "above", ")", ".", "We", "assume", "is", "complete", ",", "i.e.", ",", "any", "possible", "value", "has", "a", "corresponding", "member", "in", ".", "A", "training", "data", "set", "with", "instances", "is", "then", "denoted", "as", ".", "We", "use", "boldface", "lowercase", "letters", "like", "to", "denote", "vectors", ",", "and", "the", "-", "th", "element", "of", "is", "denoted", "as", ".", "The", "goal", "of", "DLDL", "is", "to", "directly", "learn", "a", "conditional", "probability", "mass", "function", "from", ",", "where", "is", "the", "parameters", "in", "the", "framework", ".", "subsection", ":", "Deep", "label", "distribution", "learning", "Given", "an", "instance", "with", "label", "distribution", ",", "we", "assume", "that", "is", "the", "activation", "of", "the", "last", "fully", "connected", "layer", "in", "a", "deep", "ConvNet", ".", "We", "use", "a", "softmax", "function", "to", "turn", "these", "activations", "into", "a", "probability", "distribution", ",", "that", "is", ",", "Given", "a", "training", "data", "set", ",", "the", "goal", "of", "DLDL", "is", "to", "find", "to", "generate", "a", "distribution", "that", "is", "similar", "to", ".", "There", "are", "different", "criteria", "to", "measure", "the", "similarity", "or", "distance", "between", "two", "distributions", ".", "For", "example", ",", "if", "the", "Kullback", "-", "Leibler", "(", "KL", ")", "divergence", "is", "used", "as", "the", "measurement", "of", "the", "similarity", "between", "the", "ground", "-", "truth", "and", "predicted", "label", "distribution", ",", "then", "the", "best", "parameter", "is", "determined", "by", "Thus", ",", "we", "can", "define", "the", "loss", "function", "as", ":", "Stochastic", "gradient", "descent", "is", "used", "to", "minimize", "the", "objective", "function", "Eq", ".", "[", "reference", "]", ".", "For", "any", "and", ",", "and", "the", "derivative", "of", "softmax", "(", "Eq", ".", "[", "reference", "]", ")", "is", "well", "known", ",", "as", "where", "is", "1", "if", ",", "and", "0", "otherwise", ".", "According", "to", "the", "chain", "rule", ",", "for", "any", "fixed", ",", "we", "have", "Thus", ",", "the", "derivative", "of", "with", "respect", "to", "is", "Once", "is", "learned", ",", "the", "label", "distribution", "of", "any", "new", "instance", "can", "be", "generated", "by", "a", "forward", "run", "of", "the", "network", ".", "If", "the", "expected", "class", "label", "is", "a", "single", "one", ",", "DLDL", "outputs", ",", "where", "Prediction", "with", "multiple", "labels", "is", "also", "allowed", ",", "which", "could", "be", "a", "set", "where", "is", "a", "predefined", "threshold", ".", "If", "the", "expected", "output", "is", "a", "real", "number", ",", "DLDL", "predicts", "the", "expectation", "of", ",", "as", "where", ".", "This", "indicates", "that", "DLDL", "is", "suitable", "for", "both", "classification", "and", "regression", "tasks", ".", "subsection", ":", "Label", "distribution", "construction", "The", "ground", "-", "truth", "label", "distribution", "is", "not", "available", "in", "most", "existing", "datasets", ",", "which", "must", "be", "generated", "under", "proper", "assumptions", ".", "A", "desirable", "label", "distribution", "must", "satisfy", "some", "basic", "principles", ":", "(", "1", ")", "should", "be", "a", "probability", "distribution", ".", "Thus", ",", "we", "have", "and", ".", "(", "2", ")", "The", "probability", "values", "should", "have", "difference", "among", "all", "possible", "labels", "associated", "with", "an", "image", ".", "In", "other", "words", ",", "a", "less", "ambiguous", "category", "must", "be", "assigned", "high", "probability", "and", "those", "more", "ambiguous", "labels", "must", "have", "low", "probabilities", ".", "In", "this", "section", ",", "we", "propose", "the", "way", "to", "construct", "label", "distributions", "for", "age", "estimation", ",", "head", "pose", "estimation", ",", "multi", "-", "label", "classification", "and", "semantic", "segmentation", ".", "For", "age", "estimation", ",", "we", "assume", "that", "the", "probabilities", "should", "concentrate", "around", "the", "ground", "-", "truth", "age", ".", "Thus", ",", "we", "quantize", "to", "get", "using", "a", "normal", "distribution", ".", "For", "example", ",", "the", "apparent", "age", "of", "a", "face", "is", "labeled", "by", "hundreds", "of", "users", ".", "The", "ground", "-", "truth", "(", "including", "a", "mean", "and", "a", "standard", "deviation", ")", "is", "calculated", "from", "all", "the", "votes", ".", "For", "this", "problem", ",", "we", "find", "the", "range", "of", "the", "target", "(", "e.g.", ",", ")", ",", "quantize", "it", "into", "a", "complete", "and", "ordered", "label", "set", ",", "where", "is", "the", "label", "set", "size", "and", "are", "all", "possible", "predictions", "for", ".", "A", "label", "distribution", "is", "then", ",", "where", "is", "the", "probability", "that", "(", "i.e.", ",", "for", ")", ".", "Since", "we", "use", "equal", "step", "size", "in", "quantizing", ",", "the", "normal", "p.d.f", ".", "(", "probability", "density", "function", ")", "is", "a", "natural", "choice", "to", "generate", "the", "ground", "-", "truth", "from", "and", ":", "where", ".", "Fig", ".", "[", "reference", "]", "shows", "a", "face", "and", "its", "corresponding", "label", "distribution", ".", "For", "problems", "where", "is", "unknown", ",", "we", "will", "show", "that", "a", "reasonably", "chosen", "also", "works", "well", "in", "DLDL", ".", "For", "head", "pose", "estimation", ",", "we", "need", "to", "jointly", "estimate", "pitch", "and", "yaw", "angles", ".", "Thus", ",", "learning", "joint", "distribution", "is", "also", "necessary", "in", "DLDL", ".", "Suppose", "the", "label", "set", "is", ",", "where", "is", "a", "pair", "of", "values", ".", "That", "is", ",", "we", "want", "to", "learn", "the", "joint", "distribution", "of", "two", "variables", ".", "Then", ",", "the", "label", "distribution", "can", "be", "represented", "by", "an", "matrix", ",", "whose", "-", "th", "element", "is", ".", "For", "example", ",", "when", "we", "use", "two", "angles", "(", "pitch", "and", "yaw", ")", "to", "describe", "a", "head", "pose", ",", "is", "a", "pair", "of", "pitch", "and", "yaw", "angles", ".", "Given", "an", "instance", "with", "ground", "-", "truth", "mean", "and", "covariance", "matrix", ",", "we", "calculate", "its", "label", "distribution", "as", "where", ".", "In", "the", "above", ",", "we", "assume", ",", "that", "is", ",", "the", "covariance", "matrix", "is", "diagonal", ".", "Fig", ".", "[", "reference", "]", "shows", "a", "joint", "label", "distribution", "with", "head", "pose", "and", ".", "For", "multi", "-", "label", "classification", ",", "a", "multi", "-", "label", "image", "always", "contains", "at", "least", "one", "object", "of", "the", "class", "of", "interest", ".", "There", "are", "usually", "multiple", "labels", "for", "an", "image", ".", "These", "labels", "are", "grouped", "into", "three", "different", "levels", ",", "including", "Positive", ",", "Negative", "and", "Difficult", "in", "the", "PASCAL", "VOC", "dataset", ".", "A", "label", "is", "Positive", "means", "an", "image", "contains", "objects", "from", "that", "category", ",", "and", "Negative", "otherwise", ".", "Difficult", "indicates", "that", "an", "object", "is", "clearly", "visible", "but", "difficult", "to", "recognize", ".", "Existing", "multi", "-", "label", "methods", "often", "view", "Difficult", "as", "Negative", ",", "which", "leads", "to", "the", "loss", "of", "useful", "information", ".", "It", "is", "not", "reasonable", "either", "if", "we", "simply", "treat", "Difficult", "as", "Positive", ".", "Therefore", ",", "a", "nature", "choice", "is", "to", "use", "label", "ambiguity", ".", "We", "define", "different", "probabilities", "for", "different", "types", "of", "labels", ",", "as", "for", "Positive", ",", "Difficult", "and", "Negative", "labels", ",", "respectively", ".", "Furthermore", ",", "an", "normalization", "is", "applied", "to", "ensure", ":", "where", "equals", ",", "or", "if", "the", "label", "is", "Positive", ",", "Difficult", "or", "Negative", ",", "respectively", ".", "The", "label", "distribution", "is", "shown", "for", "a", "multi", "-", "label", "image", "in", "Fig", ".", "[", "reference", "]", ".", "For", "semantic", "segmentation", ",", "we", "need", "to", "label", "a", "pixel", "as", "belonging", "to", "one", "class", "if", "it", "is", "a", "pixel", "inside", "an", "object", "of", "that", "class", ",", "or", "as", "the", "background", "otherwise", ".", "Let", "denote", "the", "annotation", "of", "the", "-", "th", "pixel", ",", "where", "(", "assuming", "there", "are", "categories", "and", "0", "for", "background", ")", ".", "Fully", "Convolutional", "Networks", "(", "FCN", ")", "have", "been", "an", "effective", "solution", "to", "this", "task", ".", "In", "FCN", ",", "a", "ground", "-", "truth", "label", "means", "that", "and", "for", "all", ".", "However", ",", "it", "is", "very", "difficult", "to", "specify", "ground", "-", "truth", "labels", "for", "pixels", "close", "to", "object", "boundaries", ",", "because", "labels", "of", "these", "pixels", "are", "inherently", "ambiguous", ".", "We", "propose", "a", "mechanism", "to", "describe", "the", "label", "ambiguity", "in", "the", "boundaries", ".", "Considering", "a", "Gaussian", "kernel", "matrix", ",", "we", "replace", "the", "original", "label", "distribution", "with", ",", "as", "where", ",", ",", "is", "the", "kernel", "size", ",", "and", "are", "padding", "and", "stride", "sizes", ".", "In", "our", "experiment", ",", "we", "set", ",", "and", ",", "and", "the", "generated", "label", "distribution", "is", "Fig", ".", "[", "reference", "]", "gives", "the", "semantic", "label", "distribution", "for", "a", "bird", "image", "which", "shows", "that", "the", "ambiguity", "is", "encoded", "in", "the", "label", "distributions", ".", "subsection", ":", "The", "DLDL", "architecture", "and", "training", "details", "We", "use", "a", "deep", "ConvNet", "and", "a", "training", "set", "to", "learn", "a", "as", "the", "estimation", "of", ".", "The", "structure", "of", "our", "network", "is", "based", "on", "popular", "deep", "models", "such", "as", "ZF", "-", "Net", "and", "VGG", "-", "Nets", ".", "The", "ZF", "-", "Net", "consists", "five", "convolution", "layers", ",", "followed", "by", "three", "fully", "connected", "layers", ".", "The", "VGG", "-", "Nets", "architecture", "includes", "16", "or", "19", "layers", ".", "We", "modify", "the", "last", "fully", "connected", "layer", "’s", "output", "based", "on", "the", "task", "and", "replace", "the", "original", "softmax", "loss", "function", "with", "the", "KL", "loss", "function", ".", "In", "addition", ",", "we", "use", "the", "parameter", "ReLU", "for", "ZF", "-", "Net", ".", "In", "our", "network", ",", "the", "input", "is", "an", "order", "three", "tensor", "and", "the", "output", "may", "be", "a", "vector", "(", "age", "estimation", "and", "multi", "-", "label", "classification", ")", ",", "a", "matrix", "(", "head", "pose", "estimation", ")", "or", "a", "tensor", "(", "semantic", "segmentation", ")", ".", "In", "this", "paper", ",", "we", "train", "the", "deep", "models", "in", "two", "ways", ":", "Training", "from", "scratch", ".", "For", "ZF", "-", "Net", ",", "the", "initialization", "is", "performed", "randomly", ",", "based", "on", "a", "Gaussian", "distribution", "with", "zero", "mean", "and", "0.01", "standard", "deviation", ",", "and", "biases", "are", "initialized", "to", "zero", ".", "The", "coefficient", "of", "the", "parameter", "ReLU", "is", "initialized", "to", "0.25", ".", "The", "dropout", "is", "applied", "to", "the", "last", "two", "fully", "connected", "layers", "with", "rate", "0.5", ".", "The", "coefficient", "of", "weight", "decay", "is", "set", "to", ".", "Optimization", "is", "done", "by", "Stochastic", "Gradient", "Descent", "(", "SGD", ")", "using", "mini", "-", "batches", "of", "128", "and", "the", "momentum", "coefficient", "is", "0.9", ".", "The", "initial", "learning", "rate", "is", "set", "to", "0.01", ".", "The", "total", "number", "of", "epochs", "is", "about", "20", ".", "Fine", "-", "tuning", ".", "Three", "pre", "-", "trained", "models", "including", "VGG", "-", "Nets", "(", "16", "-", "layers", "and", "19", "-", "layers", ")", "and", "VGG", "-", "Face", "(", "16", "-", "layers", ")", "are", "used", "to", "fine", "-", "tune", "for", "different", "tasks", ".", "We", "remove", "these", "pre", "-", "trained", "models", "’", "classification", "layer", "and", "loss", "layer", ",", "and", "put", "in", "our", "label", "distribution", "layer", "which", "is", "initialized", "by", "the", "Gaussian", "distribution", "and", "the", "KL", "loss", "layer", ".", "The", "learning", "rates", "of", "the", "convolutional", "layers", ",", "the", "first", "two", "fully", "-", "connected", "layers", "and", "the", "label", "distribution", "layer", "are", "initialized", "as", "0.001", ",", "0.001", "and", "0.01", ",", "respectively", ".", "We", "fine", "-", "tune", "all", "layers", "by", "back", "propagation", "through", "the", "whole", "net", "using", "mini", "-", "batches", "of", "32", ".", "The", "total", "number", "of", "epochs", "is", "about", "10", "for", "age", "estimation", "and", "20", "for", "multi", "-", "label", "classification", ".", "section", ":", "Experiments", "We", "evaluate", "DLDL", "on", "four", "tasks", ",", "i.e.", ",", "age", "estimation", ",", "head", "pose", "estimation", ",", "multi", "-", "label", "classification", "and", "semantic", "segmentation", ".", "Our", "implementation", "is", "based", "on", "MatConvNet", ".", "All", "our", "experiments", "are", "carried", "out", "on", "a", "NVIDIA", "K40", "GPU", "with", "12", "GB", "of", "onboard", "memory", ".", "subsection", ":", "Age", "estimation", "Datasets", ".", "Two", "age", "estimation", "datasets", "are", "used", "in", "our", "experiments", ".", "The", "first", "is", "Morph", ",", "which", "is", "one", "of", "the", "largest", "publicly", "available", "age", "datasets", ".", "There", "are", "55", ",", "134", "face", "images", "from", "more", "than", "13", ",", "000", "subjects", ".", "Ages", "range", "from", "16", "to", "77", ".", "Since", "no", "TRAIN", "/", "TEST", "split", "is", "provided", ",", "10", "-", "fold", "cross", "-", "validation", "is", "used", "for", "Morph", ".", "The", "second", "dataset", "is", "from", "the", "apparent", "age", "estimation", "competition", ",", "the", "first", "competition", "track", "of", "the", "ICCV", "ChaLearn", "LAP", "2015", "workshop", ".", "Compared", "with", "Morph", ",", "this", "dataset", "(", "ChaLearn", ")", "consists", "of", "images", "collected", "in", "the", "wild", ",", "without", "any", "position", ",", "illumination", "or", "quality", "restriction", ".", "The", "only", "condition", "is", "that", "each", "image", "contains", "only", "one", "face", ".", "The", "dataset", "has", "4", ",", "699", "images", ",", "and", "is", "split", "into", "2", ",", "476", "training", "(", "TRAIN", ")", ",", "1", ",", "136", "validation", "(", "VAL", ")", "and", "1", ",", "087", "testing", "(", "TEST", ")", "images", ".", "The", "apparent", "age", "(", "i.e.", ",", "how", "old", "does", "this", "person", "look", "like", ")", "of", "each", "image", "is", "labeled", "by", "multiple", "individuals", ".", "The", "age", "of", "face", "images", "range", "from", "3", "to", "85", ".", "For", "each", "image", ",", "its", "mean", "age", "and", "the", "corresponding", "standard", "deviation", "are", "given", ".", "Since", "the", "ground", "-", "truth", "for", "TEST", "images", "are", "not", "published", ",", "we", "train", "on", "the", "TRAIN", "split", "and", "evaluate", "on", "the", "VAL", "split", "of", "ChaLearn", "images", ".", "Baselines", ".", "To", "demonstrate", "the", "effectiveness", "of", "DLDL", ",", "we", "firstly", "consider", "two", "related", "methods", "as", "baselines", ":", "ConvNet", "+", "LS", "(", "KL", ")", "and", "ConvNet", "+", "LD", "(", "-", "div", ")", ".", "The", "former", "uses", "label", "smoothing", "(", "LS", ")", "as", "ground", "-", "truth", "and", "KL", "divergence", "as", "loss", "function", ".", "The", "latter", "uses", "label", "distribution", "(", "LD", ")", "as", "ground", "-", "truth", "and", "divergence", "as", "loss", "function", ",", "which", "is", "In", "addition", ",", "we", "also", "compare", "DLDL", "with", "the", "following", "baseline", "methods", ":", "BFGS", "-", "LDL", "Geng", "et", "al", ".", "proposed", "the", "label", "distribution", "learning", "approach", "(", "IIS", "-", "LLD", ")", "for", "age", "and", "head", "pose", "estimation", ".", "They", "used", "traditional", "image", "features", ".", "To", "further", "improve", "IIS", "-", "LLD", ",", "Geng", "et", "al", ".", "proposed", "a", "BFGS", "-", "LDL", "algorithm", "by", "using", "the", "effective", "quasi", "-", "Newton", "optimization", "method", "BFGS", ".", "C", "-", "ConvNet", "Classification", "ConvNets", "have", "obtained", "very", "competitive", "performance", "in", "various", "computer", "vision", "tasks", ".", "ZF", "-", "Net", "and", "VGG", "-", "Net", "are", "popular", "models", "which", "use", "the", "softmax", "loss", ".", "We", "replace", "the", "ImageNet", "-", "specific", "1000", "-", "way", "classification", "in", "these", "modes", "with", "the", "label", "set", ".", "R", "-", "ConvNet", "ConvNets", "are", "also", "successively", "trained", "for", "regression", "tasks", ".", "In", "R", "-", "ConvNet", ",", "the", "ground", "-", "truth", "label", "(", "age", "and", "pose", "angle", ")", "is", "projected", "into", "the", "range", "by", "the", "mapping", ",", "where", "and", "are", "the", "maximum", "and", "minimum", "values", "in", "the", "training", "label", "set", ".", "During", "prediction", ",", "the", "R", "-", "ConvNet", "regression", "result", "is", "reverse", "mapped", "to", "get", ".", "To", "speed", "up", "convergence", ",", "the", "last", "fully", "connected", "layer", "is", "followed", "a", "hyperbolic", "tangent", "activation", "function", ",", "which", "maps", "to", ".", "The", "squared", ",", "and", "-", "ins", "loss", "functions", "are", "used", "in", "R", "-", "ConvNet", ".", "Implementation", "details", ".", "We", "use", "the", "same", "preprocessing", "pipeline", "for", "all", "compared", "methods", ",", "including", "face", "detection", ",", "facial", "key", "points", "detection", "and", "face", "alignment", ",", "as", "shown", "in", "Fig", "[", "reference", "]", ".", "We", "employ", "the", "DPM", "model", "to", "detect", "the", "main", "facial", "region", ".", "Then", ",", "the", "detected", "face", "is", "fed", "into", "cascaded", "convolution", "networks", "to", "get", "the", "five", "facial", "key", "points", ",", "including", "the", "left", "/", "right", "eye", "centers", ",", "nose", "tip", "and", "left", "/", "right", "mouth", "corners", ".", "Finally", ",", "based", "on", "these", "facial", "points", ",", "we", "align", "the", "face", "to", "the", "upright", "pose", ".", "Data", "augmentation", "are", "only", "applied", "to", "the", "training", "images", "for", "ChaLearn", ".", "For", "one", "color", "input", "training", "image", ",", "we", "generate", "its", "gray", "-", "scale", "version", ",", "and", "left", "-", "right", "flip", "both", "color", "and", "gray", "-", "scale", "versions", ".", "Thus", ",", "every", "training", "image", "turns", "into", "4", "images", ".", "[", "Input", "]", "[", "Detection", "]", "[", "Facial", "points", "]", "[", "Alignment", "]", "We", "define", "for", "both", "datasets", ".", "The", "label", "distribution", "of", "each", "image", "is", "generated", "using", "Eq", ".", "[", "reference", "]", ".", "The", "mean", "is", "provided", "in", "both", "Morph", "and", "ChaLearn", ".", "The", "standard", "deviation", ",", "however", ",", "is", "provided", "in", "ChaLearn", "but", "not", "in", "Morph", ".", "We", "simply", "set", "in", "Morph", ".", "Experiments", "for", "different", "methods", "are", "conducted", "under", "the", "same", "data", "splits", ".", "1", "Used", "80", "%", "of", "Morph", "images", "for", "training", "and", "20", "%", "for", "evaluation", ";", "2", "Used", "additional", "external", "face", "images", "(", "i.e.", ",", "IMDB", "-", "WIKI", ")", ";", "3", "Used", "pre", "-", "trained", "model", "(", "i.e.", ",", "VGG", "-", "Nets", "or", "VGG", "-", "Face", ")", ".", "Evaluation", "criteria", ".", "Mean", "Absolute", "Error", "(", "MAE", ")", "and", "Cumulative", "Score", "(", "CS", ")", "are", "used", "to", "evaluate", "the", "performance", "of", "age", "estimation", ".", "MAE", "is", "the", "average", "difference", "between", "the", "predicted", "and", "the", "real", "age", ":", "where", "and", "are", "the", "estimated", "and", "ground", "-", "truth", "age", "of", "the", "-", "th", "testing", "image", ",", "respectively", ".", "CS", "is", "defined", "as", "the", "accuracy", "rate", "of", "correct", "estimation", ":", "where", "is", "the", "number", "of", "correct", "estimation", ",", "i.e.", ",", "testing", "images", "that", "satisfy", ".", "In", "our", "experiment", ",", ".", "In", "addition", ",", "a", "special", "measurement", "(", "named", "-", "error", ")", "is", "defined", "by", "the", "ChaLearn", "competition", ",", "computed", "as", "Results", ".", "Table", "[", "reference", "]", "lists", "results", "on", "both", "datasets", ".", "The", "upper", "part", "shows", "results", "in", "the", "literature", ".", "The", "middle", "part", "shows", "the", "baseline", "results", ".", "The", "lower", "part", "shows", "the", "results", "of", "the", "proposed", "approach", ".", "The", "first", "term", "in", "the", "parenthesis", "behind", "each", "method", "is", "the", "loss", "function", "corresponding", "to", "the", "method", ".", "Max", "or", "Exp", "represent", "predicting", "according", "to", "Eq", ".", "[", "reference", "]", "or", "[", "reference", "]", ",", "respectively", ".", "Since", "cross", "-", "validation", "is", "used", "in", "Morph", ",", "we", "also", "provide", "its", "standard", "deviations", ".", "[", "ChaLearn", "]", "[", "Morph", "]", "[", "AFLW", "]", "[", "subfigure", "]", "labelformat", "=", "empty", "[", "40", "]", "[", "19", "]", "[", "62", "]", "[", "23", "]", "[", "38", "]", "[", "24", "]", "[", "26", "]", "[", "66", "]", "[", "52", "]", "[", "22", "]", "[", "red39.69", "]", "[", "red19.29", "]", "[", "red61.61", "]", "[", "red22.94", "]", "[", "red37.87", "]", "[", "red24.27", "]", "[", "red25.40", "]", "[", "blue60.17", "]", "[", "blue35.06", "]", "[", "blue28.55", "]", "From", "Table", "[", "reference", "]", ",", "we", "can", "see", "that", "DLDL", "consistently", "outperforms", "baselines", "and", "other", "published", "methods", ".", "The", "difference", "between", "DLDL", "(", "KL", ",", "Max", ")", "and", "its", "competitor", "C", "-", "ConvNet", "(", "softmax", ",", "Max", ")", "is", "0.51", "on", "Morph", ".", "This", "gap", "is", "more", "than", "6", "times", "the", "sum", "of", "their", "standard", "deviations", "(", "0.03", "+", "0.05", ")", ",", "showing", "statistically", "significant", "differences", ".", "The", "advantage", "of", "DLDL", "over", "R", "-", "ConvNet", ",", "C", "-", "ConvNet", "and", "ConvNet", "+", "LS", "suggests", "that", "learning", "label", "distribution", "is", "advantageous", "in", "deep", "end", "-", "to", "-", "end", "models", ".", "DLDL", "has", "much", "better", "results", "than", "BFGS", "-", "LDL", ",", "which", "shows", "that", "the", "learned", "deep", "features", "are", "more", "powerful", "than", "manually", "designed", "ones", ".", "Compared", "to", "ConvNet", "+", "LD", "(", "-", "div", ")", ",", "DLDL", "(", "KL", ")", "achieves", "lower", "MAE", "on", "both", "datasets", ".", "It", "indicates", "that", "KL", "-", "divergence", "is", "better", "than", "-", "divergence", "for", "measuring", "the", "similarity", "of", "two", "distributions", "in", "this", "context", ".", "We", "find", "that", "C", "-", "ConvNet", "and", "R", "-", "ConvNet", "are", "not", "stable", ".", "The", "R", "-", "ConvNet", "(", ")", "method", ",", "although", "being", "the", "second", "best", "method", "for", "ChaLearn", ",", "is", "inferior", "to", "C", "-", "ConvNet", "(", "softmax", ",", "Exp", ")", "for", "Morph", ".", "In", "addition", ",", "we", "also", "find", "that", "Eq", ".", "[", "reference", "]", "is", "better", "than", "Eq", ".", "[", "reference", "]", "in", "many", "cases", ",", "which", "suggests", "that", "Eq", ".", "[", "reference", "]", "is", "more", "suitable", "than", "Eq", ".", "[", "reference", "]", "for", "age", "estimation", ".", "Fine", "-", "tuning", "DLDL", ".", "Instead", "of", "training", "DLDL", "from", "scratch", ",", "we", "also", "fine", "-", "tune", "the", "network", "of", "VGG", "-", "Face", ".", "On", "the", "small", "scale", "ChaLearn", "dataset", ",", "the", "MAE", "of", "DLDL", "is", "reduced", "from", "5.34", "to", "3.51", ",", "yielding", "a", "significant", "improvement", ".", "The", "-", "error", "of", "DLDL", "is", "reduced", "from", "0.44", "to", "0.31", ",", "which", "is", "close", "to", "the", "best", "competition", "result", "0.28", "on", "the", "validation", "set", ".", "In", ",", "external", "training", "images", "(", "260", ",", "282", "additional", "external", "training", "images", "with", "real", "age", "annotation", ")", "were", "used", ".", "DLDL", "only", "uses", "the", "ChaLearn", "dataset", "’s", "2", ",", "476", "training", "images", "and", "is", "the", "best", "among", "ChaLearn", "teams", "that", "do", "not", "use", "external", "data", ".", "In", "the", "competition", ",", "the", "best", "external", "-", "data", "-", "free", "-", "error", "is", "0.48", ",", "which", "is", "worse", "than", "DLDL", "’s", ".", "However", ",", "the", "idea", "in", "to", "use", "external", "data", "is", "useful", "for", "further", "reducing", "DLDL", "’s", "estimation", "error", ".", "Fig", ".", "[", "reference", "]", "and", "Fig", ".", "[", "reference", "]", "show", "the", "CS", "curves", "on", "ChaLearn", "and", "Morph", "datasets", ".", "At", "every", "error", "level", ",", "our", "DLDL", "fine", "-", "tuned", "VGG", "-", "Face", "always", "achieves", "the", "best", "accuracy", "among", "all", "methods", ".", "It", "is", "noteworthy", "that", "the", "CS", "curves", "of", "DLDL", "(", "KL", ",", "Max", ")", "and", "ConvNet", "(", "-", "div", ",", "Max", ")", "are", "very", "close", "to", "that", "of", "the", "DLDL", "+", "VGG", "-", "Face", "(", "KL", ",", "Max", ")", "on", "Morph", "even", "without", "lots", "of", "external", "data", "and", "very", "deep", "model", ".", "This", "observation", "supports", "the", "idea", "that", "using", "DLDL", "can", "achieve", "competitive", "performance", "even", "with", "limited", "training", "samples", ".", "In", "Fig", ".", "[", "reference", "]", ",", "we", "show", "some", "examples", "of", "face", "images", "from", "the", "ChaLearn", "validation", "set", "and", "predicted", "label", "distributions", "by", "DLDL", "(", "KL", ",", "Exp", ")", ".", "In", "many", "cases", ",", "our", "solution", "is", "able", "to", "accurately", "predict", "the", "apparent", "age", "of", "faces", ".", "Failures", "may", "come", "from", "two", "causes", ".", "The", "first", "is", "the", "failure", "to", "detect", "or", "align", "the", "face", ".", "The", "second", "is", "some", "extreme", "conditions", "of", "face", "images", "such", "as", "occlusion", ",", "low", "resolution", ",", "heavy", "makeup", "and", "old", "photos", ".", "subsection", ":", "Head", "pose", "estimation", "Datasets", ".", "We", "use", "three", "datasets", "in", "head", "pose", "estimation", ":", "Pointing’04", ",", "BJUT", "-", "3D", "and", "Annotated", "Facial", "Landmarks", "in", "the", "Wild", "(", "AFLW", ")", ".", "In", "them", ",", "head", "pose", "is", "determined", "by", "two", "angles", ":", "pitch", "and", "yaw", ".", "Pointing’04", "discretizes", "the", "pitch", "into", "9", "angles", "and", "the", "yaw", "into", "13", "angles", ".", "When", "the", "pitch", "angel", "is", "or", ",", "the", "yaw", "angle", "is", "always", "set", "to", ".", "Thus", ",", "there", "are", "93", "poses", "in", "total", ".", "The", "head", "images", "are", "taken", "from", "15", "different", "human", "subjects", "in", "two", "different", "time", "periods", ",", "resulting", "in", "images", ".", "BJUT", "-", "3D", "contains", "500", "3D", "faces", "(", "250", "male", "and", "250", "female", "people", ")", ",", "acquired", "by", "a", "CyberWare", "Laser", "Scanner", "in", "an", "engineered", "environment", ".", "9", "pitch", "angles", "and", "13", "yaw", "angles", "are", "used", ".", "There", "are", "in", "total", "93", "poses", "in", "this", "dataset", ",", "similar", "to", "that", "in", "Pointing’04", ".", "Therefore", ",", "face", "images", "are", "obtained", ".", "Unlike", "Pointing’04", "and", "BJUT", "-", "3D", ",", "the", "AFLW", "is", "a", "real", "-", "world", "face", "database", ".", "Head", "pose", "is", "coarsely", "obtained", "by", "fitting", "a", "mean", "3D", "face", "with", "the", "POSIT", "algorithm", ".", "The", "dataset", "contains", "about", "24k", "faces", "in", "real", "-", "world", "images", ".", "We", "select", "23", ",", "409", "faces", "to", "ensure", "pitch", "and", "yaw", "angles", "within", ".", "Implementation", "details", ".", "The", "head", "region", "is", "provided", "by", "bounding", "box", "annotations", "in", "Pointing’04", "and", "AFLW", ".", "The", "BJUT", "-", "3D", "does", "not", "contain", "background", "regions", ".", "Therefore", ",", "we", "will", "not", "perform", "any", "preprocessing", ".", "In", "DLDL", ",", "we", "set", "in", "Pointing’04", "and", "in", "BJUT", "-", "3D", "for", "constructing", "label", "distributions", ".", "For", "AFLW", ",", "ground", "-", "truth", "of", "head", "pose", "angles", "are", "given", "as", "real", "numbers", ".", "Ground", "-", "truth", "(", "pitch", "and", "yaw", ")", "angles", "are", "divided", "from", "to", "in", "steps", "of", ",", "so", "we", "get", "(", "pitch", ",", "yaw", ")", "pair", "category", "labels", ".", "We", "set", "for", "AFLW", ".", "Since", "the", "discrete", "Jeffrey", "’s", "divergence", "is", "used", "in", "LDL", ",", "we", "implement", "BFGS", "-", "LDL", "with", "the", "Kullback", "-", "Leibler", "divergence", ".", "All", "experiments", "are", "performed", "under", "the", "same", "setting", ",", "including", "data", "splits", ",", "input", "size", "and", "network", "architecture", ".", "To", "validate", "the", "effectiveness", "of", "DLDL", "for", "head", "pose", "estimation", ",", "we", "use", "the", "same", "baselines", "as", "age", "estimation", ".", "Our", "experiments", "show", "that", "Eq", ".", "[", "reference", "]", "has", "lower", "accuracy", "than", "Eq", ".", "[", "reference", "]", ".", "Hence", ",", "we", "use", "Eq", ".", "[", "reference", "]", "in", "this", "section", ".", "Evaluation", "criteria", ".", "Three", "types", "of", "prediction", "values", "are", "evaluated", ":", "pitch", ",", "yaw", ",", "and", "pitch", "+", "yaw", ",", "where", "pitch", "+", "yaw", "jointly", "estimates", "the", "pitch", "and", "yaw", "angles", ".", "Two", "different", "measurements", "are", "used", ",", "which", "is", "MAE", "(", "Eq", ".", "[", "reference", "]", ")", "and", "classification", "accuracy", "(", "Acc", ")", ".", "When", "we", "treat", "different", "poses", "as", "different", "classes", ",", "Acc", "measures", "the", "pose", "class", "classification", "accuracy", ".", "In", "particular", ",", "the", "MAE", "of", "pitch", "+", "yaw", "is", "calculated", "as", "the", "Euclidean", "distance", "between", "the", "predicted", "(", "pitch", ",", "yaw", ")", "pair", "and", "the", "ground", "-", "truth", "pair", ";", "the", "Acc", "of", "pitch", "+", "yaw", "is", "calculated", "by", "regarding", "each", "(", "pitch", ",", "yaw", ")", "pair", "as", "a", "class", ".", "For", "R", "-", "ConvNet", ",", "we", "only", "report", "its", "MAE", "but", "not", "Acc", ",", "because", "its", "predicted", "value", "are", "continuous", "real", "numbers", ".", "All", "methods", "are", "tested", "with", "5", "-", "fold", "cross", "validation", "for", "Pointing’04", "and", "BJUT", "-", "3D", "following", ".", "For", "AFLW", ",", "15", ",", "561", "face", "images", "are", "randomly", "chosen", "for", "training", ",", "and", "the", "remaining", "7", ",", "848", "for", "evaluation", ".", "The", "setup", "is", "similar", "to", "the", "recent", "literature", "(", "14", ",", "000", "images", "for", "training", "and", "the", "rest", "7", ",", "041", "images", "for", "testing", ")", ".", "[", "subfigure", "]", "labelformat", "=", "empty", "[", "(", "+", "77", ",", "-", "4", ")", "]", "[", "(", "-", "16", ",", "-", "1", ")", "]", "[", "(", "-", "1", ",", "-", "30", ")", "]", "[", "(", "+", "30", ",", "+", "8", ")", "]", "[", "(", "+", "4", ",", "-", "4", ")", "]", "[", "(", "-", "36", ",", "+", "13", ")", "]", "[", "(", "-", "87", ",", "-", "3", ")", "]", "[", "(", "-", "61", ",", "-", "58", ")", "]", "[", "(", "+", "63", ",", "+", "12", ")", "]", "[", "(", "+", "80", ",", "-", "27", ")", "]", "[", "red", "(+", "75", ",", "-", "3", ")", "]", "[", "red", "(-", "15", ",", "0", ")", "]", "[", "red", "(-", "3", ",", "-", "27", ")", "]", "[", "red", "(+", "27", ",", "+", "6", ")", "]", "[", "red", "(+", "6", ",", "-", "3", "]", "[", "red", "(-", "39", ",", "+", "15", "]", "[", "red", "(-", "87", ",", "0", ")", "]", "[", "blue", "(-", "3", ",", "-", "12", ")", "]", "[", "blue", "(+", "21", ",", "+", "18", "]", "[", "blue", "(+", "45", ",", "-", "15", ")", "]", "Results", ".", "Tables", "[", "reference", "]", ",", "[", "reference", "]", "and", "[", "reference", "]", "show", "results", "on", "Pointing’04", ",", "BJUT", "-", "3D", "and", "AFLW", ",", "respectively", ".", "Pointing’04", "is", "small", "scale", "with", "only", "2", ",", "790", "images", ".", "We", "observe", "that", "BFGS", "-", "LDL", "(", "with", "hand", "-", "crafted", "features", ")", "has", "much", "lower", "MAE", "and", "much", "higher", "accuracy", "than", "deep", "learning", "methods", "C", "-", "ConvNet", ",", "R", "-", "ConvNet", "and", "ConvNet", "+", "LS", ".", "One", "reasonable", "conjecture", "is", "that", "C", "-", "ConvNet", ",", "R", "-", "ConvNet", "and", "ConvNet", "+", "LS", "are", "not", "well", "-", "learned", "with", "only", "small", "number", "of", "training", "images", ".", "DLDL", ",", "however", ",", "successfully", "learns", "the", "head", "pose", ".", "For", "example", ",", "its", "accuracy", "for", "pitch", "+", "yaw", "is", "73.15", "(", "and", "C", "-", "ConvNet", "is", "only", "42.97", ")", ".", "That", "is", ",", "DLDL", "is", "able", "to", "perform", "deep", "learning", "with", "few", "training", "images", ",", "while", "C", "-", "ConvNet", "R", "-", "ConvNet", "and", "ConvNet", "+", "LS", "have", "failed", "for", "this", "task", ".", "On", "BJUT", "-", "3D", "and", "AFLW", "which", "have", "enough", "training", "data", ",", "we", "observe", "that", "many", "deep", "learning", "methods", "show", "higher", "performance", "than", "BFGS", "-", "LDL", ".", "DLDL", "achieves", "the", "best", "performance", ":", "it", "has", "much", "lower", "MAE", "and", "higher", "accuracy", "than", "other", "methods", ".", "Another", "observation", "is", "also", "worth", "mentioning", ".", "Although", "R", "-", "ConvNet", "is", "better", "than", "C", "-", "ConvNet", "when", "label", "is", "dense", "such", "as", "age", "estimation", "and", "head", "pose", "estimation", "on", "AFLW", ",", "it", "is", "obviously", "worse", "than", "C", "-", "ConvNet", "on", "BJUT", "-", "3D", "and", "pointing’04", "for", "head", "pose", "estimation", "which", "have", "sparse", "labels", ".", "In", "other", "words", ",", "the", "performance", "of", "C", "-", "ConvNet", "and", "R", "-", "ConvNet", "are", "not", "very", "robust", ",", "while", "the", "proposed", "method", "consistently", "achieves", "excellent", "performance", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "pitch", "+", "yaw", "CS", "curves", "on", "the", "AFLW", "dataset", ".", "There", "is", "an", "obvious", "gap", "between", "DLDL", "and", "baseline", "methods", "at", "every", "error", "level", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "predicted", "label", "distributions", "for", "different", "head", "poses", "on", "the", "AFLW", "testing", "set", "using", "the", "DLDL", "model", ".", "Our", "approach", "can", "estimate", "head", "pose", "with", "low", "errors", "but", "may", "fail", "under", "some", "extreme", "conditions", ".", "It", "is", "noteworthy", "that", "DLDL", "may", "produce", "more", "incorrect", "estimations", "when", "both", "yaw", "and", "pitch", "are", "large", "(", "e.g.", ",", ")", ".", "The", "reason", "might", "be", "that", "there", "are", "much", "fewer", "training", "examples", "for", "large", "angles", "than", "for", "other", "angles", ".", "subsection", ":", "Multi", "-", "label", "classification", "Datasets", ".", "We", "evaluate", "our", "approach", "for", "multi", "-", "label", "classification", "on", "the", "PASCAL", "VOC", "dataset", ":", "PASCAL", "VOC2007", "and", "VOC2012", ".", "There", "are", "9", ",", "963", "and", "22", ",", "531", "images", "in", "them", ",", "respectively", ".", "Each", "image", "is", "annotated", "with", "one", "or", "several", "labels", ",", "corresponding", "to", "20", "object", "categories", ".", "These", "images", "are", "divided", "into", "three", "subsets", "including", "TRAIN", ",", "VAL", "and", "TEST", "sets", ".", "We", "train", "on", "the", "TRAINVAL", "set", "and", "evaluate", "on", "the", "TEST", "set", ".", "The", "evaluation", "metric", "is", "average", "precision", "(", "AP", ")", "and", "mean", "average", "precision", "(", "mAP", ")", ",", "complying", "with", "the", "PASCAL", "challenge", "protocols", ".", "We", "denote", "our", "methods", "as", "Images", "-", "Fine", "-", "tuning", "-", "DLDL", "(", "IF", "-", "DLDL", ")", "and", "Proposals", "-", "Fine", "-", "tuning", "-", "DLDL", "(", "PF", "-", "DLDL", ")", "when", "ConvNets", "are", "fine", "-", "tuned", "by", "images", "and", "proposals", "of", "images", ",", "respectively", ".", "Details", "of", "these", "two", "variants", "are", "explained", "later", "in", "this", "section", ".", "We", "compare", "the", "proposed", "approaches", "with", "the", "following", "methods", ":", "VGG", "+", "SVM", "[", "]", ".", "This", "method", "densely", "extracted", "4", ",", "096", "dimensional", "ConvNet", "features", "at", "the", "penultimate", "layer", "of", "VGG", "-", "Nets", "pre", "-", "trained", "on", "ImageNet", ".", "These", "features", "from", "different", "scales", "(", "smallest", "image", "side", ")", "were", "aggregated", "by", "average", "pooling", ".", "Then", ",", "these", "averaged", "features", "from", "two", "networks", "(", "“", "Net", "-", "D", "”", "containing", "16", "layers", "and", "“", "Net", "-", "E", "”", "containing", "19", "layers", ")", "were", "further", "fused", "by", "stacking", ".", "Finally", ",", "normalized", "the", "resulting", "image", "features", "and", "used", "these", "features", "to", "train", "a", "linear", "SVM", "classifier", "for", "multi", "-", "label", "classification", ".", "HCP", "[", "]", ".", "HCP", "proposed", "to", "solve", "the", "multi", "-", "label", "object", "recognition", "task", "by", "extracting", "object", "proposals", "from", "the", "images", ".", "The", "method", "used", "image", "label", "and", "square", "loss", "to", "fine", "-", "tune", "a", "pre", "-", "trained", "ConvNet", ".", "Then", ",", "BING", "or", "EdgeBoxes", "was", "used", "to", "extract", "object", "proposals", ",", "which", "were", "used", "to", "fine", "-", "tune", "the", "ConvNet", "again", ".", "Finally", ",", "scores", "of", "these", "proposals", "were", "max", "-", "pooled", "to", "obtain", "the", "prediction", ".", "Fev", "+", "Lv", "[", "]", ".", "This", "approach", "transformed", "the", "multi", "-", "label", "object", "recognition", "problem", "into", "a", "multi", "-", "class", "multi", "-", "instance", "learning", "problem", ".", "Two", "views", "(", "label", "view", "and", "feature", "view", ")", "were", "extracted", "for", "each", "proposal", "of", "images", ".", "Then", ",", "these", "two", "views", "were", "encoded", "by", "a", "Fisher", "vector", "for", "each", "image", ".", "IF", "-", "VGG", "-", "ℓ2", "and", "IF", "-", "VGG", "-", "KL", ".", "We", "fine", "-", "tune", "the", "VGG", "-", "Nets", "with", "square", "loss", "and", "multi", "-", "label", "cross", "-", "entropy", "loss", "and", "use", "them", "as", "our", "IF", "-", "DLDL", "’s", "baselines", ".", "They", "are", "trained", "using", "the", "same", "setting", ".", "Implementation", "details", ".", "According", "to", "the", "ground", "-", "truth", "labels", ",", "we", "set", "different", "probabilities", "for", "all", "possible", "labels", "on", "PASCAL", "VOC", "dataset", ".", "In", "our", "experiments", ",", ",", ",", ".", "Finally", ",", "similar", "to", "label", "smoothing", ",", "a", "uniform", "distribution", "is", "added", "to", ",", "where", ".", "IF", "-", "DLDL", ".", "Following", ",", "each", "training", "image", "is", "individually", "rescaled", "by", "randomly", "sampling", "in", "the", "range", "[", "256", ",", "512", "]", ".", "We", "randomly", "crop", "patches", "from", "these", "resized", "images", ".", "We", "also", "adjust", "the", "pooling", "kernel", "in", "the", "pool5", "layer", "from", "to", ".", "Max", "-", "pooling", "and", "Avg", "-", "pooling", "are", "used", "at", "pool5", "to", "train", "two", "ConvNets", ".", "We", "obtain", "four", "ConvNet", "models", "thought", "fine", "-", "tuning", "“", "Net", "-", "D", "”", "and", "“", "Net", "-", "E", "”", ".", "At", "the", "prediction", "stage", ",", "the", "smaller", "side", "of", "each", "image", "is", "scaled", "to", "a", "fixed", "length", ".", "Each", "scaled", "image", "is", "fed", "to", "the", "fine", "-", "tuned", "ConvNets", "to", "obtain", "the", "20", "-", "dim", "probability", "outputs", ".", "These", "probability", "outputs", "from", "different", "scales", "and", "different", "models", "are", "averaged", "to", "form", "the", "final", "prediction", ".", "PF", "-", "DLDL", ".", "Following", ",", "we", "further", "fine", "-", "tune", "IF", "-", "DLDL", "models", "with", "proposals", "of", "images", "to", "boost", "performance", ".", "For", "each", "training", "image", ",", "we", "employ", "EdgeBoxes", "to", "produce", "a", "set", "of", "proposal", "bounding", "boxes", "which", "are", "grouped", "into", "clusters", "by", "the", "normalized", "cut", "algorithm", ".", "For", "each", "cluster", ",", "the", "top", "proposals", "with", "higher", "predictive", "scores", "generated", "by", "EdgeBoxes", "are", "resized", "into", "square", "shapes", "(", "i.e.", ",", ")", ".", "As", "a", "result", ",", "we", "can", "obtain", "proposals", "for", "an", "image", ".", "Finally", ",", "these", "resized", "proposals", "are", "fed", "into", "a", "fine", "-", "tuned", "IF", "-", "DLDL", "model", "to", "obtain", "prediction", "scores", "and", "these", "scores", "are", "fused", "by", "max", "-", "pooling", "to", "form", "the", "prediction", "distribution", "of", "the", "image", ".", "This", "process", "can", "be", "learned", "by", "using", "an", "end", "-", "to", "-", "end", "way", ".", "In", "our", "implementation", ",", "we", "set", "and", "at", "the", "training", "and", "the", "prediction", "stage", ",", "respectively", ".", "Similar", "to", "IF", "-", "DLDL", ",", "we", "also", "average", "fuse", "prediction", "scores", "of", "different", "models", "to", "generate", "the", "final", "prediction", ".", "Results", ".", "In", "Table", "[", "reference", "]", ",", "we", "compare", "single", "model", "results", "(", "average", "AP", "of", "all", "classes", ")", "on", "VOC2007", ".", "Our", "PF", "-", "DLDL", "defeats", "all", "the", "other", "methods", ".", "Compared", "with", "Fev", "+", "Lv", ",", "1.7", "%", "improvement", "can", "be", "achieved", "by", "PF", "-", "DLDL", "even", "without", "using", "the", "bounding", "box", "annotation", ".", "Compared", "with", "HCP", "-", "VGG", ",", "our", "PF", "-", "DLDL", "can", "achieve", "92.3", "%", "mAP", ",", "which", "is", "significantly", "higher", "than", "their", "90.9", "%", ".", "This", "further", "indicates", "that", "it", "is", "very", "important", "to", "learn", "a", "label", "distribution", ".", "Table", "[", "reference", "]", "and", "[", "reference", "]", "report", "details", "of", "all", "experimental", "results", "on", "VOC2007", "and", "VOC2012", ",", "respectively", ".", "It", "can", "be", "seen", "that", "IF", "-", "DLDL", "outperforms", "IF", "-", "VGG", "-", "by", "1.1", "%", "for", "VOC2007", "and", "1.3", "%", "for", "VOC2012", ",", "which", "indicates", "that", "the", "KL", "loss", "function", "is", "more", "suitable", "than", "loss", "for", "measuring", "the", "similarity", "of", "two", "label", "distributions", ".", "Furthermore", ",", "IF", "-", "DLDL", "improves", "IF", "-", "VGG", "-", "KL", "for", "about", "0.2–0.3", "points", "in", "mAP", ",", "which", "suggests", "that", "learning", "a", "label", "distribution", "is", "beneficial", ".", "More", "importantly", ",", "PF", "-", "DLDL", "can", "achieve", "93.4", "%", "for", "VOC2007", "and", "92.4", "%", "for", "VOC2012", "in", "mAP", "when", "we", "average", "fuse", "output", "scores", "of", "four", "PF", "-", "DLDL", "models", ".", "Our", "framework", "shows", "good", "performance", "especially", "for", "scene", "categories", "such", "as", "“", "chair", "”", ",", "‘", "table", "”", "and", "“", "sofa", "”", ".", "Although", "PF", "-", "DLDL", "significantly", "outperforms", "IF", "-", "DLDL", "in", "mAP", ",", "PF", "-", "DLDL", "has", "higher", "computational", "cost", "than", "IF", "-", "DLDL", "on", "both", "training", "and", "testing", "stages", ".", "Since", "IF", "-", "DLDL", "does", "not", "need", "region", "proposals", "or", "bounding", "box", "information", ",", "it", "may", "be", "effectively", "and", "efficiently", "implemented", "for", "practical", "multi", "-", "label", "application", "such", "as", "multi", "-", "label", "image", "retrieval", ".", "It", "is", "also", "possible", "that", "by", "adopting", "new", "techniques", "(", "such", "as", "the", "region", "proposal", "method", "using", "gated", "unit", "in", ",", "which", "has", "higher", "accuracy", "that", "ours", "on", "VOC", "tasks", ")", ",", "the", "accuracy", "of", "our", "DLDL", "methods", "can", "be", "further", "improved", ".", "[", "ChaLearn", "]", "[", "Morph", "]", "[", "Pointing’04", "]", "[", "AFLW", "pitch", "]", "[", "AFLW", "yaw", "]", "BFGS", "-", "LDL", "DLDL", "subsection", ":", "Semantic", "segmentation", "Datasets", ".", "We", "employ", "the", "PASCAL", "VOC2011", "segmentation", "dataset", "and", "the", "Semantic", "Boundaries", "Dataset", "(", "SBD", ")", "for", "training", "the", "proposed", "DLDL", ".", "There", "are", "2", ",", "224", "images", "(", "1", ",", "112", "for", "training", "and", "1", ",", "112", "for", "testing", ")", "with", "pixel", "labels", "for", "20", "semantic", "categories", "in", "VOC2011", ".", "SBD", "contains", "11", ",", "355", "annotated", "images", "(", "8", ",", "984", "for", "training", "and", "2", ",", "371", "for", "testing", ")", "from", "Hariharan", "et", "al", ".", ".", "Following", "FCN", ",", "we", "train", "DLDL", "using", "the", "union", "set", "(", "8", ",", "825", "images", ")", "of", "SBD", "and", "VOC2011", "training", "images", ".", "We", "evaluate", "the", "proposed", "approach", "on", "VOC2011", "(", "1", ",", "112", ")", "and", "VOC2012", "(", "1", ",", "456", ")", "test", "images", ".", "Evaluation", "criteria", ".", "The", "performance", "is", "measured", "in", "terms", "of", "mean", "IU", "(", "intersection", "over", "union", ")", ",", "which", "is", "the", "most", "widely", "used", "metric", "in", "semantic", "segmentation", ".", "We", "keep", "the", "same", "settings", "as", "FCN", "including", "training", "images", "and", "model", "structure", ".", "The", "main", "change", "is", "that", "we", "employ", "KL", "divergence", "as", "the", "loss", "function", "based", "on", "label", "distribution", "(", "Eq", ".", "[", "reference", "]", ")", ".", "Note", "that", "although", "we", "transform", "the", "ground", "-", "truth", "to", "label", "distribution", "in", "the", "training", "process", ",", "our", "evaluation", "rely", "only", "on", "ground", "-", "truth", "label", ".", "Recently", ",", "Conditional", "Random", "Field", "(", "CRF", ")", "has", "been", "broadly", "used", "in", "many", "state", "-", "of", "-", "the", "-", "art", "semantic", "segmentation", "systems", ".", "We", "optionally", "employ", "a", "fully", "connected", "CRF", "to", "refine", "the", "predicted", "category", "score", "maps", "using", "the", "default", "parameters", "of", ".", "Results", ".", "Table", "[", "reference", "]", "gives", "the", "performance", "of", "DLDL", "-", "8s", "and", "DLDL", "-", "8s", "-", "CRF", "on", "the", "test", "images", "of", "VOC2011", "and", "VOC2012", "and", "compares", "it", "to", "the", "well", "-", "known", "FCN", "-", "8s", ".", "DLDL", "-", "8s", "improves", "the", "mean", "IU", "of", "FCN", "-", "8s", "form", "62.7", "%", "to", "64.9", "%", "on", "VOC2011", ".", "On", "VOC2012", ",", "DLDL", "-", "8s", "leads", "to", "an", "improvement", "of", "2.3", "points", "in", "mean", "IU", ".", "DLDL", "achieves", "better", "results", "than", "FCN", ",", "which", "suggests", "it", "is", "important", "to", "improve", "the", "segmentation", "performance", "using", "label", "ambiguity", ".", "In", "addition", ",", "the", "CRF", "further", "improve", "performance", "of", "DLDL", "-", "8s", ",", "offering", "a", "2.6", "%", "absolute", "increase", "in", "mean", "IU", "both", "on", "VOC2011", "and", "VOC2012", ".", "Fig", ".", "[", "reference", "]", "shows", "four", "semantic", "segmentation", "examples", "from", "the", "VOC2011", "validation", "images", "using", "FCN", "-", "8s", ",", "DLDL", "-", "8s", "and", "DLDL", "-", "8s", "-", "CRF", ".", "We", "can", "see", "that", "DLDL", "-", "8s", "can", "successfully", "segment", "some", "small", "objects", "(", "e.g.", ",", "car", "and", "bicycle", ")", "and", "particularly", "improve", "the", "segmentation", "of", "object", "boundaries", "(", "e.g.", ",", "horse", "’s", "leg", "and", "plant", "’s", "leaves", ")", ",", "but", "FCN", "-", "8s", "does", "not", ".", "DLDL", "-", "8s", "may", "fail", ",", "e.g.", ",", "it", "sees", "a", "flowerpot", "as", "a", "potted", "plant", "in", "the", "fourth", "row", "in", "Fig", ".", "[", "reference", "]", ".", "Furthermore", ",", "compared", "to", "DLDL", "-", "8s", ",", "DLDL", "-", "8s", "-", "CRF", "is", "able", "to", "refine", "coarse", "pixel", "-", "level", "label", "predictions", "to", "produce", "sharp", "boundaries", "and", "fine", "-", "grained", "segmentations", "(", "e.g.", ",", "plant", "’s", "leaves", ")", ".", "[", "subfloat", "]", "labelformat", "=", "empty", "position", "=", "top", "[", "Image", "]", "[", "FCN", "-", "8s", "]", "[", "DLDL", "-", "8s", "]", "[", "DLDL", "-", "8s", "+", "CRF", "]", "[", "Ground", "-", "truth", "]", "[", "ChaLearn", "]", "[", "Morph", "]", "[", "BJUT", "-", "3D", "]", "[", "AFLW", "]", "section", ":", "Discussions", "In", "this", "section", ",", "we", "try", "to", "understand", "the", "generalization", "performance", "of", "DLDL", "through", "feature", "visualization", ",", "and", "to", "analyze", "why", "DLDL", "can", "achieve", "high", "accuracy", "with", "limited", "training", "data", ".", "In", "addition", ",", "a", "study", "of", "the", "hyper", "-", "parameter", "is", "also", "provided", ".", "Feature", "visualization", ".", "We", "visualize", "the", "model", "features", "in", "a", "low", "-", "dimensional", "space", ".", "Early", "layers", "learn", "low", "-", "level", "features", "(", "e.g.", ",", "edge", "and", "corner", ")", "and", "latter", "layers", "learn", "high", "level", "features", "(", "e.g.", ",", "shapes", "and", "objects", ")", "in", "a", "deep", "ConvNet", ".", "Hence", ",", "we", "extract", "the", "penultimate", "layer", "features", "(", "4", ",", "096", "-", "dimensional", ")", "on", "Morph", ",", "ChaLearn", ",", "Pointing’04", "and", "AFLW", "validation", "sets", ".", "To", "obtain", "the", "2", "-", "dimensional", "embeddings", "of", "the", "extracted", "high", "dimensional", "features", ",", "we", "employ", "a", "popular", "dimension", "reduction", "algorithm", "t", "-", "SNE", ".", "The", "low", "-", "dimensional", "embeddings", "of", "validation", "images", "from", "the", "above", "four", "datasets", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "The", "first", "row", "shows", "the", "2", "-", "dim", "embeddings", "of", "hand", "-", "crafted", "features", "(", "BIF", "for", "Morph", "and", "Chalearn", ",", "HOG", "for", "Pointing’04", "and", "AFLW", ")", "and", "the", "second", "row", "shows", "that", "of", "the", "DLDL", "features", ".", "These", "figures", "are", "colored", "by", "their", "semantic", "category", ".", "It", "can", "be", "observed", "that", "clear", "semantic", "clusterings", "(", "old", "or", "young", "for", "age", "datasets", ",", "left", "or", "right", ",", "up", "or", "down", "for", "head", "pose", "datasets", ")", "appear", "in", "deep", "features", "but", "do", "not", "in", "hand", "-", "crafted", "features", ".", "Reduce", "over", "-", "fitting", ".", "DLDL", "can", "effectively", "reduce", "over", "-", "fitting", "when", "the", "training", "set", "is", "small", ".", "This", "effect", "can", "be", "explained", "by", "the", "label", "ambiguity", ".", "Considering", "an", "input", "sample", "with", "one", "single", "label", ".", "In", "traditional", "deep", "ConvNet", ",", "and", "for", "all", ".", "In", "DLDL", ",", "the", "label", "distribution", "contains", "many", "non", "zeros", "elements", ".", "The", "diversity", "of", "labels", "helps", "reduce", "over", "-", "fitting", ".", "Moreover", ",", "the", "objective", "function", "(", "Eq", ".", "[", "reference", "]", ")", "of", "DLDL", "can", "be", "rewritten", "as", "In", "Eq", ".", "[", "reference", "]", ",", "the", "first", "term", "is", "the", "tradition", "ConvNet", "loss", "function", ".", "The", "second", "term", "maximize", "the", "log", "-", "likelihood", "of", "the", "ambiguous", "labels", ".", "Unlike", "existing", "data", "augmentation", "techniques", "such", "as", "random", "cropping", "on", "the", "images", ",", "DLDL", "augments", "data", "on", "the", "label", "side", ".", "In", "Fig", ".", "[", "reference", "]", ",", "MAE", "is", "shown", "as", "a", "function", "of", "the", "number", "of", "epochs", "on", "two", "age", "datasets", "(", "ChaLearn", "and", "Morph", ")", "and", "two", "head", "pose", "datasets", "(", "BJUT", "-", "3D", "and", "AFLW", ")", ".", "On", "ChaLearn", "and", "AFLW", ",", "C", "-", "ConveNet", "(", "softmax", ")", "achieves", "the", "lowest", "training", "MAE", ",", "but", "produces", "the", "highest", "validation", "MAE", ".", "In", "particular", ",", "the", "validation", "MAE", "increases", "after", "the", "8th", "epoch", "on", "ChaLearn", ".", "Similar", "phenomenon", "is", "observed", "on", "AFLW", ".", "This", "fact", "shows", "that", "over", "-", "fitting", "happens", "in", "C", "-", "ConvNet", "when", "the", "number", "of", "training", "images", "is", "small", ".", "Although", "there", "are", "15", ",", "561", "training", "images", "in", "AFLW", ",", "each", "category", "contains", "on", "averagely", "4", "training", "images", "since", "there", "are", "3", ",", "721", "categories", ".", "Accelerate", "convergence", ".", "We", "further", "analyze", "the", "convergence", "performance", "of", "DLDL", ",", "C", "-", "ConvNet", "and", "R", "-", "ConvNet", ".", "We", "can", "observe", "that", "the", "training", "MAE", "is", "reduced", "very", "slowly", "at", "the", "beginning", "of", "training", "using", "C", "-", "ConvNet", "and", "R", "-", "ConveNet", "in", "many", "cases", "as", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "On", "the", "contrary", ",", "the", "MAE", "of", "DLDL", "reduces", "quickly", ".", "Robust", "performance", ".", "One", "notable", "observation", "is", "that", "C", "-", "ConvNet", "and", "R", "-", "ConveNet", "is", "unstable", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "MAE", "for", "pitch", "+", "yaw", ",", "a", "complicated", "estimation", "of", "the", "joint", "distribution", ".", "This", "is", "a", "very", "sparse", "label", "set", "because", "the", "interval", "of", "adjacent", "class", "(", "pitch", "or", "yaw", ")", "is", ".", "R", "-", "ConvNet", "has", "difficulty", "in", "estimating", "this", "output", ",", "yielding", "errors", "that", "are", "roughly", "20", "times", "higher", "than", "DLDL", "and", "C", "-", "ConvNet", ".", "On", "the", "other", "hand", ",", "C", "-", "ConvNet", "easily", "fall", "into", "over", "-", "fitting", "when", "there", "are", "not", "enough", "training", "data", "(", "e.g", ",", "Fig", ".", "[", "reference", "]", "and", "Fig", ".", "[", "reference", "]", ")", ".", "The", "proposed", "DLDL", "is", "more", "amenable", "to", "small", "datasets", "or", "sparse", "labels", "than", "C", "-", "ConvNet", "and", "R", "-", "ConvNet", ".", "Analyze", "the", "hyper", "-", "parameter", ".", "DLDL", "’s", "performance", "may", "be", "affected", "by", "the", "label", "distribution", ".", "Here", ",", "we", "take", "age", "estimation", "(", "Morph", ")", "and", "head", "pose", "estimation", "(", "Pointing’04", ")", "for", "examples", ".", "is", "a", "common", "hyper", "-", "parameter", "in", "these", "tasks", "if", "it", "is", "not", "provided", "in", "the", "ground", "-", "truth", ".", "We", "have", "empirically", "set", "in", "Morph", ",", "and", "in", "Pointing’04", "in", "our", "experiments", ".", "In", "order", "to", "study", "the", "impact", "of", ",", "we", "test", "DLDL", "with", "different", "values", ",", "changing", "from", "0", "to", "3", "with", "0.5", "interval", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "MAE", "performance", "on", "Morph", "and", "Pointing’04", "with", "different", ".", "We", "can", "see", "that", "a", "proper", "is", "important", "for", "low", "MAE", ".", "But", "generally", "speaking", ",", "a", "value", "that", "is", "close", "to", "the", "interval", "between", "neighboring", "labels", "is", "a", "good", "choice", ".", "Because", "the", "shape", "of", "all", "curves", "are", "V", "-", "shape", "like", ",", "it", "is", "also", "very", "convenient", "to", "find", "an", "optimal", "value", "using", "the", "cross", "-", "validation", "strategy", ".", "section", ":", "Conclusion", "We", "observe", "that", "current", "deep", "ConvNets", "can", "not", "successfully", "learn", "good", "models", "when", "there", "are", "not", "enough", "training", "data", "and", "/", "or", "the", "labels", "are", "ambiguous", ".", "We", "propose", "DLDL", ",", "a", "deep", "label", "distribution", "learning", "framework", "to", "solve", "this", "issue", "by", "exploiting", "label", "ambiguity", ".", "In", "DLDL", ",", "each", "image", "is", "labeled", "by", "a", "label", "distribution", ",", "which", "can", "utilize", "label", "ambiguity", "in", "both", "feature", "learning", "and", "classifier", "learning", ".", "DLDL", "consistently", "improves", "the", "network", "training", "process", "in", "our", "experiments", ",", "by", "preventing", "it", "from", "over", "-", "fitting", "when", "the", "training", "set", "is", "small", ".", "We", "empirically", "showed", "that", "DLDL", "produces", "robust", "and", "competitive", "performances", "than", "traditional", "classification", "or", "regression", "deep", "models", "on", "several", "popular", "visual", "recognition", "tasks", ".", "However", ",", "constructing", "a", "reasonable", "label", "distribution", "is", "still", "challenging", "due", "to", "the", "diversity", "of", "label", "space", "for", "different", "recognition", "tasks", ".", "It", "is", "an", "interesting", "direction", "to", "extend", "DLDL", "to", "more", "recognition", "problems", "by", "constructing", "different", "label", "distributions", ".", "bibliography", ":", "References", "[", "]", "Bin", "-", "Bin", "Gao", "received", "the", "B.S.", "and", "M.S.", "degrees", "in", "applied", "mathematics", "in", "2010", "and", "2013", ",", "respectively", ".", "He", "is", "currently", "pursuing", "the", "Ph.D.", "degree", "in", "the", "Department", "of", "Computer", "Science", "and", "Technology", ",", "Nanjing", "University", ",", "China", ".", "His", "research", "interests", "include", "computer", "vision", "and", "machine", "learning", ".", "{", "IEEEbiography}", "[]", "Chao", "Xing", "received", "the", "B.S.", "degree", "in", "software", "engineering", "from", "Southeast", "University", ",", "China", ",", "in", "2014", ".", "He", "is", "currently", "a", "postgraduate", "student", "in", "the", "School", "of", "Computer", "Science", "and", "Engineering", "at", "Southeast", "University", ",", "China", ".", "His", "research", "interests", "include", "pattern", "recognition", ",", "machine", "learning", ",", "and", "data", "mining", ".", "{", "IEEEbiography}", "[]", "Chen", "-", "Wei", "Xie", "received", "his", "B.S.", "degree", "from", "Southeast", "University", ",", "China", ",", "in", "2015", ".", "He", "is", "currently", "a", "postgraduate", "student", "in", "the", "Department", "of", "Computer", "Science", "and", "Technology", ",", "Nanjing", "University", ",", "China", ".", "His", "research", "interests", "include", "computer", "vision", "and", "machine", "learning", ".", "{", "IEEEbiography}", "[]", "Jianxin", "Wu", "(", "M’09", ")", "received", "the", "B.S.", "and", "M.S.", "degrees", "in", "computer", "science", "from", "Nanjing", "University", ",", "and", "the", "Ph.D.", "degree", "in", "computer", "science", "from", "the", "Georgia", "Institute", "of", "Technology", ".", "He", "was", "an", "Assistant", "Professor", "with", "the", "Nanyang", "Technological", "University", ",", "Singapore", ".", "He", "is", "currently", "a", "Professor", "with", "the", "Department", "of", "Computer", "Science", "and", "Technology", ",", "Nanjing", "University", ",", "China", ",", "and", "is", "associated", "with", "the", "National", "Key", "Laboratory", "for", "Novel", "Software", "Technology", ",", "China", ".", "His", "current", "research", "interests", "include", "computer", "vision", "and", "machine", "learning", ".", "He", "has", "served", "as", "an", "Area", "Chair", "for", "CVPR", "2017", "and", "ICCV", "2015", ",", "a", "Senior", "PC", "Member", "for", "AAAI", "2017", "and", "AAAI", "2016", ",", "and", "an", "Associate", "Editor", "of", "Pattern", "Recognition", "Journal", ".", "{", "IEEEbiography}", "[]", "Xin", "Geng", "(", "M’13", ")", "received", "the", "B.S.", "and", "M.S.", "degrees", "in", "computer", "science", "from", "Nanjing", "University", ",", "China", ",", "in", "2001", "and", "2004", ",", "respectively", ",", "and", "the", "Ph", ".", "D", "degree", "from", "Deakin", "University", ",", "Australia", "in", "2008", ".", "He", "joined", "the", "School", "of", "Computer", "Science", "and", "Engineering", "at", "Southeast", "University", ",", "China", ",", "in", "2008", ",", "and", "is", "currently", "a", "professor", "and", "vice", "dean", "of", "the", "school", ".", "He", "has", "authored", "over", "50", "refereed", "papers", ",", "and", "he", "holds", "five", "patents", "in", "these", "areas", ".", "His", "research", "interests", "include", "pattern", "recognition", ",", "machine", "learning", ",", "and", "computer", "vision", "."], "entity": [[2, 6, ["Deep", "Label", "Distribution", "Learning"], "Method"], [7, 9, ["Label", "Ambiguity"], "Task"], [9, 12, ["Convolutional", "Neural", "Networks"], "Method"], [13, 14, ["ConvNets"], "Method"], [18, 19, ["recognition"], "Task"], [22, 25, ["visual", "recognition", "tasks"], "Task"], [61, 63, ["age", "estimation"], "Task"], [64, 67, ["head", "pose", "estimation"], "Task"], [68, 72, ["multi", "-", "label", "classification"], "Task"], [73, 75, ["semantic", "segmentation"], "Task"], [92, 93, ["classification"], "Task"], [134, 136, ["deep", "ConvNets"], "Method"], [139, 140, ["DLDL"], "Method"], [141, 145, ["Deep", "Label", "Distribution", "Learning"], "Method"], [154, 156, ["feature", "learning"], "Method"], [157, 159, ["classifier", "learning"], "Method"], [198, 200, ["age", "estimation"], "Task"], [201, 204, ["head", "pose", "estimation"], "Task"], [213, 214, ["recognition"], "Task"], [216, 224, ["multi", "-", "label", "classification", "and", "semantic", "segmentation", "tasks"], "Task"], [225, 227, ["Label", "distribution"], "Task"], [228, 230, ["deep", "learning"], "Task"], [231, 233, ["age", "estimation"], "Task"], [234, 237, ["head", "pose", "estimation"], "Task"], [238, 240, ["semantic", "segmentation"], "Task"], [244, 247, ["Convolutional", "Neural", "Networks"], "Method"], [248, 249, ["ConvNets"], "Method"], [262, 265, ["visual", "recognition", "tasks"], "Task"], [267, 269, ["image", "classification"], "Task"], [273, 275, ["semantic", "segmentation"], "Task"], [313, 315, ["age", "estimation"], "Task"], [316, 319, ["head", "pose", "estimation"], "Task"], [320, 324, ["multi", "-", "label", "classification"], "Task"], [325, 327, ["semantic", "segmentation"], "Task"], [331, 334, ["deep", "learning", "methods"], "Method"], [396, 397, ["annotators"], "Task"], [398, 400, ["semantic", "segmentation"], "Task"], [404, 406, ["pixel", "labeling"], "Task"], [491, 494, ["head", "pose", "estimation"], "Task"], [534, 537, ["image", "classification", "tasks"], "Task"], [549, 551, ["label", "ambiguity"], "Task"], [651, 653, ["semantic", "segmentation"], "Task"], [659, 661, ["label", "ambiguity"], "Task"], [680, 685, ["multi", "-", "label", "classification", "task"], "Task"], [706, 715, ["PASCAL", "Visual", "Object", "Classes", "(", "VOC", ")", "classification", "challenge"], "Task"], [740, 743, ["Head", "pose", "estimation"], "Task"], [745, 749, ["Multi", "-", "label", "classification"], "Task"], [751, 753, ["Semantic", "segmentation"], "Task"], [760, 762, ["labeling", "methods"], "Method"], [763, 767, ["single", "-", "label", "recognition"], "Method"], [768, 769, ["SLR"], "Method"], [771, 775, ["multi", "-", "label", "recognition"], "Method"], [776, 777, ["MLR"], "Method"], [779, 780, ["SLR"], "Method"], [789, 790, ["MLR"], "Method"], [803, 804, ["SLR"], "Method"], [805, 806, ["MLR"], "Method"], [836, 838, ["Label", "ambiguity"], "Task"], [841, 842, ["recognition"], "Task"], [878, 885, ["label", "distribution", "learning", "(", "LDL", ")", "approach"], "Method"], [886, 888, ["age", "estimation"], "Task"], [889, 892, ["head", "pose", "estimation"], "Task"], [898, 899, ["LDL"], "Method"], [911, 912, ["LDLogitBoost"], "Method"], [913, 916, ["AOSO", "-", "LDLogitBoost"], "Method"], [918, 920, ["general", "models"], "Method"], [923, 926, ["maximum", "entropy", "model"], "Method"], [942, 945, ["weighted", "linear", "combination"], "Method"], [974, 976, ["classifier", "learning"], "Task"], [982, 984, ["visual", "representations"], "Task"], [985, 987, ["Deep", "ConvNets"], "Method"], [991, 993, ["feature", "learning"], "Task"], [995, 997, ["ConvNet", "frameworks"], "Method"], [1001, 1005, ["classification", "and", "regression", "models"], "Method"], [1008, 1011, ["optimization", "objective", "functions"], "Metric"], [1020, 1021, ["loss"], "Method"], [1024, 1027, ["deep", "ConvNet", "models"], "Method"], [1028, 1032, ["classification", "and", "regression", "problems"], "Task"], [1084, 1086, ["image", "classification"], "Task"], [1087, 1090, ["human", "pose", "estimation"], "Task"], [1097, 1100, ["deep", "learning", "methods"], "Method"], [1120, 1121, ["ConvNet"], "Method"], [1140, 1141, ["SLR"], "Method"], [1142, 1144, ["MLR", "problems"], "Task"], [1145, 1149, ["label", "distribution", "learning", "problems"], "Task"], [1180, 1184, ["deep", "label", "distribution", "learning"], "Method"], [1217, 1219, ["label", "distribution"], "Task"], [1221, 1223, ["recognition", "tasks"], "Task"], [1229, 1231, ["deep", "ConvNet"], "Method"], [1238, 1240, ["feature", "learning"], "Task"], [1241, 1243, ["classifier", "learning"], "Task"], [1247, 1249, ["label", "distribution"], "Task"], [1250, 1252, ["deep", "ConvNets"], "Method"], [1257, 1258, ["DLDL"], "Method"], [1259, 1263, ["Deep", "Label", "Distribution", "Learning"], "Method"], [1267, 1268, ["DLDL"], "Method"], [1273, 1274, ["DLDL"], "Method"], [1276, 1283, ["end", "-", "to", "-", "end", "learning", "framework"], "Method"], [1290, 1292, ["feature", "learning"], "Method"], [1293, 1295, ["classifier", "learning"], "Method"], [1296, 1297, ["DLDL"], "Method"], [1305, 1309, ["classification", "and", "regression", "methods"], "Method"], [1339, 1341, ["predicting", "faces"], "Task"], [1347, 1348, ["DLDL"], "Method"], [1350, 1352, ["single", "model"], "Method"], [1353, 1354, ["ensemble"], "Method"], [1369, 1370, ["age"], "Task"], [1371, 1375, ["head", "pose", "estimation", "tasks"], "Task"], [1376, 1377, ["DLDL"], "Method"], [1382, 1389, ["multi", "-", "label", "classification", "and", "semantic", "segmentation"], "Task"], [1432, 1435, ["label", "distribution", "construction"], "Method"], [1487, 1489, ["visual", "recognition"], "Task"], [1493, 1495, ["image", "classification"], "Task"], [1499, 1501, ["semantic", "segmentation"], "Task"], [1503, 1505, ["facial", "attribute"], "Task"], [1507, 1508, ["age"], "Task"], [1540, 1543, ["deep", "learning", "methods"], "Method"], [1571, 1573, ["feature", "extraction"], "Task"], [1580, 1581, ["recognition"], "Task"], [1582, 1583, ["detection"], "Task"], [1584, 1585, ["estimation"], "Task"], [1589, 1590, ["SVM"], "Method"], [1591, 1593, ["random", "forest"], "Method"], [1594, 1596, ["neural", "networks"], "Method"], [1602, 1604, ["learning", "stage"], "Task"], [1614, 1618, ["label", "distribution", "learning", "approach"], "Method"], [1631, 1633, ["age", "estimation"], "Task"], [1634, 1637, ["head", "pose", "estimation"], "Task"], [1649, 1653, ["hand", "-", "crafted", "features"], "Method"], [1661, 1662, ["age"], "Task"], [1663, 1666, ["head", "pose", "estimation"], "Task"], [1670, 1673, ["learning", "feature", "representation"], "Method"], [1695, 1697, ["age", "estimation"], "Task"], [1698, 1700, ["Deep", "learning"], "Method"], [1713, 1715, ["image", "classification"], "Task"], [1719, 1721, ["semantic", "segmentation"], "Task"], [1724, 1726, ["vision", "tasks"], "Task"], [1737, 1739, ["deep", "models"], "Method"], [1740, 1741, ["classification"], "Task"], [1743, 1744, ["classification"], "Task"], [1745, 1747, ["deep", "ConvNets"], "Method"], [1752, 1754, ["regression", "tasks"], "Task"], [1756, 1759, ["head", "pose", "estimation"], "Task"], [1765, 1767, ["regression", "problems"], "Task"], [1769, 1771, ["training", "procedure"], "Method"], [1786, 1790, ["Tukey", "’s", "biweight", "function"], "Method"], [1791, 1794, ["human", "pose", "estimation"], "Task"], [1798, 1800, ["model", "architecture"], "Method"], [1801, 1804, ["deep", "ConvNet", "models"], "Method"], [1806, 1808, ["deeper", "architecture"], "Method"], [1810, 1812, ["convolution", "filters"], "Method"], [1815, 1818, ["VGG", "-", "Nets"], "Method"], [1819, 1822, ["VGG", "-", "Face"], "Method"], [1830, 1833, ["deep", "learning", "methods"], "Method"], [1870, 1872, ["label", "smoothing"], "Method"], [1873, 1874, ["LS"], "Method"], [1903, 1904, ["classifier"], "Method"], [1907, 1908, ["LS"], "Method"], [1974, 1976, ["deep", "ConvNets"], "Task"], [1977, 1978, ["Age"], "Task"], [1979, 1982, ["head", "pose", "estimation"], "Task"], [2003, 2007, ["multi", "-", "label", "classification"], "Task"], [2008, 2010, ["semantic", "segmentation"], "Task"], [2044, 2046, ["construction", "methods"], "Method"], [2047, 2049, ["label", "distribution"], "Task"], [2051, 2053, ["recognition", "tasks"], "Task"], [2069, 2074, ["deep", "label", "distribution", "learning", "problem"], "Method"], [2185, 2189, ["label", "distribution", "learning", "approach"], "Method"], [2194, 2197, ["machine", "learning", "phase"], "Task"], [2221, 2225, ["end", "deep", "learning", "method"], "Method"], [2232, 2234, ["feature", "learning"], "Method"], [2235, 2237, ["classifier", "learning"], "Method"], [2242, 2243, ["DLDL"], "Method"], [2268, 2269, ["DLDL"], "Method"], [2283, 2285, ["age", "estimation"], "Task"], [2319, 2321, ["age", "estimation"], "Task"], [2325, 2326, ["DLDL"], "Method"], [2331, 2333, ["label", "distribution"], "Task"], [2352, 2354, ["label", "distribution"], "Task"], [2356, 2359, ["deep", "learning", "machine"], "Method"], [2400, 2401, ["DLDL"], "Method"], [2474, 2475, ["DLDL"], "Method"], [2496, 2500, ["Deep", "label", "distribution", "learning"], "Method"], [2516, 2519, ["fully", "connected", "layer"], "Method"], [2521, 2523, ["deep", "ConvNet"], "Method"], [2550, 2551, ["DLDL"], "Method"], [2586, 2589, ["KL", ")", "divergence"], "Metric"], [2624, 2627, ["Stochastic", "gradient", "descent"], "Method"], [2707, 2709, ["forward", "run"], "Method"], [2727, 2728, ["Prediction"], "Task"], [2755, 2756, ["DLDL"], "Method"], [2767, 2768, ["DLDL"], "Method"], [2772, 2776, ["classification", "and", "regression", "tasks"], "Task"], [2779, 2782, ["Label", "distribution", "construction"], "Method"], [2806, 2808, ["label", "distribution"], "Method"], [2820, 2822, ["probability", "distribution"], "Method"], [2883, 2885, ["age", "estimation"], "Task"], [2886, 2889, ["head", "pose", "estimation"], "Task"], [2890, 2894, ["multi", "-", "label", "classification"], "Task"], [2895, 2897, ["semantic", "segmentation"], "Task"], [2899, 2901, ["age", "estimation"], "Task"], [2924, 2926, ["normal", "distribution"], "Method"], [3003, 3005, ["label", "distribution"], "Method"], [3026, 3027, ["quantizing"], "Task"], [3029, 3031, ["normal", "p.d.f"], "Method"], [3033, 3036, ["probability", "density", "function"], "Method"], [3083, 3084, ["DLDL"], "Method"], [3086, 3089, ["head", "pose", "estimation"], "Task"], [3102, 3105, ["learning", "joint", "distribution"], "Task"], [3109, 3110, ["DLDL"], "Method"], [3199, 3201, ["label", "distribution"], "Method"], [3236, 3240, ["multi", "-", "label", "classification"], "Task"], [3319, 3323, ["multi", "-", "label", "methods"], "Method"], [3386, 3387, ["normalization"], "Method"], [3409, 3411, ["label", "distribution"], "Task"], [3427, 3429, ["semantic", "segmentation"], "Task"], [3481, 3484, ["Fully", "Convolutional", "Networks"], "Method"], [3485, 3486, ["FCN"], "Method"], [3497, 3498, ["FCN"], "Method"], [3553, 3556, ["Gaussian", "kernel", "matrix"], "Method"], [3633, 3635, ["deep", "ConvNet"], "Method"], [3656, 3658, ["deep", "models"], "Method"], [3660, 3663, ["ZF", "-", "Net"], "Method"], [3664, 3667, ["VGG", "-", "Nets"], "Method"], [3669, 3672, ["ZF", "-", "Net"], "Method"], [3674, 3676, ["convolution", "layers"], "Method"], [3680, 3683, ["fully", "connected", "layers"], "Method"], [3685, 3689, ["VGG", "-", "Nets", "architecture"], "Method"], [3717, 3720, ["KL", "loss", "function"], "Method"], [3728, 3729, ["ReLU"], "Method"], [3730, 3733, ["ZF", "-", "Net"], "Task"], [3753, 3755, ["age", "estimation"], "Task"], [3756, 3760, ["multi", "-", "label", "classification"], "Task"], [3765, 3768, ["head", "pose", "estimation"], "Task"], [3771, 3772, ["tensor"], "Method"], [3773, 3775, ["semantic", "segmentation"], "Task"], [3784, 3786, ["deep", "models"], "Method"], [3795, 3798, ["ZF", "-", "Net"], "Task"], [3808, 3810, ["Gaussian", "distribution"], "Method"], [3830, 3831, ["ReLU"], "Method"], [3837, 3838, ["dropout"], "Method"], [3860, 3861, ["Optimization"], "Task"], [3864, 3867, ["Stochastic", "Gradient", "Descent"], "Method"], [3868, 3869, ["SGD"], "Method"], [3885, 3887, ["learning", "rate"], "Metric"], [3901, 3904, ["Fine", "-", "tuning"], "Task"], [3906, 3910, ["pre", "-", "trained", "models"], "Method"], [3911, 3914, ["VGG", "-", "Nets"], "Method"], [3915, 3918, ["16", "-", "layers"], "Method"], [3924, 3927, ["VGG", "-", "Face"], "Method"], [3928, 3931, ["16", "-", "layers"], "Method"], [3950, 3952, ["classification", "layer"], "Method"], [3953, 3955, ["loss", "layer"], "Method"], [3960, 3963, ["label", "distribution", "layer"], "Method"], [3968, 3970, ["Gaussian", "distribution"], "Method"], [3972, 3975, ["KL", "loss", "layer"], "Method"], [3977, 3979, ["learning", "rates"], "Metric"], [3981, 3983, ["convolutional", "layers"], "Method"], [3987, 3991, ["fully", "-", "connected", "layers"], "Method"], [3993, 3996, ["label", "distribution", "layer"], "Method"], [4014, 4016, ["back", "propagation"], "Method"], [4036, 4038, ["age", "estimation"], "Task"], [4041, 4045, ["multi", "-", "label", "classification"], "Task"], [4051, 4052, ["DLDL"], "Method"], [4058, 4060, ["age", "estimation"], "Task"], [4061, 4064, ["head", "pose", "estimation"], "Task"], [4065, 4069, ["multi", "-", "label", "classification"], "Task"], [4070, 4072, ["semantic", "segmentation"], "Task"], [4078, 4079, ["MatConvNet"], "Method"], [4088, 4091, ["NVIDIA", "K40", "GPU"], "Method"], [4105, 4107, ["age", "estimation"], "Task"], [4117, 4118, ["Morph"], "Material"], [4161, 4167, ["10", "-", "fold", "cross", "-", "validation"], "Method"], [4170, 4171, ["Morph"], "Material"], [4190, 4193, ["ChaLearn", "LAP", "2015"], "Material"], [4197, 4198, ["Morph"], "Material"], [4202, 4203, ["ChaLearn"], "Material"], [4354, 4355, ["DLDL"], "Method"], [4365, 4370, ["ConvNet", "+", "LS", "(", "KL"], "Method"], [4383, 4385, ["label", "smoothing"], "Method"], [4386, 4387, ["LS"], "Method"], [4405, 4406, ["LD"], "Method"], [4425, 4426, ["DLDL"], "Method"], [4432, 4433, ["BFGS"], "Method"], [4441, 4445, ["label", "distribution", "learning", "approach"], "Method"], [4446, 4449, ["IIS", "-", "LLD"], "Method"], [4451, 4452, ["age"], "Task"], [4453, 4456, ["head", "pose", "estimation"], "Task"], [4466, 4469, ["IIS", "-", "LLD"], "Method"], [4484, 4490, ["quasi", "-", "Newton", "optimization", "method", "BFGS"], "Method"], [4491, 4496, ["C", "-", "ConvNet", "Classification", "ConvNets"], "Method"], [4503, 4506, ["computer", "vision", "tasks"], "Task"], [4507, 4510, ["ZF", "-", "Net"], "Method"], [4511, 4514, ["VGG", "-", "Net"], "Method"], [4526, 4533, ["ImageNet", "-", "specific", "1000", "-", "way", "classification"], "Method"], [4541, 4545, ["R", "-", "ConvNet", "ConvNets"], "Method"], [4550, 4552, ["regression", "tasks"], "Task"], [4554, 4557, ["R", "-", "ConvNet"], "Task"], [4593, 4594, ["prediction"], "Task"], [4614, 4617, ["fully", "connected", "layer"], "Method"], [4640, 4643, ["R", "-", "ConvNet"], "Method"], [4651, 4653, ["preprocessing", "pipeline"], "Method"], [4667, 4669, ["face", "alignment"], "Task"], [4681, 4683, ["DPM", "model"], "Method"], [4698, 4701, ["cascaded", "convolution", "networks"], "Method"], [4743, 4745, ["Data", "augmentation"], "Task"], [4753, 4754, ["ChaLearn"], "Material"], [4765, 4769, ["gray", "-", "scale", "version"], "Method"], [4833, 4834, ["Morph"], "Material"], [4835, 4836, ["ChaLearn"], "Material"], [4846, 4847, ["ChaLearn"], "Material"], [4850, 4851, ["Morph"], "Material"], [4856, 4857, ["Morph"], "Material"], [4908, 4911, ["VGG", "-", "Nets"], "Method"], [4912, 4915, ["VGG", "-", "Face"], "Method"], [4917, 4919, ["Evaluation", "criteria"], "Metric"], [4920, 4923, ["Mean", "Absolute", "Error"], "Metric"], [4924, 4925, ["MAE"], "Metric"], [4927, 4931, ["Cumulative", "Score", "(", "CS"], "Metric"], [4939, 4941, ["age", "estimation"], "Task"], [4942, 4943, ["MAE"], "Metric"], [4974, 4975, ["CS"], "Method"], [5103, 5106, ["cross", "-", "validation"], "Method"], [5109, 5110, ["Morph"], "Material"], [5119, 5120, ["ChaLearn"], "Material"], [5122, 5123, ["Morph"], "Material"], [5203, 5204, ["DLDL"], "Method"], [5215, 5216, ["DLDL"], "Method"], [5224, 5227, ["C", "-", "ConvNet"], "Method"], [5228, 5229, ["softmax"], "Method"], [5235, 5236, ["Morph"], "Material"], [5264, 5265, ["DLDL"], "Method"], [5266, 5269, ["R", "-", "ConvNet"], "Method"], [5270, 5273, ["C", "-", "ConvNet"], "Method"], [5274, 5277, ["ConvNet", "+", "LS"], "Method"], [5279, 5282, ["learning", "label", "distribution"], "Task"], [5285, 5292, ["deep", "end", "-", "to", "-", "end", "models"], "Task"], [5293, 5294, ["DLDL"], "Method"], [5334, 5335, ["MAE"], "Metric"], [5342, 5345, ["KL", "-", "divergence"], "Metric"], [5353, 5357, ["similarity", "of", "two", "distributions"], "Task"], [5364, 5367, ["C", "-", "ConvNet"], "Method"], [5368, 5371, ["R", "-", "ConvNet"], "Method"], [5376, 5382, ["R", "-", "ConvNet", "(", ")", "method"], "Method"], [5390, 5391, ["ChaLearn"], "Material"], [5395, 5398, ["C", "-", "ConvNet"], "Method"], [5399, 5400, ["softmax"], "Method"], [5401, 5402, ["Exp"], "Method"], [5404, 5405, ["Morph"], "Material"], [5448, 5450, ["age", "estimation"], "Task"], [5459, 5460, ["DLDL"], "Method"], [5469, 5474, ["network", "of", "VGG", "-", "Face"], "Method"], [5502, 5503, ["DLDL"], "Method"], [5545, 5546, ["DLDL"], "Method"], [5562, 5563, ["ChaLearn"], "Material"], [5629, 5630, ["ChaLearn"], "Material"], [5631, 5632, ["Morph"], "Material"], [5651, 5652, ["accuracy"], "Metric"], [5661, 5662, ["CS"], "Metric"], [5664, 5665, ["DLDL"], "Method"], [5671, 5672, ["ConvNet"], "Method"], [5696, 5697, ["Morph"], "Material"], [5705, 5707, ["deep", "model"], "Method"], [5715, 5716, ["DLDL"], "Method"], [5750, 5751, ["DLDL"], "Method"], [5817, 5820, ["Head", "pose", "estimation"], "Task"], [5827, 5830, ["head", "pose", "estimation"], "Task"], [5939, 5942, ["CyberWare", "Laser", "Scanner"], "Method"], [5971, 5972, ["Pointing’04"], "Task"], [5981, 5982, ["Pointing’04"], "Method"], [5983, 5986, ["BJUT", "-", "3D"], "Method"], [5988, 5989, ["AFLW"], "Method"], [6005, 6008, ["mean", "3D", "face"], "Method"], [6010, 6012, ["POSIT", "algorithm"], "Method"], [6076, 6077, ["DLDL"], "Method"], [6089, 6091, ["label", "distributions"], "Task"], [6151, 6152, ["LDL"], "Method"], [6160, 6164, ["Kullback", "-", "Leibler", "divergence"], "Method"], [6181, 6183, ["network", "architecture"], "Method"], [6189, 6190, ["DLDL"], "Method"], [6191, 6194, ["head", "pose", "estimation"], "Task"], [6201, 6203, ["age", "estimation"], "Task"], [6215, 6216, ["accuracy"], "Metric"], [6236, 6238, ["Evaluation", "criteria"], "Metric"], [6242, 6244, ["prediction", "values"], "Metric"], [6276, 6277, ["MAE"], "Metric"], [6285, 6287, ["classification", "accuracy"], "Metric"], [6288, 6289, ["Acc"], "Metric"], [6300, 6301, ["Acc"], "Metric"], [6303, 6307, ["pose", "class", "classification", "accuracy"], "Metric"], [6312, 6313, ["MAE"], "Metric"], [6340, 6341, ["Acc"], "Metric"], [6361, 6364, ["R", "-", "ConvNet"], "Task"], [6369, 6370, ["MAE"], "Metric"], [6372, 6373, ["Acc"], "Metric"], [6388, 6393, ["5", "-", "fold", "cross", "validation"], "Method"], [6394, 6395, ["Pointing’04"], "Task"], [6396, 6400, ["BJUT", "-", "3D", "following"], "Task"], [6402, 6403, ["AFLW"], "Method"], [6658, 6659, ["Pointing’04"], "Task"], [6685, 6686, ["MAE"], "Metric"], [6689, 6690, ["accuracy"], "Metric"], [6691, 6697, ["deep", "learning", "methods", "C", "-", "ConvNet"], "Method"], [6698, 6701, ["R", "-", "ConvNet"], "Method"], [6702, 6705, ["ConvNet", "+", "LS"], "Method"], [6711, 6714, ["C", "-", "ConvNet"], "Method"], [6715, 6718, ["R", "-", "ConvNet"], "Method"], [6719, 6722, ["ConvNet", "+", "LS"], "Method"], [6735, 6736, ["DLDL"], "Method"], [6749, 6750, ["accuracy"], "Metric"], [6751, 6754, ["pitch", "+", "yaw"], "Task"], [6758, 6761, ["C", "-", "ConvNet"], "Method"], [6769, 6770, ["DLDL"], "Method"], [6774, 6776, ["deep", "learning"], "Task"], [6782, 6788, ["C", "-", "ConvNet", "R", "-", "ConvNet"], "Method"], [6789, 6792, ["ConvNet", "+", "LS"], "Method"], [6814, 6817, ["deep", "learning", "methods"], "Method"], [6825, 6826, ["DLDL"], "Method"], [6835, 6836, ["MAE"], "Metric"], [6838, 6839, ["accuracy"], "Metric"], [6851, 6854, ["R", "-", "ConvNet"], "Method"], [6857, 6860, ["C", "-", "ConvNet"], "Method"], [6866, 6868, ["age", "estimation"], "Task"], [6869, 6872, ["head", "pose", "estimation"], "Task"], [6873, 6874, ["AFLW"], "Task"], [6880, 6883, ["C", "-", "ConvNet"], "Method"], [6884, 6887, ["BJUT", "-", "3D"], "Method"], [6888, 6889, ["pointing’04"], "Method"], [6890, 6893, ["head", "pose", "estimation"], "Task"], [6905, 6908, ["C", "-", "ConvNet"], "Method"], [6909, 6912, ["R", "-", "ConvNet"], "Method"], [6949, 6950, ["DLDL"], "Method"], [7003, 7004, ["DLDL"], "Method"], [7042, 7046, ["Multi", "-", "label", "classification"], "Task"], [7053, 7057, ["multi", "-", "label", "classification"], "Task"], [7128, 7130, ["evaluation", "metric"], "Metric"], [7131, 7133, ["average", "precision"], "Metric"], [7134, 7135, ["AP"], "Metric"], [7156, 7163, ["Images", "-", "Fine", "-", "tuning", "-", "DLDL"], "Method"], [7164, 7167, ["IF", "-", "DLDL"], "Method"], [7169, 7176, ["Proposals", "-", "Fine", "-", "tuning", "-", "DLDL"], "Method"], [7177, 7180, ["PF", "-", "DLDL"], "Method"], [7182, 7183, ["ConvNets"], "Method"], [7218, 7221, ["VGG", "+", "SVM"], "Method"], [7236, 7238, ["penultimate", "layer"], "Method"], [7239, 7242, ["VGG", "-", "Nets"], "Method"], [7261, 7263, ["average", "pooling"], "Method"], [7274, 7278, ["Net", "-", "D", "”"], "Method"], [7295, 7296, ["stacking"], "Method"], [7311, 7314, ["linear", "SVM", "classifier"], "Method"], [7315, 7319, ["multi", "-", "label", "classification"], "Task"], [7320, 7321, ["HCP"], "Method"], [7324, 7325, ["HCP"], "Method"], [7329, 7335, ["multi", "-", "label", "object", "recognition", "task"], "Task"], [7356, 7360, ["pre", "-", "trained", "ConvNet"], "Method"], [7381, 7382, ["ConvNet"], "Method"], [7397, 7398, ["prediction"], "Task"], [7409, 7415, ["multi", "-", "label", "object", "recognition", "problem"], "Task"], [7417, 7425, ["multi", "-", "class", "multi", "-", "instance", "learning", "problem"], "Task"], [7464, 7469, ["IF", "-", "VGG", "-", "KL"], "Method"], [7475, 7478, ["VGG", "-", "Nets"], "Method"], [7479, 7481, ["square", "loss"], "Method"], [7482, 7489, ["multi", "-", "label", "cross", "-", "entropy", "loss"], "Method"], [7494, 7499, ["IF", "-", "DLDL", "’s", "baselines"], "Method"], [7543, 7545, ["label", "smoothing"], "Method"], [7555, 7558, ["IF", "-", "DLDL"], "Method"], [7601, 7604, ["Max", "-", "pooling"], "Method"], [7605, 7608, ["Avg", "-", "pooling"], "Method"], [7615, 7616, ["ConvNets"], "Method"], [7620, 7622, ["ConvNet", "models"], "Method"], [7627, 7631, ["Net", "-", "D", "”"], "Method"], [7640, 7642, ["prediction", "stage"], "Task"], [7663, 7667, ["fine", "-", "tuned", "ConvNets"], "Method"], [7693, 7696, ["PF", "-", "DLDL"], "Method"], [7704, 7708, ["IF", "-", "DLDL", "models"], "Method"], [7739, 7742, ["normalized", "cut", "algorithm"], "Method"], [7788, 7795, ["fine", "-", "tuned", "IF", "-", "DLDL", "model"], "Method"], [7797, 7799, ["prediction", "scores"], "Metric"], [7805, 7808, ["max", "-", "pooling"], "Method"], [7844, 7846, ["prediction", "stage"], "Task"], [7851, 7854, ["IF", "-", "DLDL"], "Method"], [7858, 7861, ["fuse", "prediction", "scores"], "Metric"], [7885, 7886, ["AP"], "Metric"], [7894, 7897, ["PF", "-", "DLDL"], "Method"], [7905, 7908, ["Fev", "+", "Lv"], "Method"], [7916, 7919, ["PF", "-", "DLDL"], "Method"], [7923, 7926, ["bounding", "box", "annotation"], "Method"], [7929, 7932, ["HCP", "-", "VGG"], "Method"], [7934, 7937, ["PF", "-", "DLDL"], "Method"], [7941, 7942, ["mAP"], "Metric"], [7992, 7995, ["IF", "-", "DLDL"], "Method"], [7996, 8000, ["IF", "-", "VGG", "-"], "Method"], [8015, 8018, ["KL", "loss", "function"], "Method"], [8034, 8037, ["IF", "-", "DLDL"], "Method"], [8038, 8043, ["IF", "-", "VGG", "-", "KL"], "Method"], [8048, 8049, ["mAP"], "Metric"], [8063, 8066, ["PF", "-", "DLDL"], "Method"], [8078, 8079, ["mAP"], "Metric"], [8082, 8085, ["fuse", "output", "scores"], "Metric"], [8087, 8091, ["PF", "-", "DLDL", "models"], "Method"], [8116, 8119, ["PF", "-", "DLDL"], "Method"], [8121, 8124, ["IF", "-", "DLDL"], "Method"], [8125, 8126, ["mAP"], "Metric"], [8127, 8130, ["PF", "-", "DLDL"], "Method"], [8132, 8134, ["computational", "cost"], "Metric"], [8135, 8138, ["IF", "-", "DLDL"], "Method"], [8140, 8144, ["training", "and", "testing", "stages"], "Task"], [8146, 8149, ["IF", "-", "DLDL"], "Method"], [8168, 8172, ["multi", "-", "label", "application"], "Task"], [8174, 8179, ["multi", "-", "label", "image", "retrieval"], "Task"], [8193, 8196, ["region", "proposal", "method"], "Method"], [8197, 8200, ["gated", "unit", "in"], "Method"], [8204, 8205, ["accuracy"], "Metric"], [8208, 8210, ["VOC", "tasks"], "Task"], [8213, 8214, ["accuracy"], "Metric"], [8224, 8225, ["ChaLearn"], "Material"], [8227, 8228, ["Morph"], "Material"], [8240, 8244, ["BFGS", "-", "LDL", "DLDL"], "Method"], [8246, 8248, ["Semantic", "segmentation"], "Task"], [8269, 8270, ["DLDL"], "Method"], [8327, 8328, ["FCN"], "Method"], [8331, 8332, ["DLDL"], "Method"], [8371, 8373, ["Evaluation", "criteria"], "Metric"], [8381, 8383, ["mean", "IU"], "Metric"], [8384, 8387, ["intersection", "over", "union"], "Metric"], [8397, 8399, ["semantic", "segmentation"], "Task"], [8406, 8407, ["FCN"], "Method"], [8453, 8455, ["training", "process"], "Task"], [8468, 8474, ["Conditional", "Random", "Field", "(", "CRF", ")"], "Method"], [8487, 8490, ["semantic", "segmentation", "systems"], "Task"], [8495, 8498, ["fully", "connected", "CRF"], "Method"], [8521, 8524, ["DLDL", "-", "8s"], "Method"], [8525, 8530, ["DLDL", "-", "8s", "-", "CRF"], "Method"], [8546, 8549, ["FCN", "-", "8s"], "Method"], [8550, 8553, ["DLDL", "-", "8s"], "Method"], [8555, 8557, ["mean", "IU"], "Metric"], [8573, 8576, ["DLDL", "-", "8s"], "Method"], [8584, 8586, ["mean", "IU"], "Metric"], [8587, 8588, ["DLDL"], "Method"], [8592, 8593, ["FCN"], "Method"], [8602, 8603, ["segmentation"], "Task"], [8612, 8613, ["CRF"], "Method"], [8617, 8620, ["DLDL", "-", "8s"], "Method"], [8628, 8630, ["mean", "IU"], "Metric"], [8643, 8645, ["semantic", "segmentation"], "Task"], [8652, 8655, ["FCN", "-", "8s"], "Method"], [8656, 8659, ["DLDL", "-", "8s"], "Method"], [8660, 8665, ["DLDL", "-", "8s", "-", "CRF"], "Method"], [8670, 8673, ["DLDL", "-", "8s"], "Method"], [8690, 8694, ["segmentation", "of", "object", "boundaries"], "Task"], [8707, 8710, ["FCN", "-", "8s"], "Method"], [8713, 8716, ["DLDL", "-", "8s"], "Method"], [8744, 8747, ["DLDL", "-", "8s"], "Method"], [8748, 8753, ["DLDL", "-", "8s", "-", "CRF"], "Method"], [8768, 8772, ["fine", "-", "grained", "segmentations"], "Task"], [8793, 8796, ["FCN", "-", "8s"], "Method"], [8798, 8801, ["DLDL", "-", "8s"], "Method"], [8803, 8808, ["DLDL", "-", "8s", "+", "CRF"], "Method"], [8815, 8816, ["ChaLearn"], "Material"], [8818, 8819, ["Morph"], "Material"], [8840, 8842, ["generalization", "performance"], "Metric"], [8843, 8844, ["DLDL"], "Method"], [8845, 8847, ["feature", "visualization"], "Method"], [8852, 8853, ["DLDL"], "Method"], [8856, 8857, ["accuracy"], "Metric"], [8876, 8878, ["Feature", "visualization"], "Task"], [8921, 8923, ["deep", "ConvNet"], "Method"], [8940, 8941, ["Morph"], "Material"], [8942, 8943, ["ChaLearn"], "Material"], [8968, 8974, ["dimension", "reduction", "algorithm", "t", "-", "SNE"], "Method"], [8976, 8980, ["low", "-", "dimensional", "embeddings"], "Method"], [9014, 9015, ["Morph"], "Material"], [9016, 9017, ["Chalearn"], "Material"], [9020, 9021, ["Pointing’04"], "Task"], [9086, 9089, ["over", "-", "fitting"], "Task"], [9090, 9091, ["DLDL"], "Method"], [9094, 9097, ["over", "-", "fitting"], "Task"], [9125, 9127, ["deep", "ConvNet"], "Method"], [9133, 9134, ["DLDL"], "Method"], [9150, 9153, ["over", "-", "fitting"], "Task"], [9167, 9168, ["DLDL"], "Method"], [9184, 9188, ["tradition", "ConvNet", "loss", "function"], "Method"], [9204, 9207, ["data", "augmentation", "techniques"], "Method"], [9209, 9211, ["random", "cropping"], "Method"], [9215, 9216, ["DLDL"], "Method"], [9230, 9231, ["MAE"], "Metric"], [9246, 9247, ["ChaLearn"], "Material"], [9248, 9249, ["Morph"], "Material"], [9264, 9265, ["ChaLearn"], "Material"], [9268, 9274, ["C", "-", "ConveNet", "(", "softmax", ")"], "Method"], [9299, 9300, ["ChaLearn"], "Material"], [9306, 9307, ["AFLW"], "Method"], [9312, 9315, ["over", "-", "fitting"], "Task"], [9317, 9320, ["C", "-", "ConvNet"], "Method"], [9356, 9358, ["Accelerate", "convergence"], "Task"], [9363, 9365, ["convergence", "performance"], "Metric"], [9366, 9367, ["DLDL"], "Method"], [9368, 9371, ["C", "-", "ConvNet"], "Method"], [9372, 9375, ["R", "-", "ConvNet"], "Method"], [9393, 9396, ["C", "-", "ConvNet"], "Method"], [9397, 9400, ["R", "-", "ConveNet"], "Method"], [9417, 9418, ["MAE"], "Metric"], [9419, 9420, ["DLDL"], "Method"], [9433, 9434, ["ConvNet"], "Method"], [9435, 9438, ["R", "-", "ConveNet"], "Method"], [9448, 9449, ["MAE"], "Metric"], [9450, 9453, ["pitch", "+", "yaw"], "Task"], [9459, 9461, ["joint", "distribution"], "Method"], [9482, 9485, ["R", "-", "ConvNet"], "Method"], [9501, 9502, ["DLDL"], "Method"], [9512, 9515, ["C", "-", "ConvNet"], "Method"], [9546, 9547, ["DLDL"], "Method"], [9557, 9560, ["C", "-", "ConvNet"], "Method"], [9561, 9564, ["R", "-", "ConvNet"], "Method"], [9571, 9572, ["DLDL"], "Method"], [9586, 9588, ["age", "estimation"], "Task"], [9589, 9590, ["Morph"], "Material"], [9592, 9595, ["head", "pose", "estimation"], "Task"], [9596, 9597, ["Pointing’04"], "Task"], [9626, 9627, ["Morph"], "Material"], [9645, 9646, ["DLDL"], "Method"], [9666, 9667, ["MAE"], "Metric"], [9669, 9670, ["Morph"], "Material"], [9671, 9672, ["Pointing’04"], "Task"], [9731, 9735, ["cross", "-", "validation", "strategy"], "Method"], [9743, 9745, ["deep", "ConvNets"], "Method"], [9768, 9769, ["DLDL"], "Method"], [9771, 9776, ["deep", "label", "distribution", "learning", "framework"], "Method"], [9786, 9787, ["DLDL"], "Method"], [9804, 9806, ["feature", "learning"], "Task"], [9807, 9809, ["classifier", "learning"], "Task"], [9810, 9811, ["DLDL"], "Method"], [9814, 9817, ["network", "training", "process"], "Task"], [9825, 9828, ["over", "-", "fitting"], "Task"], [9839, 9840, ["DLDL"], "Method"], [9847, 9852, ["classification", "or", "regression", "deep", "models"], "Method"], [9855, 9858, ["visual", "recognition", "tasks"], "Task"], [9878, 9880, ["recognition", "tasks"], "Task"], [9888, 9889, ["DLDL"], "Method"], [9891, 9893, ["recognition", "problems"], "Task"], [9949, 9951, ["computer", "vision"], "Task"], [9952, 9954, ["machine", "learning"], "Task"], [10000, 10002, ["pattern", "recognition"], "Task"], [10003, 10005, ["machine", "learning"], "Task"], [10007, 10009, ["data", "mining"], "Task"], [10054, 10056, ["computer", "vision"], "Task"], [10057, 10059, ["machine", "learning"], "Task"], [10147, 10149, ["computer", "vision"], "Task"], [10150, 10152, ["machine", "learning"], "Task"], [10282, 10284, ["pattern", "recognition"], "Task"], [10285, 10287, ["machine", "learning"], "Task"], [10289, 10291, ["computer", "vision"], "Task"], [271, 272, ["detection"], "Task"], [428, 429, ["age"], "Task"], [463, 464, ["age"], "Task"], [478, 479, ["age"], "Task"], [519, 520, ["age"], "Task"], [582, 583, ["age"], "Task"], [737, 738, ["estimation"], "Task"], [928, 929, ["LDL"], "Method"], [938, 939, ["age"], "Task"], [1092, 1093, ["detection"], "Task"], [1342, 1343, ["age"], "Task"], [1420, 1421, ["DLDL"], "Method"], [1425, 1426, ["DLDL"], "Method"], [1429, 1430, ["DLDL"], "Method"], [1497, 1498, ["detection"], "Task"], [1512, 1513, ["estimation"], "Task"], [1717, 1718, ["detection"], "Task"], [1762, 1763, ["detection"], "Task"], [2015, 2016, ["DLDL"], "Method"], [2028, 2029, ["DLDL"], "Method"], [2036, 2037, ["DLDL"], "Method"], [2060, 2061, ["DLDL"], "Method"], [2090, 2091, ["age"], "Task"], [2161, 2162, ["age"], "Task"], [2210, 2211, ["DLDL"], "Method"], [2723, 2724, ["DLDL"], "Method"], [2914, 2915, ["age"], "Task"], [2932, 2933, ["age"], "Task"], [3625, 3626, ["DLDL"], "Method"], [3644, 3645, ["estimation"], "Task"], [4101, 4102, ["estimation"], "Task"], [4127, 4128, ["age"], "Task"], [4179, 4180, ["age"], "Task"], [4180, 4181, ["estimation"], "Task"], [4272, 4273, ["age"], "Task"], [4294, 4295, ["age"], "Task"], [4310, 4311, ["age"], "Task"], [4344, 4345, ["ChaLearn"], "Material"], [4374, 4375, ["LD"], "Method"], [4434, 4435, ["LDL"], "Method"], [4478, 4479, ["LDL"], "Method"], [4564, 4565, ["age"], "Task"], [4660, 4661, ["detection"], "Task"], [4665, 4666, ["detection"], "Task"], [4875, 4876, ["Morph"], "Material"], [4953, 4954, ["age"], "Task"], [4964, 4965, ["age"], "Task"], [4983, 4984, ["estimation"], "Task"], [4991, 4992, ["estimation"], "Task"], [5020, 5021, ["ChaLearn"], "Material"], [5301, 5302, ["LDL"], "Method"], [5322, 5323, ["LD"], "Method"], [5328, 5329, ["DLDL"], "Method"], [5454, 5455, ["DLDL"], "Method"], [5479, 5480, ["ChaLearn"], "Material"], [5483, 5484, ["MAE"], "Metric"], [5485, 5486, ["DLDL"], "Method"], [5539, 5540, ["age"], "Task"], [5549, 5550, ["ChaLearn"], "Material"], [5591, 5592, ["DLDL"], "Method"], [5608, 5609, ["DLDL"], "Method"], [5610, 5611, ["estimation"], "Task"], [5640, 5641, ["DLDL"], "Method"], [5685, 5686, ["DLDL"], "Method"], [5742, 5743, ["ChaLearn"], "Material"], [5770, 5771, ["age"], "Task"], [6157, 6158, ["LDL"], "Method"], [6674, 6675, ["LDL"], "Method"], [6823, 6824, ["LDL"], "Method"], [6979, 6980, ["DLDL"], "Method"], [7141, 7142, ["mAP"], "Metric"], [8216, 8217, ["DLDL"], "Method"], [9032, 9033, ["DLDL"], "Method"], [9057, 9058, ["age"], "Task"], [9243, 9244, ["age"], "Task"], [9278, 9279, ["MAE"], "Metric"], [9285, 9286, ["MAE"], "Metric"], [9292, 9293, ["MAE"], "Metric"], [9382, 9383, ["MAE"], "Metric"], [9456, 9457, ["estimation"], "Task"], [9685, 9686, ["MAE"], "Metric"]]} {"doc_id": "02b3d1d162080d9aefd3fc30a0bcc9a843073b5d", "text": ["In", "this", "work", "we", "explore", "recent", "advances", "in", "Recurrent", "Neural", "Networks", "for", "large", "scale", "Language", "Modeling", ",", "a", "task", "central", "to", "language", "understanding", ".", "We", "extend", "current", "models", "to", "deal", "with", "two", "key", "challenges", "present", "in", "this", "task", ":", "corpora", "and", "vocabulary", "sizes", ",", "and", "complex", ",", "long", "term", "structure", "of", "language", ".", "We", "perform", "an", "exhaustive", "study", "on", "techniques", "such", "as", "character", "Convolutional", "Neural", "Networks", "or", "Long", "-", "Short", "Term", "Memory", ",", "on", "the", "One", "Billion", "Word", "Benchmark", ".", "Our", "best", "single", "model", "significantly", "improves", "state", "-", "of", "-", "the", "-", "art", "perplexity", "from", "51.3", "down", "to", "30.0", "(", "whilst", "reducing", "the", "number", "of", "parameters", "by", "a", "factor", "of", "20", ")", ",", "while", "an", "ensemble", "of", "models", "sets", "a", "new", "record", "by", "improving", "perplexity", "from", "41.0", "down", "to", "23.7", ".", "We", "also", "release", "these", "models", "for", "the", "NLP", "and", "ML", "community", "to", "study", "and", "improve", "upon", ".", "ExploringtheLimitsofLanguageModeling", "section", ":", "Introduction", "Language", "Modeling", "(", "LM", ")", "is", "a", "task", "central", "to", "Natural", "Language", "Processing", "(", "NLP", ")", "and", "Language", "Understanding", ".", "Models", "which", "can", "accurately", "place", "distributions", "over", "sentences", "not", "only", "encode", "complexities", "of", "language", "such", "as", "grammatical", "structure", ",", "but", "also", "distill", "a", "fair", "amount", "of", "information", "about", "the", "knowledge", "that", "a", "corpora", "may", "contain", ".", "Indeed", ",", "models", "that", "are", "able", "to", "assign", "a", "low", "probability", "to", "sentences", "that", "are", "grammatically", "correct", "but", "unlikely", "may", "help", "other", "tasks", "in", "fundamental", "language", "understanding", "like", "question", "answering", ",", "machine", "translation", ",", "or", "text", "summarization", ".", "LMs", "have", "played", "a", "key", "role", "in", "traditional", "NLP", "tasks", "such", "as", "speech", "recognition", ",", "machine", "translation", ",", "or", "text", "summarization", ".", "Often", "(", "although", "not", "always", ")", ",", "training", "better", "language", "models", "improves", "the", "underlying", "metrics", "of", "the", "downstream", "task", "(", "such", "as", "word", "error", "rate", "for", "speech", "recognition", ",", "or", "BLEU", "score", "for", "translation", ")", ",", "which", "makes", "the", "task", "of", "training", "better", "LMs", "valuable", "by", "itself", ".", "Further", ",", "when", "trained", "on", "vast", "amounts", "of", "data", ",", "language", "models", "compactly", "extract", "knowledge", "encoded", "in", "the", "training", "data", ".", "For", "example", ",", "when", "trained", "on", "movie", "subtitles", ",", "these", "language", "models", "are", "able", "to", "generate", "basic", "answers", "to", "questions", "about", "object", "colors", ",", "facts", "about", "people", ",", "etc", ".", "Lastly", ",", "recently", "proposed", "sequence", "-", "to", "-", "sequence", "models", "employ", "conditional", "language", "models", "as", "their", "key", "component", "to", "solve", "diverse", "tasks", "like", "machine", "translation", "or", "video", "generation", ".", "Deep", "Learning", "and", "Recurrent", "Neural", "Networks", "(", "RNNs", ")", "have", "fueled", "language", "modeling", "research", "in", "the", "past", "years", "as", "it", "allowed", "researchers", "to", "explore", "many", "tasks", "for", "which", "the", "strong", "conditional", "independence", "assumptions", "are", "unrealistic", ".", "Despite", "the", "fact", "that", "simpler", "models", ",", "such", "as", "N", "-", "grams", ",", "only", "use", "a", "short", "history", "of", "previous", "words", "to", "predict", "the", "next", "word", ",", "they", "are", "still", "a", "key", "component", "to", "high", "quality", ",", "low", "perplexity", "LMs", ".", "Indeed", ",", "most", "recent", "work", "on", "large", "scale", "LM", "has", "shown", "that", "RNNs", "are", "great", "in", "combination", "with", "N", "-", "grams", ",", "as", "they", "may", "have", "different", "strengths", "that", "complement", "N", "-", "gram", "models", ",", "but", "worse", "when", "considered", "in", "isolation", ".", "We", "believe", "that", ",", "despite", "much", "work", "being", "devoted", "to", "small", "data", "sets", "like", "the", "Penn", "Tree", "Bank", "(", "PTB", ")", ",", "research", "on", "larger", "tasks", "is", "very", "relevant", "as", "overfitting", "is", "not", "the", "main", "limitation", "in", "current", "language", "modeling", ",", "but", "is", "the", "main", "characteristic", "of", "the", "PTB", "task", ".", "Results", "on", "larger", "corpora", "usually", "show", "better", "what", "matters", "as", "many", "ideas", "work", "well", "on", "small", "data", "sets", "but", "fail", "to", "improve", "on", "larger", "data", "sets", ".", "Further", ",", "given", "current", "hardware", "trends", "and", "vast", "amounts", "of", "text", "available", "on", "the", "Web", ",", "it", "is", "much", "more", "straightforward", "to", "tackle", "large", "scale", "modeling", "than", "it", "used", "to", "be", ".", "Thus", ",", "we", "hope", "that", "our", "work", "will", "help", "and", "motivate", "researchers", "to", "work", "on", "traditional", "LM", "beyond", "PTB", "–", "for", "this", "purpose", ",", "we", "will", "open", "-", "source", "our", "models", "and", "training", "recipes", ".", "We", "focused", "on", "a", "well", "known", ",", "large", "scale", "LM", "benchmark", ":", "the", "One", "Billion", "Word", "Benchmark", "data", "set", ".", "This", "data", "set", "is", "much", "larger", "than", "PTB", "(", "one", "thousand", "fold", ",", "800k", "word", "vocabulary", "and", "1B", "words", "training", "data", ")", "and", "far", "more", "challenging", ".", "Similar", "to", "Imagenet", ",", "which", "helped", "advance", "computer", "vision", ",", "we", "believe", "that", "releasing", "and", "working", "on", "large", "data", "sets", "and", "models", "with", "clear", "benchmarks", "will", "help", "advance", "Language", "Modeling", ".", "The", "contributions", "of", "our", "work", "are", "as", "follows", ":", "We", "explored", ",", "extended", "and", "tried", "to", "unify", "some", "of", "the", "current", "research", "on", "large", "scale", "LM", ".", "Specifically", ",", "we", "designed", "a", "Softmax", "loss", "which", "is", "based", "on", "character", "level", "CNNs", ",", "is", "efficient", "to", "train", ",", "and", "is", "as", "precise", "as", "a", "full", "Softmax", "which", "has", "orders", "of", "magnitude", "more", "parameters", ".", "Our", "study", "yielded", "significant", "improvements", "to", "the", "state", "-", "of", "-", "the", "-", "art", "on", "a", "well", "known", ",", "large", "scale", "LM", "task", ":", "from", "51.3", "down", "to", "30.0", "perplexity", "for", "single", "models", "whilst", "reducing", "the", "number", "of", "parameters", "by", "a", "factor", "of", "20", ".", "We", "show", "that", "an", "ensemble", "of", "a", "number", "of", "different", "models", "can", "bring", "down", "perplexity", "on", "this", "task", "to", "23.7", ",", "a", "large", "improvement", "compared", "to", "current", "state", "-", "of", "-", "art", ".", "We", "share", "the", "model", "and", "recipes", "in", "order", "to", "help", "and", "motivate", "further", "research", "in", "this", "area", ".", "In", "Section", "[", "reference", "]", "we", "review", "important", "concepts", "and", "previous", "work", "on", "language", "modeling", ".", "Section", "[", "reference", "]", "presents", "our", "contributions", "to", "the", "field", "of", "neural", "language", "modeling", ",", "emphasizing", "large", "scale", "recurrent", "neural", "network", "training", ".", "Sections", "[", "reference", "]", "and", "[", "reference", "]", "aim", "at", "exhaustively", "describing", "our", "experience", "and", "understanding", "throughout", "the", "project", ",", "as", "well", "as", "emplacing", "our", "work", "relative", "to", "other", "known", "approaches", ".", "section", ":", "Related", "Work", "In", "this", "section", "we", "describe", "previous", "work", "relevant", "to", "the", "approaches", "discussed", "in", "this", "paper", ".", "A", "more", "detailed", "discussion", "on", "language", "modeling", "research", "is", "provided", "in", ".", "subsection", ":", "Language", "Models", "Language", "Modeling", "(", "LM", ")", "has", "been", "a", "central", "task", "in", "NLP", ".", "The", "goal", "of", "LM", "is", "to", "learn", "a", "probability", "distribution", "over", "sequences", "of", "symbols", "pertaining", "to", "a", "language", ".", "Much", "work", "has", "been", "done", "on", "both", "parametric", "(", "e.g.", ",", "log", "-", "linear", "models", ")", "and", "non", "-", "parametric", "approaches", "(", "e.g.", ",", "count", "-", "based", "LMs", ")", ".", "Count", "-", "based", "approaches", "(", "based", "on", "statistics", "of", "N", "-", "grams", ")", "typically", "add", "smoothing", "which", "account", "for", "unseen", "(", "yet", "possible", ")", "sequences", ",", "and", "have", "been", "quite", "successful", ".", "To", "this", "extent", ",", "Kneser", "-", "Ney", "smoothed", "5", "-", "gram", "models", "are", "a", "fairly", "strong", "baseline", "which", ",", "for", "large", "amounts", "of", "training", "data", ",", "have", "challenged", "other", "parametric", "approaches", "based", "on", "Neural", "Networks", ".", "Most", "of", "our", "work", "is", "based", "on", "Recurrent", "Neural", "Networks", "(", "RNN", ")", "models", "which", "retain", "long", "term", "dependencies", ".", "To", "this", "extent", ",", "we", "used", "the", "Long", "-", "Short", "Term", "Memory", "model", "which", "uses", "a", "gating", "mechanism", "to", "ensure", "proper", "propagation", "of", "information", "through", "many", "time", "steps", ".", "Much", "work", "has", "been", "done", "on", "small", "and", "large", "scale", "RNN", "-", "based", "LMs", ".", "The", "architectures", "that", "we", "considered", "in", "this", "paper", "are", "represented", "in", "Figure", "[", "reference", "]", ".", "In", "our", "work", ",", "we", "train", "models", "on", "the", "popular", "One", "Billion", "Word", "Benchmark", ",", "which", "can", "be", "considered", "to", "be", "a", "medium", "-", "sized", "data", "set", "for", "count", "-", "based", "LMs", "but", "a", "very", "large", "data", "set", "for", "NN", "-", "based", "LMs", ".", "This", "regime", "is", "most", "interesting", "to", "us", "as", "we", "believe", "learning", "a", "very", "good", "model", "of", "human", "language", "is", "a", "complex", "task", "which", "will", "require", "large", "models", ",", "and", "thus", "large", "amounts", "of", "data", ".", "Further", "advances", "in", "data", "availability", "and", "computational", "resources", "helped", "our", "study", ".", "We", "argue", "this", "leap", "in", "scale", "enabled", "tremendous", "advances", "in", "deep", "learning", ".", "A", "clear", "example", "found", "in", "computer", "vision", "is", "Imagenet", ",", "which", "enabled", "learning", "complex", "vision", "models", "from", "large", "amounts", "of", "data", ".", "A", "crucial", "aspect", "which", "we", "discuss", "in", "detail", "in", "later", "sections", "is", "the", "size", "of", "our", "models", ".", "Despite", "the", "large", "number", "of", "parameters", ",", "we", "try", "to", "minimize", "computation", "as", "much", "as", "possible", "by", "adopting", "a", "strategy", "proposed", "in", "of", "projecting", "a", "relatively", "big", "recurrent", "state", "space", "down", "so", "that", "the", "matrices", "involved", "remain", "relatively", "small", ",", "yet", "the", "model", "has", "large", "memory", "capacity", ".", "subsection", ":", "Convolutional", "Embedding", "Models", "There", "is", "an", "increased", "interest", "in", "incorporating", "character", "-", "level", "inputs", "to", "build", "word", "embeddings", "for", "various", "NLP", "problems", ",", "including", "part", "-", "of", "-", "speech", "tagging", ",", "parsing", "and", "language", "modeling", ".", "The", "additional", "character", "information", "has", "been", "shown", "useful", "on", "relatively", "small", "benchmark", "data", "sets", ".", "The", "approach", "proposed", "in", "builds", "word", "embeddings", "using", "bidirectional", "LSTMs", "over", "the", "characters", ".", "The", "recurrent", "networks", "process", "sequences", "of", "characters", "from", "both", "sides", "and", "their", "final", "state", "vectors", "are", "concatenated", ".", "The", "resulting", "representation", "is", "then", "fed", "to", "a", "Neural", "Network", ".", "This", "model", "achieved", "very", "good", "results", "on", "a", "part", "-", "of", "-", "speech", "tagging", "task", ".", "In", ",", "the", "words", "characters", "are", "processed", "by", "a", "1", "-", "d", "CNN", "with", "max", "-", "pooling", "across", "the", "sequence", "for", "each", "convolutional", "feature", ".", "The", "resulting", "features", "are", "fed", "to", "a", "2", "-", "layer", "highway", "network", ",", "which", "allows", "the", "embedding", "to", "learn", "semantic", "representations", ".", "The", "model", "was", "evaluated", "on", "small", "-", "scale", "language", "modeling", "experiments", "for", "various", "languages", "and", "matched", "the", "best", "results", "on", "the", "PTB", "data", "set", "despite", "having", "60", "%", "fewer", "parameters", ".", "subsection", ":", "Softmax", "Over", "Large", "Vocabularies", "Assigning", "probability", "distributions", "over", "large", "vocabularies", "is", "computationally", "challenging", ".", "For", "modeling", "language", ",", "maximizing", "log", "-", "likelihood", "of", "a", "given", "word", "sequence", "leads", "to", "optimizing", "cross", "-", "entropy", "between", "the", "target", "probability", "distribution", "(", "e.g.", ",", "the", "target", "word", "we", "should", "be", "predicting", ")", ",", "and", "our", "model", "predictions", ".", "Generally", ",", "predictions", "come", "from", "a", "linear", "layer", "followed", "by", "a", "Softmax", "non", "-", "linearity", ":", "where", "is", "the", "logit", "corresponding", "to", "a", "word", ".", "The", "logit", "is", "generally", "computed", "as", "an", "inner", "product", "where", "is", "a", "context", "vector", "and", "is", "a", "“", "word", "embedding", "”", "for", ".", "The", "main", "challenge", "when", "is", "very", "large", "(", "in", "the", "order", "of", "one", "million", "in", "this", "paper", ")", "is", "the", "fact", "that", "computing", "all", "inner", "products", "between", "and", "all", "embeddings", "becomes", "prohibitively", "slow", "during", "training", "(", "even", "when", "exploiting", "matrix", "-", "matrix", "multiplications", "and", "modern", "GPUs", ")", ".", "Several", "approaches", "have", "been", "proposed", "to", "cope", "with", "the", "scaling", "issue", ":", "importance", "sampling", ",", "Noise", "Contrastive", "Estimation", "(", "NCE", ")", ",", "self", "normalizing", "partition", "functions", "or", "Hierarchical", "Softmax", "–", "they", "all", "offer", "good", "solutions", "to", "this", "problem", ".", "We", "found", "importance", "sampling", "to", "be", "quite", "effective", "on", "this", "task", ",", "and", "explain", "the", "connection", "between", "it", "and", "NCE", "in", "the", "following", "section", ",", "as", "they", "are", "closely", "related", ".", "section", ":", "Language", "Modeling", "Improvements", "Recurrent", "Neural", "Networks", "based", "LMs", "employ", "the", "chain", "rule", "to", "model", "joint", "probabilities", "over", "word", "sequences", ":", "where", "the", "context", "of", "all", "previous", "words", "is", "encoded", "with", "an", "LSTM", ",", "and", "the", "probability", "over", "words", "uses", "a", "Softmax", "(", "see", "Figure", "[", "reference", "]", "(", "a", ")", ")", ".", "subsection", ":", "Relationship", "between", "Noise", "Contrastive", "Estimation", "and", "Importance", "Sampling", "As", "discussed", "in", "Section", "[", "reference", "]", ",", "a", "large", "scale", "Softmax", "is", "necessary", "for", "training", "good", "LMs", "because", "of", "the", "vocabulary", "size", ".", "A", "Hierarchical", "Softmax", "employs", "a", "tree", "in", "which", "the", "probability", "distribution", "over", "words", "is", "decomposed", "into", "a", "product", "of", "two", "probabilities", "for", "each", "word", ",", "greatly", "reducing", "training", "and", "inference", "time", "as", "only", "the", "path", "specified", "by", "the", "hierarchy", "needs", "to", "be", "computed", "and", "updated", ".", "Choosing", "a", "good", "hierarchy", "is", "important", "for", "obtaining", "good", "results", "and", "we", "did", "not", "explore", "this", "approach", "further", "for", "this", "paper", "as", "sampling", "methods", "worked", "well", "for", "our", "setup", ".", "Sampling", "approaches", "are", "only", "useful", "during", "training", ",", "as", "they", "propose", "an", "approximation", "to", "the", "loss", "which", "is", "cheap", "to", "compute", "(", "also", "in", "a", "distributed", "setting", ")", "–", "however", ",", "at", "inference", "time", "one", "still", "has", "to", "compute", "the", "normalization", "term", "over", "all", "words", ".", "Noise", "Contrastive", "Estimation", "(", "NCE", ")", "proposes", "to", "consider", "a", "surrogate", "binary", "classification", "task", "in", "which", "a", "classifier", "is", "trained", "to", "discriminate", "between", "true", "data", ",", "or", "samples", "coming", "from", "some", "arbitrary", "distribution", ".", "If", "both", "the", "noise", "and", "data", "distributions", "were", "known", ",", "the", "optimal", "classifier", "would", "be", ":", "where", "is", "the", "binary", "random", "variable", "indicating", "whether", "comes", "from", "the", "true", "data", "distribution", ",", "is", "the", "number", "of", "negative", "samples", "per", "positive", "word", ",", "and", "and", "are", "the", "data", "and", "noise", "distribution", "respectively", "(", "we", "dropped", "any", "dependency", "on", "previous", "words", "for", "notational", "simplicity", ")", ".", "It", "is", "easy", "to", "show", "that", "if", "we", "train", "a", "logistic", "classifier", "where", "is", "the", "logistic", "function", ",", "then", ",", "is", "a", "good", "approximation", "of", "(", "is", "a", "logit", "which", "e.g.", "an", "LSTM", "LM", "computes", ")", ".", "The", "other", "technique", ",", "which", "is", "based", "on", "importance", "sampling", "(", "IS", ")", ",", "proposes", "to", "directly", "approximate", "the", "partition", "function", "(", "which", "comprises", "a", "sum", "over", "all", "words", ")", "with", "an", "estimate", "of", "it", "through", "importance", "sampling", ".", "Though", "the", "methods", "look", "superficially", "similar", ",", "we", "will", "derive", "a", "similar", "surrogate", "classification", "task", "akin", "to", "NCE", "which", "arrives", "at", "IS", ",", "showing", "a", "strong", "connection", "between", "the", "two", ".", "Suppose", "that", ",", "instead", "of", "having", "a", "binary", "task", "to", "decide", "if", "a", "word", "comes", "from", "the", "data", "or", "from", "the", "noise", "distribution", ",", "we", "want", "to", "identify", "the", "words", "coming", "from", "the", "true", "data", "distribution", "in", "a", "set", ",", "comprised", "of", "noise", "samples", "and", "one", "data", "distribution", "sample", ".", "Thus", ",", "we", "can", "train", "a", "multiclass", "loss", "over", "a", "multinomial", "random", "variable", "which", "maximizes", ",", "assuming", "w.l.o.g", ".", "that", "is", "always", "the", "word", "coming", "from", "true", "data", ".", "By", "Bayes", "rule", ",", "and", "ignoring", "terms", "that", "are", "constant", "with", "respect", "to", ",", "we", "can", "write", ":", "and", ",", "following", "a", "similar", "argument", "than", "for", "NCE", ",", "if", "we", "define", "then", "is", "a", "good", "approximation", "of", ".", "Note", "that", "the", "only", "difference", "between", "NCE", "and", "IS", "is", "that", ",", "in", "NCE", ",", "we", "define", "a", "binary", "classification", "task", "between", "true", "or", "noise", "words", "with", "a", "logistic", "loss", ",", "whereas", "in", "IS", "we", "define", "a", "multiclass", "classification", "problem", "with", "a", "Softmax", "and", "cross", "entropy", "loss", ".", "We", "hope", "that", "our", "derivation", "helps", "clarify", "the", "similarities", "and", "differences", "between", "the", "two", ".", "In", "particular", ",", "we", "observe", "that", "IS", ",", "as", "it", "optimizes", "a", "multiclass", "classification", "task", "(", "in", "contrast", "to", "solving", "a", "binary", "task", ")", ",", "may", "be", "a", "better", "choice", ".", "Indeed", ",", "the", "updates", "to", "the", "logits", "with", "IS", "are", "tied", "whereas", "in", "NCE", "they", "are", "independent", ".", "subsection", ":", "CNN", "Softmax", "The", "character", "-", "level", "features", "allow", "for", "a", "smoother", "and", "compact", "parametrization", "of", "the", "word", "embeddings", ".", "Recent", "efforts", "on", "small", "scale", "language", "modeling", "have", "used", "CNN", "character", "embeddings", "for", "the", "input", "embeddings", ".", "Although", "not", "as", "straightforward", ",", "we", "propose", "an", "extension", "to", "this", "idea", "to", "also", "reduce", "the", "number", "of", "parameters", "of", "the", "Softmax", "layer", ".", "Recall", "from", "Section", "[", "reference", "]", "that", "the", "Softmax", "computes", "a", "logit", "as", "where", "is", "a", "context", "vector", "and", "the", "word", "embedding", ".", "Instead", "of", "building", "a", "matrix", "of", "(", "whose", "rows", "correspond", "to", ")", ",", "we", "produce", "with", "a", "CNN", "over", "the", "characters", "of", "as", "–", "we", "call", "this", "a", "CNN", "Softmax", ".", "We", "used", "the", "same", "network", "architecture", "to", "dynamically", "generate", "the", "Softmax", "word", "embeddings", "without", "sharing", "the", "parameters", "with", "the", "input", "word", "-", "embedding", "sub", "-", "network", ".", "For", "inference", ",", "the", "vectors", "can", "be", "precomputed", ",", "so", "there", "is", "no", "computational", "complexity", "increase", "w.r.t", ".", "the", "regular", "Softmax", ".", "We", "note", "that", ",", "when", "using", "an", "importance", "sampling", "loss", "such", "as", "the", "one", "described", "in", "Section", "[", "reference", "]", ",", "only", "a", "few", "logits", "have", "non", "-", "zero", "gradient", "(", "those", "corresponding", "to", "the", "true", "and", "sampled", "words", ")", ".", "With", "a", "Softmax", "where", "are", "independently", "learned", "word", "embeddings", ",", "this", "is", "not", "a", "problem", ".", "But", "we", "observed", "that", ",", "when", "using", "a", "CNN", ",", "all", "the", "logits", "become", "tied", "as", "the", "function", "mapping", "from", "to", "is", "quite", "smooth", ".", "As", "a", "result", ",", "a", "much", "smaller", "learning", "rate", "had", "to", "be", "used", ".", "Even", "with", "this", ",", "the", "model", "lacks", "capacity", "to", "differentiate", "between", "words", "that", "have", "very", "different", "meanings", "but", "that", "are", "spelled", "similarly", ".", "Thus", ",", "a", "reasonable", "compromise", "was", "to", "add", "a", "small", "correction", "factor", "which", "is", "learned", "per", "word", ",", "such", "that", ":", "where", "is", "a", "matrix", "projecting", "a", "low", "-", "dimensional", "embedding", "vector", "back", "up", "to", "the", "dimensionality", "of", "the", "projected", "LSTM", "hidden", "state", "of", ".", "This", "amounts", "to", "adding", "a", "bottleneck", "linear", "layer", ",", "and", "brings", "the", "CNN", "Softmax", "much", "closer", "to", "our", "best", "result", ",", "as", "can", "be", "seen", "in", "Table", "[", "reference", "]", ",", "where", "adding", "a", "128", "-", "dim", "correction", "halves", "the", "gap", "between", "regular", "and", "the", "CNN", "Softmax", ".", "Aside", "from", "a", "big", "reduction", "in", "the", "number", "of", "parameters", "and", "incorporating", "morphological", "knowledge", "from", "words", ",", "the", "other", "benefit", "of", "this", "approach", "is", "that", "out", "-", "of", "-", "vocabulary", "(", "OOV", ")", "words", "can", "easily", "be", "scored", ".", "This", "may", "be", "useful", "for", "other", "problems", "such", "as", "Machine", "Translation", "where", "handling", "out", "-", "of", "-", "vocabulary", "words", "is", "very", "important", ".", "This", "approach", "also", "allows", "parallel", "training", "over", "various", "data", "sets", "since", "the", "model", "is", "no", "longer", "explicitly", "parametrized", "by", "the", "vocabulary", "size", "–", "or", "the", "language", ".", "This", "has", "shown", "to", "help", "when", "using", "byte", "-", "level", "input", "embeddings", "for", "named", "entity", "recognition", ",", "and", "we", "hope", "it", "will", "enable", "similar", "gains", "when", "used", "to", "map", "onto", "words", ".", "subsection", ":", "Char", "LSTM", "Predictions", "The", "CNN", "Softmax", "layer", "can", "handle", "arbitrary", "words", "and", "is", "much", "more", "efficient", "in", "terms", "of", "number", "of", "parameters", "than", "the", "full", "Softmax", "matrix", ".", "It", "is", ",", "though", ",", "still", "considerably", "slow", ",", "as", "to", "evaluate", "perplexities", "we", "need", "to", "compute", "the", "partition", "function", ".", "A", "class", "of", "models", "that", "solve", "this", "problem", "more", "efficiently", "are", "character", "-", "level", "LSTMs", ".", "They", "make", "predictions", "one", "character", "at", "a", "time", ",", "thus", "allowing", "to", "compute", "probabilities", "over", "a", "much", "smaller", "vocabulary", ".", "On", "the", "other", "hand", ",", "these", "models", "are", "more", "difficult", "to", "train", "and", "seem", "to", "perform", "worse", "even", "in", "small", "tasks", "like", "PTB", ".", "Most", "likely", "this", "is", "due", "to", "the", "sequences", "becoming", "much", "longer", "on", "average", "as", "the", "LSTM", "reads", "the", "input", "character", "by", "character", "instead", "of", "word", "by", "word", ".", "Thus", ",", "we", "combine", "the", "word", "and", "character", "-", "level", "models", "by", "feeding", "a", "word", "-", "level", "LSTM", "hidden", "state", "into", "a", "small", "LSTM", "that", "predicts", "the", "target", "word", "one", "character", "at", "a", "time", "(", "see", "Figure", "[", "reference", "]", "(", "c", ")", ")", ".", "In", "order", "to", "make", "the", "whole", "process", "reasonably", "efficient", ",", "we", "train", "the", "standard", "LSTM", "model", "until", "convergence", ",", "freeze", "its", "weights", ",", "and", "replace", "the", "standard", "word", "-", "level", "Softmax", "layer", "with", "the", "aforementioned", "character", "-", "level", "LSTM", ".", "The", "resulting", "model", "scales", "independently", "of", "vocabulary", "size", "–", "both", "for", "training", "and", "inference", ".", "However", ",", "it", "does", "seem", "to", "be", "worse", "than", "regular", "and", "CNN", "Softmax", "–", "we", "are", "hopeful", "that", "further", "research", "will", "enable", "these", "models", "to", "replace", "fixed", "vocabulary", "models", "whilst", "being", "computationally", "attractive", ".", "section", ":", "Experiments", "All", "experiments", "were", "run", "using", "the", "TensorFlow", "system", ",", "with", "the", "exception", "of", "some", "older", "models", "which", "were", "used", "in", "the", "ensemble", ".", "subsection", ":", "Data", "Set", "The", "experiments", "are", "performed", "on", "the", "1B", "Word", "Benchmark", "data", "set", "introduced", "by", ",", "which", "is", "a", "publicly", "available", "benchmark", "for", "measuring", "progress", "of", "statistical", "language", "modeling", ".", "The", "data", "set", "contains", "about", "0.8B", "words", "with", "a", "vocabulary", "of", "793471", "words", ",", "including", "sentence", "boundary", "markers", ".", "All", "the", "sentences", "are", "shuffled", "and", "the", "duplicates", "are", "removed", ".", "The", "words", "that", "are", "out", "of", "vocabulary", "(", "OOV", ")", "are", "marked", "with", "a", "special", "UNK", "token", "(", "there", "are", "approximately", "0.3", "%", "such", "words", ")", ".", "subsection", ":", "Model", "Setup", "The", "typical", "measure", "used", "for", "reporting", "progress", "in", "language", "modeling", "is", "perplexity", ",", "which", "is", "the", "average", "per", "-", "word", "log", "-", "probability", "on", "the", "holdout", "data", "set", ":", ".", "We", "follow", "the", "standard", "procedure", "and", "sum", "over", "all", "the", "words", "(", "including", "the", "end", "of", "sentence", "symbol", ")", ".", "We", "used", "the", "1B", "Word", "Benchmark", "data", "set", "without", "any", "pre", "-", "processing", ".", "Given", "the", "shuffled", "sentences", ",", "they", "are", "input", "to", "the", "network", "as", "a", "batch", "of", "independent", "streams", "of", "words", ".", "Whenever", "a", "sentence", "ends", ",", "a", "new", "one", "starts", "without", "any", "padding", "(", "thus", "maximizing", "the", "occupancy", "per", "batch", ")", ".", "For", "the", "models", "that", "consume", "characters", "as", "inputs", "or", "as", "targets", ",", "each", "word", "is", "fed", "to", "the", "model", "as", "a", "sequence", "of", "character", "IDs", "of", "preespecified", "length", "(", "see", "Figure", "[", "reference", "]", "(", "b", ")", ")", ".", "The", "words", "were", "processed", "to", "include", "special", "begin", "and", "end", "of", "word", "tokens", "and", "were", "padded", "to", "reach", "the", "expected", "length", ".", "I.e.", "if", "the", "maximum", "word", "length", "was", "10", ",", "the", "word", "“", "cat", "”", "would", "be", "transformed", "to", "“", "$", "cat^", "”", "due", "to", "the", "CNN", "model", ".", "In", "our", "experiments", "we", "found", "that", "limiting", "the", "maximum", "word", "length", "in", "training", "to", "50", "was", "sufficient", "to", "reach", "very", "good", "results", "while", "32", "was", "clearly", "insufficient", ".", "We", "used", "256", "characters", "in", "our", "vocabulary", "and", "the", "non", "-", "ascii", "symbols", "were", "represented", "as", "a", "sequence", "of", "bytes", ".", "subsection", ":", "Model", "Architecture", "We", "evaluated", "many", "variations", "of", "RNN", "LM", "architectures", ".", "These", "include", "the", "dimensionalities", "of", "the", "embedding", "layers", ",", "the", "state", ",", "projection", "sizes", ",", "and", "number", "of", "LSTM", "layers", "to", "use", ".", "Exhaustively", "trying", "all", "combinations", "would", "be", "extremely", "time", "consuming", "for", "such", "a", "large", "data", "set", ",", "but", "our", "findings", "suggest", "that", "LSTMs", "with", "a", "projection", "layer", "(", "i.e.", ",", "a", "bottleneck", "between", "hidden", "states", "as", "in", ")", "trained", "with", "truncated", "BPTT", "for", "20", "steps", "performed", "well", ".", "Following", "we", "use", "dropout", "before", "and", "after", "every", "LSTM", "layer", ".", "The", "biases", "of", "LSTM", "forget", "gate", "were", "initialized", "to", "1.0", ".", "The", "size", "of", "the", "models", "will", "be", "described", "in", "more", "detail", "in", "the", "following", "sections", ",", "and", "the", "choices", "of", "hyper", "-", "parameters", "will", "be", "released", "as", "open", "source", "upon", "publication", ".", "For", "any", "model", "using", "character", "embedding", "CNNs", ",", "we", "closely", "follow", "the", "architecture", "from", ".", "The", "only", "important", "difference", "is", "that", "we", "use", "a", "larger", "number", "of", "convolutional", "features", "of", "4096", "to", "give", "enough", "capacity", "to", "the", "model", ".", "The", "resulting", "embedding", "is", "then", "linearly", "transformed", "to", "match", "the", "LSTM", "projection", "sizes", ".", "This", "allows", "it", "to", "match", "the", "performance", "of", "regular", "word", "embeddings", "but", "only", "uses", "a", "small", "fraction", "of", "parameters", ".", "subsection", ":", "Training", "Procedure", "The", "models", "were", "trained", "until", "convergence", "with", "an", "AdaGrad", "optimizer", "using", "a", "learning", "rate", "of", "0.2", ".", "In", "all", "the", "experiments", "the", "RNNs", "were", "unrolled", "for", "20", "steps", "without", "ever", "resetting", "the", "LSTM", "states", ".", "We", "used", "a", "batch", "size", "of", "128", ".", "We", "clip", "the", "gradients", "of", "the", "LSTM", "weights", "such", "that", "their", "norm", "is", "bounded", "by", "1.0", ".", "Using", "these", "hyper", "-", "parameters", "we", "found", "large", "LSTMs", "to", "be", "relatively", "easy", "to", "train", ".", "The", "same", "learning", "rate", "was", "used", "in", "almost", "all", "of", "the", "experiments", ".", "In", "a", "few", "cases", "we", "had", "to", "reduce", "it", "by", "an", "order", "of", "magnitude", ".", "Unless", "otherwise", "stated", ",", "the", "experiments", "were", "performed", "with", "32", "GPU", "workers", "and", "asynchronous", "gradient", "updates", ".", "Further", "details", "will", "be", "fully", "specified", "with", "the", "code", "upon", "publication", ".", "Training", "a", "model", "for", "such", "large", "target", "vocabulary", "(", "793471", "words", ")", "required", "to", "be", "careful", "with", "some", "details", "about", "the", "approximation", "to", "full", "Softmax", "using", "importance", "sampling", ".", "We", "used", "a", "large", "number", "of", "negative", "(", "or", "noise", ")", "samples", ":", "8192", "such", "samples", "were", "drawn", "per", "step", ",", "but", "were", "shared", "across", "all", "the", "target", "words", "in", "the", "batch", "(", "2560", "total", ",", "i.e.", "128", "times", "20", "unrolled", "steps", ")", ".", "This", "results", "in", "multiplying", "(", "2560", "x", "1024", ")", "times", "(", "1024", "x", "(", "8192", "+", "1", ")", ")", "(", "instead", "of", "(", "2560", "x", "1024", ")", "times", "(", "1024", "x", "793471", ")", ")", ",", "i.e.", "about", "100", "-", "fold", "less", "computation", ".", "section", ":", "Results", "and", "Analysis", "In", "this", "section", "we", "summarize", "the", "results", "of", "our", "experiments", "and", "do", "an", "in", "-", "depth", "analysis", ".", "Table", "[", "reference", "]", "contains", "all", "results", "for", "our", "models", "compared", "to", "previously", "published", "work", ".", "Table", "[", "reference", "]", "shows", "previous", "and", "our", "own", "work", "on", "ensembles", "of", "models", ".", "We", "hope", "that", "our", "encouraging", "results", ",", "which", "improved", "the", "best", "perplexity", "of", "a", "single", "model", "from", "51.3", "to", "30.0", "(", "whilst", "reducing", "the", "model", "size", "considerably", ")", ",", "and", "set", "a", "new", "record", "with", "ensembles", "at", "23.7", ",", "will", "enable", "rapid", "research", "and", "progress", "to", "advance", "Language", "Modeling", ".", "For", "this", "purpose", ",", "we", "will", "release", "the", "model", "weights", "and", "recipes", "upon", "publication", ".", "subsection", ":", "Size", "Matters", "Unsurprisingly", ",", "size", "matters", ":", "when", "training", "on", "a", "very", "large", "and", "complex", "data", "set", ",", "fitting", "the", "training", "data", "with", "an", "LSTM", "is", "fairly", "challenging", ".", "Thus", ",", "the", "size", "of", "the", "LSTM", "layer", "is", "a", "very", "important", "factor", "that", "influences", "the", "results", ",", "as", "seen", "in", "Table", "[", "reference", "]", ".", "The", "best", "models", "are", "the", "largest", "we", "were", "able", "to", "fit", "into", "a", "GPU", "memory", ".", "Our", "largest", "model", "was", "a", "2", "-", "layer", "LSTM", "with", "8192", "+", "1024", "dimensional", "recurrent", "state", "in", "each", "of", "the", "layers", ".", "Increasing", "the", "embedding", "and", "projection", "size", "also", "helps", "but", "causes", "a", "large", "increase", "in", "the", "number", "of", "parameters", ",", "which", "is", "less", "desirable", ".", "Lastly", ",", "training", "an", "RNN", "instead", "of", "an", "LSTM", "yields", "poorer", "results", "(", "about", "5", "perplexity", "worse", ")", "for", "a", "comparable", "model", "size", ".", "subsection", ":", "Regularization", "Importance", "As", "shown", "in", "Table", "[", "reference", "]", ",", "using", "dropout", "improves", "the", "results", ".", "To", "our", "surprise", ",", "even", "relatively", "small", "models", "(", "e.g.", ",", "single", "layer", "LSTM", "with", "2048", "units", "projected", "to", "512", "dimensional", "outputs", ")", "can", "over", "-", "fit", "the", "training", "set", "if", "trained", "long", "enough", ",", "eventually", "yielding", "holdout", "set", "degradation", ".", "Using", "dropout", "on", "non", "-", "recurrent", "connections", "largely", "mitigates", "these", "issues", ".", "While", "over", "-", "fitting", "still", "occurs", ",", "there", "is", "no", "more", "need", "for", "early", "stopping", ".", "For", "models", "that", "had", "4096", "or", "less", "units", "in", "the", "LSTM", "layer", ",", "we", "used", "10", "%", "dropout", "probability", ".", "For", "larger", "models", ",", "25", "%", "was", "significantly", "better", ".", "Even", "with", "such", "regularization", ",", "perplexities", "on", "the", "training", "set", "can", "be", "as", "much", "as", "6", "points", "below", "test", ".", "In", "one", "experiment", "we", "tried", "to", "use", "a", "smaller", "vocabulary", "comprising", "of", "the", "100", ",", "000", "most", "frequent", "words", "and", "found", "the", "difference", "between", "train", "and", "test", "to", "be", "smaller", "–", "which", "suggests", "that", "too", "much", "capacity", "is", "given", "to", "rare", "words", ".", "This", "is", "less", "of", "an", "issue", "with", "character", "CNN", "embedding", "models", "as", "the", "embeddings", "are", "shared", "across", "all", "words", ".", "subsection", ":", "Importance", "Sampling", "is", "Data", "Efficient", "Table", "[", "reference", "]", "shows", "the", "test", "perplexities", "of", "NCE", "vs", "IS", "loss", "after", "a", "few", "epochs", "of", "2048", "unit", "LSTM", "with", "512", "projection", ".", "The", "IS", "objective", "significantly", "improves", "the", "speed", "and", "the", "overall", "performance", "of", "the", "model", "when", "compared", "to", "NCE", ".", "subsection", ":", "Word", "Embeddings", "vs", "Character", "CNN", "Replacing", "the", "embedding", "layer", "with", "a", "parametrized", "neural", "network", "that", "process", "characters", "of", "a", "given", "word", "allows", "the", "model", "to", "consume", "arbitrary", "words", "and", "is", "not", "restricted", "to", "a", "fixed", "vocabulary", ".", "This", "property", "is", "useful", "for", "data", "sets", "with", "conversational", "or", "informal", "text", "as", "well", "as", "for", "morphologically", "rich", "languages", ".", "Our", "experiments", "show", "that", "using", "character", "-", "level", "embeddings", "is", "feasible", "and", "does", "not", "degrade", "performance", "–", "in", "fact", ",", "our", "best", "single", "model", "uses", "a", "Character", "CNN", "embedding", ".", "An", "additional", "advantage", "is", "that", "the", "number", "of", "parameters", "of", "the", "input", "layer", "is", "reduced", "by", "a", "factor", "of", "11", "(", "though", "training", "speed", "is", "slightly", "worse", ")", ".", "For", "inference", ",", "the", "embeddings", "can", "be", "precomputed", "so", "there", "is", "no", "speed", "penalty", ".", "Overall", ",", "the", "embedding", "of", "the", "best", "model", "is", "parametrized", "by", "72", "M", "weights", "(", "down", "from", "820", "M", "weights", ")", ".", "Table", "[", "reference", "]", "shows", "a", "few", "examples", "of", "nearest", "neighbor", "embeddings", "for", "some", "out", "-", "of", "-", "vocabulary", "words", "when", "character", "CNNs", "are", "used", ".", "subsection", ":", "Smaller", "Models", "with", "CNN", "Softmax", "Even", "with", "character", "-", "level", "embeddings", ",", "the", "model", "is", "still", "fairly", "large", "(", "though", "much", "smaller", "than", "the", "best", "competing", "models", "from", "previous", "work", ")", ".", "Most", "of", "the", "parameters", "are", "in", "the", "linear", "layer", "before", "the", "Softmax", ":", "820", "M", "versus", "a", "total", "of", "1.04B", "parameters", ".", "In", "one", "of", "the", "experiments", "we", "froze", "the", "word", "-", "LSTM", "after", "convergence", "and", "replaced", "the", "Softmax", "layer", "with", "the", "CNN", "Softmax", "sub", "-", "network", ".", "Without", "any", "fine", "-", "tuning", "that", "model", "was", "able", "to", "reach", "39.8", "perplexity", "with", "only", "293", "M", "weights", "(", "as", "seen", "in", "Table", "[", "reference", "]", ")", ".", "As", "described", "in", "Section", "[", "reference", "]", ",", "adding", "a", "“", "correction", "”", "word", "embedding", "term", "alleviates", "the", "gap", "between", "regular", "and", "CNN", "Softmax", ".", "Indeed", ",", "we", "can", "trade", "-", "off", "model", "size", "versus", "perplexity", ".", "For", "instance", ",", "by", "adding", "100", "M", "weights", "(", "through", "a", "128", "dimensional", "bottleneck", "embedding", ")", "we", "achieve", "35.8", "perplexity", "(", "see", "Table", "[", "reference", "]", ")", ".", "To", "contrast", "with", "the", "CNN", "Softmax", ",", "we", "also", "evaluated", "a", "model", "that", "replaces", "the", "Softmax", "layer", "with", "a", "smaller", "LSTM", "that", "predicts", "one", "character", "at", "a", "time", "(", "see", "Section", "[", "reference", "]", ")", ".", "Such", "a", "model", "does", "not", "have", "to", "learn", "long", "dependencies", "because", "the", "base", "LSTM", "still", "operates", "at", "the", "word", "-", "level", "(", "see", "Figure", "[", "reference", "]", "(", "c", ")", ")", ".", "With", "a", "single", "-", "layer", "LSTM", "of", "1024", "units", "we", "reached", "49.0", "test", "perplexity", ",", "far", "below", "the", "best", "model", ".", "In", "order", "to", "make", "the", "comparisons", "more", "fair", ",", "we", "performed", "a", "very", "expensive", "marginalization", "over", "the", "words", "in", "the", "vocabulary", "(", "to", "rule", "out", "words", "not", "in", "the", "dictionary", "which", "the", "character", "LSTM", "would", "assign", "some", "probability", ")", ".", "When", "doing", "this", "marginalization", ",", "the", "perplexity", "improved", "a", "bit", "down", "to", "47.9", ".", "subsection", ":", "Training", "Speed", "We", "used", "32", "Tesla", "K40", "GPUs", "to", "train", "our", "models", ".", "The", "smaller", "version", "of", "the", "LSTM", "model", "with", "2048", "units", "and", "512", "projections", "needs", "less", "than", "10", "hours", "to", "reach", "below", "45", "perplexity", "and", "after", "only", "2", "hours", "of", "training", "the", "model", "beats", "previous", "state", "-", "of", "-", "the", "art", "on", "this", "data", "set", ".", "The", "best", "model", "needs", "about", "5", "days", "to", "get", "to", "35", "perplexity", "and", "10", "days", "to", "32.5", ".", "The", "best", "results", "were", "achieved", "after", "3", "weeks", "of", "training", ".", "See", "Table", "[", "reference", "]", "for", "more", "details", ".", "subsection", ":", "Ensembles", "We", "averaged", "several", "of", "our", "best", "models", "and", "we", "were", "able", "to", "reach", "23.7", "test", "perplexity", "(", "more", "details", "and", "results", "can", "be", "seen", "in", "Table", "[", "reference", "]", ")", ",", "which", "is", "more", "than", "40", "%", "improvement", "over", "previous", "work", ".", "Interestingly", ",", "including", "the", "best", "N", "-", "gram", "model", "reduces", "the", "perplexity", "by", "1.2", "point", "even", "though", "the", "model", "is", "rather", "weak", "on", "its", "own", "(", "67.6", "perplexity", ")", ".", "Most", "previous", "work", "had", "to", "either", "ensemble", "with", "the", "best", "N", "-", "gram", "model", "(", "as", "their", "RNN", "only", "used", "a", "limited", "output", "vocabulary", "of", "a", "few", "thousand", "words", ")", ",", "or", "use", "N", "-", "gram", "features", "as", "additional", "input", "to", "the", "RNN", ".", "Our", "results", ",", "on", "the", "contrary", ",", "suggest", "that", "N", "-", "grams", "are", "of", "limited", "benefit", ",", "and", "suggest", "that", "a", "carefully", "trained", "LSTM", "LM", "is", "the", "most", "competitive", "model", ".", "subsection", ":", "LSTMs", "are", "best", "on", "the", "tail", "words", "Figure", "[", "reference", "]", "shows", "the", "difference", "in", "log", "probabilities", "between", "our", "best", "model", "(", "at", "30.0", "perplexity", ")", "and", "the", "KN", "-", "5", ".", "As", "can", "be", "seen", "from", "the", "plot", ",", "the", "LSTM", "is", "better", "across", "all", "the", "buckets", "and", "significantly", "outperforms", "KN", "-", "5", "on", "the", "rare", "words", ".", "This", "is", "encouraging", "as", "it", "seems", "to", "suggest", "that", "LSTM", "LMs", "may", "fare", "even", "better", "for", "languages", "or", "data", "sets", "where", "the", "number", "of", "rare", "words", "is", "larger", "than", "traditional", "N", "-", "gram", "models", ".", "subsection", ":", "Samples", "from", "the", "model", "To", "qualitatively", "evaluate", "the", "model", ",", "we", "sampled", "many", "sentences", ".", "We", "discarded", "short", "and", "politically", "incorrect", "ones", ",", "but", "the", "sample", "shown", "below", "is", "otherwise", "“", "raw", "”", "(", "i.e.", ",", "not", "hand", "picked", ")", ".", "The", "samples", "are", "of", "high", "quality", "–", "which", "is", "not", "a", "surprise", ",", "given", "the", "perplexities", "attained", "–", "but", "there", "are", "still", "some", "occasional", "mistakes", ".", "Sentences", "generated", "by", "the", "ensemble", "(", "about", "26", "perplexity", ")", ":", "With", "even", "more", "new", "technologies", "coming", "onto", "the", "market", "quickly", "during", "the", "past", "three", "years", ",", "an", "increasing", "number", "of", "companies", "now", "must", "tackle", "the", "ever", "-", "changing", "and", "ever", "-", "changing", "environmental", "challenges", "online", ".", "<", "S", ">", "Check", "back", "for", "updates", "on", "this", "breaking", "news", "story", ".", "<", "S", ">", "About", "800", "people", "gathered", "at", "Hever", "Castle", "on", "Long", "Beach", "from", "noon", "to", "2", "pm", ",", "three", "to", "four", "times", "that", "of", "the", "funeral", "cortège", ".", "<", "S", ">", "We", "are", "aware", "of", "written", "instructions", "from", "the", "copyright", "holder", "not", "to", ",", "in", "any", "way", ",", "mention", "Rosenberg", "’s", "negative", "comments", "if", "they", "are", "relevant", "as", "indicated", "in", "the", "documents", ",", "”", "eBay", "said", "in", "a", "statement", ".", "<", "S", ">", "It", "is", "now", "known", "that", "coffee", "and", "cacao", "products", "can", "do", "no", "harm", "on", "the", "body", ".", "<", "S", ">", "Yuri", "Zhirkov", "was", "in", "attendance", "at", "the", "Stamford", "Bridge", "at", "the", "start", "of", "the", "second", "half", "but", "neither", "Drogba", "nor", "Malouda", "was", "able", "to", "push", "on", "through", "the", "Barcelona", "defence", ".", "section", ":", "Discussion", "and", "Conclusions", "In", "this", "paper", "we", "have", "shown", "that", "RNN", "LMs", "can", "be", "trained", "on", "large", "amounts", "of", "data", ",", "and", "outperform", "competing", "models", "including", "carefully", "tuned", "N", "-", "grams", ".", "The", "reduction", "in", "perplexity", "from", "51.3", "to", "30.0", "is", "due", "to", "several", "key", "components", "which", "we", "studied", "in", "this", "paper", ".", "Thus", ",", "a", "large", ",", "regularized", "LSTM", "LM", ",", "with", "projection", "layers", "and", "trained", "with", "an", "approximation", "to", "the", "true", "Softmax", "with", "importance", "sampling", "performs", "much", "better", "than", "N", "-", "grams", ".", "Unlike", "previous", "work", ",", "we", "do", "not", "require", "to", "interpolate", "both", "the", "RNN", "LM", "and", "the", "N", "-", "gram", ",", "and", "the", "gains", "of", "doing", "so", "are", "rather", "marginal", ".", "By", "exploring", "recent", "advances", "in", "model", "architectures", "(", "e.g.", "LSTMs", ")", ",", "exploiting", "small", "character", "CNNs", ",", "and", "by", "sharing", "our", "findings", "in", "this", "paper", "and", "accompanying", "code", "and", "models", "(", "to", "be", "released", "upon", "publication", ")", ",", "we", "hope", "to", "inspire", "research", "on", "large", "scale", "Language", "Modeling", ",", "a", "problem", "we", "consider", "crucial", "towards", "language", "understanding", ".", "We", "hope", "for", "future", "research", "to", "focus", "on", "reasonably", "sized", "datasets", "taking", "inspiration", "from", "recent", "advances", "seen", "in", "the", "computer", "vision", "community", "thanks", "to", "efforts", "such", "as", "Imagenet", ".", "section", ":", "Acknowledgements", "We", "thank", "Ciprian", "Chelba", ",", "Ilya", "Sutskever", ",", "and", "the", "Google", "Brain", "Team", "for", "their", "help", "and", "discussions", ".", "We", "also", "thank", "Koray", "Kavukcuoglu", "for", "his", "help", "with", "the", "manuscript", ".", "bibliography", ":", "References"], "entity": [[8, 11, ["Recurrent", "Neural", "Networks"], "Method"], [21, 23, ["language", "understanding"], "Task"], [62, 66, ["character", "Convolutional", "Neural", "Networks"], "Method"], [67, 72, ["Long", "-", "Short", "Term", "Memory"], "Method"], [75, 79, ["One", "Billion", "Word", "Benchmark"], "Material"], [93, 94, ["perplexity"], "Metric"], [115, 118, ["ensemble", "of", "models"], "Method"], [124, 125, ["perplexity"], "Metric"], [138, 142, ["NLP", "and", "ML", "community"], "Task"], [152, 154, ["Language", "Modeling"], "Task"], [155, 157, ["LM", ")"], "Task"], [162, 168, ["Natural", "Language", "Processing", "(", "NLP", ")"], "Task"], [169, 171, ["Language", "Understanding"], "Task"], [232, 235, ["fundamental", "language", "understanding"], "Task"], [236, 238, ["question", "answering"], "Task"], [239, 241, ["machine", "translation"], "Task"], [243, 245, ["text", "summarization"], "Task"], [246, 247, ["LMs"], "Task"], [254, 256, ["NLP", "tasks"], "Task"], [258, 260, ["speech", "recognition"], "Task"], [261, 263, ["machine", "translation"], "Task"], [265, 267, ["text", "summarization"], "Task"], [277, 279, ["language", "models"], "Method"], [282, 283, ["metrics"], "Metric"], [285, 287, ["downstream", "task"], "Task"], [290, 293, ["word", "error", "rate"], "Metric"], [294, 296, ["speech", "recognition"], "Task"], [298, 300, ["BLEU", "score"], "Metric"], [301, 302, ["translation"], "Task"], [311, 312, ["LMs"], "Task"], [326, 328, ["language", "models"], "Task"], [347, 349, ["language", "models"], "Method"], [371, 377, ["sequence", "-", "to", "-", "sequence", "models"], "Method"], [378, 381, ["conditional", "language", "models"], "Method"], [390, 392, ["machine", "translation"], "Task"], [393, 395, ["video", "generation"], "Task"], [396, 398, ["Deep", "Learning"], "Method"], [399, 402, ["Recurrent", "Neural", "Networks"], "Method"], [403, 404, ["RNNs"], "Method"], [441, 444, ["N", "-", "grams"], "Method"], [469, 472, ["low", "perplexity", "LMs"], "Metric"], [479, 482, ["large", "scale", "LM"], "Task"], [485, 486, ["RNNs"], "Method"], [491, 494, ["N", "-", "grams"], "Method"], [503, 507, ["N", "-", "gram", "models"], "Method"], [553, 555, ["language", "modeling"], "Task"], [563, 565, ["PTB", "task"], "Task"], [616, 619, ["large", "scale", "modeling"], "Task"], [641, 642, ["LM"], "Task"], [643, 644, ["PTB"], "Method"], [657, 659, ["training", "recipes"], "Method"], [687, 688, ["PTB"], "Method"], [697, 701, ["1B", "words", "training", "data"], "Material"], [714, 716, ["computer", "vision"], "Task"], [735, 737, ["Language", "Modeling"], "Task"], [761, 764, ["large", "scale", "LM"], "Task"], [770, 772, ["Softmax", "loss"], "Method"], [776, 779, ["character", "level", "CNNs"], "Method"], [791, 793, ["full", "Softmax"], "Method"], [830, 831, ["perplexity"], "Metric"], [860, 861, ["perplexity"], "Metric"], [910, 912, ["language", "modeling"], "Task"], [929, 935, ["large", "scale", "recurrent", "neural", "network", "training"], "Method"], [993, 995, ["language", "modeling"], "Task"], [1002, 1004, ["Language", "Models"], "Task"], [1004, 1006, ["Language", "Modeling"], "Task"], [1007, 1009, ["LM", ")"], "Task"], [1015, 1016, ["NLP"], "Task"], [1020, 1021, ["LM"], "Task"], [1047, 1051, ["log", "-", "linear", "models"], "Method"], [1053, 1057, ["non", "-", "parametric", "approaches"], "Method"], [1066, 1070, ["Count", "-", "based", "approaches"], "Method"], [1073, 1078, ["statistics", "of", "N", "-", "grams"], "Method"], [1081, 1082, ["smoothing"], "Method"], [1102, 1110, ["Kneser", "-", "Ney", "smoothed", "5", "-", "gram", "models"], "Method"], [1127, 1129, ["parametric", "approaches"], "Method"], [1131, 1133, ["Neural", "Networks"], "Method"], [1141, 1148, ["Recurrent", "Neural", "Networks", "(", "RNN", ")", "models"], "Method"], [1161, 1167, ["Long", "-", "Short", "Term", "Memory", "model"], "Method"], [1170, 1172, ["gating", "mechanism"], "Method"], [1224, 1228, ["One", "Billion", "Word", "Benchmark"], "Material"], [1274, 1276, ["human", "language"], "Task"], [1315, 1317, ["deep", "learning"], "Task"], [1323, 1325, ["computer", "vision"], "Task"], [1332, 1334, ["vision", "models"], "Method"], [1408, 1411, ["Convolutional", "Embedding", "Models"], "Method"], [1424, 1426, ["word", "embeddings"], "Method"], [1428, 1430, ["NLP", "problems"], "Task"], [1432, 1438, ["part", "-", "of", "-", "speech", "tagging"], "Task"], [1439, 1440, ["parsing"], "Task"], [1441, 1443, ["language", "modeling"], "Task"], [1464, 1466, ["word", "embeddings"], "Task"], [1474, 1476, ["recurrent", "networks"], "Method"], [1499, 1501, ["Neural", "Network"], "Method"], [1510, 1517, ["part", "-", "of", "-", "speech", "tagging", "task"], "Task"], [1527, 1531, ["1", "-", "d", "CNN"], "Method"], [1532, 1535, ["max", "-", "pooling"], "Method"], [1550, 1555, ["2", "-", "layer", "highway", "network"], "Method"], [1559, 1560, ["embedding"], "Method"], [1562, 1564, ["semantic", "representations"], "Method"], [1598, 1599, ["Softmax"], "Method"], [1613, 1615, ["modeling", "language"], "Task"], [1659, 1661, ["linear", "layer"], "Method"], [1664, 1668, ["Softmax", "non", "-", "linearity"], "Method"], [1679, 1680, ["logit"], "Method"], [1685, 1687, ["inner", "product"], "Method"], [1735, 1736, ["training"], "Task"], [1740, 1744, ["matrix", "-", "matrix", "multiplications"], "Method"], [1746, 1747, ["GPUs"], "Method"], [1758, 1760, ["scaling", "issue"], "Task"], [1761, 1763, ["importance", "sampling"], "Method"], [1764, 1767, ["Noise", "Contrastive", "Estimation"], "Method"], [1768, 1769, ["NCE"], "Method"], [1771, 1775, ["self", "normalizing", "partition", "functions"], "Method"], [1776, 1778, ["Hierarchical", "Softmax"], "Method"], [1790, 1792, ["importance", "sampling"], "Method"], [1807, 1808, ["NCE"], "Method"], [1821, 1823, ["Language", "Modeling"], "Task"], [1831, 1833, ["chain", "rule"], "Method"], [1852, 1853, ["LSTM"], "Method"], [1877, 1880, ["Noise", "Contrastive", "Estimation"], "Method"], [1881, 1883, ["Importance", "Sampling"], "Method"], [1894, 1895, ["Softmax"], "Method"], [1900, 1901, ["LMs"], "Task"], [1908, 1910, ["Hierarchical", "Softmax"], "Method"], [1934, 1938, ["training", "and", "inference", "time"], "Task"], [1975, 1977, ["sampling", "methods"], "Method"], [1983, 1985, ["Sampling", "approaches"], "Method"], [1989, 1990, ["training"], "Task"], [2008, 2010, ["distributed", "setting"], "Task"], [2029, 2032, ["Noise", "Contrastive", "Estimation"], "Method"], [2033, 2034, ["NCE"], "Method"], [2039, 2043, ["surrogate", "binary", "classification", "task"], "Task"], [2046, 2047, ["classifier"], "Method"], [2075, 2076, ["classifier"], "Method"], [2136, 2138, ["logistic", "classifier"], "Method"], [2158, 2160, ["LSTM", "LM"], "Method"], [2171, 2173, ["importance", "sampling"], "Method"], [2174, 2175, ["IS"], "Method"], [2199, 2201, ["importance", "sampling"], "Method"], [2214, 2217, ["surrogate", "classification", "task"], "Task"], [2219, 2220, ["NCE"], "Method"], [2240, 2242, ["binary", "task"], "Task"], [2289, 2291, ["multiclass", "loss"], "Method"], [2313, 2315, ["Bayes", "rule"], "Method"], [2338, 2339, ["NCE"], "Method"], [2356, 2357, ["NCE"], "Method"], [2363, 2364, ["NCE"], "Method"], [2368, 2371, ["binary", "classification", "task"], "Task"], [2378, 2380, ["logistic", "loss"], "Method"], [2387, 2390, ["multiclass", "classification", "problem"], "Task"], [2392, 2397, ["Softmax", "and", "cross", "entropy", "loss"], "Method"], [2425, 2428, ["multiclass", "classification", "task"], "Task"], [2434, 2436, ["binary", "task"], "Task"], [2464, 2466, ["CNN", "Softmax"], "Method"], [2480, 2482, ["word", "embeddings"], "Method"], [2492, 2495, ["CNN", "character", "embeddings"], "Method"], [2521, 2523, ["Softmax", "layer"], "Method"], [2532, 2533, ["Softmax"], "Method"], [2544, 2546, ["word", "embedding"], "Method"], [2564, 2565, ["CNN"], "Method"], [2575, 2577, ["CNN", "Softmax"], "Method"], [2582, 2584, ["network", "architecture"], "Method"], [2598, 2604, ["word", "-", "embedding", "sub", "-", "network"], "Method"], [2606, 2607, ["inference"], "Task"], [2618, 2620, ["computational", "complexity"], "Metric"], [2624, 2626, ["regular", "Softmax"], "Method"], [2634, 2637, ["importance", "sampling", "loss"], "Method"], [2670, 2671, ["Softmax"], "Method"], [2692, 2693, ["CNN"], "Method"], [2716, 2718, ["learning", "rate"], "Metric"], [2786, 2787, ["LSTM"], "Method"], [2796, 2799, ["bottleneck", "linear", "layer"], "Method"], [2803, 2805, ["CNN", "Softmax"], "Method"], [2833, 2834, ["regular"], "Method"], [2836, 2838, ["CNN", "Softmax"], "Method"], [2887, 2889, ["Machine", "Translation"], "Task"], [2890, 2897, ["handling", "out", "-", "of", "-", "vocabulary", "words"], "Task"], [2905, 2907, ["parallel", "training"], "Task"], [2935, 2940, ["byte", "-", "level", "input", "embeddings"], "Method"], [2941, 2944, ["named", "entity", "recognition"], "Task"], [2963, 2965, ["LSTM", "Predictions"], "Method"], [2966, 2969, ["CNN", "Softmax", "layer"], "Method"], [3002, 3003, ["perplexities"], "Task"], [3069, 3070, ["PTB"], "Task"], [3086, 3087, ["LSTM"], "Method"], [3104, 3110, ["word", "and", "character", "-", "level", "models"], "Method"], [3113, 3119, ["word", "-", "level", "LSTM", "hidden", "state"], "Method"], [3122, 3123, ["LSTM"], "Method"], [3158, 3160, ["LSTM", "model"], "Method"], [3171, 3176, ["word", "-", "level", "Softmax", "layer"], "Method"], [3179, 3183, ["character", "-", "level", "LSTM"], "Method"], [3195, 3196, ["training"], "Task"], [3197, 3198, ["inference"], "Task"], [3208, 3212, ["regular", "and", "CNN", "Softmax"], "Method"], [3225, 3228, ["fixed", "vocabulary", "models"], "Method"], [3242, 3244, ["TensorFlow", "system"], "Method"], [3269, 3274, ["1B", "Word", "Benchmark", "data", "set"], "Material"], [3360, 3362, ["language", "modeling"], "Task"], [3363, 3364, ["perplexity"], "Task"], [3368, 3375, ["average", "per", "-", "word", "log", "-", "probability"], "Metric"], [3405, 3410, ["1B", "Word", "Benchmark", "data", "set"], "Material"], [3543, 3545, ["CNN", "model"], "Method"], [3604, 3607, ["RNN", "LM", "architectures"], "Method"], [3626, 3627, ["LSTM"], "Method"], [3652, 3653, ["LSTMs"], "Method"], [3655, 3657, ["projection", "layer"], "Method"], [3670, 3672, ["truncated", "BPTT"], "Method"], [3686, 3688, ["LSTM", "layer"], "Method"], [3692, 3695, ["LSTM", "forget", "gate"], "Method"], [3736, 3739, ["character", "embedding", "CNNs"], "Method"], [3781, 3782, ["LSTM"], "Method"], [3793, 3796, ["regular", "word", "embeddings"], "Method"], [3817, 3819, ["AdaGrad", "optimizer"], "Method"], [3821, 3823, ["learning", "rate"], "Metric"], [3831, 3832, ["RNNs"], "Method"], [3841, 3842, ["LSTM"], "Method"], [3858, 3859, ["LSTM"], "Method"], [3877, 3878, ["LSTMs"], "Method"], [3887, 3889, ["learning", "rate"], "Metric"], [3926, 3929, ["asynchronous", "gradient", "updates"], "Method"], [3965, 3967, ["full", "Softmax"], "Task"], [3968, 3970, ["importance", "sampling"], "Method"], [4159, 4161, ["Language", "Modeling"], "Task"], [4203, 4204, ["LSTM"], "Method"], [4214, 4216, ["LSTM", "layer"], "Method"], [4255, 4259, ["2", "-", "layer", "LSTM"], "Method"], [4300, 4301, ["RNN"], "Method"], [4304, 4305, ["LSTM"], "Method"], [4317, 4319, ["model", "size"], "Metric"], [4333, 4334, ["dropout"], "Method"], [4349, 4352, ["single", "layer", "LSTM"], "Method"], [4380, 4381, ["dropout"], "Method"], [4417, 4419, ["LSTM", "layer"], "Method"], [4440, 4441, ["regularization"], "Method"], [4507, 4511, ["character", "CNN", "embedding", "models"], "Method"], [4522, 4524, ["Importance", "Sampling"], "Method"], [4545, 4548, ["2048", "unit", "LSTM"], "Method"], [4569, 4570, ["NCE"], "Method"], [4573, 4575, ["Word", "Embeddings"], "Method"], [4576, 4578, ["Character", "CNN"], "Method"], [4580, 4582, ["embedding", "layer"], "Method"], [4584, 4587, ["parametrized", "neural", "network"], "Method"], [4635, 4639, ["character", "-", "level", "embeddings"], "Method"], [4656, 4659, ["Character", "CNN", "embedding"], "Method"], [4682, 4684, ["training", "speed"], "Metric"], [4690, 4691, ["inference"], "Task"], [4735, 4738, ["nearest", "neighbor", "embeddings"], "Method"], [4747, 4749, ["character", "CNNs"], "Method"], [4754, 4756, ["Smaller", "Models"], "Method"], [4757, 4759, ["CNN", "Softmax"], "Method"], [4761, 4765, ["character", "-", "level", "embeddings"], "Method"], [4797, 4798, ["Softmax"], "Method"], [4816, 4819, ["word", "-", "LSTM"], "Method"], [4824, 4826, ["Softmax", "layer"], "Method"], [4828, 4833, ["CNN", "Softmax", "sub", "-", "network"], "Method"], [4836, 4839, ["fine", "-", "tuning"], "Method"], [4882, 4886, ["regular", "and", "CNN", "Softmax"], "Method"], [4894, 4896, ["model", "size"], "Metric"], [4910, 4914, ["128", "dimensional", "bottleneck", "embedding"], "Method"], [4931, 4933, ["CNN", "Softmax"], "Method"], [4942, 4944, ["Softmax", "layer"], "Method"], [4947, 4948, ["LSTM"], "Method"], [4976, 4977, ["LSTM"], "Method"], [4997, 5004, ["single", "-", "layer", "LSTM", "of", "1024", "units"], "Method"], [5048, 5050, ["character", "LSTM"], "Method"], [5072, 5074, ["Training", "Speed"], "Metric"], [5077, 5080, ["Tesla", "K40", "GPUs"], "Method"], [5090, 5092, ["LSTM", "model"], "Method"], [5093, 5095, ["2048", "units"], "Method"], [5218, 5222, ["N", "-", "gram", "model"], "Method"], [5253, 5257, ["N", "-", "gram", "model"], "Method"], [5260, 5261, ["RNN"], "Method"], [5285, 5286, ["RNN"], "Method"], [5310, 5312, ["LSTM", "LM"], "Method"], [5320, 5321, ["LSTMs"], "Method"], [5344, 5345, ["perplexity"], "Metric"], [5348, 5351, ["KN", "-", "5"], "Method"], [5361, 5362, ["LSTM"], "Method"], [5371, 5374, ["KN", "-", "5"], "Method"], [5388, 5390, ["LSTM", "LMs"], "Method"], [5409, 5413, ["N", "-", "gram", "models"], "Method"], [5698, 5701, ["N", "-", "grams"], "Method"], [5705, 5706, ["perplexity"], "Metric"], [5728, 5731, ["regularized", "LSTM", "LM"], "Method"], [5733, 5735, ["projection", "layers"], "Method"], [5743, 5744, ["Softmax"], "Method"], [5745, 5747, ["importance", "sampling"], "Method"], [5751, 5754, ["N", "-", "grams"], "Method"], [5767, 5769, ["RNN", "LM"], "Method"], [5771, 5774, ["N", "-", "gram"], "Method"], [5790, 5792, ["model", "architectures"], "Method"], [5794, 5795, ["LSTMs"], "Method"], [5798, 5801, ["small", "character", "CNNs"], "Method"], [5840, 5842, ["language", "understanding"], "Task"], [5862, 5865, ["computer", "vision", "community"], "Task"], [14, 16, ["Language", "Modeling"], "Task"], [407, 409, ["language", "modeling"], "Task"], [667, 670, ["large", "scale", "LM"], "Task"], [673, 677, ["One", "Billion", "Word", "Benchmark"], "Material"], [820, 823, ["large", "scale", "LM"], "Task"], [925, 927, ["language", "modeling"], "Task"], [1063, 1064, ["LMs"], "Task"], [1196, 1197, ["LMs"], "Task"], [1245, 1246, ["LMs"], "Task"], [1256, 1257, ["LMs"], "Task"], [1468, 1469, ["LSTMs"], "Method"], [1573, 1575, ["language", "modeling"], "Task"], [1828, 1829, ["LMs"], "Task"], [2488, 2490, ["language", "modeling"], "Task"], [3025, 3026, ["LSTMs"], "Method"], [3288, 3290, ["language", "modeling"], "Task"], [5681, 5682, ["LMs"], "Task"], [5831, 5833, ["Language", "Modeling"], "Task"]]} {"doc_id": "0398552184f80db111e9c28bf533b395f233ac00", "text": ["Bridging", "Saliency", "Detection", "to", "Weakly", "Supervised", "Object", "Detection", "Based", "on", "Self", "-", "paced", "Curriculum", "Learning", "section", ":", "Abstract", "Weakly", "-", "supervised", "object", "detection", "(", "WOD", ")", "is", "a", "challenging", "problems", "in", "computer", "vision", ".", "The", "key", "problem", "is", "to", "simultaneously", "infer", "the", "exact", "object", "locations", "in", "the", "training", "images", "and", "train", "the", "object", "detectors", ",", "given", "only", "the", "training", "images", "with", "weak", "image", "-", "level", "labels", ".", "Intuitively", ",", "by", "simulating", "the", "selective", "attention", "mechanism", "of", "human", "visual", "system", ",", "saliency", "detection", "technique", "can", "select", "attractive", "objects", "in", "scenes", "and", "thus", "is", "a", "potential", "way", "to", "provide", "useful", "priors", "for", "WOD", ".", "However", ",", "the", "way", "to", "adopt", "saliency", "detection", "in", "WOD", "is", "not", "trivial", "since", "the", "detected", "saliency", "region", "might", "be", "possibly", "highly", "ambiguous", "in", "complex", "cases", ".", "To", "this", "end", ",", "this", "paper", "first", "comprehensively", "analyzes", "the", "challenges", "in", "applying", "saliency", "detection", "to", "WOD", ".", "Then", ",", "we", "make", "one", "of", "the", "earliest", "efforts", "to", "bridge", "saliency", "detection", "to", "WOD", "via", "the", "self", "-", "paced", "curriculum", "learning", ",", "which", "can", "guide", "the", "learning", "procedure", "to", "gradually", "achieve", "faithful", "knowledge", "of", "multi", "-", "class", "objects", "from", "easy", "to", "hard", ".", "The", "experimental", "results", "demonstrate", "that", "the", "proposed", "approach", "can", "successfully", "bridge", "saliency", "detection", "and", "WOD", "tasks", "and", "achieve", "the", "state", "-", "of", "-", "the", "-", "art", "object", "detection", "results", "under", "the", "weak", "supervision", ".", "section", ":", "Introduction", "1", "Object", "detection", "is", "one", "of", "the", "most", "fundamental", "yetchallenging", "problems", "in", "computer", "vision", "community", ".", "The", "most", "recent", "breakthrough", "was", "achieved", "by", "Girshick", "et", "al", ".", ",", "who", "trained", "the", "Convolutional", "Neural", "Network", "(", "CNN", ")", "by", "using", "large", "amount", "of", "human", "labelled", "bounding", "boxes", "to", "learn", "the", "powerful", "feature", "representations", "and", "object", "classifiers", ".", "Despite", "their", "success", ",", "the", "problem", "of", "object", "detection", "is", "still", "under", "-", "addressed", "in", "practice", "due", "to", "the", "heavy", "burden", "of", "labeling", "the", "training", "samples", ".", "Essentially", ",", "in", "this", "big", "data", "era", ",", "humans", "more", "desire", "intelligent", "machines", "which", "are", "capable", "of", "automatically", "discovering", "the", "intrinsic", "patterns", "from", "the", "cheaply", "and", "massively", "collected", "weakly", "*", "The", "corresponding", "author", "labeled", "images", ".", "Thus", "weakly", "supervised", "object", "detection", "(", "WOD", ")", "systems", "have", "been", "gaining", "more", "interests", "recently", ".", "The", "key", "problem", "in", "WOD", "is", "how", "to", "extract", "the", "exact", "object", "localizations", "and", "train", "the", "corresponding", "object", "detectors", "from", "the", "weakly", "labelled", "training", "images", ".", "In", "such", "chicken", "-", "egg", "problem", ",", "most", "methods", "(", "including", "the", "proposed", "one", ")", "usually", "use", "the", "alternative", "learning", "strategy", "that", "first", "provides", "some", "coarse", "estimation", "to", "initialize", "the", "potential", "object", "locations", "and", "then", "gradually", "train", "the", "object", "detectors", "and", "update", "object", "locations", "jointly", ".", "In", "this", "paper", ",", "we", "leverage", "saliency", "detection", "to", "initialize", "the", "potential", "object", "locations", "due", "to", "the", "following", "reasons", ":", "1", ")", "Saliency", "detection", ";", "[", "reference", "][", "reference", "]", "aims", "at", "simulating", "the", "selective", "attention", "mechanism", "of", "human", "visual", "system", "to", "automatically", "select", "sub", "-", "regions", "(", "usually", "the", "regions", "containing", "objects", "of", "interest", ")", "in", "image", "scenes", ".", "Thus", ",", "it", "can", "be", "readily", "utilized", "to", "provide", "useful", "priors", "to", "estimate", "the", "potential", "object", "localizations", "and", "fit", "well", "to", "the", "investigated", "task", ".", "2", ")", "Some", "recent", "saliency", "detection", "methods", "such", "as", "[", "reference", "]", "and", "[", "reference", "]", "can", "process", "much", "faster", "than", "the", "priors", ",", "e.g.", ",", "intra", "-", "class", "similarity", "[", "reference", "]", ",", "inter", "-", "class", "variance", "[", "reference", "]", ",", "and", "distance", "mapping", "relation", "[", "reference", "]", ",", "adopted", "in", "the", "existing", "WOD", "systems", ".", "3", ")", "Several", "existing", "works", ",", "e.g.", ",", ",", "have", "attempted", "to", "apply", "saliency", "detection", "techniques", "to", "WOD", ".", "However", "they", "still", "have", "not", "sufficiently", "explore", "the", "intrinsic", "bridge", "between", "these", "two", "tasks", ",", "which", "motivates", "us", "to", "clarify", "the", "insightful", "relationship", "between", "these", "two", "tasks", "and", "further", "develop", "powerful", "learning", "regime", "to", "bridge", "them", ".", "Essentially", ",", "although", "it", "sounds", "reasonable", "to", "apply", "saliency", "detection", "to", "WOD", ",", "the", "way", "to", "bridge", "these", "two", "tasks", "is", "not", "trivial", ".", "The", "main", "problem", "is", "that", "saliency", "detection", "is", "formulated", "as", "category", "-", "free", "models", "which", "only", "distinguish", "attractive", "regions", "from", "the", "image", "background", "while", "irrelevant", "to", "the", "concrete", "object", "category", ".", "Thus", ",", "as", "shown", "in", "Fig", ".", "1", ",", "in", "the", "images", "only", "containing", "one", "category", "of", "objects", "(", "considered", "as", "\"", "easy", "\"", "images", ")", ",", "the", "objects", "can", "be", "captured", "by", "saliency", "detection", "methods", "easily", "and", "associated", "with", "the", "corresponding", "image", "label", "properly", ".", "Whereas", "in", "the", "images", "weakly", "labelled", "as", "containing", "multiple", "categories", "of", "objects", "(", "considered", "as", "\"", "hard", "\"", "images", ")", ",", "objects", "in", "all", "of", "categories", "will", "have", "the", "probabilities", "to", "attract", "the", "human", "attention", "and", "their", "corresponding", "locations", "are", "also", "hard", "for", "saliency", "models", "to", "identify", ",", "which", "largely", "increases", "the", "ambiguity", "when", "considering", "to", "apply", "the", "obtained", "salient", "detection", "results", "to", "initializing", "the", "training", "samples", "for", "WOD", ".", "Thus", ",", "it", "is", "unreliable", "to", "directly", "apply", "saliency", "detection", "to", "WOD", ".", "To", "alleviate", "this", "problem", ",", "we", "propose", "to", "bridge", "saliency", "detection", "to", "WOD", "via", "a", "self", "-", "paced", "curriculum", "learning", "(", "SPCL", ")", "regime", ".", "SPCL", "was", "proposed", "in", "as", "a", "general", "learning", "framework", "including", "both", "the", "curriculum", "learning", "(", "CL", ")", "and", "self", "-", "paced", "learning", "(", "SPL", ")", "components", ".", "To", "the", "best", "of", "our", "knowledge", ",", "both", "of", "these", "two", "learning", "components", "are", "critical", "in", "successfully", "bridging", "saliency", "detection", "to", "WOD", ",", "whereas", "none", "of", "the", "existing", "literature", "has", "explored", "them", "before", ".", "Specifically", ",", "CL", "was", "proposed", "by", "[", "reference", "]", ",", "which", "is", "usually", "learned", "based", "on", "the", "learning", "priorities", "derived", "by", "predetermined", "heuristics", "for", "particular", "problems", ".", "SPL", "was", "proposed", "by", "[", "reference", "]", ",", "where", "the", "learning", "pace", "is", "dynamically", "generated", "by", "the", "learner", "itself", ",", "according", "to", "which", "the", "learner", "has", "already", "learned", "from", "the", "data", ".", "Thus", ",", "the", "CL", "and", "SPL", "components", "in", "SPCL", "can", "be", "correspondingly", "used", "to", "solve", "the", "training", "sample", "initialization", "and", "object", "detector", "updating", "problems", "in", "the", "proposed", "saliency", "-", "guided", "WOD", ".", "To", "implement", "SPCL", "for", "our", "task", ",", "we", "first", "design", "a", "task", "-", "specific", "curriculum", "to", "assign", "the", "\"", "easy", "\"", "images", "with", "larger", "priority", "than", "the", "\"", "hard", "\"", "images", "during", "the", "learning", "procedure", ",", "which", "indicates", "that", "only", "the", "salient", "object", "hypotheses", "in", "the", "\"", "easy", "\"", "images", "are", "selected", "as", "the", "initial", "training", "samples", ",", "while", "the", "object", "hypotheses", "in", "the", "\"", "hard", "\"", "images", "will", "be", "gradually", "involved", "in", "the", "subsequent", "learning", "iterations", ".", "To", "guide", "the", "learner", "to", "gradually", "learn", "faithful", "knowledge", "of", "multi", "-", "class", "objects", "from", "the", "\"", "easy", "\"", "(", "high", "-", "confidence", ")", "images", "to", "the", "\"", "hard", "\"", "(", "high", "-", "ambiguity", ")", "ones", ",", "a", "novel", "self", "-", "paced", "learning", "regularizer", "is", "proposed", "to", "enforce", "the", "learner", "to", "select", "confident", "and", "diverse", "training", "hypotheses", "in", "each", "iteration", "and", "learn", "the", "object", "detectors", "of", "multiple", "categories", "simultaneously", ".", "Finally", ",", "the", "proposed", "SPCL", "regime", "can", "fit", "well", "to", "solve", "the", "problems", "in", "this", "paper", ".", "Compared", "with", "the", "SPCL", "model", "in", ",", "the", "learning", "regime", "proposed", "in", "this", "paper", "mainly", "has", "three", "differences", ":", "1", ")", "We", "design", "a", "task", "-", "specific", "learning", "curriculum", "for", "bridging", "saliency", "detection", "and", "WOD", "effectively", ".", "2", ")", "We", "introduce", "an", "additional", "term", ",", "the", "sample", "diversity", "term", ",", "in", "the", "self", "-", "paced", "regularizer", "to", "prevent", "the", "selected", "training", "hypotheses", "from", "drifting", "to", "a", "small", "collection", "of", "training", "images", ".", "3", ")", "Considering", "the", "latent", "relationship", "among", "the", "multiple", "categories", "of", "co", "-", "occurring", "objects", ",", "we", "further", "generalize", "the", "SPL", "regime", "into", "multi", "-", "class", "formulation", ",", "which", "facilitates", "the", "learning", "system", "to", "penalize", "indiscriminative", "object", "hypotheses", "predicted", "as", "belonging", "to", "multiple", "object", "categories", "at", "the", "same", "time", ".", "To", "sum", "up", ",", "there", "are", "three", "-", "fold", "contributions", "in", "this", "paper", ":", "", "We", "comprehensively", "analyze", "the", "prospect", "and", "challenges", "in", "the", "idea", "of", "bridging", "saliency", "detection", "to", "WOD", "and", "propose", "an", "effective", "way", "to", "alleviate", "the", "problem", ",", "which", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "detection", "performance", "under", "the", "weak", "supervision", ".", "", "We", "establish", "a", "novel", "SPCL", "regime", "containing", "both", "the", "task", "-", "specific", "learning", "curriculum", "and", "the", "data", "-", "driven", "self", "-", "learning", "pace", ".", "The", "regime", "is", "well", "formulated", "as", "a", "concise", "optimization", "model", ".", "", "We", "incorporate", "SPCL", "with", "a", "sample", "diversity", "term", "and", "further", "generalize", "it", "to", "work", "in", "multi", "-", "class", "scenario", ".", "section", ":", "Related", "Works", "Saliency", "detection", ":", "Most", "saliency", "detection", "methods", "highlight", "the", "attractive", "image", "regions", "by", "exploring", "some", "bottom", "-", "up", "cues", ".", "As", "one", "frequently", "explored", "cue", ",", "local", "contrast", "[", "reference", "][", "reference", "]", "is", "usually", "used", "in", "the", "saliency", "detection", "models", "to", "highlight", "the", "image", "regions", "appearing", "differently", "with", "their", "spatially", "neighbor", "regions", ".", "Another", "widely", "used", "bottom", "-", "up", "cue", "is", "the", "global", "contrast", ".", "Being", "different", "from", "local", "contrast", ",", "global", "contrast", "[", "reference", "][", "reference", "]", "is", "used", "to", "discover", "image", "regions", "which", "are", "unique", "in", "the", "entire", "image", "context", ".", "More", "recently", ",", "background", "prior", "becomes", "another", "important", "cue", "for", "saliency", "detection", ".", "This", "kind", "of", "methods", ",", "e.g.", ",", ",", "assume", "that", "regions", "near", "image", "boundaries", "are", "probably", "backgrounds", "and", "detect", "salient", "regions", "as", "Figure", "1", ":", "This", "figure", "illustrates", "the", "main", "idea", "of", "this", "paper", ".", "As", "can", "be", "seen", ",", "in", "the", "training", "images", "with", "weak", "labels", ",", "some", "of", "them", "(", "in", "blue", "frame", ")", "are", "labelled", "only", "containing", "one", "object", "category", ",", "which", "are", "considered", "as", "\"", "easy", "\"", "images", "for", "the", "saliency", "detection", "methods", ".", "While", "others", "(", "in", "pink", "frame", ")", "labelled", "as", "containing", "multiple", "object", "categories", "are", "considered", "as", "\"", "hard", "\"", "images", ".", "Due", "to", "the", "category", "free", "property", "of", "saliency", "detection", ",", "the", "objects", "in", "\"", "easy", "\"", "images", "have", "larger", "confidence", "to", "be", "extracted", "correctly", "by", "the", "saliency", "detection", "methods", ",", "whereas", "the", "objects", "in", "\"", "hard", "\"", "images", "can", "not", "be", "extracted", "successfully", ".", "To", "this", "end", ",", "we", "develop", "a", "novel", "self", "-", "paced", "curriculum", "learning", "paradigm", "to", "guide", "the", "learner", "to", "gradually", "achieve", "the", "faithful", "knowledge", "of", "the", "multiple", "object", "categories", "from", "easy", "(", "confident", ")", "to", "hard", "(", "ambiguous", ")", ".", "calculating", "the", "contrast", "to", "these", "image", "boundary", "regions", ".", "As", "can", "be", "seen", ",", "the", "bottom", "-", "up", "cues", "explored", "by", "saliency", "detection", "models", "are", "highly", "potential", "to", "provide", "helpful", "priors", "to", "the", "object", "localizations", "in", "each", "image", ".", "Weakly", "-", "supervised", "object", "detection", ":", "Two", "key", "issues", "in", "WOD", "are", "1", ")", "predict", "the", "potential", "object", "localizations", "and", "2", ")", "learn", "the", "object", "detectors", ".", "Some", "early", "WOD", "methods", "[", "reference", "]", "held", "the", "view", "that", "a", "better", "initial", "estimation", "of", "the", "object", "localizations", "is", "critical", "to", "this", "task", "as", "they", "can", "largely", "impact", "the", "subsequent", "learning", "process", ".", "Thus", ",", "they", "explored", "different", "ways", ",", "e.g.", "intra", "-", "class", "similarity", "[", "reference", "]", ",", "inter", "-", "class", "variance", "[", "reference", "]", ",", "and", "distance", "mapping", "relation", "[", "reference", "]", ",", "to", "initialize", "the", "training", "object", "hypotheses", ".", "Later", "on", ",", "some", "recent", "WOD", "methods", "started", "to", "pay", "more", "attention", "to", "the", "optimization", "procedure", "designed", "for", "better", "training", "object", "detectors", "under", "the", "weak", "supervision", ".", "For", "example", ",", "[", "reference", "][", "reference", "]", "proposed", "to", "smooth", "the", "object", "formulation", "to", "better", "obtain", "the", "optimal", "solutions", ".", "[", "reference", "]", "proposed", "to", "incorporate", "convex", "clustering", "in", "the", "learning", "procedure", ",", "which", "enforces", "the", "local", "similarity", "of", "the", "selected", "hypotheses", "during", "optimization", ".", "Essentially", ",", "both", "of", "the", "above", "mentioned", "problems", "are", "critical", "in", "WOD", "task", ".", "To", "this", "end", ",", "this", "paper", "proposes", "a", "novel", "SPCL", "model", "which", "explicitly", "encode", "both", "the", "former", "problem", "(", "with", "the", "designed", "curriculum", ")", "and", "the", "later", "(", "with", "the", "self", "-", "paced", "regularizer", ")", "into", "a", "unified", "formulation", "and", "handle", "both", "problems", "in", "a", "theoretically", "sound", "manner", ".", "section", ":", "Self", "-", "paced", "(", "curriculum", ")", "learning", ":", "Inspired", "by", "the", "learning", "process", "of", "humans", "/", "animals", ",", "the", "theory", "of", "self", "-", "paced", "(", "or", "curriculum", ")", "learning", "[", "reference", "][", "reference", "]", "is", "proposed", "lately", ".", "The", "idea", "is", "to", "learn", "the", "model", "iteratively", "from", "easy", "to", "complex", "samples", "in", "a", "self", "-", "paced", "fashion", ".", "By", "virtue", "of", "its", "generality", ",", "the", "SPL", "theory", "has", "been", "widely", "applied", "to", "various", "tasks", ",", "such", "as", "multi", "-", "view", "clustering", "[", "reference", "]", ",", "multi", "-", "label", "propagation", "[", "reference", "]", ",", "multimedia", "event", "detection", ",", "and", "co", "-", "saliency", "detection", ".", "More", "recently", ",", "introduced", "the", "pre", "-", "defined", "learning", "curriculum", "to", "the", "conventional", "self", "-", "paced", "learning", "regime", "which", "can", "take", "into", "account", "both", "the", "helpful", "prior", "knowledge", "known", "before", "training", "and", "the", "self", "-", "learning", "progress", "during", "training", ".", "Inspired", "by", "this", "work", ",", "we", "design", "a", "task", "-", "specific", "learning", "curriculum", "and", "construct", "a", "unified", "SPCL", "model", "specifically", "for", "both", "the", "saliency", "detection", "and", "WOD", "tasks", ",", "through", "which", "both", "can", "be", "naturally", "related", ".", "section", ":", "The", "Proposed", "Approach", "section", ":", "Algorithm", "Overview", "Given", "K", "training", "images", "with", "weak", "labels", "consisting", "of", "C", "categories", ",", "we", "first", "extract", "the", "bounding", "box", "object", "hypotheses", "and", "their", "corresponding", "feature", "representations", "from", "each", "image", ".", "Denote", "the", "features", "of", "each", "hypothesis", "in", "the", "k", "th", "image", "as", "{", ",", "where", ",", "(", ")", "∈", "[", "0", ",", "1", "]", ",", "indicating", "the", "labels", "and", "the", "real", "-", "valued", "importance", "weights", "of", "each", "hypothesis", ",", "respectively", ",", "which", "are", "unknown", "at", "the", "beginning", "and", "will", "be", "optimized", "during", "the", "proposed", "learning", "regime", ".", "The", "aim", "of", "the", "proposed", "approach", "is", "to", "learn", "the", "object", "detectors", "{", "W", ",", "b", "}", ",", "where", "=", "{", "}", ",", "=", "{", "}", ",", "of", "C", "object", "categories", "from", "the", "weakly", "-", "labeled", "training", "images", ",", "and", "then", "use", "them", "to", "detect", "objects", "in", "the", "test", "images", ".", "Specifically", ",", "we", "first", "design", "a", "simple", "yet", "effective", "curriculum", "to", "select", "the", "salient", "hypotheses", "in", "\"", "easy", "\"", "images", "as", "initialization", ".", "Then", ",", "the", "object", "detectors", "are", "trained", "and", "updated", "gradually", "under", "the", "guidance", "of", "the", "proposed", "self", "-", "paced", "learning", "strategy", ".", "Finally", ",", "the", "obtained", "object", "detectors", "are", "used", "to", "detect", "the", "corresponding", "objects", "in", "the", "test", "images", ".", "The", "overall", "algorithm", "is", "shown", "in", "Algorithm", "1", ".", "section", ":", "Problem", "Formulation", "Given", "object", "hypotheses", "from", "the", "training", "images", ",", "we", "propose", "a", "simple", "yet", "effective", "curriculum", "to", "initialize", "the", "learning", "procedure", ".", "Specifically", ",", "we", "first", "obtain", "the", "\"", "easy", "\"", "images", "based", "on", "the", "number", "of", "weak", "labels", "of", "each", "image", ",", "i.e.", ",", "images", "weakly", "labelled", "as", "only", "containing", "one", "object", "category", "are", "considered", "as", "\"", "easy", "\"", ".", "Then", ",", "for", "each", "\"", "easy", "\"", "image", ",", "we", "adopt", "an", "unsupervised", "saliency", "detection", "method", ",", "i.e.", ",", "RC", "[", "reference", "]", "in", "this", "paper", "due", "to", "its", "efficiency", ",", "to", "generate", "the", "corresponding", "saliency", "estimation", ".", "Finally", ",", "the", "important", "weights", "v", "are", "initialized", "as", "the", "intersection", "-", "over", "-", "union", "(", "IOU", ")", "score", "between", "each", "hypothesis", "and", "the", "salient", "region", ".", "The", "hypotheses", "with", "weights", "larger", "than", "0", "are", "selected", "as", "the", "initial", "training", "hypotheses", "and", "their", "labels", "in", "y", "are", "set", "according", "to", "the", "label", "of", "the", "images", "containing", "them", ".", "Afterwards", ",", "in", "order", "to", "gradually", "adapt", "the", "learner", "from", "the", "\"", "easy", "\"", "domain", "to", "the", "\"", "hard", "\"", "domain", "and", "finally", "capture", "the", "faithful", "knowledge", "of", "the", "objects", "of", "interest", ",", "a", "novel", "self", "-", "paced", "learning", "regularizer", "is", "proposed", "as", "follows", ":", "+", "1", "≤", "2", ",", "enforces", "that", "each", "hypothesis", "should", "belong", "to", "only", "one", "object", "category", ",", "or", "no", "class", ",", "i.e.", ",", "the", "background", "category", ".", "This", "constraint", "inherently", "penalizes", "the", "indiscriminative", "object", "hypotheses", ",", "i.e.", "the", "hypotheses", "predicted", "to", "belong", "to", "multiple", "object", "categories", ",", "when", "calculating", "their", "importance", "weight", "in", "(", "2", ")", ".", "The", "third", "one", ",", "i.e.", "∑", ",", "*", "(", ")", "+", "1", "≥", "2", ",", "means", "that", "for", "all", "object", "hypotheses", "located", "in", "the", "k", "th", "image", ",", "at", "least", "one", "should", "belong", "to", "the", "class", "which", "the", "image", "has", "been", "weakly", "annotated", ".", "This", "will", "make", "the", "learned", "result", "finely", "comply", "with", "the", "prior", "knowledge", ".", "In", "the", "proposed", "SPCL", "regime", ",", "the", "self", "-", "paced", "capability", "is", "followed", "by", "the", "involvement", "of", "the", "SPL", "regularizer", "(", ")", ";", ",", "with", "the", "following", "form", ":", "where", ",", "are", "the", "class", "-", "specific", "parameters", "imposed", "on", "the", "easiness", "term", "and", "the", "diversity", "term", ",", "respectively", ".", "The", "negative", "l", "1", "-", "norm", "term", "is", "inherited", "from", "the", "conventional", "SPL", "[", "reference", "]", ",", "which", "favors", "selecting", "easy", "over", "complex", "hypotheses", ".", "If", "we", "omit", "the", "diversity", "term", ",", "i.e.", "let", "=", "0", ",", "the", "regularizer", "degenerates", "to", "the", "traditional", "hard", "SPL", "function", "proposed", "in", "[", "reference", "]", ",", "which", "conducts", "either", "1", "or", "0", "(", "i.e.", "selected", "in", "training", "or", "not", ")", "for", "the", "weight", ",", "(", ")", "imposed", "on", "hypothesis", "(", ")", ",", "by", "judging", "whether", "its", "loss", "value", "is", "smaller", "than", "the", "pace", "parameter", "or", "not", ".", "That", "is", ",", "a", "sample", "with", "smaller", "loss", "is", "taken", "as", "an", "easy", "sample", "and", "thus", "should", "be", "learned", "preferentially", "and", "vice", "versa", ".", "Another", "regularization", "term", "favors", "selecting", "diverse", "hypotheses", "residing", "in", "more", "images", ".", "This", "can", "be", "easily", "understood", "by", "seeing", "that", "its", "negative", "leads", "to", "the", "group", "-", "wise", "sparse", "representation", "of", "v.", "Contrariwise", ",", "this", "diversity", "term", "should", "have", "a", "counter", "-", "effect", "to", "group", "-", "wise", "sparsity", ".", "That", "is", ",", "minimizing", "this", "diversity", "term", "tends", "to", "disperse", "non", "-", "zero", "elements", "of", "v", "over", "more", "images", ",", "and", "thus", "favors", "selecting", "more", "diverse", "hypotheses", ".", "Consequently", ",", "this", "anti", "-", "group", "-", "sparsity", "representation", "is", "expected", "to", "realize", "the", "desired", "diversity", ".", "Different", "from", "the", "commonly", "utilized", "l", "2", ",", "1", "norm", ",", "our", "utilized", "group", "-", "sparsity", "term", "is", "concave", ",", "leading", "to", "the", "convexity", "of", "its", "negative", ".", "This", "on", "one", "side", "simplifies", "the", "designation", "of", "the", "solving", "strategy", ",", "and", "on", "the", "other", "hand", "well", "fits", "the", "previous", "axiomatic", "definition", "for", "the", "SPL", "regularizer", ".", "section", ":", "Optimization", "Method", "The", "solution", "of", "(", "1", ")", "can", "be", "approximately", "attained", "by", "alternatively", "optimizing", "the", "involved", "parameters", "{", "W", ",", "b", "}", ",", "y", "and", "v", "as", "described", "in", "Algorithm", "1", ".", "The", "optimization", "mainly", "contains", "following", "steps", ":", "Object", "detectors", "updating", ":", "Optimize", "object", "detector", "parameters", "{", "W", ",", "b", "}", "via", "one", "-", "vs", "-", "all", "SVM", "under", "fixed", "y", "and", "v.", "In", "this", "case", ",", "(", "1", ")", "degenerates", "to", "the", "following", "form", ":", ";", ",", ",", "which", "can", "be", "equivalently", "reformulated", "as", "solving", "the", "following", "sub", "-", "optimization", "problems", "for", "each", "c=1", ",", "2", ",", "…", ",", "C", ":", ",", ".", "This", "is", "a", "standard", "one", "-", "vs", "-", "all", "(", "weighted", ")", "SVM", "model", "[", "reference", "]", ".", "section", ":", "Hypotheses", "labelling", ":", "Optimize", "under", "fixed", "{", "W", ",", "b", "}", "and", "v", ":", "The", "goal", "of", "this", "step", "is", "to", "learn", "the", "pseudo", "-", "labels", "of", "training", "hypotheses", "from", "the", "current", "object", "detectors", ".", "The", "model", "in", "this", "case", "can", "be", "reformulated", "as", ":", "This", "problem", "can", "be", "equivalently", "decomposed", "into", "sub", "-", "problems", "with", "respect", "to", "each", "=", "1", ",", "⋯", ",", ",", "i.e.", "for", "each", "image", ",", "where", "c", "*", "is", "the", "weak", "labels", "of", "the", "k", "th", "image", ":", "[", "reference", "]", "indicates", "the", "labels", "of", "the", "hypotheses", "in", "the", "k", "th", "image", ".", "Its", "global", "optimum", "can", "be", "attained", "by", "Algorithm", "2", ",", "which", "can", "be", "derived", "from", "the", "theorem", "in", ".", "Hypotheses", "re", "-", "weighting", ":", "Optimize", "v", "under", "fixed", "{", "W", ",", "b", "}", "and", "y", ":", "After", "updating", "the", "pseudo", "-", "labels", ",", "we", "aim", "to", "renew", "the", "weights", "on", "all", "hypotheses", "to", "reflect", "their", "different", "importance", "to", "learning", "of", "the", "current", "decision", "surface", ".", "In", "this", "case", ",", "(", "1", ")", "degenerates", "to", "the", "following", "form", ":", "which", "is", "equivalent", "to", "independently", "solving", "the", "following", "sub", "-", "optimization", "problem", "for", "each", "=", "1", ",", "⋯", ",", "and", "=", "1", ",", "⋯", ",", "via", ":", "We", "can", "easily", "simplify", "the", "above", "optimization", "problem", "as", ":", "This", "model", "is", "convex", "and", "according", "to", ",", "we", "can", "apdopt", "an", "effective", "algorithm", ",", "i.e.", ",", "the", "Algorithm", "3", ",", "for", "extracting", "the", "global", "optimum", "to", "it", ".", "section", ":", "Experimental", "Results", "section", ":", "Experimental", "Settings", "We", "evaluate", "our", "method", "on", "the", "Pascal", "VOC", "2007", "dataset", "[", "reference", "]", "which", "is", "widely", "used", "by", "the", "previous", "works", ".", "In", "our", "experiments", ",", "we", "follow", "the", "previous", "works", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "to", "discard", "any", "images", "that", "only", "contain", "object", "instances", "marked", "as", "\"", "difficult", "\"", "or", "\"", "truncated", "\"", "during", "the", "training", "phase", ",", "while", "all", "the", "images", "in", "the", "VOC07", "-", "Test", "are", "used", "during", "the", "test", "phase", ".", "For", "fair", "comparison", ",", "we", "follow", "the", "standard", "VOC", "procedure", "[", "reference", "]", "and", "report", "average", "precision", "(", "AP", ")", "on", "the", "Pascal", "VOC", "2007", "test", "split", ".", "Being", "consistent", "with", "the", "recently", "proposed", "WOD", "methods", "[", "reference", "][", "reference", "][", "reference", "]", ",", "we", "apply", "Selective", "Search", "[", "reference", "]", "to", "generate", "around", "1500", "bounding", "box", "object", "hypotheses", "in", "each", "image", "and", "adopt", "the", "CNN", "features", "[", "reference", "]", "pre", "-", "trained", "on", "the", "ImageNet", "2012", "to", "represent", "each", "of", "the", "extracted", "object", "hypotheses", ".", "Before", "training", ",", "and", "in", "(", "2", ")", "need", "to", "be", "set", "in", "advance", ".", "As", "suggested", "in", ",", "we", "set", "according", "to", "the", "number", "of", "the", "selected", "training", "hypotheses", "which", "is", "set", "to", "be", "2", "%", "of", "the", "total", "bounding", "box", "windows", "extracted", "from", "the", "images", "weakly", "labelled", "as", "containing", "the", "c", "th", "object", "category", ".", "Then", ",", "is", "set", "to", "be", "equal", "to", "empirically", ".", "section", ":", "Comparison", "to", "the", "State", "-", "of", "-", "the", "-", "arts", "In", "this", "section", ",", "we", "evaluate", "the", "object", "detection", "performance", "of", "our", "framework", "by", "comparing", "it", "with", "6", "state", "-", "of", "-", "the", "-", "art", "WOD", "approaches", "which", "are", "PR", "[", "reference", "]", ",", "CC", "[", "reference", "]", ",", "MDD", "[", "reference", "]", ",", "LLO", "[", "reference", "]", ",", "VPC", "[", "reference", "]", ",", "and", "MfMIL", "[", "reference", "]", ".", "For", "quantitative", "comparison", ",", "we", "report", "the", "evaluation", "results", "in", "terms", "of", "the", "AP", "score", "in", "Fig", ".", "2", ".", "As", "can", "be", "seen", ",", "the", "proposed", "approach", "obtains", "the", "highest", "score", "of", "29.96", "on", "average", ".", "According", "to", "our", "analysis", ",", "the", "proposed", "approach", "can", "obtain", "significantly", "better", "results", "than", "MDD", "and", "MfMIL", "mainly", "due", "to", "the", "better", "feature", "representation", ",", "stronger", "saliency", "prior", ",", "and", "more", "powerful", "learning", "scheme", ".", "Compared", "with", "PR", ",", "CC", ",", "LLO", ",", "and", "VPC", ",", "the", "performance", "gain", "of", "the", "proposed", "approach", "mainly", "comes", "from", "the", "core", "insight", "of", "this", "paper", ",", "i.e.", ",", "developing", "property", "way", "to", "bridge", "saliency", "detection", "to", "WOD", ",", "as", "we", "used", "the", "same", "feature", "representation", "with", "these", "methods", ".", "More", "specifically", ",", "compared", "with", "PR", "and", "CC", ",", "the", "performance", "gain", "of", "the", "proposed", "approach", "comes", "mainly", "from", "the", "idea", "to", "bridge", "saliency", "detection", "to", "WOD", "because", "they", "only", "adopted", "weak", "priors", "in", "their", "initialization", ".", "Compared", "with", "LLO", "and", "VPC", ",", "the", "performance", "gain", "of", "the", "proposed", "approach", "mainly", "comes", "from", "the", "proposed", "SPCL", "regime", "as", "these", "two", "methods", "also", "explored", "strong", "prior", "information", "for", "initializing", "the", "training", "hypotheses", "in", "their", "frameworks", ".", "Some", "examples", "of", "the", "detection", "results", "are", "also", "shown", "in", "Fig", ".", "2", ",", "which", "includes", "some", "successful", "cases", ",", "i.e.", ",", "the", "examples", "in", "the", "bus", "and", "cat", "categories", ",", "as", "well", "as", "some", "failure", "cases", ",", "i.e.", ",", "examples", "in", "the", "plant", "and", "chair", "categories", ".", "The", "successful", "cases", "subjectively", "demonstrate", "the", "effectiveness", "of", "the", "proposed", "approach", ".", "For", "the", "failure", "cases", ",", "the", "main", "problem", "is", "that", "very", "limited", "number", "of", "images", "only", "contains", "the", "objects", "like", "plant", "and", "chair", ",", "leading", "to", "the", "insufficient", "training", "hypotheses", "in", "the", "initialization", "stage", ".", "This", "problem", "can", "be", "solved", "by", "designing", "more", "proper", "learning", "curriculum", "for", "WOD", "in", "future", "works", ".", "section", ":", "Model", "Analysis", "To", "further", "analyze", "the", "proposed", "framework", "in", "this", "paper", ",", "we", "make", "more", "comprehensive", "evaluations", "in", "this", "section", "by", "comparing", "with", "five", "baseline", "models", "as", "described", "in", "Table", "1", ".", "The", "experimental", "results", "are", "shown", "in", "Fig", ".", "3", ",", "from", "which", "we", "can", "see", ":", "1", ")", "The", "performance", "gap", "between", "Sal", "+", "SVM", "and", "OURS", "demonstrates", "the", "importance", "to", "develop", "proper", "ways", "to", "bridge", "saliency", "detection", "and", "WOD", ".", "2", ")", "The", "experimental", "results", "of", "Sal", "+", "SPL", ",", "Sal", "+", "SPCL", ",", "and", "OURS", "demonstrate", "the", "better", "performance", "of", "the", "proposed", "learning", "regime", "as", "compared", "with", "some", "existing", "self", "-", "paced", "(", "curriculum", ")", "learning", "regimes", ".", "3", ")", "The", "performance", "gap", "between", "OURS", "and", "LLO", "+", "SPCL", "*", "demonstrates", "the", "saliency", "prior", "can", "provide", "more", "helpful", "information", "than", "the", "prior", "designed", "in", "LLO", ".", "4", ")", "The", "performance", "gap", "between", "LLO", "+", "SPCL", "*", "and", "LLO", "indicates", "the", "better", "capability", "of", "the", "proposed", "learning", "regime", "as", "compared", "with", "the", "learning", "model", "in", "one", "state", "-", "of", "-", "the", "-", "art", "WOD", "framework", ".", "According", "to", "the", "above", "analysis", ",", "the", "key", "insight", "of", "this", "paper", ",", "i.e.", ",", "developing", "powerful", "learning", "regime", ",", "i.e.", ",", "the", "proposed", "SPCL", ",", "can", "better", "bridge", "saliency", "detection", "to", "WOD", "and", "help", "the", "learner", "to", "capture", "the", "faithful", "knowledge", "of", "the", "object", "categories", "under", "weak", "supervision", ",", "has", "been", "demonstrated", "comprehensively", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "in", "order", "to", "address", "the", "challenging", "WOD", "problem", ",", "we", "proposed", "an", "effective", "framework", "to", "bridge", "saliency", "detection", "to", "the", "investigated", "task", "based", "on", "a", "novel", "SPCL", "regime", ".", "The", "insight", "of", "this", "paper", "is", "that", "by", "developing", "powerful", "learning", "regime", "which", "contains", "both", "the", "task", "-", "specific", "learning", "curriculum", "and", "the", "data", "-", "driven", "self", "-", "learning", "pace", ",", "saliency", "detection", "technique", "can", "be", "better", "leveraged", "to", "provide", "beneficial", "information", "for", "helping", "the", "learner", "to", "capture", "the", "faithful", "knowledge", "of", "the", "object", "categories", "under", "weak", "supervision", ".", "Experiments", "including", "comparisons", "to", "other", "state", "-", "of", "-", "the", "-", "arts", "and", "comprehensive", "analysis", "of", "the", "proposed", "framework", "on", "the", "benchmark", "dataset", "have", "demonstrated", "the", "effectiveness", "of", "our", "approach", ".", "For", "the", "future", "work", ",", "inspired", "by", "[", "reference", "]", ",", "we", "plane", "to", "enable", "the", "proposed", "method", "to", "transfer", "the", "knowledge", "that", "has", "be", "captured", "to", "new", "concepts", "via", "novel", "regularizers", ".", ".", "LLO", "Baseline", "WOD", "framework", "[", "reference", "]", ".", "LLO", "+", "SPCL", "*", "Replace", "the", "learning", "model", ",", "i.e.", ",", "SLSVM", ",", "in", "LLO", "with", "the", "proposed", "SPCL", "regime", ".", "section", ":", "section", ":", "Acknowledgments", ":", "This", "work", "was", "supported", "in", "part", "by", "the", "National", "Science", "Foundation", "of", "China", "under", "Grants", "61522207", "and", "61473231", ",", "the", "Doctorate", "Foundation", ",", "and", "the", "Excellent", "Doctorate", "Foundation", "of", "Northwestern", "Polytechnical", "University", ".", "section", ":"], "entity": [[0, 3, ["Bridging", "Saliency", "Detection"], "Task"], [4, 8, ["Weakly", "Supervised", "Object", "Detection"], "Task"], [10, 15, ["Self", "-", "paced", "Curriculum", "Learning"], "Method"], [18, 23, ["Weakly", "-", "supervised", "object", "detection"], "Task"], [31, 33, ["computer", "vision"], "Task"], [52, 54, ["object", "detectors"], "Method"], [72, 75, ["selective", "attention", "mechanism"], "Method"], [76, 79, ["human", "visual", "system"], "Method"], [100, 101, ["WOD"], "Task"], [108, 110, ["saliency", "detection"], "Task"], [111, 112, ["WOD"], "Task"], [142, 144, ["saliency", "detection"], "Task"], [145, 146, ["WOD"], "Task"], [158, 160, ["saliency", "detection"], "Task"], [161, 162, ["WOD"], "Task"], [164, 169, ["self", "-", "paced", "curriculum", "learning"], "Method"], [202, 204, ["saliency", "detection"], "Task"], [205, 207, ["WOD", "tasks"], "Task"], [217, 219, ["object", "detection"], "Task"], [229, 231, ["Object", "detection"], "Task"], [240, 243, ["computer", "vision", "community"], "Task"], [259, 265, ["Convolutional", "Neural", "Network", "(", "CNN", ")"], "Method"], [278, 280, ["feature", "representations"], "Method"], [281, 283, ["object", "classifiers"], "Method"], [291, 293, ["object", "detection"], "Task"], [322, 324, ["intelligent", "machines"], "Method"], [348, 352, ["weakly", "supervised", "object", "detection"], "Task"], [353, 354, ["WOD"], "Task"], [367, 368, ["WOD"], "Task"], [374, 376, ["object", "localizations"], "Task"], [380, 382, ["object", "detectors"], "Method"], [391, 395, ["chicken", "-", "egg", "problem"], "Task"], [408, 410, ["learning", "strategy"], "Method"], [414, 416, ["coarse", "estimation"], "Task"], [427, 429, ["object", "detectors"], "Method"], [441, 443, ["saliency", "detection"], "Task"], [457, 459, ["Saliency", "detection"], "Task"], [469, 472, ["selective", "attention", "mechanism"], "Method"], [473, 476, ["human", "visual", "system"], "Method"], [509, 512, ["potential", "object", "localizations"], "Task"], [574, 576, ["WOD", "systems"], "Task"], [594, 595, ["WOD"], "Task"], [627, 629, ["learning", "regime"], "Method"], [641, 643, ["saliency", "detection"], "Task"], [644, 645, ["WOD"], "Task"], [662, 664, ["saliency", "detection"], "Task"], [667, 671, ["category", "-", "free", "models"], "Method"], [777, 779, ["saliency", "models"], "Method"], [793, 795, ["salient", "detection"], "Task"], [802, 803, ["WOD"], "Task"], [812, 814, ["saliency", "detection"], "Task"], [815, 816, ["WOD"], "Task"], [826, 828, ["saliency", "detection"], "Task"], [829, 830, ["WOD"], "Task"], [832, 837, ["self", "-", "paced", "curriculum", "learning"], "Method"], [838, 839, ["SPCL"], "Method"], [842, 843, ["SPCL"], "Method"], [849, 851, ["learning", "framework"], "Method"], [854, 856, ["curriculum", "learning"], "Method"], [857, 858, ["CL"], "Method"], [860, 864, ["self", "-", "paced", "learning"], "Method"], [865, 866, ["SPL"], "Method"], [880, 882, ["learning", "components"], "Method"], [887, 889, ["saliency", "detection"], "Task"], [890, 891, ["WOD"], "Task"], [905, 906, ["CL"], "Method"], [924, 926, ["predetermined", "heuristics"], "Method"], [930, 931, ["SPL"], "Method"], [965, 966, ["CL"], "Method"], [967, 968, ["SPL"], "Method"], [970, 971, ["SPCL"], "Method"], [978, 981, ["training", "sample", "initialization"], "Task"], [982, 986, ["object", "detector", "updating", "problems"], "Task"], [989, 993, ["saliency", "-", "guided", "WOD"], "Task"], [996, 997, ["SPCL"], "Method"], [1027, 1029, ["learning", "procedure"], "Task"], [1135, 1137, ["object", "detectors"], "Method"], [1146, 1148, ["SPCL", "regime"], "Method"], [1162, 1164, ["SPCL", "model"], "Method"], [1167, 1169, ["learning", "regime"], "Method"], [1183, 1188, ["task", "-", "specific", "learning", "curriculum"], "Method"], [1193, 1194, ["WOD"], "Task"], [1211, 1215, ["self", "-", "paced", "regularizer"], "Method"], [1254, 1258, ["multi", "-", "class", "formulation"], "Method"], [1262, 1264, ["learning", "system"], "Method"], [1311, 1312, ["WOD"], "Task"], [1332, 1334, ["detection", "performance"], "Metric"], [1344, 1346, ["SPCL", "regime"], "Method"], [1349, 1354, ["task", "-", "specific", "learning", "curriculum"], "Method"], [1356, 1363, ["data", "-", "driven", "self", "-", "learning", "pace"], "Method"], [1372, 1374, ["optimization", "model"], "Method"], [1378, 1379, ["SPCL"], "Method"], [1391, 1395, ["multi", "-", "class", "scenario"], "Task"], [1400, 1402, ["Saliency", "detection"], "Task"], [1504, 1506, ["saliency", "detection"], "Task"], [1613, 1615, ["saliency", "detection"], "Task"], [1723, 1725, ["object", "localizations"], "Task"], [1729, 1734, ["Weakly", "-", "supervised", "object", "detection"], "Task"], [1739, 1740, ["WOD"], "Task"], [1745, 1748, ["potential", "object", "localizations"], "Task"], [1753, 1755, ["object", "detectors"], "Method"], [1773, 1775, ["object", "localizations"], "Task"], [1787, 1789, ["learning", "process"], "Task"], [1843, 1845, ["optimization", "procedure"], "Method"], [1848, 1851, ["training", "object", "detectors"], "Task"], [1868, 1870, ["object", "formulation"], "Method"], [1883, 1885, ["convex", "clustering"], "Method"], [1887, 1889, ["learning", "procedure"], "Method"], [1900, 1901, ["optimization"], "Task"], [1913, 1915, ["WOD", "task"], "Task"], [1925, 1927, ["SPCL", "model"], "Method"], [1938, 1939, ["curriculum"], "Method"], [1946, 1950, ["self", "-", "paced", "regularizer"], "Method"], [1967, 1974, ["Self", "-", "paced", "(", "curriculum", ")", "learning"], "Method"], [1978, 1980, ["learning", "process"], "Method"], [1988, 1996, ["self", "-", "paced", "(", "or", "curriculum", ")", "learning"], "Method"], [2044, 2048, ["multi", "-", "view", "clustering"], "Task"], [2052, 2056, ["multi", "-", "label", "propagation"], "Task"], [2060, 2063, ["multimedia", "event", "detection"], "Task"], [2078, 2080, ["learning", "curriculum"], "Method"], [2118, 2123, ["task", "-", "specific", "learning", "curriculum"], "Method"], [2126, 2129, ["unified", "SPCL", "model"], "Method"], [2133, 2135, ["saliency", "detection"], "Task"], [2136, 2137, ["WOD"], "Task"], [2179, 2181, ["feature", "representations"], "Method"], [2239, 2241, ["learning", "regime"], "Method"], [2252, 2254, ["object", "detectors"], "Method"], [2319, 2321, ["object", "detectors"], "Method"], [2342, 2344, ["object", "detectors"], "Method"], [2383, 2384, ["curriculum"], "Method"], [2387, 2389, ["learning", "procedure"], "Method"], [2448, 2449, ["RC"], "Method"], [2464, 2466, ["saliency", "estimation"], "Task"], [2477, 2482, ["intersection", "-", "over", "-", "union"], "Metric"], [2483, 2484, ["IOU"], "Metric"], [2687, 2689, ["SPCL", "regime"], "Method"], [2745, 2746, ["SPL"], "Method"], [2771, 2772, ["regularizer"], "Method"], [2851, 2853, ["regularization", "term"], "Method"], [2875, 2880, ["group", "-", "wise", "sparse", "representation"], "Method"], [2930, 2936, ["anti", "-", "group", "-", "sparsity", "representation"], "Method"], [2949, 2954, ["l", "2", ",", "1", "norm"], "Method"], [2957, 2961, ["group", "-", "sparsity", "term"], "Method"], [2981, 2983, ["solving", "strategy"], "Method"], [3002, 3004, ["Optimization", "Method"], "Method"], [3036, 3037, ["optimization"], "Task"], [3042, 3045, ["Object", "detectors", "updating"], "Method"], [3056, 3062, ["one", "-", "vs", "-", "all", "SVM"], "Method"], [3092, 3096, ["sub", "-", "optimization", "problems"], "Task"], [3112, 3122, ["one", "-", "vs", "-", "all", "(", "weighted", ")", "SVM", "model"], "Method"], [3128, 3130, ["Hypotheses", "labelling"], "Task"], [3160, 3162, ["object", "detectors"], "Method"], [3245, 3249, ["Hypotheses", "re", "-", "weighting"], "Task"], [3312, 3316, ["sub", "-", "optimization", "problem"], "Task"], [3337, 3339, ["optimization", "problem"], "Task"], [3384, 3388, ["Pascal", "VOC", "2007", "dataset"], "Material"], [3465, 3467, ["VOC", "procedure"], "Method"], [3472, 3474, ["average", "precision"], "Metric"], [3475, 3476, ["AP"], "Metric"], [3479, 3484, ["Pascal", "VOC", "2007", "test", "split"], "Material"], [3503, 3505, ["Selective", "Search"], "Method"], [3522, 3524, ["CNN", "features"], "Method"], [3629, 3631, ["object", "detection"], "Task"], [3647, 3649, ["WOD", "approaches"], "Method"], [3651, 3652, ["PR"], "Method"], [3656, 3657, ["CC"], "Method"], [3661, 3662, ["MDD"], "Method"], [3666, 3667, ["LLO"], "Method"], [3671, 3672, ["VPC"], "Method"], [3677, 3678, ["MfMIL"], "Method"], [3733, 3734, ["MDD"], "Method"], [3735, 3736, ["MfMIL"], "Method"], [3741, 3743, ["feature", "representation"], "Method"], [3751, 3753, ["learning", "scheme"], "Method"], [3756, 3757, ["PR"], "Method"], [3758, 3759, ["CC"], "Method"], [3760, 3761, ["LLO"], "Method"], [3763, 3764, ["VPC"], "Method"], [3789, 3791, ["saliency", "detection"], "Task"], [3792, 3793, ["WOD"], "Task"], [3799, 3801, ["feature", "representation"], "Method"], [3810, 3811, ["PR"], "Method"], [3812, 3813, ["CC"], "Method"], [3828, 3830, ["saliency", "detection"], "Task"], [3831, 3832, ["WOD"], "Task"], [3844, 3845, ["LLO"], "Method"], [3846, 3847, ["VPC"], "Method"], [3860, 3862, ["SPCL", "regime"], "Method"], [3884, 3885, ["detection"], "Task"], [3972, 3974, ["initialization", "stage"], "Method"], [3984, 3986, ["learning", "curriculum"], "Method"], [3987, 3988, ["WOD"], "Task"], [4062, 4064, ["saliency", "detection"], "Task"], [4065, 4066, ["WOD"], "Task"], [4073, 4074, ["Sal"], "Method"], [4075, 4076, ["SPL"], "Method"], [4077, 4078, ["Sal"], "Method"], [4079, 4080, ["SPCL"], "Method"], [4090, 4092, ["learning", "regime"], "Method"], [4097, 4105, ["self", "-", "paced", "(", "curriculum", ")", "learning", "regimes"], "Method"], [4114, 4115, ["LLO"], "Method"], [4116, 4117, ["SPCL"], "Method"], [4132, 4133, ["LLO"], "Method"], [4140, 4141, ["LLO"], "Method"], [4142, 4143, ["SPCL"], "Method"], [4145, 4146, ["LLO"], "Method"], [4153, 4155, ["learning", "regime"], "Method"], [4159, 4161, ["learning", "model"], "Method"], [4190, 4192, ["learning", "regime"], "Method"], [4197, 4198, ["SPCL"], "Method"], [4202, 4204, ["saliency", "detection"], "Task"], [4205, 4206, ["WOD"], "Task"], [4241, 4243, ["WOD", "problem"], "Task"], [4251, 4253, ["saliency", "detection"], "Task"], [4261, 4263, ["SPCL", "regime"], "Method"], [4274, 4276, ["learning", "regime"], "Method"], [4287, 4294, ["data", "-", "driven", "self", "-", "learning", "pace"], "Method"], [4385, 4386, ["regularizers"], "Method"], [4388, 4389, ["LLO"], "Method"], [4396, 4397, ["LLO"], "Method"], [4398, 4399, ["SPCL"], "Method"], [4402, 4404, ["learning", "model"], "Method"], [4407, 4408, ["SLSVM"], "Method"], [4410, 4411, ["LLO"], "Method"], [4414, 4416, ["SPCL", "regime"], "Method"], [24, 25, ["WOD"], "Task"], [80, 82, ["saliency", "detection"], "Task"], [524, 526, ["saliency", "detection"], "Task"], [590, 592, ["saliency", "detection"], "Task"], [721, 723, ["saliency", "detection"], "Task"], [1111, 1115, ["self", "-", "paced", "learning"], "Method"], [1190, 1192, ["saliency", "detection"], "Task"], [1251, 1252, ["SPL"], "Method"], [1308, 1310, ["saliency", "detection"], "Task"], [1404, 1406, ["saliency", "detection"], "Task"], [1438, 1440, ["saliency", "detection"], "Task"], [1581, 1583, ["saliency", "detection"], "Task"], [1632, 1634, ["saliency", "detection"], "Task"], [1658, 1663, ["self", "-", "paced", "curriculum", "learning"], "Method"], [1711, 1713, ["saliency", "detection"], "Task"], [1758, 1759, ["WOD"], "Task"], [1834, 1835, ["WOD"], "Task"], [2032, 2033, ["SPL"], "Method"], [2067, 2069, ["saliency", "detection"], "Task"], [2083, 2087, ["self", "-", "paced", "learning"], "Method"], [2332, 2336, ["self", "-", "paced", "learning"], "Method"], [2442, 2444, ["saliency", "detection"], "Task"], [2560, 2564, ["self", "-", "paced", "learning"], "Method"], [2702, 2703, ["SPL"], "Method"], [2777, 2778, ["SPL"], "Method"], [2997, 2998, ["SPL"], "Method"], [3491, 3492, ["WOD"], "Task"], [3695, 3696, ["AP"], "Metric"], [4048, 4049, ["Sal"], "Method"], [4170, 4171, ["WOD"], "Task"], [4280, 4285, ["task", "-", "specific", "learning", "curriculum"], "Method"], [4295, 4297, ["saliency", "detection"], "Task"], [4390, 4391, ["WOD"], "Task"]]} {"doc_id": "05d2700846c0323f79c1344aca5333994c7c03a5", "text": ["document", ":", "The", "IBM", "2016", "English", "Conversational", "Telephone", "Speech", "Recognition", "System", "We", "describe", "a", "collection", "of", "acoustic", "and", "language", "modeling", "techniques", "that", "lowered", "the", "word", "error", "rate", "of", "our", "English", "conversational", "telephone", "LVCSR", "system", "to", "a", "record", "6.6", "%", "on", "the", "Switchboard", "subset", "of", "the", "Hub5", "2000", "evaluation", "testset", ".", "On", "the", "acoustic", "side", ",", "we", "use", "a", "score", "fusion", "of", "three", "strong", "models", ":", "recurrent", "nets", "with", "maxout", "activations", ",", "very", "deep", "convolutional", "nets", "with", "3x3", "kernels", ",", "and", "bidirectional", "long", "short", "-", "term", "memory", "nets", "which", "operate", "on", "FMLLR", "and", "i", "-", "vector", "features", ".", "On", "the", "language", "modeling", "side", ",", "we", "use", "an", "updated", "model", "“", "M", "”", "and", "hierarchical", "neural", "network", "LMs", ".", "GeorgeSaon", ",", "TomSercu", ",", "StevenRennieandHong", "-", "KwangJ.Kuo", "IBMT.J.WatsonResearchCenter", ",", "YorktownHeights", ",", "NY", ",", "10598", "gsaon@us.ibm.com", "©", "IEEE2004", "SubmittedtoInterspeech2016", ",", "pleasedonotredistribute", "Index", "Terms", ":", "recurrent", "neural", "networks", ",", "convolutional", "neural", "networks", ",", "conversational", "speech", "recognition", "section", ":", "Introduction", "The", "landscape", "of", "neural", "network", "acoustic", "modeling", "is", "rapidly", "evolving", ".", "Spurred", "by", "the", "success", "of", "deep", "feed", "-", "forward", "neural", "nets", "for", "LVCSR", "in", "and", "inspired", "by", "other", "research", "areas", "like", "image", "classification", "and", "natural", "language", "processing", ",", "many", "speech", "groups", "have", "looked", "at", "more", "sophisticated", "architectures", "such", "as", "deep", "convolutional", "nets", ",", "deep", "recurrent", "nets", ",", "time", "-", "delay", "neural", "nets", ",", "and", "long", "-", "short", "term", "memory", "nets", ".", "The", "trend", "is", "to", "remove", "a", "lot", "of", "the", "complexity", "and", "human", "knowledge", "that", "was", "necessary", "in", "the", "past", "to", "build", "good", "ASR", "systems", "(", "e.g.", "speaker", "adaptation", ",", "phonetic", "context", "modeling", ",", "discriminative", "feature", "processing", ",", "etc", ".", ")", "and", "to", "replace", "them", "with", "a", "powerful", "neural", "network", "architecture", "that", "can", "be", "trained", "agnostically", "on", "a", "lot", "of", "data", ".", "With", "the", "advent", "of", "numerous", "neural", "network", "toolkits", "which", "can", "implement", "these", "sophisticated", "models", "out", "-", "of", "-", "the", "-", "box", "and", "powerful", "hardware", "based", "on", "GPUs", ",", "the", "barrier", "of", "entry", "for", "building", "high", "performing", "ASR", "systems", "has", "been", "lowered", "considerably", ".", "First", "case", "in", "point", ":", "front", "-", "end", "processing", "has", "been", "simplified", "considerably", "with", "the", "use", "of", "CNNs", "which", "treat", "the", "log", "-", "mel", "spectral", "representation", "as", "an", "image", "and", "do", "n’t", "require", "extra", "processing", "steps", "such", "as", "PLP", "cepstra", ",", "LDA", ",", "FMLLR", ",", "fMPE", "transforms", ",", "etc", ".", "Second", "case", "in", "point", ":", "end", "-", "to", "-", "end", "ASR", "systems", "such", "as", "bypass", "the", "need", "of", "having", "phonetic", "context", "decision", "trees", "and", "HMMs", "altogether", "and", "directly", "map", "the", "sequence", "of", "acoustic", "features", "to", "a", "sequence", "of", "characters", "or", "context", "independent", "phones", ".", "Third", "case", "in", "point", ":", "training", "algorithms", "such", "as", "connectionist", "temporal", "classification", "do", "n’t", "require", "an", "initial", "alignment", "of", "the", "training", "data", "which", "is", "typically", "done", "with", "a", "GMM", "-", "based", "baseline", "model", ".", "The", "above", "points", "beg", "the", "question", "whether", ",", "in", "this", "age", "of", "readily", "available", "NN", "toolkits", ",", "speech", "recognition", "expertise", "is", "still", "necessary", "or", "whether", "one", "can", "simply", "point", "a", "neural", "net", "to", "the", "audio", "and", "transcripts", ",", "let", "it", "train", ",", "and", "obtain", "a", "good", "acoustic", "model", ".", "While", "it", "is", "true", "that", ",", "as", "the", "amount", "of", "training", "data", "increases", ",", "the", "need", "for", "human", "ASR", "expertise", "is", "lessened", ",", "at", "the", "moment", "the", "performance", "of", "end", "-", "to", "-", "end", "systems", "ultimately", "remains", "inferior", "to", "that", "of", "more", "traditional", ",", "i.e.", "HMM", "and", "decision", "tree", "-", "based", ",", "approaches", ".", "Since", "the", "goal", "of", "this", "work", "is", "to", "obtain", "the", "lowest", "possible", "WER", "on", "the", "Switchboard", "dataset", "regardless", "of", "other", "practical", "considerations", "such", "as", "speed", "and", "/", "or", "simplicity", ",", "we", "have", "focused", "on", "the", "latter", "approaches", ".", "The", "paper", "is", "organized", "as", "follows", ":", "in", "section", "[", "reference", "]", "we", "discuss", "acoustic", "and", "language", "modeling", "improvements", "and", "in", "section", "[", "reference", "]", "we", "summarize", "our", "findings", ".", "section", ":", "System", "improvements", "In", "this", "section", "we", "describe", "three", "different", "acoustic", "models", "that", "were", "trained", "on", "2000", "hours", "of", "English", "conversational", "telephone", "speech", ":", "recurrent", "nets", "with", "maxout", "activations", "and", "annealed", "dropout", ",", "very", "deep", "convolutional", "nets", "with", "3", "3", "kernels", ",", "and", "bidirectional", "long", "short", "-", "term", "memory", "nets", "operating", "on", "FMLLR", "and", "i", "-", "vector", "features", ".", "All", "models", "are", "used", "in", "a", "hybrid", "HMM", "decoding", "scenario", "by", "subtracting", "the", "logarithm", "of", "the", "HMM", "state", "priors", "from", "the", "log", "of", "the", "softmax", "output", "scores", ".", "The", "training", "and", "test", "data", ",", "frontend", "processing", ",", "speaker", "adaptation", "are", "identical", "to", "and", "their", "description", "will", "be", "omitted", ".", "At", "the", "end", "of", "the", "section", ",", "we", "also", "provide", "an", "update", "on", "our", "vocabulary", "and", "language", "modeling", "experiments", ".", "subsection", ":", "Recurrent", "nets", "with", "maxout", "activations", "We", "remind", "the", "reader", "that", "maxout", "nets", "generalize", "ReLu", "units", "by", "employing", "non", "-", "linearities", "of", "the", "form", "where", "the", "subsets", "of", "neurons", "are", "typically", "disjoint", ".", "In", "we", "have", "shown", "that", "maxout", "DNNs", "and", "CNNs", "trained", "with", "annealed", "dropout", "outperform", "their", "sigmoid", "-", "based", "counterparts", "on", "both", "300", "hours", "and", "2000", "hours", "training", "regimes", ".", "What", "was", "missing", "there", "was", "a", "comparison", "between", "maxout", "and", "sigmoid", "for", "unfolded", "RNNs", ".", "The", "architecture", "of", "the", "maxout", "RNNs", "comprises", "one", "recurrent", "layer", "with", "2828", "units", "projected", "to", "1414", "units", "via", "non", "-", "overlapping", "maxout", "operations", ".", "This", "layer", "is", "followed", "by", "4", "non", "-", "recurrent", "layers", "with", "2828", "units", "(", "also", "projected", "to", "1414", ")", "followed", "by", "a", "bottleneck", "with", "1024", "512", "units", "and", "an", "output", "layer", "with", "32000", "neurons", "corresponding", "to", "as", "many", "context", "-", "dependent", "HMM", "states", ".", "The", "number", "of", "neurons", "for", "the", "maxout", "layers", "have", "been", "chosen", "such", "that", "the", "weight", "matrices", "have", "roughly", "the", "same", "number", "of", "parameters", "as", "the", "baseline", "sigmoid", "network", "which", "has", "2048", "units", "per", "hidden", "layer", ".", "The", "recurrent", "layer", "is", "unfolded", "backwards", "in", "time", "for", "6", "time", "steps", "and", "has", "340", "-", "dimensional", "inputs", "consisting", "of", "6", "spliced", "right", "context", "40", "-", "dimensional", "FMLLR", "frames", "(", ")", "to", "which", "we", "append", "a", "100", "-", "dimensional", "speaker", "-", "based", "ivector", ".", "The", "unfolded", "maxout", "RNN", "architecture", "is", "depicted", "in", "Figure", "[", "reference", "]", ".", "The", "network", "is", "trained", "one", "hidden", "layer", "at", "a", "time", "with", "discriminative", "pretraining", "followed", "by", "12", "epochs", "of", "SGD", "CE", "training", "on", "randomized", "minibatches", "of", "250", "samples", ".", "The", "model", "is", "refined", "with", "Hessian", "-", "free", "sequence", "discriminative", "training", "using", "the", "state", "-", "based", "MBR", "criterion", "for", "10", "iterations", ".", "In", "Table", "[", "reference", "]", "we", "report", "the", "error", "rates", "for", "sigmoid", "and", "maxout", "RNNs", "on", "the", "Switchboard", "and", "CallHome", "subsets", "of", "Hub5’00", ".", "The", "decodings", "are", "done", "with", "a", "small", "vocabulary", "of", "30", "K", "words", "and", "a", "small", "4", "-", "gram", "language", "model", "with", "4", "M", "n", "-", "grams", ".", "Note", "that", "the", "sigmoid", "RNNs", "have", "better", "error", "rates", "than", "what", "was", "reported", "in", "because", "they", "have", "been", "retrained", "after", "the", "data", "has", "been", "realigned", "with", "the", "best", "joint", "RNN", "/", "CNN", "model", ".", "We", "observe", "that", "the", "maxout", "RNNs", "are", "consistently", "better", "and", "that", ",", "by", "themselves", ",", "they", "achieve", "a", "similar", "WER", "as", "our", "previous", "best", "model", "which", "was", "the", "joint", "RNN", "/", "CNN", "with", "sigmoid", "activations", ".", "subsection", ":", "Very", "deep", "convolutional", "networks", "Very", "deep", "CNNs", "with", "small", "kernels", "have", "recently", "been", "shown", "to", "achieve", "very", "strong", "performance", "as", "acoustic", "models", "in", "hybrid", "NN", "-", "HMM", "speech", "recognition", "systems", ".", "Results", "were", "provided", "after", "cross", "-", "entropy", "training", "on", "the", "300", "hours", "switchboard", "-", "1", "dataset", "in", ",", "and", "results", "from", "sequence", "training", "on", "both", "switchboard", "-", "1", "and", "the", "2000", "hours", "switchboard", "+", "Fisher", "dataset", "are", "in", ".", "The", "very", "deep", "convolutional", "networks", "are", "inspired", "by", "the", "“", "VGG", "Net", "”", "architecture", "introduced", "in", "for", "the", "2014", "ImageNet", "classification", "challenge", ",", "with", "the", "central", "idea", "to", "replace", "large", "convolutional", "kernels", "by", "small", "kernels", ".", "By", "stacking", "many", "of", "these", "convolutional", "layers", "with", "ReLU", "nonlinearities", "before", "pooling", "layers", ",", "the", "same", "receptive", "field", "is", "created", "with", "less", "parameters", "and", "more", "nonlinearity", ".", "Figure", "[", "reference", "]", "shows", "the", "design", "of", "the", "networks", ".", "Note", "that", "as", "we", "go", "deeper", "in", "the", "network", ",", "the", "time", "and", "frequency", "resolution", "is", "reduced", "through", "pooling", "only", ",", "while", "the", "convolutions", "are", "zero", "-", "padded", "as", "to", "not", "reduce", "the", "size", "of", "the", "feature", "maps", ".", "We", "increase", "the", "number", "of", "feature", "maps", "gradually", "from", "64", "to", "512", "(", "indicated", "by", "the", "different", "colors", ")", ".", "We", "pool", "right", "before", "the", "layer", "that", "increases", "the", "number", "of", "feature", "maps", ".", "Note", "that", "the", "indication", "of", "feature", "map", "size", "on", "the", "right", "only", "applies", "to", "the", "rightmost", "2", "designs", ".", "In", "contrast", ",", "the", "classical", "CNN", "architecture", "has", "only", "two", "layers", ",", "goes", "to", "512", "feature", "maps", "directly", ",", "and", "uses", "a", "large", "kernel", "on", "the", "first", "layer", ".", "Our", "10", "-", "layer", "CNN", "has", "about", "the", "same", "number", "of", "parameters", "as", "the", "classical", "CNN", ",", "converges", "in", "5", "times", "fewer", "epochs", ",", "but", "is", "computationally", "more", "expensive", ".", "Results", "for", "3", "variations", "of", "the", "10", "-", "layer", "CNN", "are", "in", "table", "[", "reference", "]", ".", "For", "model", "combination", ",", "we", "use", "the", "version", "with", "pooling", ",", "which", "is", "the", "exact", "same", "model", "without", "modifications", "from", "the", "original", "paper", ".", "Our", "implementation", "was", "done", "in", "Torch", ".", "We", "adopt", "the", "balanced", "sampling", "from", ",", "by", "sampling", "from", "context", "dependent", "state", "with", "probability", ".", "We", "keep", "throughout", "the", "experiments", "during", "cross", "-", "entropy", "training", ".", "During", "CE", "training", ",", "we", "optimize", "with", "simple", "SGD", "or", "NAG", ",", "during", "ST", "we", "found", "NAG", "to", "be", "superior", "to", "SGD", ".", "We", "regularize", "the", "stochastic", "sequence", "training", "by", "adding", "the", "gradient", "of", "cross", "-", "entropy", "loss", ",", "as", "proposed", "in", ".", "subsection", ":", "Bidirectional", "LSTMs", "Given", "the", "recent", "popularity", "of", "LSTMs", "for", "acoustic", "modeling", ",", "we", "have", "experimented", "with", "such", "models", "on", "the", "Switchboard", "task", "using", "the", "Torch", "toolkit", ".", "We", "have", "looked", "at", "the", "effect", "of", "the", "input", "features", "on", "LSTM", "performance", ",", "the", "number", "of", "layers", "and", "whether", "start", "states", "for", "the", "recurrent", "layers", "should", "be", "reset", "or", "carried", "over", ".", "We", "use", "bidirectional", "LSTMs", "that", "are", "trained", "on", "non", "-", "overlapping", "subsequences", "of", "20", "frames", ".", "The", "subsequences", "coming", "from", "the", "same", "utterance", "are", "contiguous", "so", "that", "the", "left", "-", "to", "-", "right", "final", "states", "for", "the", "current", "subsequence", "can", "be", "copied", "to", "the", "left", "-", "to", "-", "right", "start", "states", "for", "the", "next", "subsequence", "(", "i.e.", "carried", "over", ")", ".", "For", "processing", "speed", "and", "in", "order", "to", "get", "good", "gradient", "estimates", ",", "we", "group", "subsequences", "from", "multiple", "utterances", "into", "minibatches", "of", "size", "256", ".", "Regardless", "of", "the", "number", "of", "LSTM", "layers", ",", "all", "models", "use", "a", "linear", "bottleneck", "of", "size", "256", "before", "the", "softmax", "output", "layer", "(", "of", "size", "32000", ")", ".", "In", "one", "experiment", ",", "we", "compare", "the", "effect", "of", "input", "features", "on", "model", "performance", ".", "The", "baseline", "models", "are", "trained", "on", "40", "-", "dimensional", "FMLLR", "+", "100", "-", "dimensional", "ivector", "frames", "and", "have", "1024", "(", "or", "512", ")", "LSTM", "units", "per", "layer", "and", "per", "direction", "(", "left", "-", "to", "-", "right", "and", "right", "-", "to", "-", "left", ")", ".", "The", "forward", "and", "backward", "activations", "from", "the", "previous", "LSTM", "layer", "are", "concatenated", "and", "fed", "into", "the", "next", "LSTM", "layer", ".", "The", "contrast", "model", "is", "a", "single", "layer", "bidirectional", "LSTM", "trained", "on", "128", "-", "dim", "features", "obtained", "by", "performing", "PCA", "on", "512", "-", "dimensional", "bottleneck", "features", ".", "The", "features", "are", "obtained", "from", "a", "6", "-", "layer", "DNN", "cross", "entropy", "trained", "on", "blocks", "of", "11", "consecutive", "FMLLR", "frames", "and", "100", "-", "dimensional", "i", "-", "vectors", ".", "In", "Table", "[", "reference", "]", ",", "we", "report", "recognition", "results", "on", "Hub5’00", "for", "these", "four", "models", "trained", "with", "15", "passes", "of", "cross", "-", "entropy", "SGD", "on", "the", "300", "hour", "(", "SWB", "-", "1", ")", "subset", ".", "Due", "to", "a", "bug", "that", "affected", "our", "earlier", "multi", "-", "layer", "LSTM", "results", ",", "we", "decided", "to", "go", "ahead", "with", "single", "layer", "bidirectional", "LSTMs", "on", "bottleneck", "features", "on", "the", "full", "2000", "hour", "training", "set", ".", "We", "also", "experimented", "with", "how", "to", "deal", "with", "the", "start", "states", "at", "the", "beginning", "of", "the", "left", "-", "to", "-", "right", "pass", ".", "One", "option", "is", "to", "carry", "them", "over", "from", "the", "previous", "subsequence", "and", "the", "other", "one", "is", "to", "reset", "the", "start", "states", "at", "the", "beginning", "of", "each", "subsequence", ".", "In", "Figure", "[", "reference", "]", "we", "compare", "the", "cross", "-", "entropy", "loss", "on", "held", "-", "out", "data", "between", "these", "two", "models", ".", "As", "can", "be", "seen", ",", "the", "LSTM", "model", "with", "carried", "over", "start", "states", "is", "much", "better", "at", "predicting", "the", "correct", "HMM", "state", ".", "However", ",", "when", "comparing", "word", "error", "rates", "in", "Table", "[", "reference", "]", ",", "the", "LSTM", "with", "start", "states", "that", "are", "reset", "has", "a", "better", "performance", ".", "We", "surmise", "that", "this", "is", "because", "the", "increased", "memory", "of", "the", "LSTM", "with", "carried", "over", "start", "states", "is", "in", "conflict", "with", "the", "state", "sequence", "constraints", "imposed", "by", "the", "HMM", "topology", "and", "the", "language", "model", ".", "Additionally", ",", "we", "show", "the", "WERs", "of", "the", "DNN", "used", "for", "the", "bottleneck", "features", "and", "of", "a", "4", "-", "layer", "512", "unit", "LSTM", ".", "We", "observe", "that", "the", "4", "layer", "LSTM", "is", "significantly", "better", "than", "the", "DNN", "and", "the", "two", "single", "layer", "LSTMs", "trained", "on", "bottleneck", "features", ".", "subsection", ":", "Model", "combination", "In", "Table", "[", "reference", "]", "we", "report", "the", "performance", "of", "the", "individual", "models", "(", "RNN", ",", "VGG", "and", "4", "-", "layer", "LSTM", ")", "described", "in", "the", "previous", "subsections", "as", "well", "as", "the", "results", "after", "frame", "-", "level", "score", "fusion", ".", "All", "decodings", "are", "done", "with", "a", "30", "K", "word", "vocabulary", "and", "a", "small", "4", "-", "gram", "language", "model", "with", "4", "M", "n", "-", "grams", ".", "We", "note", "that", "RNNs", "and", "VGG", "nets", "exhibit", "similar", "performance", "and", "have", "a", "strong", "complementarity", "which", "improves", "the", "WER", "by", "0.6", "%", "and", "0.9", "%", "on", "SWB", "and", "CH", ",", "respectively", ".", "subsection", ":", "Language", "modeling", "experiments", "Our", "language", "modeling", "strategy", "largely", "parallels", "that", "described", "in", ".", "For", "completeness", ",", "we", "will", "repeat", "some", "of", "the", "details", "here", ".", "The", "main", "difference", "is", "an", "increase", "in", "the", "vocabulary", "size", "from", "30", "K", "words", "to", "85", "K", "words", ".", "When", "comparing", "acoustic", "models", "in", "previous", "sections", ",", "we", "used", "a", "relatively", "small", "legacy", "language", "model", "used", "in", "previous", "publications", ":", "a", "4", "M", "n", "-", "gram", "(", "n=4", ")", "language", "model", "with", "a", "vocabulary", "of", "30.5", "K", "words", ".", "We", "wanted", "to", "increase", "the", "language", "model", "coverage", "in", "a", "manner", "that", "others", "can", "replicate", ".", "To", "this", "end", ",", "we", "increased", "the", "vocabulary", "size", "from", "30.5", "K", "words", "to", "85", "K", "words", "by", "adding", "the", "vocabulary", "of", "the", "publicly", "available", "Broadcast", "News", "task", ".", "We", "also", "added", "to", "the", "LM", "publicly", "available", "text", "data", "from", "LDC", ",", "including", "Switchboard", ",", "Fisher", ",", "Gigaword", ",", "and", "Broadcast", "News", "and", "Conversations", ".", "The", "most", "relevant", "data", "are", "the", "transcripts", "of", "the", "1975", "hour", "audio", "data", "used", "to", "train", "the", "acoustic", "model", ",", "consisting", "of", "about", "24", "M", "words", ".", "For", "each", "corpus", "we", "trained", "a", "4", "-", "gram", "model", "with", "modified", "Kneser", "-", "Ney", "smoothing", ".", "The", "component", "LMs", "are", "linearly", "interpolated", "with", "weights", "chosen", "to", "optimize", "perplexity", "on", "a", "held", "-", "out", "set", ".", "Entropy", "pruning", "was", "applied", ",", "resulting", "in", "a", "single", "4", "-", "gram", "LM", "consisting", "of", "36", "M", "n", "-", "grams", ".", "This", "new", "n", "-", "gram", "LM", "was", "used", "together", "with", "our", "best", "acoustic", "model", "to", "decode", "and", "generate", "word", "lattices", "for", "LM", "rescoring", "experiments", ".", "The", "first", "two", "lines", "of", "Table", "[", "reference", "]", "show", "the", "improvement", "using", "this", "larger", "n", "-", "gram", "LM", "with", "larger", "vocabulary", "trained", "on", "more", "data", ".", "The", "WER", "improved", "by", "1.0", "%", "for", "SWB", ".", "Part", "of", "this", "improvement", "(", "0.1", "-", "0.2", "%", ")", "was", "due", "to", "also", "using", "a", "larger", "beam", "for", "decoding", "and", "a", "change", "in", "vocabulary", "tokenization", ".", "We", "used", "two", "types", "of", "LMs", "for", "LM", "rescoring", ":", "model", "M", ",", "a", "class", "-", "based", "exponential", "model", "and", "feed", "-", "forward", "neural", "network", "LM", "(", "NNLM", ")", ".", "We", "built", "a", "model", "M", "LM", "on", "each", "corpus", "and", "interpolated", "the", "models", ",", "together", "with", "the", "36", "M", "n", "-", "gram", "LM", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "using", "model", "M", "results", "in", "an", "improvement", "of", "0.6", "%", "on", "SWB", ".", "We", "built", "two", "NNLMs", "for", "interpolation", ".", "One", "was", "trained", "on", "just", "the", "most", "relevant", "data", ":", "the", "24", "M", "word", "corpus", "(", "Switchboard", "/", "Fisher", "/", "CallHome", "acoustic", "transcripts", ")", ".", "Another", "was", "trained", "on", "a", "560", "M", "word", "subset", "of", "the", "LM", "training", "data", ":", "in", "order", "to", "speed", "up", "training", "for", "this", "larger", "set", ",", "we", "employed", "a", "hierarchical", "NNLM", "approximation", ".", "Table", "[", "reference", "]", "shows", "that", "the", "NNLMs", "provided", "an", "additional", "0.4", "%", "improvement", "over", "the", "model", "M", "result", "on", "SWB", ".", "Compared", "with", "the", "n", "-", "gram", "LM", "baseline", ",", "LM", "rescoring", "yielded", "a", "total", "improvement", "of", "1.0", "%", "on", "SWB", "(", "7.6", "%", "to", "6.6", "%", ")", "and", "1.5", "%", "on", "CH", "(", "13.7", "%", "to", "12.2", "%", ")", ".", "section", ":", "Conclusion", "In", "our", "previous", "Switchboard", "system", "paper", "we", "have", "observed", "a", "good", "complementarity", "between", "recurrent", "nets", "and", "convolutional", "nets", "and", "their", "combination", "led", "to", "significant", "accuracy", "gains", ".", "In", "this", "paper", "we", "have", "presented", "an", "improved", "unfolded", "RNN", "(", "with", "maxout", "instead", "of", "sigmoid", "activations", ")", "and", "a", "stronger", "CNN", "obtained", "by", "adding", "more", "convolutional", "layers", "with", "smaller", "kernels", "and", "ReLu", "nonlinearities", ".", "These", "improved", "models", "still", "have", "good", "complementarity", "and", "their", "frame", "-", "level", "score", "combination", "in", "conjunction", "with", "a", "multi", "-", "layer", "LSTM", "leads", "to", "a", "0.4%", "-", "0.7", "%", "decrease", "in", "WER", "over", "the", "LSTM", ".", "Multi", "-", "layer", "LSTMs", "were", "the", "strongest", "performing", "model", "followed", "closely", "by", "the", "RNN", "and", "VGG", "nets", ".", "We", "also", "believe", "that", "LSTMs", "have", "more", "potential", "for", "direct", "sequence", "-", "to", "-", "sequence", "modeling", "and", "we", "are", "actively", "exploring", "this", "area", "of", "research", ".", "On", "the", "language", "modeling", "side", ",", "we", "have", "increased", "our", "vocabulary", "from", "30", "K", "to", "85", "K", "words", "and", "updated", "our", "component", "LMs", ".", "At", "the", "moment", ",", "we", "are", "less", "than", "3", "%", "away", "from", "achieving", "human", "performance", "on", "the", "Switchboard", "data", "(", "estimated", "to", "be", "around", "4", "%", ")", ".", "Unfortunately", ",", "it", "looks", "like", "future", "improvements", "on", "this", "task", "will", "be", "considerably", "harder", "to", "get", "and", "will", "probably", "require", "a", "breakthrough", "in", "direct", "sequence", "-", "to", "-", "sequence", "modeling", "and", "a", "significant", "increase", "in", "training", "data", ".", "section", ":", "Acknowledgment", "The", "authors", "wish", "to", "thank", "E.", "Marcheret", ",", "J.", "Cui", "and", "M.", "Nussbaum", "-", "Thom", "for", "useful", "suggestions", "about", "LSTMs", ".", "bibliography", ":", "References"], "entity": [[3, 11, ["IBM", "2016", "English", "Conversational", "Telephone", "Speech", "Recognition", "System"], "Task"], [16, 21, ["acoustic", "and", "language", "modeling", "techniques"], "Method"], [24, 27, ["word", "error", "rate"], "Metric"], [29, 34, ["English", "conversational", "telephone", "LVCSR", "system"], "Method"], [41, 43, ["Switchboard", "subset"], "Material"], [62, 64, ["strong", "models"], "Method"], [65, 67, ["recurrent", "nets"], "Method"], [68, 70, ["maxout", "activations"], "Method"], [72, 75, ["deep", "convolutional", "nets"], "Method"], [76, 78, ["3x3", "kernels"], "Method"], [80, 87, ["bidirectional", "long", "short", "-", "term", "memory", "nets"], "Method"], [99, 102, ["language", "modeling", "side"], "Task"], [106, 108, ["updated", "model"], "Method"], [109, 111, ["M", "”"], "Method"], [112, 116, ["hierarchical", "neural", "network", "LMs"], "Method"], [140, 143, ["recurrent", "neural", "networks"], "Method"], [144, 147, ["convolutional", "neural", "networks"], "Method"], [148, 151, ["conversational", "speech", "recognition"], "Task"], [157, 161, ["neural", "network", "acoustic", "modeling"], "Method"], [170, 176, ["deep", "feed", "-", "forward", "neural", "nets"], "Method"], [177, 178, ["LVCSR"], "Method"], [186, 188, ["image", "classification"], "Task"], [189, 192, ["natural", "language", "processing"], "Task"], [204, 207, ["deep", "convolutional", "nets"], "Method"], [212, 217, ["time", "-", "delay", "neural", "nets"], "Method"], [219, 225, ["long", "-", "short", "term", "memory", "nets"], "Method"], [248, 250, ["ASR", "systems"], "Method"], [252, 254, ["speaker", "adaptation"], "Task"], [255, 258, ["phonetic", "context", "modeling"], "Task"], [259, 262, ["discriminative", "feature", "processing"], "Task"], [273, 276, ["neural", "network", "architecture"], "Method"], [292, 295, ["neural", "network", "toolkits"], "Method"], [313, 314, ["GPUs"], "Method"], [323, 325, ["ASR", "systems"], "Method"], [335, 339, ["front", "-", "end", "processing"], "Task"], [347, 348, ["CNNs"], "Method"], [351, 356, ["log", "-", "mel", "spectral", "representation"], "Method"], [364, 366, ["processing", "steps"], "Method"], [368, 370, ["PLP", "cepstra"], "Method"], [371, 372, ["LDA"], "Method"], [373, 374, ["FMLLR"], "Method"], [375, 377, ["fMPE", "transforms"], "Method"], [385, 392, ["end", "-", "to", "-", "end", "ASR", "systems"], "Method"], [399, 403, ["phonetic", "context", "decision", "trees"], "Method"], [404, 405, ["HMMs"], "Method"], [429, 431, ["training", "algorithms"], "Method"], [433, 436, ["connectionist", "temporal", "classification"], "Method"], [437, 438, ["n’t"], "Method"], [452, 457, ["GMM", "-", "based", "baseline", "model"], "Method"], [472, 474, ["NN", "toolkits"], "Method"], [475, 478, ["speech", "recognition", "expertise"], "Task"], [488, 490, ["neural", "net"], "Method"], [504, 506, ["acoustic", "model"], "Method"], [552, 558, ["HMM", "and", "decision", "tree", "-", "based"], "Method"], [573, 574, ["WER"], "Metric"], [576, 578, ["Switchboard", "dataset"], "Material"], [613, 617, ["acoustic", "and", "language", "modeling"], "Task"], [640, 642, ["acoustic", "models"], "Method"], [654, 656, ["recurrent", "nets"], "Method"], [657, 659, ["maxout", "activations"], "Method"], [660, 662, ["annealed", "dropout"], "Method"], [664, 667, ["deep", "convolutional", "nets"], "Method"], [669, 671, ["3", "kernels"], "Method"], [673, 680, ["bidirectional", "long", "short", "-", "term", "memory", "nets"], "Method"], [682, 683, ["FMLLR"], "Method"], [695, 699, ["hybrid", "HMM", "decoding", "scenario"], "Method"], [723, 725, ["frontend", "processing"], "Task"], [726, 728, ["speaker", "adaptation"], "Task"], [752, 756, ["vocabulary", "and", "language", "modeling"], "Task"], [760, 762, ["Recurrent", "nets"], "Method"], [763, 765, ["maxout", "activations"], "Method"], [770, 772, ["maxout", "nets"], "Method"], [773, 775, ["ReLu", "units"], "Method"], [797, 799, ["maxout", "DNNs"], "Method"], [800, 801, ["CNNs"], "Method"], [803, 805, ["annealed", "dropout"], "Method"], [807, 811, ["sigmoid", "-", "based", "counterparts"], "Method"], [829, 830, ["maxout"], "Method"], [831, 832, ["sigmoid"], "Method"], [833, 835, ["unfolded", "RNNs"], "Task"], [840, 842, ["maxout", "RNNs"], "Method"], [844, 846, ["recurrent", "layer"], "Method"], [866, 870, ["non", "-", "recurrent", "layers"], "Method"], [882, 883, ["bottleneck"], "Method"], [889, 891, ["output", "layer"], "Method"], [929, 932, ["baseline", "sigmoid", "network"], "Method"], [941, 943, ["recurrent", "layer"], "Method"], [976, 983, ["100", "-", "dimensional", "speaker", "-", "based", "ivector"], "Method"], [1002, 1004, ["hidden", "layer"], "Method"], [1008, 1010, ["discriminative", "pretraining"], "Method"], [1015, 1018, ["SGD", "CE", "training"], "Method"], [1030, 1036, ["Hessian", "-", "free", "sequence", "discriminative", "training"], "Method"], [1038, 1043, ["state", "-", "based", "MBR", "criterion"], "Method"], [1055, 1057, ["error", "rates"], "Metric"], [1058, 1062, ["sigmoid", "and", "maxout", "RNNs"], "Method"], [1064, 1065, ["Switchboard"], "Material"], [1069, 1070, ["Hub5’00"], "Material"], [1072, 1073, ["decodings"], "Task"], [1086, 1091, ["4", "-", "gram", "language", "model"], "Method"], [1093, 1097, ["M", "n", "-", "grams"], "Method"], [1101, 1103, ["sigmoid", "RNNs"], "Method"], [1105, 1107, ["error", "rates"], "Metric"], [1136, 1138, ["maxout", "RNNs"], "Method"], [1151, 1152, ["WER"], "Metric"], [1165, 1167, ["sigmoid", "activations"], "Method"], [1171, 1174, ["deep", "convolutional", "networks"], "Method"], [1178, 1180, ["small", "kernels"], "Method"], [1190, 1192, ["acoustic", "models"], "Method"], [1193, 1200, ["hybrid", "NN", "-", "HMM", "speech", "recognition", "systems"], "Task"], [1205, 1209, ["cross", "-", "entropy", "training"], "Method"], [1222, 1224, ["sequence", "training"], "Method"], [1233, 1234, ["switchboard"], "Material"], [1242, 1245, ["deep", "convolutional", "networks"], "Method"], [1250, 1254, ["VGG", "Net", "”", "architecture"], "Method"], [1259, 1262, ["ImageNet", "classification", "challenge"], "Task"], [1281, 1283, ["convolutional", "layers"], "Method"], [1284, 1286, ["ReLU", "nonlinearities"], "Method"], [1287, 1289, ["pooling", "layers"], "Method"], [1332, 1333, ["pooling"], "Method"], [1337, 1338, ["convolutions"], "Method"], [1411, 1413, ["CNN", "architecture"], "Method"], [1429, 1430, ["kernel"], "Method"], [1436, 1440, ["10", "-", "layer", "CNN"], "Method"], [1449, 1451, ["classical", "CNN"], "Method"], [1471, 1475, ["10", "-", "layer", "CNN"], "Method"], [1483, 1485, ["model", "combination"], "Task"], [1491, 1492, ["pooling"], "Method"], [1516, 1518, ["balanced", "sampling"], "Method"], [1535, 1539, ["cross", "-", "entropy", "training"], "Method"], [1541, 1543, ["CE", "training"], "Task"], [1548, 1549, ["SGD"], "Method"], [1550, 1551, ["NAG"], "Method"], [1553, 1554, ["ST"], "Method"], [1556, 1557, ["NAG"], "Method"], [1561, 1562, ["SGD"], "Method"], [1574, 1578, ["cross", "-", "entropy", "loss"], "Metric"], [1585, 1587, ["Bidirectional", "LSTMs"], "Method"], [1592, 1593, ["LSTMs"], "Method"], [1594, 1596, ["acoustic", "modeling"], "Task"], [1605, 1607, ["Switchboard", "task"], "Material"], [1609, 1611, ["Torch", "toolkit"], "Method"], [1623, 1624, ["LSTM"], "Task"], [1636, 1638, ["recurrent", "layers"], "Method"], [1647, 1649, ["bidirectional", "LSTMs"], "Method"], [1707, 1709, ["processing", "speed"], "Metric"], [1715, 1717, ["gradient", "estimates"], "Task"], [1735, 1737, ["LSTM", "layers"], "Method"], [1742, 1744, ["linear", "bottleneck"], "Method"], [1796, 1798, ["LSTM", "units"], "Method"], [1825, 1827, ["LSTM", "layer"], "Method"], [1834, 1836, ["LSTM", "layer"], "Method"], [1838, 1840, ["contrast", "model"], "Method"], [1842, 1846, ["single", "layer", "bidirectional", "LSTM"], "Method"], [1855, 1856, ["PCA"], "Method"], [1869, 1875, ["6", "-", "layer", "DNN", "cross", "entropy"], "Method"], [1899, 1900, ["recognition"], "Task"], [1902, 1903, ["Hub5’00"], "Material"], [1912, 1916, ["cross", "-", "entropy", "SGD"], "Method"], [1935, 1939, ["multi", "-", "layer", "LSTM"], "Method"], [1947, 1951, ["single", "layer", "bidirectional", "LSTMs"], "Method"], [2021, 2025, ["cross", "-", "entropy", "loss"], "Metric"], [2041, 2043, ["LSTM", "model"], "Method"], [2062, 2065, ["word", "error", "rates"], "Metric"], [2072, 2073, ["LSTM"], "Method"], [2095, 2096, ["LSTM"], "Method"], [2112, 2114, ["HMM", "topology"], "Method"], [2116, 2118, ["language", "model"], "Method"], [2124, 2125, ["WERs"], "Metric"], [2127, 2128, ["DNN"], "Method"], [2136, 2142, ["4", "-", "layer", "512", "unit", "LSTM"], "Method"], [2147, 2150, ["4", "layer", "LSTM"], "Method"], [2155, 2156, ["DNN"], "Method"], [2159, 2162, ["single", "layer", "LSTMs"], "Method"], [2169, 2171, ["Model", "combination"], "Method"], [2185, 2186, ["RNN"], "Method"], [2187, 2193, ["VGG", "and", "4", "-", "layer", "LSTM"], "Method"], [2205, 2210, ["frame", "-", "level", "score", "fusion"], "Method"], [2212, 2213, ["decodings"], "Task"], [2224, 2229, ["4", "-", "gram", "language", "model"], "Method"], [2231, 2235, ["M", "n", "-", "grams"], "Method"], [2239, 2240, ["RNNs"], "Method"], [2241, 2243, ["VGG", "nets"], "Method"], [2254, 2255, ["WER"], "Metric"], [2262, 2263, ["SWB"], "Material"], [2264, 2265, ["CH"], "Metric"], [2270, 2272, ["Language", "modeling"], "Task"], [2274, 2277, ["language", "modeling", "strategy"], "Method"], [2303, 2305, ["vocabulary", "size"], "Metric"], [2316, 2318, ["acoustic", "models"], "Method"], [2328, 2330, ["language", "model"], "Method"], [2337, 2346, ["M", "n", "-", "gram", "(", "n=4", ")", "language", "model"], "Method"], [2377, 2379, ["vocabulary", "size"], "Metric"], [2413, 2414, ["Switchboard"], "Material"], [2442, 2444, ["acoustic", "model"], "Method"], [2458, 2462, ["4", "-", "gram", "model"], "Method"], [2463, 2468, ["modified", "Kneser", "-", "Ney", "smoothing"], "Method"], [2470, 2472, ["component", "LMs"], "Method"], [2480, 2481, ["perplexity"], "Metric"], [2488, 2490, ["Entropy", "pruning"], "Method"], [2497, 2501, ["4", "-", "gram", "LM"], "Method"], [2504, 2508, ["M", "n", "-", "grams"], "Method"], [2511, 2515, ["n", "-", "gram", "LM"], "Method"], [2521, 2523, ["acoustic", "model"], "Method"], [2530, 2533, ["LM", "rescoring", "experiments"], "Task"], [2549, 2553, ["n", "-", "gram", "LM"], "Method"], [2562, 2563, ["WER"], "Metric"], [2568, 2569, ["SWB"], "Material"], [2589, 2590, ["decoding"], "Task"], [2594, 2596, ["vocabulary", "tokenization"], "Task"], [2602, 2603, ["LMs"], "Method"], [2604, 2606, ["LM", "rescoring"], "Task"], [2607, 2609, ["model", "M"], "Method"], [2611, 2616, ["class", "-", "based", "exponential", "model"], "Method"], [2617, 2626, ["feed", "-", "forward", "neural", "network", "LM", "(", "NNLM", ")"], "Method"], [2630, 2633, ["model", "M", "LM"], "Method"], [2645, 2650, ["M", "n", "-", "gram", "LM"], "Method"], [2660, 2662, ["model", "M"], "Method"], [2670, 2671, ["SWB"], "Material"], [2675, 2676, ["NNLMs"], "Method"], [2677, 2678, ["interpolation"], "Task"], [2695, 2696, ["Switchboard"], "Material"], [2724, 2725, ["training"], "Task"], [2733, 2736, ["hierarchical", "NNLM", "approximation"], "Method"], [2744, 2745, ["NNLMs"], "Method"], [2757, 2758, ["SWB"], "Material"], [2762, 2767, ["n", "-", "gram", "LM", "baseline"], "Method"], [2768, 2770, ["LM", "rescoring"], "Method"], [2778, 2779, ["SWB"], "Material"], [2790, 2791, ["CH"], "Task"], [2805, 2807, ["Switchboard", "system"], "Method"], [2815, 2817, ["recurrent", "nets"], "Method"], [2818, 2820, ["convolutional", "nets"], "Method"], [2826, 2827, ["accuracy"], "Metric"], [2841, 2842, ["maxout"], "Method"], [2850, 2851, ["CNN"], "Method"], [2855, 2857, ["convolutional", "layers"], "Method"], [2873, 2878, ["frame", "-", "level", "score", "combination"], "Method"], [2882, 2886, ["multi", "-", "layer", "LSTM"], "Method"], [2895, 2896, ["WER"], "Metric"], [2898, 2899, ["LSTM"], "Method"], [2900, 2904, ["Multi", "-", "layer", "LSTMs"], "Method"], [2913, 2914, ["RNN"], "Method"], [2915, 2917, ["VGG", "nets"], "Method"], [2922, 2923, ["LSTMs"], "Method"], [2927, 2934, ["direct", "sequence", "-", "to", "-", "sequence", "modeling"], "Task"], [2946, 2949, ["language", "modeling", "side"], "Task"], [2965, 2967, ["component", "LMs"], "Method"], [2985, 2987, ["Switchboard", "data"], "Material"], [3019, 3026, ["direct", "sequence", "-", "to", "-", "sequence", "modeling"], "Task"], [3056, 3057, ["LSTMs"], "Method"], [209, 211, ["recurrent", "nets"], "Method"], [361, 362, ["n’t"], "Method"], [987, 988, ["RNN"], "Method"], [1127, 1128, ["RNN"], "Method"], [1161, 1162, ["RNN"], "Method"], [1176, 1177, ["CNNs"], "Method"], [1213, 1214, ["switchboard"], "Material"], [1226, 1227, ["switchboard"], "Material"], [1567, 1569, ["sequence", "training"], "Method"], [1921, 1922, ["SWB"], "Material"], [2838, 2839, ["RNN"], "Method"]]} {"doc_id": "0626908dd710b91aece1a81f4ca0635f23fc47f3", "text": ["document", ":", "Rethinking", "the", "Inception", "Architecture", "for", "Computer", "Vision", "Convolutional", "networks", "are", "at", "the", "core", "of", "most", "state", "-", "of", "-", "the", "-", "art", "computer", "vision", "solutions", "for", "a", "wide", "variety", "of", "tasks", ".", "Since", "2014", "very", "deep", "convolutional", "networks", "started", "to", "become", "mainstream", ",", "yielding", "substantial", "gains", "in", "various", "benchmarks", ".", "Although", "increased", "model", "size", "and", "computational", "cost", "tend", "to", "translate", "to", "immediate", "quality", "gains", "for", "most", "tasks", "(", "as", "long", "as", "enough", "labeled", "data", "is", "provided", "for", "training", ")", ",", "computational", "efficiency", "and", "low", "parameter", "count", "are", "still", "enabling", "factors", "for", "various", "use", "cases", "such", "as", "mobile", "vision", "and", "big", "-", "data", "scenarios", ".", "Here", "we", "are", "exploring", "ways", "to", "scale", "up", "networks", "in", "ways", "that", "aim", "at", "utilizing", "the", "added", "computation", "as", "efficiently", "as", "possible", "by", "suitably", "factorized", "convolutions", "and", "aggressive", "regularization", ".", "We", "benchmark", "our", "methods", "on", "the", "ILSVRC", "2012", "classification", "challenge", "validation", "set", "demonstrate", "substantial", "gains", "over", "the", "state", "of", "the", "art", ":", "top", "-", "and", "top", "-", "error", "for", "single", "frame", "evaluation", "using", "a", "network", "with", "a", "computational", "cost", "of", "billion", "multiply", "-", "adds", "per", "inference", "and", "with", "using", "less", "than", "25", "million", "parameters", ".", "With", "an", "ensemble", "of", "models", "and", "multi", "-", "crop", "evaluation", ",", "we", "report", "top", "-", "error", "and", "top", "-", "error", ".", "section", ":", "Introduction", "Since", "the", "2012", "ImageNet", "competition", "winning", "entry", "by", "Krizhevsky", "et", "al", ",", "their", "network", "“", "AlexNet", "”", "has", "been", "successfully", "applied", "to", "a", "larger", "variety", "of", "computer", "vision", "tasks", ",", "for", "example", "to", "object", "-", "detection", ",", "segmentation", ",", "human", "pose", "estimation", ",", "video", "classification", ",", "object", "tracking", ",", "and", "superresolution", ".", "These", "successes", "spurred", "a", "new", "line", "of", "research", "that", "focused", "on", "finding", "higher", "performing", "convolutional", "neural", "networks", ".", "Starting", "in", "2014", ",", "the", "quality", "of", "network", "architectures", "significantly", "improved", "by", "utilizing", "deeper", "and", "wider", "networks", ".", "VGGNet", "and", "GoogLeNet", "yielded", "similarly", "high", "performance", "in", "the", "2014", "ILSVRC", "classification", "challenge", ".", "One", "interesting", "observation", "was", "that", "gains", "in", "the", "classification", "performance", "tend", "to", "transfer", "to", "significant", "quality", "gains", "in", "a", "wide", "variety", "of", "application", "domains", ".", "This", "means", "that", "architectural", "improvements", "in", "deep", "convolutional", "architecture", "can", "be", "utilized", "for", "improving", "performance", "for", "most", "other", "computer", "vision", "tasks", "that", "are", "increasingly", "reliant", "on", "high", "quality", ",", "learned", "visual", "features", ".", "Also", ",", "improvements", "in", "the", "network", "quality", "resulted", "in", "new", "application", "domains", "for", "convolutional", "networks", "in", "cases", "where", "AlexNet", "features", "could", "not", "compete", "with", "hand", "engineered", ",", "crafted", "solutions", ",", "e.g.", "proposal", "generation", "in", "detection", ".", "Although", "VGGNet", "has", "the", "compelling", "feature", "of", "architectural", "simplicity", ",", "this", "comes", "at", "a", "high", "cost", ":", "evaluating", "the", "network", "requires", "a", "lot", "of", "computation", ".", "On", "the", "other", "hand", ",", "the", "Inception", "architecture", "of", "GoogLeNet", "was", "also", "designed", "to", "perform", "well", "even", "under", "strict", "constraints", "on", "memory", "and", "computational", "budget", ".", "For", "example", ",", "GoogleNet", "employed", "only", "5", "million", "parameters", ",", "which", "represented", "a", "reduction", "with", "respect", "to", "its", "predecessor", "AlexNet", ",", "which", "used", "million", "parameters", ".", "Furthermore", ",", "VGGNet", "employed", "about", "3x", "more", "parameters", "than", "AlexNet", ".", "The", "computational", "cost", "of", "Inception", "is", "also", "much", "lower", "than", "VGGNet", "or", "its", "higher", "performing", "successors", ".", "This", "has", "made", "it", "feasible", "to", "utilize", "Inception", "networks", "in", "big", "-", "data", "scenarios", ",", ",", "where", "huge", "amount", "of", "data", "needed", "to", "be", "processed", "at", "reasonable", "cost", "or", "scenarios", "where", "memory", "or", "computational", "capacity", "is", "inherently", "limited", ",", "for", "example", "in", "mobile", "vision", "settings", ".", "It", "is", "certainly", "possible", "to", "mitigate", "parts", "of", "these", "issues", "by", "applying", "specialized", "solutions", "to", "target", "memory", "use", ",", "or", "by", "optimizing", "the", "execution", "of", "certain", "operations", "via", "computational", "tricks", ".", "However", ",", "these", "methods", "add", "extra", "complexity", ".", "Furthermore", ",", "these", "methods", "could", "be", "applied", "to", "optimize", "the", "Inception", "architecture", "as", "well", ",", "widening", "the", "efficiency", "gap", "again", ".", "Still", ",", "the", "complexity", "of", "the", "Inception", "architecture", "makes", "it", "more", "difficult", "to", "make", "changes", "to", "the", "network", ".", "If", "the", "architecture", "is", "scaled", "up", "naively", ",", "large", "parts", "of", "the", "computational", "gains", "can", "be", "immediately", "lost", ".", "Also", ",", "does", "not", "provide", "a", "clear", "description", "about", "the", "contributing", "factors", "that", "lead", "to", "the", "various", "design", "decisions", "of", "the", "GoogLeNet", "architecture", ".", "This", "makes", "it", "much", "harder", "to", "adapt", "it", "to", "new", "use", "-", "cases", "while", "maintaining", "its", "efficiency", ".", "For", "example", ",", "if", "it", "is", "deemed", "necessary", "to", "increase", "the", "capacity", "of", "some", "Inception", "-", "style", "model", ",", "the", "simple", "transformation", "of", "just", "doubling", "the", "number", "of", "all", "filter", "bank", "sizes", "will", "lead", "to", "a", "4x", "increase", "in", "both", "computational", "cost", "and", "number", "of", "parameters", ".", "This", "might", "prove", "prohibitive", "or", "unreasonable", "in", "a", "lot", "of", "practical", "scenarios", ",", "especially", "if", "the", "associated", "gains", "are", "modest", ".", "In", "this", "paper", ",", "we", "start", "with", "describing", "a", "few", "general", "principles", "and", "optimization", "ideas", "that", "that", "proved", "to", "be", "useful", "for", "scaling", "up", "convolution", "networks", "in", "efficient", "ways", ".", "Although", "our", "principles", "are", "not", "limited", "to", "Inception", "-", "type", "networks", ",", "they", "are", "easier", "to", "observe", "in", "that", "context", "as", "the", "generic", "structure", "of", "the", "Inception", "style", "building", "blocks", "is", "flexible", "enough", "to", "incorporate", "those", "constraints", "naturally", ".", "This", "is", "enabled", "by", "the", "generous", "use", "of", "dimensional", "reduction", "and", "parallel", "structures", "of", "the", "Inception", "modules", "which", "allows", "for", "mitigating", "the", "impact", "of", "structural", "changes", "on", "nearby", "components", ".", "Still", ",", "one", "needs", "to", "be", "cautious", "about", "doing", "so", ",", "as", "some", "guiding", "principles", "should", "be", "observed", "to", "maintain", "high", "quality", "of", "the", "models", ".", "section", ":", "General", "Design", "Principles", "Here", "we", "will", "describe", "a", "few", "design", "principles", "based", "on", "large", "-", "scale", "experimentation", "with", "various", "architectural", "choices", "with", "convolutional", "networks", ".", "At", "this", "point", ",", "the", "utility", "of", "the", "principles", "below", "are", "speculative", "and", "additional", "future", "experimental", "evidence", "will", "be", "necessary", "to", "assess", "their", "accuracy", "and", "domain", "of", "validity", ".", "Still", ",", "grave", "deviations", "from", "these", "principles", "tended", "to", "result", "in", "deterioration", "in", "the", "quality", "of", "the", "networks", "and", "fixing", "situations", "where", "those", "deviations", "were", "detected", "resulted", "in", "improved", "architectures", "in", "general", ".", "Avoid", "representational", "bottlenecks", ",", "especially", "early", "in", "the", "network", ".", "Feed", "-", "forward", "networks", "can", "be", "represented", "by", "an", "acyclic", "graph", "from", "the", "input", "layer", "(", "s", ")", "to", "the", "classifier", "or", "regressor", ".", "This", "defines", "a", "clear", "direction", "for", "the", "information", "flow", ".", "For", "any", "cut", "separating", "the", "inputs", "from", "the", "outputs", ",", "one", "can", "access", "the", "amount", "of", "information", "passing", "though", "the", "cut", ".", "One", "should", "avoid", "bottlenecks", "with", "extreme", "compression", ".", "In", "general", "the", "representation", "size", "should", "gently", "decrease", "from", "the", "inputs", "to", "the", "outputs", "before", "reaching", "the", "final", "representation", "used", "for", "the", "task", "at", "hand", ".", "Theoretically", ",", "information", "content", "can", "not", "be", "assessed", "merely", "by", "the", "dimensionality", "of", "the", "representation", "as", "it", "discards", "important", "factors", "like", "correlation", "structure", ";", "the", "dimensionality", "merely", "provides", "a", "rough", "estimate", "of", "information", "content", ".", "Higher", "dimensional", "representations", "are", "easier", "to", "process", "locally", "within", "a", "network", ".", "Increasing", "the", "activations", "per", "tile", "in", "a", "convolutional", "network", "allows", "for", "more", "disentangled", "features", ".", "The", "resulting", "networks", "will", "train", "faster", ".", "Spatial", "aggregation", "can", "be", "done", "over", "lower", "dimensional", "embeddings", "without", "much", "or", "any", "loss", "in", "representational", "power", ".", "For", "example", ",", "before", "performing", "a", "more", "spread", "out", "(", "e.g.", ")", "convolution", ",", "one", "can", "reduce", "the", "dimension", "of", "the", "input", "representation", "before", "the", "spatial", "aggregation", "without", "expecting", "serious", "adverse", "effects", ".", "We", "hypothesize", "that", "the", "reason", "for", "that", "is", "the", "strong", "correlation", "between", "adjacent", "unit", "results", "in", "much", "less", "loss", "of", "information", "during", "dimension", "reduction", ",", "if", "the", "outputs", "are", "used", "in", "a", "spatial", "aggregation", "context", ".", "Given", "that", "these", "signals", "should", "be", "easily", "compressible", ",", "the", "dimension", "reduction", "even", "promotes", "faster", "learning", ".", "Balance", "the", "width", "and", "depth", "of", "the", "network", ".", "Optimal", "performance", "of", "the", "network", "can", "be", "reached", "by", "balancing", "the", "number", "of", "filters", "per", "stage", "and", "the", "depth", "of", "the", "network", ".", "Increasing", "both", "the", "width", "and", "the", "depth", "of", "the", "network", "can", "contribute", "to", "higher", "quality", "networks", ".", "However", ",", "the", "optimal", "improvement", "for", "a", "constant", "amount", "of", "computation", "can", "be", "reached", "if", "both", "are", "increased", "in", "parallel", ".", "The", "computational", "budget", "should", "therefore", "be", "distributed", "in", "a", "balanced", "way", "between", "the", "depth", "and", "width", "of", "the", "network", ".", "Although", "these", "principles", "might", "make", "sense", ",", "it", "is", "not", "straightforward", "to", "use", "them", "to", "improve", "the", "quality", "of", "networks", "out", "of", "box", ".", "The", "idea", "is", "to", "use", "them", "judiciously", "in", "ambiguous", "situations", "only", ".", "section", ":", "Factorizing", "Convolutions", "with", "Large", "Filter", "Size", "Much", "of", "the", "original", "gains", "of", "the", "GoogLeNet", "network", "arise", "from", "a", "very", "generous", "use", "of", "dimension", "reduction", ".", "This", "can", "be", "viewed", "as", "a", "special", "case", "of", "factorizing", "convolutions", "in", "a", "computationally", "efficient", "manner", ".", "Consider", "for", "example", "the", "case", "of", "a", "convolutional", "layer", "followed", "by", "a", "convolutional", "layer", ".", "In", "a", "vision", "network", ",", "it", "is", "expected", "that", "the", "outputs", "of", "near", "-", "by", "activations", "are", "highly", "correlated", ".", "Therefore", ",", "we", "can", "expect", "that", "their", "activations", "can", "be", "reduced", "before", "aggregation", "and", "that", "this", "should", "result", "in", "similarly", "expressive", "local", "representations", ".", "Here", "we", "explore", "other", "ways", "of", "factorizing", "convolutions", "in", "various", "settings", ",", "especially", "in", "order", "to", "increase", "the", "computational", "efficiency", "of", "the", "solution", ".", "Since", "Inception", "networks", "are", "fully", "convolutional", ",", "each", "weight", "corresponds", "to", "one", "multiplication", "per", "activation", ".", "Therefore", ",", "any", "reduction", "in", "computational", "cost", "results", "in", "reduced", "number", "of", "parameters", ".", "This", "means", "that", "with", "suitable", "factorization", ",", "we", "can", "end", "up", "with", "more", "disentangled", "parameters", "and", "therefore", "with", "faster", "training", ".", "Also", ",", "we", "can", "use", "the", "computational", "and", "memory", "savings", "to", "increase", "the", "filter", "-", "bank", "sizes", "of", "our", "network", "while", "maintaining", "our", "ability", "to", "train", "each", "model", "replica", "on", "a", "single", "computer", ".", "subsection", ":", "Factorization", "into", "smaller", "convolutions", "Convolutions", "with", "larger", "spatial", "filters", "(", "e.g.", "or", ")", "tend", "to", "be", "disproportionally", "expensive", "in", "terms", "of", "computation", ".", "For", "example", ",", "a", "convolution", "with", "filters", "over", "a", "grid", "with", "filters", "is", "25", "/", "9", "=", "2.78", "times", "more", "computationally", "expensive", "than", "a", "convolution", "with", "the", "same", "number", "of", "filters", ".", "Of", "course", ",", "a", "filter", "can", "capture", "dependencies", "between", "signals", "between", "activations", "of", "units", "further", "away", "in", "the", "earlier", "layers", ",", "so", "a", "reduction", "of", "the", "geometric", "size", "of", "the", "filters", "comes", "at", "a", "large", "cost", "of", "expressiveness", ".", "However", ",", "we", "can", "ask", "whether", "a", "convolution", "could", "be", "replaced", "by", "a", "multi", "-", "layer", "network", "with", "less", "parameters", "with", "the", "same", "input", "size", "and", "output", "depth", ".", "If", "we", "zoom", "into", "the", "computation", "graph", "of", "the", "convolution", ",", "we", "see", "that", "each", "output", "looks", "like", "a", "small", "fully", "-", "connected", "network", "sliding", "over", "tiles", "over", "its", "input", "(", "see", "Figure", "[", "reference", "]", ")", ".", "Since", "we", "are", "constructing", "a", "vision", "network", ",", "it", "seems", "natural", "to", "exploit", "translation", "invariance", "again", "and", "replace", "the", "fully", "connected", "component", "by", "a", "two", "layer", "convolutional", "architecture", ":", "the", "first", "layer", "is", "a", "convolution", ",", "the", "second", "is", "a", "fully", "connected", "layer", "on", "top", "of", "the", "output", "grid", "of", "the", "first", "layer", "(", "see", "Figure", "[", "reference", "]", ")", ".", "Sliding", "this", "small", "network", "over", "the", "input", "activation", "grid", "boils", "down", "to", "replacing", "the", "convolution", "with", "two", "layers", "of", "convolution", "(", "compare", "Figure", "[", "reference", "]", "with", "[", "reference", "]", ")", ".", "This", "setup", "clearly", "reduces", "the", "parameter", "count", "by", "sharing", "the", "weights", "between", "adjacent", "tiles", ".", "To", "analyze", "the", "expected", "computational", "cost", "savings", ",", "we", "will", "make", "a", "few", "simplifying", "assumptions", "that", "apply", "for", "the", "typical", "situations", ":", "We", "can", "assume", "that", ",", "that", "is", "that", "we", "want", "to", "change", "the", "number", "of", "activations", "/", "unit", "by", "a", "constant", "alpha", "factor", ".", "Since", "the", "convolution", "is", "aggregating", ",", "is", "typically", "slightly", "larger", "than", "one", "(", "around", "1.5", "in", "the", "case", "of", "GoogLeNet", ")", ".", "Having", "a", "two", "layer", "replacement", "for", "the", "layer", ",", "it", "seems", "reasonable", "to", "reach", "this", "expansion", "in", "two", "steps", ":", "increasing", "the", "number", "of", "filters", "by", "in", "both", "steps", ".", "In", "order", "to", "simplify", "our", "estimate", "by", "choosing", "(", "no", "expansion", ")", ",", "If", "we", "would", "naivly", "slide", "a", "network", "without", "reusing", "the", "computation", "between", "neighboring", "grid", "tiles", ",", "we", "would", "increase", "the", "computational", "cost", ".", "sliding", "this", "network", "can", "be", "represented", "by", "two", "convolutional", "layers", "which", "reuses", "the", "activations", "between", "adjacent", "tiles", ".", "This", "way", ",", "we", "end", "up", "with", "a", "net", "reduction", "of", "computation", ",", "resulting", "in", "a", "relative", "gain", "of", "by", "this", "factorization", ".", "The", "exact", "same", "saving", "holds", "for", "the", "parameter", "count", "as", "each", "parameter", "is", "used", "exactly", "once", "in", "the", "computation", "of", "the", "activation", "of", "each", "unit", ".", "Still", ",", "this", "setup", "raises", "two", "general", "questions", ":", "Does", "this", "replacement", "result", "in", "any", "loss", "of", "expressiveness", "?", "If", "our", "main", "goal", "is", "to", "factorize", "the", "linear", "part", "of", "the", "computation", ",", "would", "it", "not", "suggest", "to", "keep", "linear", "activations", "in", "the", "first", "layer", "?", "We", "have", "ran", "several", "control", "experiments", "(", "for", "example", "see", "figure", "[", "reference", "]", ")", "and", "using", "linear", "activation", "was", "always", "inferior", "to", "using", "rectified", "linear", "units", "in", "all", "stages", "of", "the", "factorization", ".", "We", "attribute", "this", "gain", "to", "the", "enhanced", "space", "of", "variations", "that", "the", "network", "can", "learn", "especially", "if", "we", "batch", "-", "normalize", "the", "output", "activations", ".", "One", "can", "see", "similar", "effects", "when", "using", "linear", "activations", "for", "the", "dimension", "reduction", "components", ".", "subsection", ":", "Spatial", "Factorization", "into", "Asymmetric", "Convolutions", "The", "above", "results", "suggest", "that", "convolutions", "with", "filters", "larger", "a", "might", "not", "be", "generally", "useful", "as", "they", "can", "always", "be", "reduced", "into", "a", "sequence", "of", "convolutional", "layers", ".", "Still", "we", "can", "ask", "the", "question", "whether", "one", "should", "factorize", "them", "into", "smaller", ",", "for", "example", "convolutions", ".", "However", ",", "it", "turns", "out", "that", "one", "can", "do", "even", "better", "than", "by", "using", "asymmetric", "convolutions", ",", "e.g.", ".", "For", "example", "using", "a", "convolution", "followed", "by", "a", "convolution", "is", "equivalent", "to", "sliding", "a", "two", "layer", "network", "with", "the", "same", "receptive", "field", "as", "in", "a", "convolution", "(", "see", "figure", "[", "reference", "]", ")", ".", "Still", "the", "two", "-", "layer", "solution", "is", "cheaper", "for", "the", "same", "number", "of", "output", "filters", ",", "if", "the", "number", "of", "input", "and", "output", "filters", "is", "equal", ".", "By", "comparison", ",", "factorizing", "a", "convolution", "into", "a", "two", "convolution", "represents", "only", "a", "saving", "of", "computation", ".", ".", "In", "theory", ",", "we", "could", "go", "even", "further", "and", "argue", "that", "one", "can", "replace", "any", "convolution", "by", "a", "convolution", "followed", "by", "a", "convolution", "and", "the", "computational", "cost", "saving", "increases", "dramatically", "as", "grows", "(", "see", "figure", "6", ")", ".", "In", "practice", ",", "we", "have", "found", "that", "employing", "this", "factorization", "does", "not", "work", "well", "on", "early", "layers", ",", "but", "it", "gives", "very", "good", "results", "on", "medium", "grid", "-", "sizes", "(", "On", "feature", "maps", ",", "where", "ranges", "between", "and", ")", ".", "On", "that", "level", ",", "very", "good", "results", "can", "be", "achieved", "by", "using", "convolutions", "followed", "by", "convolutions", ".", "section", ":", "Utility", "of", "Auxiliary", "Classifiers", "has", "introduced", "the", "notion", "of", "auxiliary", "classifiers", "to", "improve", "the", "convergence", "of", "very", "deep", "networks", ".", "The", "original", "motivation", "was", "to", "push", "useful", "gradients", "to", "the", "lower", "layers", "to", "make", "them", "immediately", "useful", "and", "improve", "the", "convergence", "during", "training", "by", "combating", "the", "vanishing", "gradient", "problem", "in", "very", "deep", "networks", ".", "Also", "Lee", "et", "al", "argues", "that", "auxiliary", "classifiers", "promote", "more", "stable", "learning", "and", "better", "convergence", ".", "Interestingly", ",", "we", "found", "that", "auxiliary", "classifiers", "did", "not", "result", "in", "improved", "convergence", "early", "in", "the", "training", ":", "the", "training", "progression", "of", "network", "with", "and", "without", "side", "head", "looks", "virtually", "identical", "before", "both", "models", "reach", "high", "accuracy", ".", "Near", "the", "end", "of", "training", ",", "the", "network", "with", "the", "auxiliary", "branches", "starts", "to", "overtake", "the", "accuracy", "of", "the", "network", "without", "any", "auxiliary", "branch", "and", "reaches", "a", "slightly", "higher", "plateau", ".", "Also", "used", "two", "side", "-", "heads", "at", "different", "stages", "in", "the", "network", ".", "The", "removal", "of", "the", "lower", "auxiliary", "branch", "did", "not", "have", "any", "adverse", "effect", "on", "the", "final", "quality", "of", "the", "network", ".", "Together", "with", "the", "earlier", "observation", "in", "the", "previous", "paragraph", ",", "this", "means", "that", "original", "the", "hypothesis", "of", "that", "these", "branches", "help", "evolving", "the", "low", "-", "level", "features", "is", "most", "likely", "misplaced", ".", "Instead", ",", "we", "argue", "that", "the", "auxiliary", "classifiers", "act", "as", "regularizer", ".", "This", "is", "supported", "by", "the", "fact", "that", "the", "main", "classifier", "of", "the", "network", "performs", "better", "if", "the", "side", "branch", "is", "batch", "-", "normalized", "or", "has", "a", "dropout", "layer", ".", "This", "also", "gives", "a", "weak", "supporting", "evidence", "for", "the", "conjecture", "that", "batch", "normalization", "acts", "as", "a", "regularizer", ".", "section", ":", "Efficient", "Grid", "Size", "Reduction", "Traditionally", ",", "convolutional", "networks", "used", "some", "pooling", "operation", "to", "decrease", "the", "grid", "size", "of", "the", "feature", "maps", ".", "In", "order", "to", "avoid", "a", "representational", "bottleneck", ",", "before", "applying", "maximum", "or", "average", "pooling", "the", "activation", "dimension", "of", "the", "network", "filters", "is", "expanded", ".", "For", "example", ",", "starting", "a", "grid", "with", "filters", ",", "if", "we", "would", "like", "to", "arrive", "at", "a", "grid", "with", "filters", ",", "we", "first", "need", "to", "compute", "a", "stride", "-", "1", "convolution", "with", "filters", "and", "then", "apply", "an", "additional", "pooling", "step", ".", "This", "means", "that", "the", "overall", "computational", "cost", "is", "dominated", "by", "the", "expensive", "convolution", "on", "the", "larger", "grid", "using", "operations", ".", "One", "possibility", "would", "be", "to", "switch", "to", "pooling", "with", "convolution", "and", "therefore", "resulting", "in", "reducing", "the", "computational", "cost", "by", "a", "quarter", ".", "However", ",", "this", "creates", "a", "representational", "bottlenecks", "as", "the", "overall", "dimensionality", "of", "the", "representation", "drops", "to", "resulting", "in", "less", "expressive", "networks", "(", "see", "Figure", "[", "reference", "]", ")", ".", "Instead", "of", "doing", "so", ",", "we", "suggest", "another", "variant", "the", "reduces", "the", "computational", "cost", "even", "further", "while", "removing", "the", "representational", "bottleneck", ".", "(", "see", "Figure", "[", "reference", "]", ")", ".", "We", "can", "use", "two", "parallel", "stride", "2", "blocks", ":", "and", ".", "is", "a", "pooling", "layer", "(", "either", "average", "or", "maximum", "pooling", ")", "the", "activation", ",", "both", "of", "them", "are", "stride", "the", "filter", "banks", "of", "which", "are", "concatenated", "as", "in", "figure", "[", "reference", "]", ".", "section", ":", "Inception", "-", "v2", "Here", "we", "are", "connecting", "the", "dots", "from", "above", "and", "propose", "a", "new", "architecture", "with", "improved", "performance", "on", "the", "ILSVRC", "2012", "classification", "benchmark", ".", "The", "layout", "of", "our", "network", "is", "given", "in", "table", "[", "reference", "]", ".", "Note", "that", "we", "have", "factorized", "the", "traditional", "convolution", "into", "three", "convolutions", "based", "on", "the", "same", "ideas", "as", "described", "in", "section", "[", "reference", "]", ".", "For", "the", "Inception", "part", "of", "the", "network", ",", "we", "have", "traditional", "inception", "modules", "at", "the", "with", "filters", "each", ".", "This", "is", "reduced", "to", "a", "grid", "with", "filters", "using", "the", "grid", "reduction", "technique", "described", "in", "section", "[", "reference", "]", ".", "This", "is", "is", "followed", "by", "instances", "of", "the", "factorized", "inception", "modules", "as", "depicted", "in", "figure", "[", "reference", "]", ".", "This", "is", "reduced", "to", "a", "grid", "with", "the", "grid", "reduction", "technique", "depicted", "in", "figure", "[", "reference", "]", ".", "At", "the", "coarsest", "level", ",", "we", "have", "two", "Inception", "modules", "as", "depicted", "in", "figure", "[", "reference", "]", ",", "with", "a", "concatenated", "output", "filter", "bank", "size", "of", "2048", "for", "each", "tile", ".", "The", "detailed", "structure", "of", "the", "network", ",", "including", "the", "sizes", "of", "filter", "banks", "inside", "the", "Inception", "modules", ",", "is", "given", "in", "the", "supplementary", "material", ",", "given", "in", "the", "model.txt", "that", "is", "in", "the", "tar", "-", "file", "of", "this", "submission", ".", "However", ",", "we", "have", "observed", "that", "the", "quality", "of", "the", "network", "is", "relatively", "stable", "to", "variations", "as", "long", "as", "the", "principles", "from", "Section", "[", "reference", "]", "are", "observed", ".", "Although", "our", "network", "is", "layers", "deep", ",", "our", "computation", "cost", "is", "only", "about", "higher", "than", "that", "of", "GoogLeNet", "and", "it", "is", "still", "much", "more", "efficient", "than", "VGGNet", ".", "section", ":", "Model", "Regularization", "via", "Label", "Smoothing", "Here", "we", "propose", "a", "mechanism", "to", "regularize", "the", "classifier", "layer", "by", "estimating", "the", "marginalized", "effect", "of", "label", "-", "dropout", "during", "training", ".", "For", "each", "training", "example", ",", "our", "model", "computes", "the", "probability", "of", "each", "label", ":", ".", "Here", ",", "are", "the", "logits", "or", "unnormalized", "log", "-", "probabilities", ".", "Consider", "the", "ground", "-", "truth", "distribution", "over", "labels", "for", "this", "training", "example", ",", "normalized", "so", "that", ".", "For", "brevity", ",", "let", "us", "omit", "the", "dependence", "of", "and", "on", "example", ".", "We", "define", "the", "loss", "for", "the", "example", "as", "the", "cross", "entropy", ":", ".", "Minimizing", "this", "is", "equivalent", "to", "maximizing", "the", "expected", "log", "-", "likelihood", "of", "a", "label", ",", "where", "the", "label", "is", "selected", "according", "to", "its", "ground", "-", "truth", "distribution", ".", "Cross", "-", "entropy", "loss", "is", "differentiable", "with", "respect", "to", "the", "logits", "and", "thus", "can", "be", "used", "for", "gradient", "training", "of", "deep", "models", ".", "The", "gradient", "has", "a", "rather", "simple", "form", ":", ",", "which", "is", "bounded", "between", "and", ".", "Consider", "the", "case", "of", "a", "single", "ground", "-", "truth", "label", ",", "so", "that", "and", "for", "all", ".", "In", "this", "case", ",", "minimizing", "the", "cross", "entropy", "is", "equivalent", "to", "maximizing", "the", "log", "-", "likelihood", "of", "the", "correct", "label", ".", "For", "a", "particular", "example", "with", "label", ",", "the", "log", "-", "likelihood", "is", "maximized", "for", ",", "where", "is", "Dirac", "delta", ",", "which", "equals", "for", "and", "otherwise", ".", "This", "maximum", "is", "not", "achievable", "for", "finite", "but", "is", "approached", "if", "for", "all", "–", "that", "is", ",", "if", "the", "logit", "corresponding", "to", "the", "ground", "-", "truth", "label", "is", "much", "great", "than", "all", "other", "logits", ".", "This", ",", "however", ",", "can", "cause", "two", "problems", ".", "First", ",", "it", "may", "result", "in", "over", "-", "fitting", ":", "if", "the", "model", "learns", "to", "assign", "full", "probability", "to", "the", "ground", "-", "truth", "label", "for", "each", "training", "example", ",", "it", "is", "not", "guaranteed", "to", "generalize", ".", "Second", ",", "it", "encourages", "the", "differences", "between", "the", "largest", "logit", "and", "all", "others", "to", "become", "large", ",", "and", "this", ",", "combined", "with", "the", "bounded", "gradient", ",", "reduces", "the", "ability", "of", "the", "model", "to", "adapt", ".", "Intuitively", ",", "this", "happens", "because", "the", "model", "becomes", "too", "confident", "about", "its", "predictions", ".", "We", "propose", "a", "mechanism", "for", "encouraging", "the", "model", "to", "be", "less", "confident", ".", "While", "this", "may", "not", "be", "desired", "if", "the", "goal", "is", "to", "maximize", "the", "log", "-", "likelihood", "of", "training", "labels", ",", "it", "does", "regularize", "the", "model", "and", "makes", "it", "more", "adaptable", ".", "The", "method", "is", "very", "simple", ".", "Consider", "a", "distribution", "over", "labels", ",", "independent", "of", "the", "training", "example", "x", ",", "and", "a", "smoothing", "parameter", ".", "For", "a", "training", "example", "with", "ground", "-", "truth", "label", ",", "we", "replace", "the", "label", "distribution", "with", "which", "is", "a", "mixture", "of", "the", "original", "ground", "-", "truth", "distribution", "and", "the", "fixed", "distribution", ",", "with", "weights", "and", ",", "respectively", ".", "This", "can", "be", "seen", "as", "the", "distribution", "of", "the", "label", "obtained", "as", "follows", ":", "first", ",", "set", "it", "to", "the", "ground", "-", "truth", "label", ";", "then", ",", "with", "probability", ",", "replace", "with", "a", "sample", "drawn", "from", "the", "distribution", ".", "We", "propose", "to", "use", "the", "prior", "distribution", "over", "labels", "as", ".", "In", "our", "experiments", ",", "we", "used", "the", "uniform", "distribution", ",", "so", "that", "We", "refer", "to", "this", "change", "in", "ground", "-", "truth", "label", "distribution", "as", "label", "-", "smoothing", "regularization", ",", "or", "LSR", ".", "Note", "that", "LSR", "achieves", "the", "desired", "goal", "of", "preventing", "the", "largest", "logit", "from", "becoming", "much", "larger", "than", "all", "others", ".", "Indeed", ",", "if", "this", "were", "to", "happen", ",", "then", "a", "single", "would", "approach", "while", "all", "others", "would", "approach", ".", "This", "would", "result", "in", "a", "large", "cross", "-", "entropy", "with", "because", ",", "unlike", ",", "all", "have", "a", "positive", "lower", "bound", ".", "Another", "interpretation", "of", "LSR", "can", "be", "obtained", "by", "considering", "the", "cross", "entropy", ":", "Thus", ",", "LSR", "is", "equivalent", "to", "replacing", "a", "single", "cross", "-", "entropy", "loss", "with", "a", "pair", "of", "such", "losses", "and", ".", "The", "second", "loss", "penalizes", "the", "deviation", "of", "predicted", "label", "distribution", "from", "the", "prior", ",", "with", "the", "relative", "weight", ".", "Note", "that", "this", "deviation", "could", "be", "equivalently", "captured", "by", "the", "KL", "divergence", ",", "since", "and", "is", "fixed", ".", "When", "is", "the", "uniform", "distribution", ",", "is", "a", "measure", "of", "how", "dissimilar", "the", "predicted", "distribution", "is", "to", "uniform", ",", "which", "could", "also", "be", "measured", "(", "but", "not", "equivalently", ")", "by", "negative", "entropy", ";", "we", "have", "not", "experimented", "with", "this", "approach", ".", "In", "our", "ImageNet", "experiments", "with", "classes", ",", "we", "used", "and", ".", "For", "ILSVRC", "2012", ",", "we", "have", "found", "a", "consistent", "improvement", "of", "about", "absolute", "both", "for", "top", "-", "error", "and", "the", "top", "-", "error", "(", "cf", ".", "Table", "[", "reference", "]", ")", ".", "section", ":", "Training", "Methodology", "We", "have", "trained", "our", "networks", "with", "stochastic", "gradient", "utilizing", "the", "TensorFlow", "distributed", "machine", "learning", "system", "using", "replicas", "running", "each", "on", "a", "NVidia", "Kepler", "GPU", "with", "batch", "size", "for", "epochs", ".", "Our", "earlier", "experiments", "used", "momentum", "with", "a", "decay", "of", ",", "while", "our", "best", "models", "were", "achieved", "using", "RMSProp", "with", "decay", "of", "and", ".", "We", "used", "a", "learning", "rate", "of", ",", "decayed", "every", "two", "epoch", "using", "an", "exponential", "rate", "of", ".", "In", "addition", ",", "gradient", "clipping", "with", "threshold", "was", "found", "to", "be", "useful", "to", "stabilize", "the", "training", ".", "Model", "evaluations", "are", "performed", "using", "a", "running", "average", "of", "the", "parameters", "computed", "over", "time", ".", "section", ":", "Performance", "on", "Lower", "Resolution", "Input", "A", "typical", "use", "-", "case", "of", "vision", "networks", "is", "for", "the", "the", "post", "-", "classification", "of", "detection", ",", "for", "example", "in", "the", "Multibox", "context", ".", "This", "includes", "the", "analysis", "of", "a", "relative", "small", "patch", "of", "the", "image", "containing", "a", "single", "object", "with", "some", "context", ".", "The", "tasks", "is", "to", "decide", "whether", "the", "center", "part", "of", "the", "patch", "corresponds", "to", "some", "object", "and", "determine", "the", "class", "of", "the", "object", "if", "it", "does", ".", "The", "challenge", "is", "that", "objects", "tend", "to", "be", "relatively", "small", "and", "low", "-", "resolution", ".", "This", "raises", "the", "question", "of", "how", "to", "properly", "deal", "with", "lower", "resolution", "input", ".", "The", "common", "wisdom", "is", "that", "models", "employing", "higher", "resolution", "receptive", "fields", "tend", "to", "result", "in", "significantly", "improved", "recognition", "performance", ".", "However", "it", "is", "important", "to", "distinguish", "between", "the", "effect", "of", "the", "increased", "resolution", "of", "the", "first", "layer", "receptive", "field", "and", "the", "effects", "of", "larger", "model", "capacitance", "and", "computation", ".", "If", "we", "just", "change", "the", "resolution", "of", "the", "input", "without", "further", "adjustment", "to", "the", "model", ",", "then", "we", "end", "up", "using", "computationally", "much", "cheaper", "models", "to", "solve", "more", "difficult", "tasks", ".", "Of", "course", ",", "it", "is", "natural", ",", "that", "these", "solutions", "loose", "out", "already", "because", "of", "the", "reduced", "computational", "effort", ".", "In", "order", "to", "make", "an", "accurate", "assessment", ",", "the", "model", "needs", "to", "analyze", "vague", "hints", "in", "order", "to", "be", "able", "to", "“", "hallucinate", "”", "the", "fine", "details", ".", "This", "is", "computationally", "costly", ".", "The", "question", "remains", "therefore", ":", "how", "much", "does", "higher", "input", "resolution", "helps", "if", "the", "computational", "effort", "is", "kept", "constant", ".", "One", "simple", "way", "to", "ensure", "constant", "effort", "is", "to", "reduce", "the", "strides", "of", "the", "first", "two", "layer", "in", "the", "case", "of", "lower", "resolution", "input", ",", "or", "by", "simply", "removing", "the", "first", "pooling", "layer", "of", "the", "network", ".", "For", "this", "purpose", "we", "have", "performed", "the", "following", "three", "experiments", ":", "receptive", "field", "with", "stride", "and", "maximum", "pooling", "after", "the", "first", "layer", ".", "receptive", "field", "with", "stride", "and", "maximum", "pooling", "after", "the", "first", "layer", ".", "receptive", "field", "with", "stride", "and", "without", "pooling", "after", "the", "first", "layer", ".", "All", "three", "networks", "have", "almost", "identical", "computational", "cost", ".", "Although", "the", "third", "network", "is", "slightly", "cheaper", ",", "the", "cost", "of", "the", "pooling", "layer", "is", "marginal", "and", "(", "within", "of", "the", "total", "cost", "of", "the", ")", "network", ".", "In", "each", "case", ",", "the", "networks", "were", "trained", "until", "convergence", "and", "their", "quality", "was", "measured", "on", "the", "validation", "set", "of", "the", "ImageNet", "ILSVRC", "2012", "classification", "benchmark", ".", "The", "results", "can", "be", "seen", "in", "table", "[", "reference", "]", ".", "Although", "the", "lower", "-", "resolution", "networks", "take", "longer", "to", "train", ",", "the", "quality", "of", "the", "final", "result", "is", "quite", "close", "to", "that", "of", "their", "higher", "resolution", "counterparts", ".", "However", ",", "if", "one", "would", "just", "naively", "reduce", "the", "network", "size", "according", "to", "the", "input", "resolution", ",", "then", "network", "would", "perform", "much", "more", "poorly", ".", "However", "this", "would", "an", "unfair", "comparison", "as", "we", "would", "are", "comparing", "a", "16", "times", "cheaper", "model", "on", "a", "more", "difficult", "task", ".", "Also", "these", "results", "of", "table", "[", "reference", "]", "suggest", ",", "one", "might", "consider", "using", "dedicated", "high", "-", "cost", "low", "resolution", "networks", "for", "smaller", "objects", "in", "the", "R", "-", "CNN", "context", ".", "section", ":", "Experimental", "Results", "and", "Comparisons", "Inception", "-", "v2", "RMSProp", "Inception", "-", "v2", "Label", "Smoothing", "Inception", "-", "v2", "Factorized", "Table", "[", "reference", "]", "shows", "the", "experimental", "results", "about", "the", "recognition", "performance", "of", "our", "proposed", "architecture", "(", "Inception", "-", "v2", ")", "as", "described", "in", "Section", "[", "reference", "]", ".", "Each", "Inception", "-", "v2", "line", "shows", "the", "result", "of", "the", "cumulative", "changes", "including", "the", "highlighted", "new", "modification", "plus", "all", "the", "earlier", "ones", ".", "Label", "Smoothing", "refers", "to", "method", "described", "in", "Section", "[", "reference", "]", ".", "Factorized", "includes", "a", "change", "that", "factorizes", "the", "first", "convolutional", "layer", "into", "a", "sequence", "of", "convolutional", "layers", ".", "BN", "-", "auxiliary", "refers", "to", "the", "version", "in", "which", "the", "fully", "connected", "layer", "of", "the", "auxiliary", "classifier", "is", "also", "batch", "-", "normalized", ",", "not", "just", "the", "convolutions", ".", "We", "are", "referring", "to", "the", "model", "in", "last", "row", "of", "Table", "[", "reference", "]", "as", "Inception", "-", "v3", "and", "evaluate", "its", "performance", "in", "the", "multi", "-", "crop", "and", "ensemble", "settings", ".", "All", "our", "evaluations", "are", "done", "on", "the", "48238", "non", "-", "blacklisted", "examples", "on", "the", "ILSVRC", "-", "2012", "validation", "set", ",", "as", "suggested", "by", ".", "We", "have", "evaluated", "all", "the", "50000", "examples", "as", "well", "and", "the", "results", "were", "roughly", "0.1", "%", "worse", "in", "top", "-", "5", "error", "and", "around", "0.2", "%", "in", "top", "-", "1", "error", ".", "In", "the", "upcoming", "version", "of", "this", "paper", ",", "we", "will", "verify", "our", "ensemble", "result", "on", "the", "test", "set", ",", "but", "at", "the", "time", "of", "our", "last", "evaluation", "of", "BN", "-", "Inception", "in", "spring", "indicates", "that", "the", "test", "and", "validation", "set", "error", "tends", "to", "correlate", "very", "well", ".", "section", ":", "Conclusions", "We", "have", "provided", "several", "design", "principles", "to", "scale", "up", "convolutional", "networks", "and", "studied", "them", "in", "the", "context", "of", "the", "Inception", "architecture", ".", "This", "guidance", "can", "lead", "to", "high", "performance", "vision", "networks", "that", "have", "a", "relatively", "modest", "computation", "cost", "compared", "to", "simpler", ",", "more", "monolithic", "architectures", ".", "Our", "highest", "quality", "version", "of", "Inception", "-", "v3", "reaches", ",", "top", "-", "and", "top", "-", "5", "error", "for", "single", "crop", "evaluation", "on", "the", "ILSVR", "2012", "classification", ",", "setting", "a", "new", "state", "of", "the", "art", ".", "This", "is", "achieved", "with", "relatively", "modest", "(", ")", "increase", "in", "computational", "cost", "compared", "to", "the", "network", "described", "in", "Ioffe", "et", "al", ".", "Still", "our", "solution", "uses", "much", "less", "computation", "than", "the", "best", "published", "results", "based", "on", "denser", "networks", ":", "our", "model", "outperforms", "the", "results", "of", "He", "et", "al", "–", "cutting", "the", "top", "-", "(", "top", "-", ")", "error", "by", "(", ")", "relative", ",", "respectively", "–", "while", "being", "six", "times", "cheaper", "computationally", "and", "using", "at", "least", "five", "times", "less", "parameters", "(", "estimated", ")", ".", "Our", "ensemble", "of", "four", "Inception", "-", "v3", "models", "reaches", "with", "multi", "-", "crop", "evaluation", "reaches", "top", "-", "error", "which", "represents", "an", "over", "reduction", "to", "the", "best", "published", "results", "and", "is", "almost", "half", "of", "the", "error", "of", "ILSVRC", "2014", "winining", "GoogLeNet", "ensemble", ".", "We", "have", "also", "demonstrated", "that", "high", "quality", "results", "can", "be", "reached", "with", "receptive", "field", "resolution", "as", "low", "as", ".", "This", "might", "prove", "to", "be", "helpful", "in", "systems", "for", "detecting", "relatively", "small", "objects", ".", "We", "have", "studied", "how", "factorizing", "convolutions", "and", "aggressive", "dimension", "reductions", "inside", "neural", "network", "can", "result", "in", "networks", "with", "relatively", "low", "computational", "cost", "while", "maintaining", "high", "quality", ".", "The", "combination", "of", "lower", "parameter", "count", "and", "additional", "regularization", "with", "batch", "-", "normalized", "auxiliary", "classifiers", "and", "label", "-", "smoothing", "allows", "for", "training", "high", "quality", "networks", "on", "relatively", "modest", "sized", "training", "sets", ".", "bibliography", ":", "References"], "entity": [[4, 6, ["Inception", "Architecture"], "Method"], [7, 9, ["Computer", "Vision"], "Task"], [9, 11, ["Convolutional", "networks"], "Method"], [24, 27, ["computer", "vision", "solutions"], "Method"], [37, 40, ["deep", "convolutional", "networks"], "Method"], [54, 56, ["model", "size"], "Metric"], [57, 59, ["computational", "cost"], "Metric"], [82, 84, ["computational", "efficiency"], "Metric"], [85, 88, ["low", "parameter", "count"], "Metric"], [98, 100, ["mobile", "vision"], "Task"], [101, 105, ["big", "-", "data", "scenarios"], "Task"], [130, 132, ["factorized", "convolutions"], "Method"], [133, 135, ["aggressive", "regularization"], "Method"], [142, 148, ["ILSVRC", "2012", "classification", "challenge", "validation", "set"], "Material"], [158, 160, ["top", "-"], "Metric"], [161, 164, ["top", "-", "error"], "Metric"], [165, 168, ["single", "frame", "evaluation"], "Task"], [173, 175, ["computational", "cost"], "Metric"], [181, 182, ["inference"], "Task"], [193, 196, ["ensemble", "of", "models"], "Method"], [197, 201, ["multi", "-", "crop", "evaluation"], "Task"], [204, 207, ["top", "-", "error"], "Metric"], [208, 211, ["top", "-", "error"], "Metric"], [217, 219, ["2012", "ImageNet"], "Material"], [230, 232, ["AlexNet", "”"], "Method"], [241, 244, ["computer", "vision", "tasks"], "Task"], [248, 251, ["object", "-", "detection"], "Task"], [252, 253, ["segmentation"], "Task"], [254, 257, ["human", "pose", "estimation"], "Task"], [258, 260, ["video", "classification"], "Task"], [261, 263, ["object", "tracking"], "Task"], [265, 266, ["superresolution"], "Task"], [281, 284, ["convolutional", "neural", "networks"], "Method"], [292, 294, ["network", "architectures"], "Method"], [303, 304, ["VGGNet"], "Method"], [305, 306, ["GoogLeNet"], "Method"], [313, 316, ["ILSVRC", "classification", "challenge"], "Task"], [325, 327, ["classification", "performance"], "Metric"], [348, 351, ["deep", "convolutional", "architecture"], "Method"], [360, 363, ["computer", "vision", "tasks"], "Task"], [380, 382, ["network", "quality"], "Metric"], [388, 390, ["convolutional", "networks"], "Method"], [406, 408, ["proposal", "generation"], "Task"], [409, 410, ["detection"], "Task"], [412, 413, ["VGGNet"], "Method"], [443, 445, ["Inception", "architecture"], "Method"], [446, 447, ["GoogLeNet"], "Method"], [466, 467, ["GoogleNet"], "Method"], [482, 483, ["AlexNet"], "Method"], [491, 492, ["VGGNet"], "Method"], [498, 499, ["AlexNet"], "Method"], [501, 503, ["computational", "cost"], "Metric"], [504, 505, ["Inception"], "Method"], [510, 511, ["VGGNet"], "Method"], [524, 526, ["Inception", "networks"], "Method"], [527, 531, ["big", "-", "data", "scenarios"], "Task"], [559, 562, ["mobile", "vision", "settings"], "Task"], [575, 577, ["specialized", "solutions"], "Method"], [591, 593, ["computational", "tricks"], "Method"], [612, 614, ["Inception", "architecture"], "Method"], [629, 631, ["Inception", "architecture"], "Method"], [654, 656, ["computational", "gains"], "Metric"], [682, 684, ["GoogLeNet", "architecture"], "Method"], [717, 721, ["Inception", "-", "style", "model"], "Method"], [732, 735, ["filter", "bank", "sizes"], "Method"], [743, 745, ["computational", "cost"], "Metric"], [784, 786, ["optimization", "ideas"], "Method"], [795, 797, ["convolution", "networks"], "Method"], [808, 812, ["Inception", "-", "type", "networks"], "Method"], [827, 831, ["Inception", "style", "building", "blocks"], "Method"], [848, 850, ["dimensional", "reduction"], "Method"], [855, 857, ["Inception", "modules"], "Method"], [907, 909, ["design", "principles"], "Method"], [920, 922, ["convolutional", "networks"], "Method"], [946, 947, ["accuracy"], "Metric"], [986, 988, ["representational", "bottlenecks"], "Method"], [995, 999, ["Feed", "-", "forward", "networks"], "Method"], [1015, 1018, ["classifier", "or", "regressor"], "Method"], [1026, 1028, ["information", "flow"], "Task"], [1057, 1058, ["compression"], "Task"], [1120, 1123, ["Higher", "dimensional", "representations"], "Method"], [1139, 1141, ["convolutional", "network"], "Method"], [1154, 1156, ["Spatial", "aggregation"], "Task"], [1169, 1171, ["representational", "power"], "Metric"], [1184, 1185, ["convolution"], "Method"], [1197, 1199, ["spatial", "aggregation"], "Method"], [1227, 1229, ["dimension", "reduction"], "Task"], [1251, 1253, ["dimension", "reduction"], "Method"], [1255, 1257, ["faster", "learning"], "Task"], [1386, 1388, ["Factorizing", "Convolutions"], "Method"], [1390, 1392, ["Filter", "Size"], "Method"], [1399, 1401, ["GoogLeNet", "network"], "Method"], [1408, 1410, ["dimension", "reduction"], "Method"], [1420, 1422, ["factorizing", "convolutions"], "Method"], [1435, 1437, ["convolutional", "layer"], "Method"], [1440, 1442, ["convolutional", "layer"], "Method"], [1445, 1447, ["vision", "network"], "Method"], [1475, 1476, ["aggregation"], "Task"], [1483, 1486, ["expressive", "local", "representations"], "Method"], [1493, 1495, ["factorizing", "convolutions"], "Method"], [1505, 1507, ["computational", "efficiency"], "Metric"], [1512, 1514, ["Inception", "networks"], "Method"], [1516, 1517, ["convolutional"], "Method"], [1532, 1534, ["computational", "cost"], "Metric"], [1546, 1547, ["factorization"], "Method"], [1560, 1561, ["training"], "Task"], [1602, 1603, ["Convolutions"], "Method"], [1619, 1620, ["computation"], "Task"], [1625, 1628, ["convolution", "with", "filters"], "Method"], [1630, 1633, ["grid", "with", "filters"], "Method"], [1645, 1646, ["convolution"], "Method"], [1657, 1658, ["filter"], "Method"], [1679, 1681, ["geometric", "size"], "Metric"], [1690, 1691, ["expressiveness"], "Metric"], [1699, 1700, ["convolution"], "Method"], [1705, 1709, ["multi", "-", "layer", "network"], "Method"], [1730, 1731, ["convolution"], "Method"], [1741, 1745, ["fully", "-", "connected", "network"], "Method"], [1764, 1766, ["vision", "network"], "Method"], [1778, 1781, ["fully", "connected", "component"], "Method"], [1784, 1787, ["layer", "convolutional", "architecture"], "Method"], [1793, 1794, ["convolution"], "Method"], [1799, 1802, ["fully", "connected", "layer"], "Method"], [1834, 1835, ["convolution"], "Method"], [1839, 1840, ["convolution"], "Method"], [1871, 1874, ["computational", "cost", "savings"], "Metric"], [1915, 1916, ["convolution"], "Method"], [1917, 1918, ["aggregating"], "Task"], [1938, 1940, ["layer", "replacement"], "Method"], [1998, 2000, ["computational", "cost"], "Metric"], [2009, 2011, ["convolutional", "layers"], "Method"], [2028, 2031, ["reduction", "of", "computation"], "Task"], [2138, 2141, ["rectified", "linear", "units"], "Method"], [2146, 2147, ["factorization"], "Task"], [2180, 2182, ["linear", "activations"], "Method"], [2184, 2187, ["dimension", "reduction", "components"], "Method"], [2190, 2192, ["Spatial", "Factorization"], "Method"], [2193, 2195, ["Asymmetric", "Convolutions"], "Method"], [2200, 2201, ["convolutions"], "Method"], [2220, 2222, ["convolutional", "layers"], "Method"], [2239, 2240, ["convolutions"], "Method"], [2255, 2257, ["asymmetric", "convolutions"], "Method"], [2264, 2265, ["convolution"], "Method"], [2268, 2269, ["convolution"], "Method"], [2274, 2277, ["two", "layer", "network"], "Method"], [2296, 2300, ["two", "-", "layer", "solution"], "Method"], [2326, 2327, ["convolution"], "Method"], [2330, 2331, ["convolution"], "Method"], [2354, 2355, ["convolution"], "Method"], [2357, 2358, ["convolution"], "Method"], [2361, 2362, ["convolution"], "Method"], [2364, 2367, ["computational", "cost", "saving"], "Metric"], [2429, 2430, ["convolutions"], "Method"], [2432, 2433, ["convolutions"], "Method"], [2438, 2440, ["Auxiliary", "Classifiers"], "Method"], [2445, 2447, ["auxiliary", "classifiers"], "Method"], [2450, 2455, ["convergence", "of", "very", "deep", "networks"], "Task"], [2478, 2479, ["training"], "Task"], [2482, 2485, ["vanishing", "gradient", "problem"], "Task"], [2487, 2489, ["deep", "networks"], "Method"], [2496, 2498, ["auxiliary", "classifiers"], "Method"], [2501, 2502, ["learning"], "Task"], [2504, 2505, ["convergence"], "Metric"], [2511, 2513, ["auxiliary", "classifiers"], "Method"], [2542, 2543, ["accuracy"], "Metric"], [2560, 2561, ["accuracy"], "Metric"], [2647, 2649, ["auxiliary", "classifiers"], "Method"], [2651, 2652, ["regularizer"], "Method"], [2662, 2663, ["classifier"], "Method"], [2679, 2681, ["dropout", "layer"], "Method"], [2693, 2695, ["batch", "normalization"], "Task"], [2698, 2699, ["regularizer"], "Method"], [2703, 2706, ["Grid", "Size", "Reduction"], "Task"], [2708, 2710, ["convolutional", "networks"], "Method"], [2712, 2714, ["pooling", "operation"], "Method"], [2734, 2738, ["maximum", "or", "average", "pooling"], "Method"], [2743, 2745, ["network", "filters"], "Method"], [2775, 2781, ["stride", "-", "1", "convolution", "with", "filters"], "Method"], [2786, 2788, ["pooling", "step"], "Method"], [2794, 2796, ["computational", "cost"], "Metric"], [2801, 2802, ["convolution"], "Method"], [2816, 2817, ["pooling"], "Method"], [2818, 2819, ["convolution"], "Method"], [2825, 2827, ["computational", "cost"], "Metric"], [2872, 2874, ["computational", "cost"], "Metric"], [2903, 2905, ["pooling", "layer"], "Method"], [2907, 2911, ["average", "or", "maximum", "pooling"], "Method"], [2982, 2983, ["convolution"], "Method"], [2985, 2986, ["convolutions"], "Method"], [3010, 3012, ["inception", "modules"], "Method"], [3023, 3026, ["grid", "with", "filters"], "Method"], [3028, 3031, ["grid", "reduction", "technique"], "Method"], [3046, 3049, ["factorized", "inception", "modules"], "Method"], [3065, 3068, ["grid", "reduction", "technique"], "Method"], [3083, 3085, ["Inception", "modules"], "Method"], [3117, 3119, ["filter", "banks"], "Method"], [3121, 3123, ["Inception", "modules"], "Method"], [3183, 3185, ["computation", "cost"], "Metric"], [3192, 3193, ["GoogLeNet"], "Method"], [3201, 3202, ["VGGNet"], "Method"], [3205, 3207, ["Model", "Regularization"], "Task"], [3208, 3210, ["Label", "Smoothing"], "Method"], [3218, 3220, ["classifier", "layer"], "Method"], [3223, 3229, ["marginalized", "effect", "of", "label", "-", "dropout"], "Method"], [3297, 3299, ["cross", "entropy"], "Metric"], [3329, 3333, ["Cross", "-", "entropy", "loss"], "Metric"], [3346, 3351, ["gradient", "training", "of", "deep", "models"], "Task"], [3481, 3484, ["over", "-", "fitting"], "Task"], [3740, 3744, ["label", "-", "smoothing", "regularization"], "Method"], [3746, 3747, ["LSR"], "Method"], [3750, 3751, ["LSR"], "Method"], [3793, 3796, ["cross", "-", "entropy"], "Metric"], [3811, 3812, ["LSR"], "Method"], [3823, 3824, ["LSR"], "Method"], [3922, 3923, ["ImageNet"], "Material"], [3932, 3934, ["ILSVRC", "2012"], "Material"], [3946, 3949, ["top", "-", "error"], "Metric"], [3951, 3954, ["top", "-", "error"], "Metric"], [3973, 3975, ["stochastic", "gradient"], "Method"], [3977, 3982, ["TensorFlow", "distributed", "machine", "learning", "system"], "Method"], [4014, 4015, ["RMSProp"], "Method"], [4023, 4025, ["learning", "rate"], "Metric"], [4052, 4053, ["training"], "Task"], [4082, 4084, ["vision", "networks"], "Method"], [4088, 4093, ["post", "-", "classification", "of", "detection"], "Task"], [4098, 4100, ["Multibox", "context"], "Task"], [4194, 4196, ["recognition", "performance"], "Metric"], [4224, 4225, ["computation"], "Task"], [4274, 4276, ["computational", "effort"], "Metric"], [4324, 4326, ["computational", "effort"], "Metric"], [4360, 4363, ["first", "pooling", "layer"], "Method"], [4383, 4385, ["maximum", "pooling"], "Method"], [4395, 4397, ["maximum", "pooling"], "Method"], [4408, 4409, ["pooling"], "Method"], [4420, 4422, ["computational", "cost"], "Metric"], [4435, 4437, ["pooling", "layer"], "Method"], [4472, 4477, ["ImageNet", "ILSVRC", "2012", "classification", "benchmark"], "Material"], [4491, 4495, ["lower", "-", "resolution", "networks"], "Method"], [4579, 4585, ["high", "-", "cost", "low", "resolution", "networks"], "Method"], [4590, 4594, ["R", "-", "CNN", "context"], "Task"], [4601, 4602, ["Inception"], "Method"], [4604, 4606, ["RMSProp", "Inception"], "Method"], [4609, 4611, ["Smoothing", "Inception"], "Method"], [4624, 4625, ["recognition"], "Task"], [4631, 4632, ["Inception"], "Method"], [4666, 4668, ["Label", "Smoothing"], "Task"], [4678, 4679, ["Factorized"], "Method"], [4692, 4694, ["convolutional", "layers"], "Method"], [4695, 4698, ["BN", "-", "auxiliary"], "Method"], [4710, 4712, ["auxiliary", "classifier"], "Method"], [4721, 4722, ["convolutions"], "Method"], [4738, 4741, ["Inception", "-", "v3"], "Method"], [4747, 4753, ["multi", "-", "crop", "and", "ensemble", "settings"], "Task"], [4768, 4773, ["ILSVRC", "-", "2012", "validation", "set"], "Material"], [4796, 4800, ["top", "-", "5", "error"], "Metric"], [4805, 4809, ["top", "-", "1", "error"], "Metric"], [4838, 4839, ["BN"], "Method"], [4846, 4851, ["test", "and", "validation", "set", "error"], "Metric"], [4869, 4871, ["convolutional", "networks"], "Method"], [4879, 4881, ["Inception", "architecture"], "Method"], [4889, 4891, ["vision", "networks"], "Task"], [4896, 4898, ["computation", "cost"], "Metric"], [4903, 4905, ["monolithic", "architectures"], "Method"], [4911, 4914, ["Inception", "-", "v3"], "Method"], [4919, 4923, ["top", "-", "5", "error"], "Metric"], [4924, 4927, ["single", "crop", "evaluation"], "Task"], [4929, 4932, ["ILSVR", "2012", "classification"], "Material"], [4951, 4953, ["computational", "cost"], "Metric"], [4977, 4979, ["denser", "networks"], "Method"], [5025, 5032, ["ensemble", "of", "four", "Inception", "-", "v3", "models"], "Method"], [5034, 5038, ["multi", "-", "crop", "evaluation"], "Method"], [5039, 5042, ["top", "-", "error"], "Metric"], [5060, 5065, ["ILSVRC", "2014", "winining", "GoogLeNet", "ensemble"], "Method"], [5094, 5098, ["detecting", "relatively", "small", "objects"], "Task"], [5103, 5105, ["factorizing", "convolutions"], "Method"], [5106, 5109, ["aggressive", "dimension", "reductions"], "Method"], [5110, 5112, ["neural", "network"], "Method"], [5118, 5121, ["low", "computational", "cost"], "Metric"], [5124, 5125, ["quality"], "Metric"], [5134, 5135, ["regularization"], "Method"], [5136, 5141, ["batch", "-", "normalized", "auxiliary", "classifiers"], "Method"], [5142, 5145, ["label", "-", "smoothing"], "Method"], [2957, 2959, ["ILSVRC", "2012"], "Material"], [4916, 4918, ["top", "-"], "Metric"], [4992, 4994, ["top", "-"], "Metric"], [4995, 4997, ["top", "-"], "Metric"]]} {"doc_id": "0678a8abea82793993cd89383319da75f6dc4be3", "text": ["document", ":", "ProNet", ":", "Learning", "to", "Propose", "Object", "-", "specific", "Boxes", "for", "Cascaded", "Neural", "Networks", "This", "paper", "aims", "to", "classify", "and", "locate", "objects", "accurately", "and", "efficiently", ",", "without", "using", "bounding", "box", "annotations", ".", "It", "is", "challenging", "as", "objects", "in", "the", "wild", "could", "appear", "at", "arbitrary", "locations", "and", "in", "different", "scales", ".", "In", "this", "paper", ",", "we", "propose", "a", "novel", "classification", "architecture", "ProNet", "based", "on", "convolutional", "neural", "networks", ".", "It", "uses", "computationally", "efficient", "neural", "networks", "to", "propose", "image", "regions", "that", "are", "likely", "to", "contain", "objects", ",", "and", "applies", "more", "powerful", "but", "slower", "networks", "on", "the", "proposed", "regions", ".", "The", "basic", "building", "block", "is", "a", "multi", "-", "scale", "fully", "-", "convolutional", "network", "which", "assigns", "object", "confidence", "scores", "to", "boxes", "at", "different", "locations", "and", "scales", ".", "We", "show", "that", "such", "networks", "can", "be", "trained", "effectively", "using", "image", "-", "level", "annotations", ",", "and", "can", "be", "connected", "into", "cascades", "or", "trees", "for", "efficient", "object", "classification", ".", "ProNet", "outperforms", "previous", "state", "-", "of", "-", "the", "-", "art", "significantly", "on", "PASCAL", "VOC", "2012", "and", "MS", "COCO", "datasets", "for", "object", "classification", "and", "point", "-", "based", "localization", ".", "section", ":", "Introduction", "We", "address", "the", "problem", "of", "object", "classification", "and", "localization", "in", "natural", "images", ".", "As", "objects", "could", "be", "small", "and", "appear", "at", "arbitrary", "locations", ",", "several", "frameworks", "rely", "on", "bounding", "boxes", "to", "train", "object", "-", "centric", "classifiers", ",", "and", "apply", "the", "classifiers", "by", "searching", "over", "different", "locations", "of", "the", "images", ".", "However", ",", "the", "annotation", "process", "for", "object", "bounding", "boxes", "is", "usually", "resource", "intensive", "and", "difficult", "to", "scale", "up", ".", "In", "light", "of", "this", ",", "we", "aim", "to", "simultaneously", "classify", "and", "locate", "objects", "given", "only", "image", "-", "level", "annotations", "for", "training", ".", "To", "cope", "with", "the", "lack", "of", "object", "-", "level", "annotations", ",", "several", "methods", "extract", "feature", "activations", "from", "convolutional", "neural", "networks", "(", "CNN", ")", "by", "scanning", "over", "different", "image", "regions", ".", "They", "then", "aggregate", "the", "extracted", "features", "into", "image", "-", "level", "representations", "for", "classification", "purpose", ".", "Under", "this", "scheme", ",", "regions", "that", "belong", "to", "the", "background", "are", "considered", "as", "important", "as", "regions", "that", "contain", "objects", ".", "Such", "global", "approaches", "tend", "to", "be", "sensitive", "to", "background", ",", "and", "can", "not", "be", "used", "directly", "for", "localization", ".", "We", "choose", "to", "use", "the", "fully", "-", "convolutional", "network", "(", "FCN", ")", "architecture", "for", "simultaneous", "object", "classification", "and", "localization", ".", "It", "replaces", "the", "fully", "-", "connected", "layers", "of", "a", "standard", "CNN", "(", "e.g.", "AlexNet", ")", "with", "convolutional", "layers", ".", "This", "enables", "an", "FCN", "to", "take", "images", "of", "arbitrary", "sizes", ",", "and", "generate", "classification", "score", "maps", "efficiently", ".", "Each", "element", "in", "a", "score", "map", "corresponds", "to", "a", "rectangular", "box", "(", "receptive", "field", ")", "in", "the", "original", "image", ".", "The", "score", "maps", "can", "then", "be", "used", "for", "classification", "and", "localization", ".", "The", "sampling", "strides", "and", "box", "sizes", "are", "determined", "by", "the", "FCN", "’s", "network", "architecture", ".", "As", "box", "sizes", "are", "fixed", ",", "FCN", "might", "face", "difficulty", "dealing", "with", "objects", "of", "different", "scales", ".", "We", "address", "this", "problem", "by", "using", "a", "multi", "-", "stream", "multi", "-", "scale", "architecture", ".", "All", "streams", "share", "the", "same", "parameters", ",", "but", "take", "input", "images", "of", "different", "scales", ".", "To", "train", "the", "multi", "-", "scale", "FCN", "without", "object", "-", "level", "annotations", ",", "we", "generate", "image", "-", "level", "scores", "by", "pooling", "the", "score", "maps", "over", "multiple", "-", "scales", ",", "and", "compute", "the", "losses", "with", "image", "-", "level", "labels", "for", "back", "-", "propagation", ".", "Once", "a", "multi", "-", "scale", "FCN", "is", "trained", ",", "it", "can", "be", "used", "for", "classification", "and", "localization", "directly", ".", "From", "another", "perspective", ",", "it", "also", "proposes", "a", "set", "of", "promising", "boxes", "that", "are", "likely", "to", "contain", "objects", ".", "We", "can", "then", "build", "a", "cascade", "architecture", "by", "zooming", "onto", "those", "promising", "boxes", ",", "and", "train", "new", "classifiers", "to", "verify", "them", ".", "The", "cascade", "allows", "the", "system", "to", "balance", "accuracy", "and", "speed", ":", "each", "stage", "filters", "out", "parts", "of", "image", "regions", "that", "are", "unlikely", "to", "contain", "objects", ".", "We", "name", "this", "propose", "and", "zoom", "pipeline", "as", "ProNet", ".", "Figure", "[", "reference", "]", "provides", "the", "high", "-", "level", "intuition", "behind", "ProNet", ":", "three", "boxes", "are", "proposed", "for", "bird", ",", "potted", "plant", "and", "cat", "categories", ".", "The", "boxes", "are", "cropped", "out", "and", "verified", "further", ",", "until", "a", "certain", "decision", "is", "made", ".", "To", "train", "the", "later", "classifiers", "in", "ProNet", ",", "we", "sample", "hard", "negatives", "based", "on", "image", "-", "level", "labels", ".", "For", "positives", ",", "as", "no", "object", "-", "level", "annotations", "are", "available", ",", "it", "is", "impossible", "to", "tell", "objects", "from", "background", ".", "To", "avoid", "over", "-", "fitting", ",", "we", "randomly", "sample", "positive", "boxes", "above", "a", "relative", "low", "threshold", ".", "Different", "positive", "boxes", "from", "the", "same", "image", "can", "be", "sampled", "at", "different", "iterations", "of", "the", "stochastic", "gradient", "descent", "training", "process", ".", "At", "test", "time", ",", "only", "a", "small", "subset", "of", "boxes", "(", "10", "to", "20", "per", "image", ")", "with", "highest", "object", "confidence", "scores", "are", "fed", "to", "the", "later", "classifiers", ".", "This", "allows", "us", "to", "utilize", "CNNs", "that", "have", "stronger", "representation", "power", "with", "little", "computational", "overhead", ".", "ProNet", "is", "highly", "configurable", ":", "for", "example", ",", "one", "could", "set", "a", "list", "of", "important", "object", "categories", ",", "and", "only", "verify", "the", "proposed", "boxes", "for", "those", "categories", ".", "Moreover", ",", "apart", "from", "a", "traditional", "chain", "-", "structured", "cascade", ",", "we", "show", "that", "it", "is", "also", "possible", "to", "build", "tree", "-", "structured", "cascades", ",", "where", "each", "branch", "handles", "categories", "from", "a", "particular", "domain", "(", "set", "of", "vehicles", "or", "animals", ")", ".", "In", "summary", ",", "our", "paper", "makes", "the", "following", "contributions", ":", "We", "propose", "ProNet", ",", "a", "cascaded", "neural", "network", "framework", "that", "zooms", "onto", "promising", "object", "-", "specific", "boxes", "for", "efficient", "object", "classification", "and", "localization", ".", "We", "introduce", "strategies", "to", "train", "ProNet", "with", "image", "-", "level", "annotations", "effectively", ";", "and", "demonstrate", "the", "implementations", "of", "chain", "-", "and", "tree", "-", "structured", "cascades", ".", "We", "show", "that", "ProNet", "outperforms", "previous", "state", "-", "of", "-", "the", "-", "art", "significantly", "on", "the", "object", "classification", "and", "point", "-", "based", "localization", "tasks", "of", "the", "PASCAL", "VOC", "2012", "dataset", "and", "the", "recently", "released", "MS", "COCO", "dataset", ".", "section", ":", "Related", "Work", "Object", "classification", "is", "a", "fundamental", "problem", "in", "Computer", "Vision", ".", "Earlier", "work", "focused", "on", "classification", "from", "object", "-", "centric", "images", ".", "They", "usually", "extract", "hand", "-", "crafted", "low", "-", "level", "features", "and", "aggregate", "the", "features", "into", "image", "-", "level", "feature", "vectors", ".", "More", "challenging", "datasets", "have", "since", "been", "collected", ".", "They", "are", "of", "larger", "scale", ",", "and", "contain", "smaller", "objects", "which", "could", "be", "partially", "occluded", ".", "Recently", ",", "deep", "convolutional", "neural", "networks", "(", "CNN", ")", "have", "achieved", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "a", "wide", "range", "of", "visual", "recognition", "tasks", ",", "including", "object", "classification", "and", "detection", ".", "Although", "CNNs", "require", "large", "amount", "of", "data", "for", "training", ",", "it", "has", "been", "shown", "that", "they", "are", "able", "to", "learn", "representations", "that", "generalize", "to", "other", "tasks", ".", "Such", "representations", "can", "be", "adapted", "to", "image", "classification", "by", "fine", "-", "tuning", ",", "or", "extracted", "as", "holistic", "features", "for", "classification", "with", "linear", "SVMs", ".", "When", "used", "as", "generic", "feature", "extractors", ",", "feature", "aggregation", "techniques", "designed", "for", "hand", "-", "crafted", "features", "can", "also", "work", "with", "CNN", "embeddings", "and", "achieve", "competitive", "performance", ".", "An", "alternative", "approach", "for", "object", "classification", "is", "via", "detection", ".", "Among", "those", "utilizing", "bounding", "box", "annotations", ",", "RCNN", "achieves", "competitive", "performance", "by", "directly", "representing", "image", "boxes", "with", "CNN", "features", "and", "learning", "classifiers", "on", "top", "of", "the", "features", ".", "Object", "proposal", "techniques", "are", "used", "to", "sample", "the", "image", "patches", "for", "classification", ".", "A", "recent", "framework", ",", "fast", "RCNN", ",", "uses", "fully", "-", "convolutional", "networks", "(", "FCN", ")", "to", "generate", "box", "-", "level", "features", "in", "batch", ",", "and", "is", "thus", "more", "computational", "efficient", ".", "Object", "localization", "with", "image", "-", "level", "annotations", "is", "a", "weakly", "-", "supervised", "problem", ".", "It", "can", "be", "formulated", "as", "a", "multiple", "instance", "learning", "problem", ",", "and", "has", "been", "addressed", "to", "learn", "concept", "detectors", "from", "Internet", "data", ".", "It", "has", "also", "been", "studied", "for", "object", "detection", "and", "segmentation", ".", "For", "object", "classification", ",", "Wei", "et", "al", ".", "treat", "images", "as", "bags", "of", "patches", ",", "where", "the", "patches", "are", "selected", "using", "objectness", "criteria", ".", "They", "then", "use", "max", "pooling", "to", "fine", "-", "tune", "CNNs", "based", "on", "image", "-", "level", "annotations", ".", "Oquab", "et", "al", ".", "follow", "a", "similar", "approach", ",", "but", "make", "the", "training", "process", "end", "-", "to", "-", "end", "by", "converting", "CNNs", "into", "FCNs", ".", "The", "proposal", "generation", "network", "in", "ProNet", "is", "also", "based", "on", "FCN", ",", "but", "uses", "a", "multi", "-", "stream", "architecture", "and", "cross", "-", "scale", "LSE", "pooling", "to", "achieve", "scale", "-", "awareness", ".", "Cascaded", "classifiers", "are", "a", "well", "-", "studied", "technique", "in", "Computer", "Vision", ".", "Cascades", "with", "CNNs", "have", "been", "explored", "for", "facial", "point", "detection", ",", "face", "detection", "and", "pose", "estimation", ".", "However", ",", "such", "methods", "require", "fully", "annotated", "training", "examples", ".", "ProNet", "adopts", "the", "cascade", "philosophy", "to", "balance", "speed", "and", "accuracy", ",", "but", "does", "not", "require", "object", "bounding", "boxes", "for", "training", ".", "Since", "ProNet", "is", "a", "general", "object", "classifier", ",", "it", "can", "also", "be", "extended", "to", "have", "tree", "structure", ",", "where", "each", "leaf", "is", "a", "domain", "expert", ".", "section", ":", "ProNet", "Framework", "ProNet", "has", "two", "basic", "components", ":", "an", "object", "-", "specific", "box", "proposal", "unit", ",", "and", "a", "verification", "unit", ".", "For", "each", "image", ",", "for", "each", "object", "category", ",", "the", "box", "proposal", "unit", "generates", "a", "list", "of", "confidence", "scores", "of", "the", "presence", "of", "the", "object", "instances", ",", "and", "the", "coordinates", "indicating", "the", "locations", "of", "the", "objects", ".", "ProNet", "then", "zooms", "onto", "image", "boxes", "with", "higher", "scores", "to", "further", "verify", "if", "they", "are", "positive", "or", "hard", "negatives", ".", "The", "verification", "units", "can", "either", "take", "all", "boxes", ",", "which", "forms", "a", "chain", "structure", ";", "or", "a", "subset", "of", "boxes", "corresponding", "to", "certain", "domains", "(", "animal", ")", ",", "which", "forms", "a", "tree", "structure", ".", "We", "implement", "these", "two", "units", "with", "convolutional", "neural", "networks", ".", "Figure", "[", "reference", "]", "illustrates", "the", "overall", "ProNet", "framework", ".", "subsection", ":", "Proposal", "Generation", "The", "first", "stage", "in", "our", "framework", "is", "to", "generate", "object", "-", "specific", "box", "proposals", "with", "CNNs", ".", "For", "an", "input", "image", "and", "object", "category", ",", "we", "want", "to", "learn", "a", "proposal", "scoring", "function", "where", "corresponds", "to", "the", "location", "of", "a", "rectangular", "image", "region", "denoted", "by", "its", "top", "left", "and", "bottom", "right", "corners", ".", "A", "typical", "CNN", "architecture", "for", "image", "classification", "task", "(", "e.g.", "AlexNet", ")", "involves", "a", "hierarchy", "of", "convolutional", "layers", "and", "fully", "connected", "layers", ".", "The", "convolutional", "layers", "operate", "on", "local", "image", "patches", "to", "extract", "feature", "representations", ".", "For", "a", "color", "image", "with", "3", "channels", ",", "the", "convolutional", "layers", "generate", "a", "feature", "map", "of", "elements", ",", "where", "is", "the", "output", "feature", "dimension", ".", "and", "correspond", "to", "the", "width", "and", "height", "of", "the", "feature", "map", ",", "they", "are", "controlled", "by", "input", "image", "size", ",", "as", "well", "as", "the", "kernel", "size", ",", "sampling", "step", "and", "padding", "size", "of", "the", "convolutional", "layers", ".", "The", "fully", "connected", "layers", "serve", "as", "classifiers", "which", "take", "fixed", "-", "size", "inputs", ",", "thus", "require", "the", "width", "and", "height", "of", "input", "images", "to", "be", "fixed", ".", "Therefore", ",", "one", "possible", "way", "to", "compute", "is", "to", "enumerate", "locations", "and", "scales", "in", "a", "sliding", "window", "fashion", "or", "with", "bounding", "box", "proposals", ",", "and", "feed", "such", "image", "regions", "to", "CNNs", ".", "We", "take", "an", "alternative", "approach", "based", "on", "fully", "convolutional", "networks", "(", "e.g.", "OverFeat", ")", ".", "Fully", "convolutional", "networks", "(", "FCN", ")", "do", "not", "contain", "fully", "-", "connected", "layers", ".", "Rather", ",", "they", "use", "only", "the", "convolutional", "layers", ",", "which", "allows", "them", "to", "process", "images", "of", "arbitrary", "sizes", ".", "The", "outputs", "of", "FCNs", "are", "in", "the", "form", "of", "feature", "maps", ",", "where", "is", "the", "number", "of", "categories", ".", "Each", "element", "in", "a", "feature", "map", "corresponds", "to", "the", "activation", "response", "for", "a", "particular", "category", "over", "a", "certain", "region", ".", "Such", "regions", "are", "called", "receptive", "fields", "for", "the", "activations", ".", "Compared", "with", "region", "sampling", "with", "sliding", "windows", "or", "bounding", "box", "proposals", ",", "FCNs", "offer", "a", "seamless", "solution", "for", "end", "-", "to", "-", "end", "training", "under", "the", "CNN", "framework", ",", "and", "also", "naturally", "allow", "the", "sharing", "of", "intermediate", "features", "over", "overlapping", "image", "regions", ".", "Scale", "adaptation", "with", "multi", "-", "stream", "FCNs", ".", "One", "issue", "in", "use", "of", "FCNs", "is", "that", "the", "sizes", "of", "receptive", "fields", "are", "typically", "fixed", ",", "while", "the", "object", "scales", "may", "vary", "a", "lot", ".", "We", "address", "this", "problem", "by", "using", "a", "multi", "-", "stream", "architecture", ".", "Assume", "an", "FCN", "has", "been", "trained", "with", "inputs", "where", "objects", "have", "been", "resized", "to", "the", "same", "scale", ".", "We", "expand", "the", "network", "into", "streams", ",", "where", "every", "stream", "shares", "the", "same", "parameters", "as", "the", "pre", "-", "trained", "one", ".", "Given", "an", "image", ",", "we", "scale", "it", "to", "different", "sizes", "and", "feed", "to", "the", "-", "stream", "FCN", ".", "The", "output", "feature", "map", "of", "each", "stream", "corresponds", "to", "a", "different", "scale", "in", "the", "original", "image", ".", "Training", "with", "image", "-", "level", "annotations", ".", "When", "object", "bounding", "boxes", "are", "available", ",", "training", "FCNs", "is", "straight", "-", "forward", ":", "one", "could", "either", "crop", "images", "with", "the", "bounding", "boxes", ",", "or", "use", "a", "loss", "function", "which", "operates", "directly", "on", "feature", "maps", "and", "takes", "the", "object", "locations", "into", "account", ".", "As", "such", "supervision", "is", "absent", ",", "we", "need", "to", "aggregate", "local", "responses", "into", "global", "ones", "so", "that", "image", "-", "level", "labels", "can", "be", "used", "for", "training", ".", "We", "use", "the", "log", "-", "sum", "-", "exp", "(", "LSE", ")", "pooling", "function", "applied", "by", "for", "semantic", "segmentation", ":", "where", "is", "the", "category", ",", "corresponds", "to", "the", "-", "th", "stream", "of", "FCN", ",", "correspond", "to", "location", "in", "the", "feature", "map", ",", "is", "the", "total", "number", "of", "such", "elements", "and", "is", "a", "hyper", "parameter", ".", "The", "function", "’s", "output", "is", "close", "to", "average", "when", "is", "small", "and", "maximum", "when", "is", "large", ".", "Setting", "larger", "makes", "the", "aggregation", "focus", "on", "a", "smaller", "subset", "of", "image", "boxes", ",", "and", "has", "the", "potential", "to", "handle", "smaller", "objects", "better", ".", "LSE", "pooling", "function", "can", "be", "implemented", "as", "a", "layer", "in", "a", "neural", "network", ".", "As", "illustrated", "in", "Figure", "[", "reference", "]", ",", "it", "is", "connected", "to", "the", "final", "layers", "of", "all", "-", "stream", "FCNs", "and", "produces", "a", "dimensional", "vector", "for", "each", "image", ".", "We", "then", "compute", "the", "loss", "for", "each", "category", "and", "back", "-", "propagate", "the", "error", "gradients", "to", "the", "earlier", "layers", ".", "Computing", "proposal", "scores", ".", "Once", "the", "FCNs", "have", "been", "trained", ",", "we", "compute", "proposal", "scores", "from", "the", "feature", "maps", ".", "Specifically", ",", "for", "every", "neuron", "in", "the", "final", "layer", "of", "single", "-", "stream", "FCN", ",", "we", "compute", "its", "receptive", "field", "and", "use", "it", "as", "the", "location", ";", "the", "corresponding", "activation", "of", "the", "neuron", "is", "used", "as", "proposal", "score", ".", "Although", "the", "exact", "receptive", "field", "may", "vary", "due", "to", "different", "padding", "strategies", ",", "we", "use", "a", "simple", "estimation", "which", "has", "been", "reported", "to", "work", "well", "in", "practice", ".", "Denote", "the", "sampling", "stride", "of", "a", "spatial", "convolutional", "layer", "as", "and", "the", "kernel", "size", "of", "a", "max", "pooling", "layer", "as", ",", "the", "overall", "sampling", "stride", "is", "given", "by", "where", "is", "the", "collection", "of", "all", "convolutional", "layers", "and", "is", "the", "collection", "of", "all", "max", "pooling", "layers", ".", "Implementation", ".", "Our", "-", "stream", "FCNs", "are", "implemented", "with", "Torch", ".", "For", "each", "stream", ",", "we", "use", "the", "CNN", "-", "M", "2048", "architecture", "proposed", "in", ".", "It", "has", "5", "convolutional", "layers", "and", "3", "fully", "-", "connected", "layers", ".", "It", "achieves", "higher", "accuracy", "on", "ImageNet", "than", "AlexNet", ",", "while", "being", "faster", "and", "less", "memory", "consuming", "than", "very", "deep", "CNNs", ".", "We", "use", "the", "model", "parameters", "released", "by", "the", "authors", ",", "which", "were", "pre", "-", "trained", "from", "ImageNet", "dataset", "with", "1", ",", "000", "categories", ".", "We", "convert", "the", "model", "into", "an", "FCN", "by", "replacing", "the", "three", "fully", "-", "connected", "layers", "with", "convolutional", "layers", ".", "The", "first", "convolutional", "layer", "has", "512", "input", "planes", ",", "4096", "output", "planes", "and", "kernel", "size", "of", "6", ".", "The", "second", "has", "4096", "input", "planes", ",", "2048", "output", "planes", "and", "kernel", "size", "of", "1", ".", "Since", "the", "final", "layer", "is", "task", "-", "specific", ",", "it", "is", "initialized", "from", "scratch", "with", "2048", "input", "planes", ",", "output", "planes", "and", "kernel", "size", "of", "1", ".", "To", "adapt", "the", "model", "parameters", "for", "object", "classification", "on", "different", "datasets", ",", "we", "only", "fine", "-", "tune", "the", "final", "two", "layers", "and", "freeze", "the", "model", "parameters", "from", "previous", "layers", ".", "The", "sampling", "stride", "of", "feature", "maps", "is", "32", "pixels", ",", "and", "the", "window", "size", "is", "223", "pixels", ".", "We", "set", "the", "number", "of", "streams", "to", "be", "3", ".", "During", "training", ",", "all", "three", "streams", "share", "the", "same", "set", "of", "parameters", ".", "To", "facilitate", "training", "with", "mini", "-", "batches", ",", "every", "image", "is", "rescaled", "to", ",", "and", "pixels", ".", "As", "the", "aspect", "ratios", "of", "images", "could", "be", "different", ",", "we", "rescale", "the", "longer", "edge", "to", "300", ",", "500", "and", "700", "respectively", ",", "and", "fill", "the", "empty", "pixels", "by", "mirroring", "the", "images", ".", "Traditional", "cross", "entropy", "loss", "for", "multi", "-", "class", "classification", "introduces", "competition", "between", "different", "classes", ",", "thus", "it", "is", "not", "suitable", "for", "images", "with", "multiple", "labels", ".", "We", "compute", "the", "loss", "with", "binary", "cross", "entropy", "criteria", "for", "each", "class", "separately", ",", "and", "sum", "up", "the", "error", "gradients", "from", "losses", "of", "all", "classes", "for", "back", "-", "propagation", ".", "subsection", ":", "Cascade", "-", "style", "Proposal", "Verification", "By", "setting", "thresholds", "on", "proposal", "scores", ",", "a", "small", "subset", "of", "image", "boxes", "which", "might", "contain", "objects", "are", "selected", ".", "Similar", "to", "object", "detection", "frameworks", ",", "we", "run", "CNN", "classifiers", "on", "the", "selected", "boxes", ".", "The", "proposal", "step", "also", "serves", "as", "a", "filter", "whose", "goal", "is", "to", "preserve", "the", "object", "boxes", "with", "high", "recall", "rate", ",", "while", "removing", "the", "easy", "negatives", ".", "The", "verification", "classifiers", "then", "address", "a", "more", "focused", "problem", "on", "a", "smaller", "set", "of", "instances", ".", "Connecting", "the", "two", "steps", "is", "essentially", "the", "same", "as", "training", "a", "cascade", "of", "classifiers", ".", "Verification", "network", "architecture", ".", "As", "a", "later", "classifier", "in", "the", "cascade", ",", "accuracy", "is", "more", "important", "than", "speed", ".", "We", "choose", "the", "VGG", "-", "16", "network", "architecture", ".", "Compared", "with", "AlexNet", "variants", ",", "it", "offers", "better", "accuracy", "for", "most", "visual", "recognition", "tasks", ",", "but", "is", "also", "slower", "and", "more", "memory", "demanding", ".", "We", "use", "the", "VGG", "-", "16", "model", "parameters", "released", "by", "the", "authors", ",", "which", "was", "trained", "on", "1", ",", "000", "ImageNet", "categories", ".", "We", "use", "the", "same", "binary", "cross", "entropy", "criterion", "to", "compute", "losses", ".", "To", "make", "the", "training", "process", "faster", ",", "we", "only", "fine", "-", "tune", "the", "final", "two", "fully", "-", "connected", "layers", "and", "freeze", "all", "previous", "layers", ".", "InputInput", "OutputOutput", "Training", "images", "with", "proposal", "scores", ",", "batch", "size", ",", "threshold", "stopping", "criteria", "not", "met", "Randomly", "select", "images", "from", ";", "Initialize", "mini", "-", "batch", ";", "has", "proposal", "with", "score", "Randomly", "sample", "a", "proposal", "where", ";", "Set", "the", "sample", "’s", "active", "class", "to", ";", "Add", "proposed", "region", "to", ";", "Resize", "and", "add", "full", "image", "to", ";", "Set", "all", "classes", "as", "active", ";", "Forward", "pass", "with", ";", "Compute", "loss", "for", "the", "active", "class", "of", "each", "sample", ";", "Update", "model", "parameters", ".", "Mini", "-", "batch", "sampling", "algorithm", "for", "training", "cascade", "classifier", "with", "stochastic", "gradient", "descent", ".", "Training", "strategy", "for", "the", "cascade", ".", "Ideally", ",", "we", "want", "the", "verification", "network", "to", "handle", "hard", "examples", "from", "both", "positive", "and", "negative", "data", ".", "When", "a", "proposed", "region", "from", "an", "image", "not", "containing", "a", "given", "label", "has", "a", "high", "score", "of", "that", "class", ",", "we", "know", "it", "is", "a", "hard", "negative", ".", "However", ",", "it", "is", "impossible", "to", "tell", "a", "hard", "positive", "from", "background", "without", "using", "bounding", "box", "annotations", ".", "We", "attempt", "to", "avoid", "using", "background", "by", "selecting", "only", "the", "top", "scoring", "image", "region", "for", "each", "positive", "class", ".", "This", "results", "in", "significant", "over", "-", "fitting", "and", "poor", "generalizability", "for", "the", "trained", "verification", "net", ".", "The", "main", "problem", "with", "the", "above", "sampling", "strategy", "is", "that", "for", "positive", "instances", ",", "only", "easy", "examples", "which", "have", "been", "learned", "well", "are", "preserved", ".", "To", "fix", "this", ",", "we", "use", "a", "random", "sampling", "strategy", "as", "described", "in", "Algorithm", "[", "reference", "]", ".", "For", "each", "image", ",", "we", "randomly", "select", "an", "image", "box", "whose", "proposal", "score", "is", "higher", "than", "threshold", "for", "class", ".", "In", "practice", ",", "the", "threshold", "is", "set", "to", "a", "relative", "low", "value", "(", ")", ".", "If", "is", "labeled", "as", "positive", "for", "the", "image", ",", "we", "treat", "the", "box", "as", "a", "positive", "instance", "(", "though", "it", "might", "belong", "to", "background", ")", ",", "and", "otherwise", "negative", ".", "Note", "that", "the", "sampled", "box", "could", "be", "easy", "negatives", "for", "classes", "beyond", ".", "To", "avoid", "oversampling", "the", "easy", "negatives", ",", "we", "set", "as", "the", "active", "class", "during", "back", "-", "propagation", "and", "only", "compute", "the", "loss", "for", "the", "active", "class", ".", "Inference", "with", "cascade", ".", "During", "inference", ",", "an", "image", "is", "passed", "to", "the", "proposal", "generation", "FCN", "to", "compute", "proposal", "scores", ".", "A", "small", "subset", "of", "proposed", "boxes", "with", "high", "scores", "are", "then", "passed", "to", "the", "verification", "network", ".", "For", "each", "class", ",", "we", "select", "the", "top", "scoring", "proposals", "if", "the", "scores", "are", "higher", "than", "threshold", ".", "We", "then", "use", "the", "following", "equation", "to", "combine", "the", "outputs", "from", "both", "networks", ":", "where", "is", "the", "set", "of", "selected", "proposals", "for", "class", ",", "is", "the", "score", "of", "class", "from", "the", "proposal", "network", "after", "LSE", "pooling", ",", "and", "is", "the", "verification", "network", "’s", "output", "for", "class", "on", "region", ".", "When", "no", "proposal", "is", "selected", ",", "we", "preserve", "scores", "from", "the", "proposal", "network", "without", "calibration", "as", "they", "are", "typically", "low", ".", "Discussion", ".", "Decomposing", "classification", "into", "cascade", "of", "proposal", "and", "verification", "networks", "allows", "the", "system", "to", "achieve", "high", "accuracy", "while", "maintaining", "a", "reasonable", "computational", "cost", ".", "It", "is", "also", "a", "flexible", "framework", "for", "different", "design", "choices", ".", "For", "example", ",", "one", "could", "decide", "to", "verify", "a", "subset", "of", "object", "classes", "which", "require", "higher", "accuracy", ".", "With", "the", "cascade", "training", "algorithm", ",", "we", "can", "build", "tree", "-", "structured", "cascaded", "neural", "networks", ",", "where", "each", "branch", "focuses", "on", "a", "subset", "of", "categories", ".", "We", "can", "also", "extend", "the", "cascade", "to", "have", "more", "stages", ",", "and", "train", "the", "new", "stages", "with", "newly", "annotated", "training", "data", ".", "Figure", "[", "reference", "]", "illustrates", "these", "structures", ".", "section", ":", "Experiments", "Experimental", "setup", ".", "We", "work", "with", "the", "PASCAL", "VOC", "2012", "dataset", "and", "the", "MS", "COCO", "dataset", ".", "VOC", "2012", "has", "5", ",", "000", "images", "for", "training", ",", "5", ",", "000", "for", "validation", "and", "10", ",", "000", "for", "testing", ".", "There", "are", "20", "object", "classes", "in", "total", ".", "COCO", "has", "80", ",", "000", "images", "for", "training", "and", "40", ",", "000", "images", "for", "validation", ".", "It", "has", "80", "object", "classes", "in", "12", "super", "-", "categories", ".", "We", "evaluated", "ProNet", "on", "object", "classification", "and", "point", "-", "based", "object", "localization", "tasks", ".", "For", "object", "classification", ",", "we", "use", "the", "average", "precision", "metric", ".", "We", "used", "VOC", "’s", "result", "server", "to", "compute", "average", "precisions", "on", "the", "VOC", "2012", "dataset", ".", "For", "point", "-", "based", "object", "localization", ",", "we", "use", "the", "criteria", "introduced", "in", ".", "For", "every", "image", "and", "every", "class", ",", "we", "output", "a", "location", "with", "maximum", "response", "for", "that", "class", ".", "The", "location", "is", "deemed", "correct", "if", "it", "falls", "into", "any", "bounding", "box", "associated", "with", "that", "class", ",", "with", "a", "tolerance", "of", "18", "pixels", "as", "used", "in", ".", "This", "information", "is", "then", "used", "to", "compute", "average", "precision", ".", "Although", "object", "extent", "is", "not", "evaluated", ",", "the", "metric", "remains", "challenging", "as", "shown", "by", ".", "To", "generate", "localization", "coordinates", "for", "evaluation", ",", "we", "kept", "track", "of", "the", "image", "boxes", "which", "give", "highest", "responses", "at", "each", "stage", ",", "and", "used", "the", "center", "point", "of", "the", "selected", "boxes", ".", "We", "tried", "different", "values", "of", "hyper", "-", "parameter", "for", "LSE", "pooling", ",", "and", "found", "that", "generally", "gave", "good", "performance", ".", "We", "fixed", "in", "all", "the", "following", "experiments", ".", "We", "used", "the", "stochastic", "gradient", "descent", "algorithm", "for", "training", ".", "To", "train", "proposal", "network", ",", "the", "learning", "rate", "was", "set", "to", "0.01", ";", "to", "train", "verification", "network", ",", "the", "learning", "rate", "was", "set", "to", "0.001", ".", "We", "set", "the", "filtering", "threshold", "for", "cascade", "to", "0.1", ".", "Which", "pooling", "method", "is", "better", "?", "We", "compare", "maximum", "pooling", ",", "average", "pooling", "and", "LSE", "pooling", "methods", "to", "train", "proposal", "network", "with", "image", "-", "level", "supervision", ".", "Table", "[", "reference", "]", "lists", "the", "classification", "and", "localization", "performance", "of", "the", "three", "different", "pooling", "methods", ".", "We", "can", "see", "that", "LSE", "achieves", "the", "best", "classification", "mAP", ".", "Average", "pooling", "is", "3.7", "%", "worse", "than", "LSE", ",", "which", "we", "believe", "is", "because", "it", "assigns", "equal", "importance", "to", "foreground", "and", "background", ".", "Max", "pooling", "is", "1.4", "%", "worse", ";", "compared", "with", "LSE", "pooling", ",", "it", "only", "uses", "a", "single", "patch", "to", "generate", "image", "-", "level", "score", ",", "thus", "is", "more", "sensitive", "to", "noise", "and", "model", "initialization", "during", "training", ".", "We", "also", "generated", "visualizations", "to", "study", "the", "impact", "of", "pooling", "method", "on", "trained", "models", ".", "Figure", "[", "reference", "]", "shows", "heat", "maps", "of", "the", "class", "train", "when", "different", "models", "are", "applied", "to", "the", "same", "image", ".", "We", "can", "see", "that", "the", "model", "trained", "by", "average", "pooling", "has", "high", "activations", "not", "only", "on", "the", "train", "but", "also", "on", "part", "of", "the", "background", ".", "For", "max", "pooling", ",", "only", "the", "wheel", "of", "the", "train", "has", "high", "response", ",", "presumably", "because", "it", "is", "the", "most", "discriminative", "for", "the", "train", ".", "Model", "trained", "by", "LSE", "pooling", "has", "high", "response", "on", "the", "train", ",", "but", "not", "on", "the", "background", ".", "Does", "cascade", "help", "?", "We", "study", "the", "impact", "of", "adding", "cascaded", "classifiers", "on", "classification", "and", "localization", "performance", ".", "We", "first", "use", "a", "single", "level", "of", "cascade", "with", "one", "multi", "-", "scale", "FCN", "and", "one", "verification", "network", ".", "For", "each", "image", "and", "each", "class", ",", "we", "selected", "the", "top", "3", "regions", "per", "scale", "if", "their", "scores", "are", "higher", "than", "0.1", ".", "The", "average", "number", "of", "regions", "to", "be", "verified", "is", "24", "per", "image", ".", "In", "Table", "[", "reference", "]", ",", "we", "can", "see", "that", "on", "PASCAL", "VOC", "2012", ",", "using", "a", "cascade", "helps", "improve", "classification", "mAP", "by", "3.3", "%", "and", "localization", "mAP", "by", "2.9", "%", ".", "Is", "a", "longer", "cascade", "better", "?", "We", "are", "interested", "in", "observing", "how", "the", "performance", "changes", "with", "more", "levels", "of", "cascade", ".", "For", "this", "purpose", ",", "we", "first", "trained", "another", "set", "of", "proposal", "and", "verification", "networks", "using", "PASCAL", "VOC", "data", "alone", ",", "but", "found", "that", "the", "network", "overfitted", "easily", ".", "Since", "the", "training", "set", "of", "VOC", "2012", "has", "only", "5", ",", "000", "images", ",", "we", "found", "that", "the", "first", "set", "of", "proposal", "and", "verification", "networks", "“", "perfectly", "solved", "”", "this", "training", "set", ",", "leaving", "little", "room", "to", "improve", "its", "generalizability", ".", "In", "light", "of", "this", ",", "we", "used", "the", "80", ",", "000", "images", "from", "COCO", "training", "set", "as", "complementary", "data", "source", ".", "It", "covers", "the", "20", "categories", "used", "in", "VOC", "but", "also", "has", "60", "other", "categories", ".", "Rather", "than", "re", "-", "training", "all", "the", "networks", "by", "combining", "VOC", "and", "COCO", "data", ",", "we", "take", "that", "the", "previous", "CNNs", "in", "the", "cascade", "have", "already", "been", "trained", "and", "fixed", ",", "and", "only", "train", "new", "CNNs", "with", "the", "extra", "data", ".", "Note", "that", "our", "cascade", "architecture", "offers", "a", "natural", "way", "to", "select", "the", "challenging", "instances", "from", "such", "incoming", "images", ".", "The", "final", "row", "in", "Table", "[", "reference", "]", "shows", "the", "mAPs", "after", "adding", "a", "new", "set", "of", "cascades", "trained", "from", "COCO", "images", ".", "We", "can", "see", "that", "it", "offers", "another", "1", "%", "improvement", "over", "the", "previous", "cascade", ",", "which", "indicates", "that", "it", "is", "desirable", "to", "train", "a", "longer", "cascade", "when", "more", "training", "data", "becomes", "available", ".", "Expanding", "cascades", "into", "trees", ".", "We", "also", "investigated", "the", "effect", "of", "building", "tree", "-", "structured", "cascades", ".", "COCO", "dataset", "is", "used", "for", "evaluation", "as", "it", "has", "3", "times", "more", "categories", "than", "VOC", ".", "We", "trained", "12", "verification", "networks", "corresponding", "to", "the", "12", "super", "-", "categories", "of", "COCO", ".", "Each", "network", "focuses", "on", "a", "single", "super", "-", "category", ",", "and", "processes", "the", "sampled", "boxes", "whose", "active", "classes", "belong", "to", "that", "super", "-", "category", ".", "At", "test", "time", ",", "each", "proposed", "box", "only", "goes", "through", "a", "single", "root", "to", "leaf", "path", "in", "the", "tree", ".", "The", "final", "row", "of", "Table", "[", "reference", "]", "shows", "its", "classification", "and", "localization", "performance", ".", "We", "can", "see", "that", "compared", "with", "the", "chain", "structured", "cascade", ",", "tree", "-", "structured", "cascade", "achieves", "better", "performance", ",", "probably", "because", "it", "trains", "the", "neural", "networks", "to", "be", "focused", "on", "a", "small", "subset", "of", "similar", "categories", ".", "Comparison", "with", "detection", "based", "approaches", ".", "We", "compare", "our", "proposed", "framework", "with", "two", "recent", "state", "-", "of", "-", "the", "-", "art", "object", "detection", "methods", ":", "RCNN", "and", "Fast", "RCNN", ".", "Unlike", "our", "framework", ",", "they", "require", "bounding", "box", "annotations", "for", "training", ".", "Both", "methods", "use", "selective", "search", "to", "generate", "object", "proposals", "and", "CNNs", "for", "classification", ".", "RCNN", "uses", "AlexNet", "pre", "-", "trained", "from", "ImageNet", ",", "while", "fast", "RCNN", "uses", "VGG", "-", "16", "pre", "-", "trained", "from", "ImageNet", ".", "To", "generate", "classification", "and", "localization", "results", ",", "for", "each", "class", "we", "select", "the", "detection", "output", "with", "maximum", "confidence", "score", ",", "and", "use", "the", "center", "of", "the", "detected", "bounding", "box", "for", "localization", "evaluation", ".", "We", "first", "fix", "the", "number", "of", "window", "proposals", "to", "1000", "for", "RCNN", "and", "fast", "RCNN", ".", "Table", "[", "reference", "]", "shows", "the", "performance", "comparison", ".", "We", "can", "see", "that", "for", "classification", ",", "our", "proposed", "framework", "outperforms", "both", "RCNN", "and", "fast", "RCNN", ".", "For", "localization", ",", "our", "proposed", "framework", "outperforms", "RCNN", ",", "but", "is", "4", "%", "worse", "than", "fast", "RCNN", ".", "We", "also", "study", "the", "impact", "of", "number", "of", "proposed", "boxes", "on", "our", "system", "’s", "performance", ".", "For", "this", "purpose", ",", "we", "let", "the", "proposal", "network", "select", "top", "regions", "per", "scale", "for", "each", "class", ",", "and", "compute", "the", "average", "number", "of", "proposed", "boxes", "per", "image", ".", "For", "comparison", ",", "we", "ask", "fast", "RCNN", "to", "use", "up", "to", "10", ",", "50", ",", "500", "and", "1000", "selective", "search", "proposals", "per", "image", ".", "Table", "[", "reference", "]", "shows", "the", "classification", "and", "localization", "performances", "respectively", ".", "We", "can", "see", "that", "ProNet", "is", "quite", "robust", "to", "the", "number", "of", "proposed", "boxes", ",", "and", "achieves", "reasonably", "good", "performance", "with", "only", "9", "boxes", "on", "average", ".", "This", "confirms", "that", "ProNet", "offers", "better", "accuracy", "with", "relatively", "small", "computational", "overhead", ".", "Meanwhile", ",", "fast", "RCNN", "requires", "many", "more", "proposals", "to", "reach", "peak", "performance", ",", "presumably", "because", "the", "selective", "search", "proposals", "are", "for", "general", "objectness", "and", "not", "optimized", "for", "object", "classification", "in", "cascade", "fashion", ".", "Comparison", "with", "other", "weakly", "-", "supervised", "methods", ".", "We", "compare", "ProNet", "with", "several", "state", "-", "of", "-", "the", "-", "art", "object", "classification", "frameworks", ".", "Classification", "and", "localization", "performance", "on", "PASCAL", "VOC", "2012", "are", "shown", "in", "Table", "[", "reference", "]", "and", "Table", "[", "reference", "]", "respectively", ".", "Table", "[", "reference", "]", "and", "Figure", "[", "reference", "]", "show", "results", "and", "localization", "examples", "on", "COCO", "dataset", ".", "Among", "the", "compared", "systems", ",", "Oquab", "et", "al", ".", "and", "NUS", "-", "HCP", "use", "CNNs", "pre", "-", "trained", "on", "the", "expanded", "ImageNet", "data", "with", "more", "than", "1500", "categories", ",", "which", "has", "been", "shown", "to", "be", "useful", "for", "classification", ".", "Since", "ProNet", "uses", "cascades", "or", "trees", "of", "CNNs", ",", "it", "can", "apply", "a", "more", "powerful", "CNN", "model", "VGG", "-", "16", "with", "small", "computational", "overhead", ".", "This", "helps", "our", "system", "outperform", "most", "of", "the", "previous", "state", "-", "of", "-", "the", "-", "art", "systems", "significantly", "on", "both", "datasets", ".", "ProNet", "is", "also", "slightly", "better", "than", "Simonyan", "et", "al", ".", "which", "extracts", "VGG", "-", "16", "features", "at", "three", "different", "scales", "over", "full", "images", ".", "Their", "system", "is", "3x", "to", "6x", "slower", "than", "our", "cascade", "at", "test", "time", ".", "Limitation", ".", "We", "evaluate", "ProNet", "using", "the", "standard", "IOU", "metric", ",", "which", "considers", "object", "extent", "as", "well", "as", "location", ".", "Since", "the", "boxes", "generated", "by", "our", "proposal", "CNN", "have", "fixed", "aspect", "ratios", ",", "we", "follow", "to", "aggregate", "the", "heat", "maps", "over", "1000", "bounding", "box", "proposals", "generated", "by", "selective", "search", "per", "image", ".", "No", "bounding", "box", "regression", "is", "conducted", ".", "Cascade", "CNN", "is", "then", "used", "to", "verify", "the", "high", "-", "scoring", "proposals", ".", "On", "PASCAL", "VOC", "2012", "validation", "set", ",", "our", "proposal", "CNN", "has", "an", "mAP", "of", "13.0", "%", "when", "overlap", "threshold", "is", "0.5", ".", "The", "cascade", "CNN", "improves", "the", "mAP", "to", "15.5", "%", ".", "Although", "both", "results", "are", "higher", "than", "11.7", "%", "as", "reported", "by", ",", "there", "is", "still", "a", "huge", "gap", "between", "the", "state", "-", "of", "-", "the", "-", "art", "object", "detection", "pipelines", ".", "Our", "proposal", "network", "tends", "to", "select", "the", "most", "discriminative", "/", "confusing", "parts", "of", "objects", ",", "which", "is", "good", "for", "cascade", "classification", "but", "bad", "for", "getting", "full", "object", "extents", ".", "Separating", "and", "counting", "multiple", "objects", "are", "also", "challenging", "issues", ".", "section", ":", "Conclusion", "We", "proposed", "ProNet", ",", "a", "cascaded", "neural", "network", "for", "object", "classification", "and", "localization", ".", "ProNet", "learns", "to", "propose", "object", "-", "specific", "boxes", "by", "multi", "-", "scale", "FCNs", "trained", "from", "image", "-", "level", "annotations", ".", "It", "then", "sends", "a", "small", "subset", "of", "promising", "boxes", "to", "latter", "CNNs", "for", "verification", ".", "Detailed", "experimental", "evaluations", "have", "shown", "the", "effectiveness", "of", "ProNet", "on", "the", "challenging", "PASCAL", "VOC", "2012", "dataset", "and", "MS", "COCO", "dataset", ".", "Acknowledgement", ":", "We", "would", "like", "to", "thank", "Sergey", "Zagoruyko", "for", "help", "with", "fast", "RCNN", "experiments", ";", "Pedro", "O.", "Pinheiro", ",", "Bolei", "Zhou", ",", "Maxime", "Oquab", ",", "Joël", "Legrand", ",", "Yuandong", "Tian", ",", "Léon", "Bottou", "and", "Florent", "Perronnin", "for", "valuable", "discussions", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["ProNet"], "Method"], [4, 15, ["Learning", "to", "Propose", "Object", "-", "specific", "Boxes", "for", "Cascaded", "Neural", "Networks"], "Method"], [59, 62, ["classification", "architecture", "ProNet"], "Method"], [64, 67, ["convolutional", "neural", "networks"], "Method"], [72, 74, ["neural", "networks"], "Method"], [103, 110, ["multi", "-", "scale", "fully", "-", "convolutional", "network"], "Method"], [143, 144, ["cascades"], "Method"], [145, 146, ["trees"], "Method"], [148, 150, ["object", "classification"], "Task"], [151, 152, ["ProNet"], "Method"], [167, 170, ["MS", "COCO", "datasets"], "Material"], [171, 173, ["object", "classification"], "Task"], [187, 189, ["object", "classification"], "Task"], [190, 191, ["localization"], "Task"], [214, 218, ["object", "-", "centric", "classifiers"], "Method"], [235, 237, ["annotation", "process"], "Task"], [238, 241, ["object", "bounding", "boxes"], "Task"], [287, 289, ["feature", "activations"], "Method"], [290, 293, ["convolutional", "neural", "networks"], "Method"], [294, 295, ["CNN"], "Method"], [310, 314, ["image", "-", "level", "representations"], "Method"], [315, 317, ["classification", "purpose"], "Task"], [339, 341, ["global", "approaches"], "Method"], [355, 356, ["localization"], "Task"], [362, 370, ["fully", "-", "convolutional", "network", "(", "FCN", ")", "architecture"], "Method"], [380, 384, ["fully", "-", "connected", "layers"], "Method"], [387, 388, ["CNN"], "Method"], [390, 391, ["AlexNet"], "Method"], [393, 395, ["convolutional", "layers"], "Method"], [399, 400, ["FCN"], "Method"], [418, 420, ["score", "map"], "Method"], [435, 437, ["score", "maps"], "Method"], [442, 443, ["classification"], "Task"], [444, 445, ["localization"], "Task"], [456, 460, ["FCN", "’s", "network", "architecture"], "Method"], [467, 468, ["FCN"], "Method"], [485, 492, ["multi", "-", "stream", "multi", "-", "scale", "architecture"], "Method"], [511, 515, ["multi", "-", "scale", "FCN"], "Task"], [547, 550, ["back", "-", "propagation"], "Method"], [553, 557, ["multi", "-", "scale", "FCN"], "Method"], [594, 596, ["cascade", "architecture"], "Method"], [606, 607, ["classifiers"], "Method"], [612, 613, ["cascade"], "Method"], [618, 619, ["accuracy"], "Metric"], [620, 621, ["speed"], "Metric"], [642, 644, ["zoom", "pipeline"], "Method"], [645, 646, ["ProNet"], "Method"], [658, 659, ["ProNet"], "Method"], [670, 672, ["cat", "categories"], "Task"], [693, 694, ["classifiers"], "Method"], [695, 696, ["ProNet"], "Method"], [731, 734, ["over", "-", "fitting"], "Task"], [761, 766, ["stochastic", "gradient", "descent", "training", "process"], "Method"], [794, 795, ["classifiers"], "Method"], [801, 802, ["CNNs"], "Method"], [809, 811, ["computational", "overhead"], "Metric"], [812, 813, ["ProNet"], "Method"], [846, 850, ["chain", "-", "structured", "cascade"], "Method"], [860, 864, ["tree", "-", "structured", "cascades"], "Method"], [894, 895, ["ProNet"], "Method"], [897, 901, ["cascaded", "neural", "network", "framework"], "Method"], [911, 913, ["object", "classification"], "Task"], [914, 915, ["localization"], "Task"], [921, 922, ["ProNet"], "Method"], [934, 941, ["chain", "-", "and", "tree", "-", "structured", "cascades"], "Method"], [945, 946, ["ProNet"], "Method"], [958, 960, ["object", "classification"], "Task"], [976, 979, ["MS", "COCO", "dataset"], "Material"], [984, 986, ["Object", "classification"], "Task"], [991, 993, ["Computer", "Vision"], "Task"], [998, 999, ["classification"], "Task"], [1052, 1056, ["deep", "convolutional", "neural", "networks"], "Method"], [1057, 1058, ["CNN"], "Method"], [1074, 1077, ["visual", "recognition", "tasks"], "Task"], [1079, 1081, ["object", "classification"], "Task"], [1082, 1083, ["detection"], "Task"], [1085, 1086, ["CNNs"], "Method"], [1117, 1119, ["image", "classification"], "Task"], [1120, 1123, ["fine", "-", "tuning"], "Task"], [1130, 1131, ["classification"], "Task"], [1132, 1134, ["linear", "SVMs"], "Method"], [1138, 1141, ["generic", "feature", "extractors"], "Method"], [1142, 1145, ["feature", "aggregation", "techniques"], "Method"], [1155, 1157, ["CNN", "embeddings"], "Method"], [1166, 1168, ["object", "classification"], "Task"], [1170, 1171, ["detection"], "Task"], [1179, 1180, ["RCNN"], "Method"], [1193, 1194, ["classifiers"], "Method"], [1211, 1212, ["classification"], "Task"], [1217, 1219, ["fast", "RCNN"], "Method"], [1221, 1225, ["fully", "-", "convolutional", "networks"], "Method"], [1226, 1227, ["FCN"], "Method"], [1247, 1251, ["image", "-", "level", "annotations"], "Task"], [1253, 1257, ["weakly", "-", "supervised", "problem"], "Task"], [1264, 1268, ["multiple", "instance", "learning", "problem"], "Task"], [1275, 1277, ["concept", "detectors"], "Task"], [1287, 1289, ["object", "detection"], "Task"], [1290, 1291, ["segmentation"], "Task"], [1293, 1295, ["object", "classification"], "Task"], [1319, 1321, ["max", "pooling"], "Method"], [1325, 1326, ["CNNs"], "Method"], [1354, 1355, ["CNNs"], "Method"], [1356, 1357, ["FCNs"], "Method"], [1363, 1364, ["ProNet"], "Method"], [1368, 1369, ["FCN"], "Method"], [1373, 1377, ["multi", "-", "stream", "architecture"], "Method"], [1378, 1383, ["cross", "-", "scale", "LSE", "pooling"], "Method"], [1385, 1388, ["scale", "-", "awareness"], "Task"], [1389, 1391, ["Cascaded", "classifiers"], "Method"], [1398, 1400, ["Computer", "Vision"], "Task"], [1401, 1402, ["Cascades"], "Method"], [1403, 1404, ["CNNs"], "Method"], [1415, 1417, ["pose", "estimation"], "Task"], [1428, 1429, ["ProNet"], "Method"], [1431, 1433, ["cascade", "philosophy"], "Method"], [1435, 1436, ["speed"], "Metric"], [1437, 1438, ["accuracy"], "Metric"], [1450, 1451, ["ProNet"], "Method"], [1453, 1456, ["general", "object", "classifier"], "Method"], [1477, 1479, ["ProNet", "Framework"], "Method"], [1479, 1480, ["ProNet"], "Method"], [1495, 1497, ["verification", "unit"], "Method"], [1515, 1517, ["confidence", "scores"], "Metric"], [1535, 1536, ["ProNet"], "Method"], [1556, 1558, ["verification", "units"], "Method"], [1595, 1598, ["convolutional", "neural", "networks"], "Method"], [1606, 1608, ["ProNet", "framework"], "Method"], [1611, 1613, ["Proposal", "Generation"], "Task"], [1622, 1627, ["object", "-", "specific", "box", "proposals"], "Task"], [1628, 1629, ["CNNs"], "Method"], [1635, 1637, ["object", "category"], "Task"], [1668, 1670, ["CNN", "architecture"], "Method"], [1671, 1674, ["image", "classification", "task"], "Task"], [1676, 1677, ["AlexNet"], "Method"], [1680, 1684, ["hierarchy", "of", "convolutional", "layers"], "Method"], [1685, 1688, ["fully", "connected", "layers"], "Method"], [1690, 1692, ["convolutional", "layers"], "Method"], [1699, 1701, ["feature", "representations"], "Method"], [1711, 1713, ["convolutional", "layers"], "Method"], [1761, 1763, ["convolutional", "layers"], "Method"], [1765, 1768, ["fully", "connected", "layers"], "Method"], [1770, 1771, ["classifiers"], "Method"], [1811, 1814, ["bounding", "box", "proposals"], "Method"], [1821, 1822, ["CNNs"], "Method"], [1830, 1833, ["fully", "convolutional", "networks"], "Method"], [1835, 1836, ["OverFeat"], "Method"], [1838, 1841, ["Fully", "convolutional", "networks"], "Method"], [1842, 1844, ["FCN", ")"], "Method"], [1847, 1851, ["fully", "-", "connected", "layers"], "Method"], [1858, 1860, ["convolutional", "layers"], "Method"], [1874, 1875, ["FCNs"], "Method"], [1922, 1924, ["region", "sampling"], "Method"], [1925, 1927, ["sliding", "windows"], "Method"], [1928, 1931, ["bounding", "box", "proposals"], "Method"], [1932, 1933, ["FCNs"], "Method"], [1938, 1944, ["end", "-", "to", "-", "end", "training"], "Task"], [1946, 1948, ["CNN", "framework"], "Method"], [1963, 1965, ["Scale", "adaptation"], "Method"], [1966, 1970, ["multi", "-", "stream", "FCNs"], "Method"], [1976, 1977, ["FCNs"], "Method"], [2004, 2008, ["multi", "-", "stream", "architecture"], "Method"], [2011, 2012, ["FCN"], "Method"], [2062, 2065, ["-", "stream", "FCN"], "Method"], [2098, 2099, ["FCNs"], "Method"], [2117, 2119, ["loss", "function"], "Method"], [2135, 2136, ["supervision"], "Task"], [2158, 2159, ["training"], "Task"], [2163, 2170, ["log", "-", "sum", "-", "exp", "(", "LSE"], "Method"], [2171, 2173, ["pooling", "function"], "Method"], [2191, 2192, ["FCN"], "Method"], [2255, 2258, ["LSE", "pooling", "function"], "Method"], [2263, 2264, ["layer"], "Method"], [2266, 2268, ["neural", "network"], "Method"], [2285, 2289, ["all", "-", "stream", "FCNs"], "Method"], [2324, 2325, ["FCNs"], "Method"], [2348, 2352, ["single", "-", "stream", "FCN"], "Method"], [2387, 2389, ["padding", "strategies"], "Method"], [2411, 2414, ["spatial", "convolutional", "layer"], "Method"], [2421, 2424, ["max", "pooling", "layer"], "Method"], [2439, 2441, ["convolutional", "layers"], "Method"], [2447, 2450, ["max", "pooling", "layers"], "Method"], [2453, 2457, ["Our", "-", "stream", "FCNs"], "Method"], [2460, 2461, ["Torch"], "Method"], [2469, 2474, ["CNN", "-", "M", "2048", "architecture"], "Method"], [2480, 2482, ["convolutional", "layers"], "Method"], [2484, 2488, ["fully", "-", "connected", "layers"], "Method"], [2492, 2493, ["accuracy"], "Metric"], [2496, 2497, ["AlexNet"], "Method"], [2507, 2509, ["deep", "CNNs"], "Method"], [2540, 2541, ["FCN"], "Method"], [2545, 2549, ["fully", "-", "connected", "layers"], "Method"], [2550, 2552, ["convolutional", "layers"], "Method"], [2555, 2557, ["convolutional", "layer"], "Method"], [2620, 2622, ["object", "classification"], "Task"], [2689, 2692, ["mini", "-", "batches"], "Method"], [2736, 2739, ["cross", "entropy", "loss"], "Method"], [2740, 2744, ["multi", "-", "class", "classification"], "Task"], [2764, 2765, ["loss"], "Metric"], [2766, 2770, ["binary", "cross", "entropy", "criteria"], "Metric"], [2787, 2790, ["back", "-", "propagation"], "Method"], [2793, 2798, ["Cascade", "-", "style", "Proposal", "Verification"], "Method"], [2826, 2828, ["CNN", "classifiers"], "Method"], [2840, 2841, ["filter"], "Method"], [2851, 2853, ["recall", "rate"], "Metric"], [2861, 2863, ["verification", "classifiers"], "Method"], [2887, 2890, ["cascade", "of", "classifiers"], "Method"], [2891, 2894, ["Verification", "network", "architecture"], "Method"], [2898, 2899, ["classifier"], "Method"], [2901, 2902, ["cascade"], "Method"], [2903, 2904, ["accuracy"], "Metric"], [2908, 2909, ["speed"], "Metric"], [2913, 2918, ["VGG", "-", "16", "network", "architecture"], "Method"], [2921, 2923, ["AlexNet", "variants"], "Method"], [2927, 2928, ["accuracy"], "Metric"], [2930, 2933, ["visual", "recognition", "tasks"], "Task"], [2946, 2951, ["VGG", "-", "16", "model", "parameters"], "Method"], [2970, 2974, ["binary", "cross", "entropy", "criterion"], "Metric"], [2993, 2997, ["fully", "-", "connected", "layers"], "Method"], [3080, 3082, ["model", "parameters"], "Method"], [3083, 3088, ["Mini", "-", "batch", "sampling", "algorithm"], "Method"], [3090, 3092, ["cascade", "classifier"], "Method"], [3093, 3096, ["stochastic", "gradient", "descent"], "Method"], [3097, 3099, ["Training", "strategy"], "Method"], [3101, 3102, ["cascade"], "Method"], [3108, 3110, ["verification", "network"], "Method"], [3199, 3201, ["verification", "net"], "Method"], [3208, 3210, ["sampling", "strategy"], "Method"], [3234, 3237, ["random", "sampling", "strategy"], "Method"], [3337, 3340, ["back", "-", "propagation"], "Method"], [3350, 3351, ["Inference"], "Method"], [3352, 3353, ["cascade"], "Method"], [3355, 3356, ["inference"], "Method"], [3363, 3366, ["proposal", "generation", "FCN"], "Method"], [3385, 3387, ["verification", "network"], "Method"], [3437, 3439, ["proposal", "network"], "Method"], [3440, 3442, ["LSE", "pooling"], "Method"], [3446, 3448, ["verification", "network"], "Method"], [3466, 3468, ["proposal", "network"], "Method"], [3469, 3470, ["calibration"], "Method"], [3479, 3480, ["classification"], "Task"], [3483, 3484, ["proposal"], "Method"], [3485, 3487, ["verification", "networks"], "Method"], [3493, 3494, ["accuracy"], "Metric"], [3498, 3500, ["computational", "cost"], "Metric"], [3528, 3529, ["accuracy"], "Metric"], [3532, 3535, ["cascade", "training", "algorithm"], "Method"], [3539, 3545, ["tree", "-", "structured", "cascaded", "neural", "networks"], "Method"], [3561, 3562, ["cascade"], "Method"], [3602, 3605, ["MS", "COCO", "dataset"], "Material"], [3620, 3621, ["validation"], "Task"], [3636, 3637, ["COCO"], "Material"], [3650, 3651, ["validation"], "Task"], [3665, 3666, ["ProNet"], "Method"], [3667, 3669, ["object", "classification"], "Task"], [3678, 3680, ["object", "classification"], "Task"], [3684, 3687, ["average", "precision", "metric"], "Metric"], [3696, 3698, ["average", "precisions"], "Metric"], [3770, 3772, ["average", "precision"], "Metric"], [3829, 3831, ["LSE", "pooling"], "Method"], [3851, 3855, ["stochastic", "gradient", "descent", "algorithm"], "Method"], [3860, 3862, ["proposal", "network"], "Method"], [3864, 3866, ["learning", "rate"], "Metric"], [3873, 3875, ["verification", "network"], "Method"], [3877, 3879, ["learning", "rate"], "Metric"], [3895, 3897, ["pooling", "method"], "Method"], [3902, 3904, ["maximum", "pooling"], "Method"], [3905, 3907, ["average", "pooling"], "Method"], [3908, 3911, ["LSE", "pooling", "methods"], "Method"], [3913, 3915, ["proposal", "network"], "Method"], [3916, 3920, ["image", "-", "level", "supervision"], "Task"], [3935, 3937, ["pooling", "methods"], "Method"], [3942, 3943, ["LSE"], "Method"], [3949, 3951, ["Average", "pooling"], "Method"], [3956, 3957, ["LSE"], "Method"], [3972, 3974, ["Max", "pooling"], "Method"], [3981, 3983, ["LSE", "pooling"], "Method"], [4004, 4006, ["model", "initialization"], "Method"], [4007, 4008, ["training"], "Task"], [4018, 4020, ["pooling", "method"], "Method"], [4021, 4023, ["trained", "models"], "Method"], [4053, 4055, ["average", "pooling"], "Method"], [4072, 4074, ["max", "pooling"], "Method"], [4099, 4101, ["LSE", "pooling"], "Method"], [4124, 4126, ["cascaded", "classifiers"], "Method"], [4139, 4140, ["cascade"], "Method"], [4142, 4146, ["multi", "-", "scale", "FCN"], "Method"], [4148, 4150, ["verification", "network"], "Method"], [4204, 4205, ["cascade"], "Method"], [4250, 4251, ["proposal"], "Method"], [4252, 4254, ["verification", "networks"], "Method"], [4289, 4290, ["proposal"], "Method"], [4291, 4293, ["verification", "networks"], "Method"], [4322, 4325, ["COCO", "training", "set"], "Material"], [4337, 4338, ["VOC"], "Task"], [4357, 4359, ["COCO", "data"], "Material"], [4365, 4366, ["CNNs"], "Method"], [4380, 4381, ["CNNs"], "Method"], [4389, 4391, ["cascade", "architecture"], "Method"], [4422, 4423, ["cascades"], "Method"], [4425, 4427, ["COCO", "images"], "Material"], [4478, 4480, ["COCO", "dataset"], "Material"], [4497, 4499, ["verification", "networks"], "Method"], [4507, 4508, ["COCO"], "Material"], [4564, 4565, ["classification"], "Task"], [4566, 4567, ["localization"], "Task"], [4576, 4579, ["chain", "structured", "cascade"], "Method"], [4580, 4584, ["tree", "-", "structured", "cascade"], "Method"], [4593, 4595, ["neural", "networks"], "Method"], [4631, 4632, ["RCNN"], "Method"], [4633, 4635, ["Fast", "RCNN"], "Method"], [4651, 4653, ["selective", "search"], "Method"], [4658, 4659, ["CNNs"], "Method"], [4660, 4661, ["classification"], "Task"], [4662, 4663, ["RCNN"], "Method"], [4664, 4665, ["AlexNet"], "Method"], [4669, 4670, ["ImageNet"], "Method"], [4672, 4674, ["fast", "RCNN"], "Method"], [4675, 4678, ["VGG", "-", "16"], "Method"], [4697, 4698, ["detection"], "Task"], [4728, 4729, ["RCNN"], "Method"], [4730, 4732, ["fast", "RCNN"], "Method"], [4747, 4748, ["classification"], "Task"], [4754, 4755, ["RCNN"], "Method"], [4756, 4758, ["fast", "RCNN"], "Method"], [4760, 4761, ["localization"], "Task"], [4766, 4767, ["RCNN"], "Method"], [4774, 4776, ["fast", "RCNN"], "Method"], [4800, 4802, ["proposal", "network"], "Method"], [4828, 4829, ["RCNN"], "Method"], [4862, 4863, ["ProNet"], "Method"], [4888, 4889, ["ProNet"], "Method"], [4891, 4892, ["accuracy"], "Metric"], [4895, 4897, ["computational", "overhead"], "Metric"], [4901, 4902, ["RCNN"], "Method"], [4925, 4927, ["object", "classification"], "Task"], [4928, 4930, ["cascade", "fashion"], "Task"], [4934, 4938, ["weakly", "-", "supervised", "methods"], "Method"], [4941, 4942, ["ProNet"], "Method"], [4955, 4956, ["Classification"], "Task"], [4957, 4958, ["localization"], "Task"], [4989, 4990, ["localization"], "Task"], [4992, 4994, ["COCO", "dataset"], "Material"], [5005, 5008, ["NUS", "-", "HCP"], "Method"], [5009, 5010, ["CNNs"], "Method"], [5032, 5033, ["classification"], "Task"], [5035, 5036, ["ProNet"], "Method"], [5037, 5038, ["cascades"], "Method"], [5039, 5042, ["trees", "of", "CNNs"], "Method"], [5049, 5054, ["CNN", "model", "VGG", "-", "16"], "Method"], [5056, 5058, ["computational", "overhead"], "Metric"], [5081, 5082, ["ProNet"], "Method"], [5114, 5115, ["cascade"], "Method"], [5123, 5124, ["ProNet"], "Method"], [5127, 5129, ["IOU", "metric"], "Metric"], [5161, 5164, ["bounding", "box", "proposals"], "Method"], [5166, 5168, ["selective", "search"], "Method"], [5172, 5175, ["bounding", "box", "regression"], "Method"], [5178, 5180, ["Cascade", "CNN"], "Method"], [5203, 5204, ["mAP"], "Metric"], [5208, 5210, ["overlap", "threshold"], "Metric"], [5214, 5216, ["cascade", "CNN"], "Method"], [5218, 5219, ["mAP"], "Metric"], [5255, 5257, ["proposal", "network"], "Method"], [5273, 5275, ["cascade", "classification"], "Task"], [5279, 5282, ["full", "object", "extents"], "Task"], [5283, 5284, ["Separating"], "Task"], [5285, 5288, ["counting", "multiple", "objects"], "Task"], [5298, 5299, ["ProNet"], "Method"], [5301, 5304, ["cascaded", "neural", "network"], "Method"], [5305, 5307, ["object", "classification"], "Task"], [5308, 5309, ["localization"], "Task"], [5310, 5311, ["ProNet"], "Method"], [5319, 5323, ["multi", "-", "scale", "FCNs"], "Method"], [5341, 5342, ["CNNs"], "Method"], [5343, 5344, ["verification"], "Task"], [5353, 5354, ["ProNet"], "Method"], [5362, 5365, ["MS", "COCO", "dataset"], "Material"], [5379, 5380, ["RCNN"], "Method"], [164, 165, ["VOC"], "Task"], [177, 178, ["localization"], "Task"], [372, 374, ["object", "classification"], "Task"], [375, 376, ["localization"], "Task"], [567, 568, ["localization"], "Task"], [964, 965, ["localization"], "Task"], [969, 970, ["VOC"], "Task"], [1201, 1202, ["proposal"], "Method"], [1245, 1246, ["localization"], "Task"], [1359, 1360, ["proposal"], "Method"], [1410, 1411, ["detection"], "Task"], [1413, 1414, ["detection"], "Task"], [1490, 1491, ["proposal"], "Method"], [1509, 1510, ["proposal"], "Method"], [1643, 1644, ["proposal"], "Method"], [2177, 2178, ["segmentation"], "Task"], [2319, 2320, ["proposal"], "Method"], [2331, 2332, ["proposal"], "Method"], [2374, 2375, ["proposal"], "Method"], [2802, 2803, ["proposal"], "Method"], [2821, 2822, ["detection"], "Task"], [2834, 2835, ["proposal"], "Method"], [3008, 3009, ["proposal"], "Method"], [3030, 3031, ["proposal"], "Method"], [3036, 3037, ["proposal"], "Method"], [3256, 3257, ["proposal"], "Method"], [3368, 3369, ["proposal"], "Method"], [3457, 3458, ["proposal"], "Method"], [3597, 3598, ["VOC"], "Task"], [3606, 3607, ["VOC"], "Task"], [3674, 3675, ["localization"], "Task"], [3690, 3691, ["VOC"], "Task"], [3700, 3701, ["VOC"], "Task"], [3709, 3710, ["localization"], "Task"], [3790, 3791, ["localization"], "Task"], [3929, 3930, ["localization"], "Task"], [3947, 3948, ["mAP"], "Metric"], [4129, 4130, ["localization"], "Task"], [4199, 4200, ["VOC"], "Task"], [4208, 4209, ["mAP"], "Metric"], [4213, 4214, ["localization"], "Task"], [4214, 4215, ["mAP"], "Metric"], [4256, 4257, ["VOC"], "Task"], [4273, 4274, ["VOC"], "Task"], [4355, 4356, ["VOC"], "Task"], [4492, 4493, ["VOC"], "Task"], [4608, 4609, ["detection"], "Task"], [4628, 4629, ["detection"], "Task"], [4688, 4689, ["localization"], "Task"], [4714, 4715, ["localization"], "Task"], [4854, 4855, ["localization"], "Task"], [4951, 4953, ["object", "classification"], "Task"], [4961, 4962, ["VOC"], "Task"], [5145, 5146, ["proposal"], "Method"], [5193, 5194, ["VOC"], "Task"], [5199, 5200, ["proposal"], "Method"], [5251, 5252, ["detection"], "Task"], [5358, 5359, ["VOC"], "Task"]]} {"doc_id": "081531984770a74e87dbd68907061b4b0f3631bf", "text": ["document", ":", "Real", "-", "Time", "Video", "Super", "-", "Resolution", "with", "Spatio", "-", "Temporal", "Networks", "and", "Motion", "Compensation", "Convolutional", "neural", "networks", "have", "enabled", "accurate", "image", "super", "-", "resolution", "in", "real", "-", "time", ".", "However", ",", "recent", "attempts", "to", "benefit", "from", "temporal", "correlations", "in", "video", "super", "-", "resolution", "have", "been", "limited", "to", "naive", "or", "inefficient", "architectures", ".", "In", "this", "paper", ",", "we", "introduce", "spatio", "-", "temporal", "sub", "-", "pixel", "convolution", "networks", "that", "effectively", "exploit", "temporal", "redundancies", "and", "improve", "reconstruction", "accuracy", "while", "maintaining", "real", "-", "time", "speed", ".", "Specifically", ",", "we", "discuss", "the", "use", "of", "early", "fusion", ",", "slow", "fusion", "and", "3D", "convolutions", "for", "the", "joint", "processing", "of", "multiple", "consecutive", "video", "frames", ".", "We", "also", "propose", "a", "novel", "joint", "motion", "compensation", "and", "video", "super", "-", "resolution", "algorithm", "that", "is", "orders", "of", "magnitude", "more", "efficient", "than", "competing", "methods", ",", "relying", "on", "a", "fast", "multi", "-", "resolution", "spatial", "transformer", "module", "that", "is", "end", "-", "to", "-", "end", "trainable", ".", "These", "contributions", "provide", "both", "higher", "accuracy", "and", "temporally", "more", "consistent", "videos", ",", "which", "we", "confirm", "qualitatively", "and", "quantitatively", ".", "Relative", "to", "single", "-", "frame", "models", ",", "spatio", "-", "temporal", "networks", "can", "either", "reduce", "the", "computational", "cost", "by", "30", "%", "whilst", "maintaining", "the", "same", "quality", "or", "provide", "a", "0.2dB", "gain", "for", "a", "similar", "computational", "cost", ".", "Results", "on", "publicly", "available", "datasets", "demonstrate", "that", "the", "proposed", "algorithms", "surpass", "current", "state", "-", "of", "-", "the", "-", "art", "performance", "in", "both", "accuracy", "and", "efficiency", ".", "SRSRsuper", "-", "resolution", "LRLRlow", "-", "resolution", "HRHRhigh", "-", "resolution", "CNNCNNconvolutionalneuralnetwork", "TVTVtotalvariation", "HDHDhighdefinition", "MSEMSEmeansquarederror", "PSNRPSNRpeaksignal", "-", "to", "-", "noiseratio", "SSIMSSIMstructuralsimilarity", "section", ":", "Introduction", "Image", "and", "video", "SR", "are", "long", "-", "standing", "challenges", "of", "signal", "processing", ".", "SR", "aims", "at", "recovering", "a", "HR", "image", "or", "video", "from", "its", "LR", "version", ",", "and", "finds", "direct", "applications", "ranging", "from", "medical", "imaging", "to", "satellite", "imaging", ",", "as", "well", "as", "facilitating", "tasks", "such", "as", "face", "recognition", ".", "The", "reconstruction", "of", "HR", "data", "from", "a", "LR", "input", "is", "however", "a", "highly", "ill", "-", "posed", "problem", "that", "requires", "additional", "constraints", "to", "be", "solved", ".", "While", "those", "constraints", "are", "often", "application", "-", "dependent", ",", "they", "usually", "rely", "on", "data", "redundancy", ".", "In", "single", "image", "SR", ",", "where", "only", "one", "LR", "image", "is", "provided", ",", "methods", "exploit", "inherent", "image", "redundancy", "in", "the", "form", "of", "local", "correlations", "to", "recover", "lost", "high", "-", "frequency", "details", "by", "imposing", "sparsity", "constraints", "or", "assuming", "other", "types", "of", "image", "statistics", "such", "as", "multi", "-", "scale", "patch", "recurrence", ".", "In", "multi", "-", "image", "SR", "it", "is", "assumed", "that", "different", "observations", "of", "the", "same", "scene", "are", "available", ",", "hence", "the", "shared", "explicit", "redundancy", "can", "be", "used", "to", "constrain", "the", "problem", "and", "attempt", "to", "invert", "the", "downscaling", "process", "directly", ".", "Transitioning", "from", "images", "to", "videos", "implies", "an", "additional", "data", "dimension", "(", "time", ")", "with", "a", "high", "degree", "of", "correlation", "that", "can", "also", "be", "exploited", "to", "improve", "performance", "in", "terms", "of", "accuracy", "as", "well", "as", "efficiency", ".", "subsection", ":", "Related", "work", "Video", "SR", "methods", "have", "mainly", "emerged", "as", "adaptations", "of", "image", "SR", "techniques", ".", "Kernel", "regression", "methods", "have", "been", "shown", "to", "be", "applicable", "to", "videos", "using", "3D", "kernels", "instead", "of", "2D", "ones", ".", "Dictionary", "learning", "approaches", ",", "which", "define", "LR", "images", "as", "a", "sparse", "linear", "combination", "of", "dictionary", "atoms", "coupled", "to", "a", "HR", "dictionary", ",", "have", "also", "been", "adapted", "from", "images", "to", "videos", ".", "Another", "approach", "is", "example", "-", "based", "patch", "recurrence", ",", "which", "assumes", "patches", "in", "a", "single", "image", "or", "video", "obey", "multi", "-", "scale", "relationships", ",", "and", "therefore", "missing", "high", "-", "frequency", "content", "at", "a", "given", "scale", "can", "be", "inferred", "from", "coarser", "scale", "patches", ".", "This", "was", "successfully", "presented", "by", "Glasner", "et", "al", ".", "for", "image", "SR", "and", "has", "later", "been", "extended", "to", "videos", ".", "When", "adapting", "a", "method", "from", "images", "to", "videos", "it", "is", "usually", "beneficial", "to", "incorporate", "the", "prior", "knowledge", "that", "frames", "of", "the", "same", "scene", "of", "a", "video", "can", "be", "approximated", "by", "a", "single", "image", "and", "a", "motion", "pattern", ".", "Estimating", "and", "compensating", "motion", "is", "a", "powerful", "mechanism", "to", "further", "constrain", "the", "problem", "and", "expose", "temporal", "correlations", ".", "It", "is", "therefore", "very", "common", "to", "find", "video", "SR", "methods", "that", "explicitly", "model", "motion", "through", "frames", ".", "A", "natural", "choice", "has", "been", "to", "preprocess", "input", "frames", "by", "compensating", "inter", "-", "frame", "motion", "using", "displacement", "fields", "obtained", "from", "off", "-", "the", "-", "shelf", "optical", "flow", "algorithms", ".", "This", "nevertheless", "requires", "frame", "preprocessing", "and", "is", "usually", "expensive", ".", "Alternatively", ",", "motion", "compensation", "can", "also", "be", "performed", "jointly", "with", "the", "SR", "task", ",", "as", "done", "in", "the", "Bayesian", "approach", "of", "Liu", "et", "al", ".", "by", "iteratively", "estimating", "motion", "as", "part", "of", "its", "wider", "modeling", "of", "the", "downscaling", "process", ".", "The", "advent", "of", "neural", "network", "techniques", "that", "can", "be", "trained", "from", "data", "to", "approximate", "complex", "nonlinear", "functions", "has", "set", "new", "performance", "standards", "in", "many", "applications", "including", "SR", ".", "Dong", "et", "al", ".", "proposed", "to", "use", "a", "CNN", "architecture", "for", "single", "image", "SR", "that", "was", "later", "extended", "by", "Kappeler", "et", "al", ".", "in", "a", "video", "SR", "network", "(", "VSRnet", ")", "which", "jointly", "processes", "multiple", "input", "frames", ".", "Additionally", ",", "compensating", "the", "motion", "of", "input", "images", "with", "a", "TV", "-", "based", "optical", "flow", "algorithm", "showed", "an", "improved", "accuracy", ".", "Joint", "motion", "compensation", "for", "SR", "with", "neural", "networks", "has", "also", "been", "studied", "through", "recurrent", "bidirectional", "networks", ".", "The", "common", "paradigm", "for", "CNN", "based", "approaches", "has", "been", "to", "upscale", "the", "LR", "image", "with", "bicubic", "interpolation", "before", "attempting", "to", "solve", "the", "SR", "problem", ".", "However", ",", "increasing", "input", "image", "size", "through", "interpolation", "considerably", "impacts", "the", "computational", "burden", "for", "CNN", "processing", ".", "A", "solution", "was", "proposed", "by", "Shi", "et", "al", ".", "with", "an", "efficient", "sub", "-", "pixel", "convolution", "network", "(", "ESPCN", ")", ",", "where", "an", "upscaling", "operation", "directly", "mapping", "from", "LR", "to", "HR", "space", "is", "learnt", "by", "the", "network", ".", "This", "technique", "reduces", "runtime", "by", "an", "order", "of", "magnitude", "and", "enables", "real", "-", "time", "video", "SR", "by", "independently", "processing", "frames", "with", "a", "single", "frame", "model", ".", "Similar", "solutions", "to", "improve", "efficiency", "have", "also", "been", "proposed", "based", "on", "transposed", "convolutions", ".", "subsection", ":", "Motivation", "and", "contributions", "[", "b", "]", "0.32", "[", "b", "]", "0.32", "[", "b", "]", "0.32", "Existing", "solutions", "for", "HD", "video", "SR", "have", "not", "been", "able", "to", "effectively", "exploit", "temporal", "correlations", "while", "performing", "in", "real", "-", "time", ".", "On", "the", "one", "hand", ",", "ESPCN", "leverages", "sub", "-", "pixel", "convolution", "for", "a", "very", "efficient", "operation", ",", "but", "its", "naive", "extension", "to", "videos", "treating", "frames", "independently", "fails", "to", "exploit", "inter", "-", "frame", "redundancies", "and", "does", "not", "enforce", "a", "temporally", "consistent", "result", ".", "VSRnet", ",", "on", "the", "other", "hand", ",", "can", "improve", "reconstruction", "quality", "by", "jointly", "processing", "multiple", "input", "frames", ".", "However", ",", "the", "preprocessing", "of", "LR", "images", "with", "bicubic", "upscaling", "and", "the", "use", "of", "an", "inefficient", "motion", "compensation", "mechanism", "slows", "runtime", "to", "about", "frames", "per", "second", "even", "on", "videos", "smaller", "than", "standard", "definition", "resolution", ".", "Spatial", "transformer", "networks", "provide", "a", "means", "to", "infer", "parameters", "for", "a", "spatial", "mapping", "between", "two", "images", ".", "These", "are", "differentiable", "networks", "that", "can", "be", "seamlessly", "combined", "and", "jointly", "trained", "with", "networks", "targeting", "other", "objectives", "to", "enhance", "their", "performance", ".", "For", "instance", ",", "spatial", "transformer", "networks", "were", "initially", "shown", "to", "facilitate", "image", "classification", "by", "transforming", "images", "onto", "the", "same", "frame", "of", "reference", ".", "Recently", ",", "it", "has", "been", "shown", "how", "spatial", "transformers", "can", "encode", "optical", "flow", "features", "with", "unsupervised", "training", ",", "but", "they", "have", "nevertheless", "not", "yet", "been", "investigated", "for", "video", "motion", "compensation", ".", "Related", "approaches", "have", "emerged", "for", "view", "synthesis", "assuming", "rigid", "transformations", ".", "In", "this", "paper", ",", "we", "combine", "the", "efficiency", "of", "sub", "-", "pixel", "convolution", "with", "the", "performance", "of", "spatio", "-", "temporal", "networks", "and", "motion", "compensation", "to", "obtain", "a", "fast", "and", "accurate", "video", "SR", "algorithm", ".", "We", "study", "different", "treatments", "of", "the", "temporal", "dimension", "with", "early", "fusion", ",", "slow", "fusion", "and", "3D", "convolutions", ",", "which", "have", "been", "previously", "suggested", "to", "extend", "classification", "from", "images", "to", "videos", ".", "Additionally", ",", "we", "build", "a", "motion", "compensation", "scheme", "based", "on", "spatial", "transformers", ",", "which", "is", "combined", "with", "spatio", "-", "temporal", "models", "to", "lead", "to", "a", "very", "efficient", "solution", "for", "video", "SR", "with", "motion", "compensation", "that", "is", "end", "-", "to", "-", "end", "trainable", ".", "A", "high", "-", "level", "diagram", "of", "the", "proposed", "approach", "is", "show", "in", "fig", ":", "network", ".", "The", "main", "contributions", "of", "this", "paper", "are", ":", "Presenting", "a", "real", "-", "time", "approach", "for", "video", "SR", "based", "on", "sub", "-", "pixel", "convolution", "and", "spatio", "-", "temporal", "networks", "that", "improves", "accuracy", "and", "temporal", "consistency", ".", "Comparing", "early", "fusion", ",", "slow", "fusion", "and", "3D", "convolutions", "as", "alternative", "architectures", "for", "discovering", "spatio", "-", "temporal", "correlations", ".", "Proposing", "an", "efficient", "method", "for", "dense", "inter", "-", "frame", "motion", "compensation", "based", "on", "a", "multi", "-", "scale", "spatial", "transformer", "network", ".", "Combining", "the", "proposed", "motion", "compensation", "technique", "with", "spatio", "-", "temporal", "models", "to", "provide", "an", "efficient", ",", "end", "-", "to", "-", "end", "trainable", "motion", "compensated", "video", "SR", "algorithm", ".", "section", ":", "Methods", "Our", "starting", "point", "is", "the", "real", "-", "time", "image", "SR", "method", "ESPCN", ".", "We", "restrict", "our", "analysis", "to", "standard", "architectural", "choices", "and", "do", "not", "further", "investigate", "potentially", "beneficial", "extensions", "such", "as", "recurrence", ",", "residual", "connections", "or", "training", "networks", "based", "on", "perceptual", "loss", "functions", ".", "Throughout", "the", "paper", "we", "assume", "all", "image", "processing", "is", "performed", "on", "the", "y", "-", "channel", "in", "colour", "space", ",", "and", "thus", "we", "represent", "all", "images", "as", "2D", "matrices", ".", "subsection", ":", "Sub", "-", "pixel", "convolution", "SR", "For", "a", "given", "LR", "image", "which", "is", "assumed", "to", "be", "the", "result", "of", "low", "-", "pass", "filtering", "and", "downscaling", "by", "a", "factor", "the", "HR", "image", ",", "the", "CNN", "super", "-", "resolved", "solution", "can", "be", "expressed", "as", "Here", ",", "are", "model", "parameters", "and", "represents", "the", "mapping", "function", "from", "LR", "to", "HR", ".", "A", "convolutional", "network", "models", "this", "function", "as", "a", "concatenation", "of", "layers", "defined", "by", "sets", "of", "weights", "and", "biases", ",", "each", "followed", "by", "non", "-", "linearities", ",", "with", ".", "Formally", ",", "the", "output", "of", "each", "layer", "is", "written", "as", "with", ".", "We", "assume", "the", "shape", "of", "filtering", "weights", "to", "be", ",", "where", "and", "represent", "the", "number", "and", "size", "of", "filters", "in", "layer", ",", "with", "the", "single", "frame", "input", "meaning", ".", "Model", "parameters", "are", "optimised", "minimising", "a", "loss", "given", "a", "set", "of", "LR", "and", "HR", "example", "image", "pairs", ",", "commonly", "MSE", ":", "Methods", "preprocessing", "with", "bicubic", "upsampling", "before", "mapping", "from", "LR", "to", "HR", "impose", "that", "the", "output", "number", "of", "filters", "is", ".", "Using", "sub", "-", "pixel", "convolution", "allows", "to", "process", "directly", "in", "the", "LR", "space", "and", "then", "use", "output", "filters", "to", "obtain", "an", "HR", "output", "tensor", "with", "shape", "that", "can", "be", "reordered", "to", "obtain", ".", "This", "implies", "that", "if", "there", "exists", "an", "upscaling", "operation", "that", "is", "better", "suited", "for", "the", "problem", "than", "bicubic", "upsampling", ",", "the", "network", "can", "learn", "it", ".", "Moreover", ",", "and", "most", "importantly", ",", "all", "convolutional", "processing", "is", "performed", "in", "LR", "space", ",", "making", "this", "approach", "very", "efficient", ".", "subsection", ":", "Spatio", "-", "temporal", "networks", "Spatio", "-", "temporal", "networks", "assume", "input", "data", "to", "be", "a", "block", "of", "spatio", "-", "temporal", "information", ",", "such", "that", "instead", "of", "a", "single", "input", "frame", ",", "a", "sequence", "of", "consecutive", "frames", "is", "considered", ".", "This", "can", "be", "represented", "in", "the", "network", "by", "introducing", "an", "additional", "dimension", "for", "temporal", "depth", ",", "with", "the", "input", "depth", "representing", "an", "odd", "number", "of", "consecutive", "input", "frames", ".", "If", "we", "denote", "the", "temporal", "radius", "of", "a", "spatio", "-", "temporal", "block", "to", "be", ",", "we", "define", "the", "group", "of", "input", "frames", "centered", "at", "time", "as", ",", "and", "the", "problem", "in", "eq", ":", "image", "-", "sr", "becomes", "The", "shape", "of", "weighting", "filters", "is", "also", "extended", "by", "their", "temporal", "size", ",", "and", "their", "tensor", "shape", "becomes", ".", "We", "note", "that", "it", "is", "possible", "to", "consider", "solutions", "that", "aim", "to", "jointly", "reconstruct", "more", "than", "a", "single", "output", "frame", ",", "which", "could", "have", "advantages", "at", "least", "in", "terms", "of", "computational", "efficiency", ".", "However", ",", "in", "this", "work", "we", "focus", "on", "the", "reconstruction", "of", "only", "a", "single", "output", "frame", ".", "subsubsection", ":", "Early", "fusion", "One", "of", "the", "most", "straightforward", "approaches", "for", "a", "CNN", "to", "process", "videos", "is", "to", "match", "the", "temporal", "depth", "of", "the", "input", "layer", "to", "the", "number", "of", "frames", ".", "This", "will", "collapse", "all", "temporal", "information", "in", "the", "first", "layer", "and", "the", "remaining", "operations", "are", "identical", "to", "those", "in", "a", "single", "image", "SR", "network", ",", "meaning", ".", "An", "illustration", "of", "early", "fusion", "is", "shown", "in", "fig", ":", "early", "-", "fusion", "for", ",", "where", "the", "temporal", "dimension", "has", "been", "colour", "coded", "and", "the", "output", "mapping", "to", "2D", "space", "is", "omitted", ".", "This", "design", "has", "been", "studied", "for", "video", "classification", "and", "action", "recognition", ",", "and", "was", "also", "one", "of", "the", "architectures", "proposed", "in", "VSRnet", ".", "However", ",", "VSRnet", "requires", "bicubic", "upsampling", "as", "opposed", "to", "sub", "-", "pixel", "convolution", ",", "making", "the", "framework", "computationally", "much", "less", "efficient", "in", "comparison", ".", "subsubsection", ":", "Slow", "fusion", "Another", "option", "is", "to", "partially", "merge", "temporal", "information", "in", "a", "hierarchical", "structure", ",", "so", "it", "is", "slowly", "fused", "as", "information", "progresses", "through", "the", "network", ".", "In", "this", "case", ",", "the", "temporal", "depth", "of", "network", "layers", "is", "configured", "to", "be", ",", "and", "therefore", "some", "layers", "also", "have", "a", "temporal", "extent", "until", "all", "information", "has", "been", "merged", "and", "the", "depth", "of", "the", "network", "reduces", "to", ".", "This", "architecture", ",", "termed", "slow", "fusion", ",", "has", "shown", "better", "performance", "than", "early", "fusion", "for", "video", "classification", ".", "In", "fig", ":", "slow", "-", "fusion", "we", "show", "a", "slow", "fusion", "network", "where", "and", "the", "rate", "of", "fusion", "is", "defined", "by", "for", "or", "otherwise", ",", "meaning", "that", "at", "each", "layer", "only", "two", "consecutive", "frames", "or", "filter", "activations", "are", "merged", "until", "the", "network", "’s", "temporal", "depth", "shrinks", "to", ".", "Note", "that", "early", "fusion", "is", "an", "special", "case", "of", "slow", "fusion", ".", "subsubsection", ":", "3D", "convolutions", "Another", "variation", "of", "slow", "fusion", "is", "to", "force", "layer", "weights", "to", "be", "shared", "across", "the", "temporal", "dimension", ",", "which", "has", "computational", "advantages", ".", "Assuming", "an", "online", "processing", "of", "frames", ",", "when", "a", "new", "frame", "becomes", "available", "the", "result", "of", "some", "layers", "for", "the", "previous", "frame", "can", "be", "reused", ".", "For", "instance", ",", "refering", "to", "the", "diagram", "in", "fig", ":", "slow", "-", "fusion", "and", "assuming", "the", "bottom", "frame", "to", "be", "the", "latest", "frame", "received", ",", "all", "activations", "above", "the", "dashed", "line", "are", "readily", "available", "because", "they", "were", "required", "for", "processing", "the", "previous", "frame", ".", "This", "architecture", "is", "equivalent", "to", "using", "3D", "convolutions", ",", "initially", "proposed", "as", "an", "effective", "tool", "to", "learn", "spatio", "-", "temporal", "features", "that", "can", "help", "for", "video", "action", "recognition", ".", "An", "illustration", "of", "this", "design", "from", "a", "3D", "convolution", "perspective", "is", "shown", "in", "fig:3dconv", ",", "where", "the", "arrangement", "of", "the", "temporal", "and", "filter", "features", "is", "swapped", "relative", "to", "fig", ":", "slow", "-", "fusion", ".", "subsection", ":", "Spatial", "transformer", "motion", "compensation", "We", "propose", "the", "use", "of", "an", "efficient", "spatial", "transformer", "network", "to", "compensate", "the", "motion", "between", "frames", "fed", "to", "the", "SR", "network", ".", "It", "has", "been", "shown", "how", "spatial", "transformers", "can", "effectively", "encode", "optical", "flow", "to", "describe", "motion", ",", "and", "are", "therefore", "suitable", "for", "motion", "compensation", ".", "We", "will", "compensate", "blocks", "of", "three", "consecutive", "frames", "to", "combine", "the", "compensation", "module", "with", "the", "SR", "network", "as", "shown", "in", "fig", ":", "network", ",", "but", "for", "simplicity", "we", "first", "introduce", "motion", "compensation", "between", "two", "frames", ".", "Notice", "that", "the", "data", "used", "contains", "inherent", "motion", "blur", "and", "(", "dis", ")", "occlusions", ",", "and", "even", "though", "an", "explicit", "modelling", "for", "these", "effects", "is", "not", "used", "it", "could", "potentially", "improve", "results", ".", "The", "task", "is", "to", "find", "the", "best", "optical", "flow", "representation", "relating", "a", "new", "frame", "with", "a", "reference", "current", "frame", ".", "The", "flow", "is", "assumed", "pixel", "-", "wise", "dense", ",", "allowing", "to", "displace", "each", "pixel", "to", "a", "new", "position", ",", "and", "the", "resulting", "pixel", "arrangement", "requires", "interpolation", "back", "onto", "a", "regular", "grid", ".", "We", "use", "bilinear", "interpolation", "as", "it", "is", "much", "more", "efficient", "than", "the", "thin", "-", "plate", "spline", "interpolation", "originally", "proposed", "in", ".", "Optical", "flow", "is", "a", "function", "of", "parameters", "and", "is", "represented", "with", "two", "feature", "maps", "corresponding", "to", "displacements", "for", "the", "and", "dimensions", ",", "thus", "a", "compensated", "image", "can", "be", "expressed", "as", ",", "or", "more", "concisely", "We", "adopt", "a", "multi", "-", "scale", "design", "to", "represent", "the", "flow", ",", "which", "has", "been", "shown", "to", "be", "effective", "in", "classical", "methods", "and", "also", "in", "more", "recently", "proposed", "spatial", "transformer", "techniques", ".", "A", "schematic", "of", "the", "design", "is", "shown", "in", "fig", ":", "transformer", "and", "flow", "estimation", "modules", "are", "detailed", "in", "tab", ":", "transformer", ".", "First", ",", "a", "coarse", "estimate", "of", "the", "flow", "is", "obtained", "by", "early", "fusing", "the", "two", "input", "frames", "and", "downscaling", "spatial", "dimensions", "with", "strided", "convolutions", ".", "The", "estimated", "flow", "is", "upscaled", "with", "sub", "-", "pixel", "convolution", "and", "the", "result", "is", "applied", "to", "warp", "the", "target", "frame", "producing", ".", "The", "warped", "image", "is", "then", "processed", "together", "with", "the", "coarse", "flow", "and", "the", "original", "images", "through", "a", "fine", "flow", "estimation", "module", ".", "This", "uses", "a", "single", "strided", "convolution", "with", "stride", "and", "a", "final", "upscaling", "stage", "to", "obtain", "a", "finer", "flow", "map", ".", "The", "final", "motion", "compensated", "frame", "is", "obtained", "by", "warping", "the", "target", "frame", "with", "the", "total", "flow", ".", "Output", "activations", "use", "tanh", "to", "represent", "pixel", "displacement", "in", "normalised", "space", ",", "such", "that", "a", "displacement", "of", "means", "maximum", "displacement", "from", "the", "center", "to", "the", "border", "of", "the", "image", ".", "To", "train", "the", "spatial", "transformer", "to", "perform", "motion", "compensation", "we", "optimise", "its", "parameters", "to", "minimise", "the", "MSE", "between", "the", "transformed", "frame", "and", "the", "reference", "frame", ".", "Similary", "to", "classical", "optical", "flow", "methods", ",", "we", "found", "that", "it", "is", "generally", "helpful", "to", "constrain", "the", "flow", "to", "behave", "smoothly", "in", "space", ",", "and", "so", "we", "penalise", "the", "Huber", "loss", "of", "the", "flow", "map", "gradients", ",", "namely", "In", "practice", "we", "approximate", "the", "Huber", "loss", "with", ",", "where", ".", "This", "function", "has", "a", "smooth", "behaviour", "near", "the", "origin", "and", "is", "sparsity", "promoting", "far", "from", "it", ".", "The", "spatial", "transformer", "module", "is", "advantageous", "relative", "to", "other", "motion", "compensation", "mechanisms", "as", "it", "is", "straightforward", "to", "combine", "with", "a", "SR", "network", "to", "perform", "joint", "motion", "compensation", "and", "video", "SR", ".", "Referring", "to", "fig", ":", "network", ",", "the", "same", "parameters", "can", "be", "used", "to", "model", "motion", "of", "the", "outer", "two", "frames", "relative", "to", "the", "central", "frame", ".", "The", "spatial", "transformer", "and", "SR", "modules", "are", "both", "differentiable", "and", "therefore", "end", "-", "to", "-", "end", "trainable", ".", "As", "a", "result", ",", "they", "can", "be", "jointly", "optimised", "to", "minimise", "a", "composite", "loss", "combining", "the", "accuracy", "of", "the", "reconstruction", "in", "eq", ":", "image", "-", "sr", "-", "objective", "with", "the", "fidelity", "of", "motion", "compensation", "in", "eq", ":", "motion", "-", "compensation", "-", "objective", ",", "namely", "section", ":", "Experiments", "and", "results", "In", "this", "section", ",", "we", "first", "analyse", "spatio", "-", "temporal", "networks", "for", "video", "SR", "in", "isolation", "and", "later", "evaluate", "the", "benefits", "of", "introducing", "motion", "compensation", ".", "We", "restrict", "our", "experiments", "to", "tackle", "and", "upscaling", "of", "full", "HD", "video", "resolution", "(", ")", ",", "and", "no", "compression", "is", "applied", ".", "To", "ensure", "a", "fair", "comparison", "of", "methods", ",", "the", "number", "of", "network", "parameters", "need", "to", "be", "comparable", "so", "that", "gains", "in", "performance", "can", "be", "attributed", "to", "specific", "choices", "of", "network", "resource", "allocation", "and", "not", "to", "a", "trivial", "increase", "in", "capacity", ".", "For", "a", "layer", ",", "the", "number", "of", "floating", "-", "point", "operations", "to", "reconstruct", "a", "frame", "is", "approximated", "by", "In", "measuring", "the", "complexity", "of", "slow", "fusion", "networks", "with", "weight", "sharing", "we", "look", "at", "steady", "-", "state", "operation", "where", "the", "output", "of", "some", "layers", "is", "reused", "from", "one", "frame", "to", "the", "following", ".", "We", "note", "that", "the", "analysis", "of", "VSRnet", "variants", "in", "does", "not", "take", "into", "account", "model", "complexity", ".", "subsection", ":", "Experimental", "setup", "subsubsection", ":", "Data", "We", "use", "the", "CDVL", "database", ",", "which", "contains", "uncompressed", "full", "HD", "videos", "excluding", "repeated", "videos", ",", "and", "choose", "a", "subset", "of", "videos", "for", "training", ".", "The", "videos", "are", "downscaled", "and", "random", "samples", "are", "extracted", "from", "each", "HR", "-", "LR", "video", "pair", "to", "obtain", "training", "samples", ",", "of", "which", "are", "used", "for", "validation", ".", "Depending", "on", "the", "network", "architecture", ",", "we", "refer", "to", "a", "sample", "as", "a", "single", "input", "-", "output", "frame", "pair", "for", "single", "frame", "networks", ",", "or", "as", "a", "block", "of", "consecutive", "LR", "input", "frames", "and", "the", "corresponding", "central", "HR", "frame", "for", "spatio", "-", "temporal", "networks", ".", "The", "remaining", "videos", "are", "used", "for", "testing", ".", "Although", "the", "total", "number", "of", "training", "frames", "is", "large", ",", "we", "foresee", "that", "the", "methods", "presented", "could", "benefit", "from", "a", "richer", ",", "more", "diverse", "set", "of", "videos", ".", "Additionally", ",", "we", "present", "a", "benchmark", "against", "various", "SR", "methods", "on", "publicly", "available", "videos", "that", "are", "recurrently", "used", "in", "the", "literature", "and", "we", "refer", "to", "as", "Vid4", ".", "subsubsection", ":", "Network", "training", "and", "parameters", "All", "SR", "models", "are", "trained", "following", "the", "same", "protocol", "and", "share", "similar", "hyperparameters", ".", "Filter", "sizes", "are", "set", "to", ",", "and", "all", "non", "-", "linearities", "are", "rectified", "linear", "units", "except", "for", "the", "output", "layer", ",", "which", "uses", "a", "linear", "activation", ".", "Biases", "are", "initialised", "to", "and", "weights", "use", "orthogonal", "initialisation", "with", "gain", "following", "recommendations", "in", ".", "All", "hidden", "layers", "are", "set", "to", "have", "the", "same", "number", "of", "features", ".", "Video", "samples", "are", "broken", "into", "non", "-", "overlapping", "sub", "-", "samples", "of", "spatial", "dimensions", ",", "which", "are", "randomly", "grouped", "in", "batches", "for", "stochastic", "optimisation", ".", "We", "employ", "Adam", "with", "a", "learning", "rate", "and", "an", "initial", "batch", "size", ".", "Every", "epochs", "the", "batch", "size", "is", "doubled", "until", "it", "reaches", "a", "maximum", "size", "of", ".", "We", "choose", "for", "layers", "where", "the", "network", "temporal", "depth", "is", "(", "layers", "in", "gray", "in", "fig", ":", "early", "-", "fusion", ",", "fig", ":", "slow", "-", "fusion", ",", "fig:3dconv", ")", ",", "and", "to", "maintain", "comparable", "network", "sizes", "we", "choose", ".", "This", "ensures", "that", "the", "number", "of", "features", "per", "hidden", "layer", "in", "early", "and", "slow", "fusion", "networks", "is", "always", "the", "same", ".", "For", "instance", ",", "the", "network", "shown", "in", "fig", ":", "slow", "-", "fusion", ",", "for", "which", "and", "for", ",", "the", "number", "of", "features", "in", "a", "layer", "network", "for", "SR", "would", "be", "6", ",", "8", ",", "12", ",", "24", ",", "24", ",", ".", "subsection", ":", "Spatio", "-", "temporal", "video", "SR", "subsubsection", ":", "Single", "vs", "multi", "frame", "early", "fusion", "First", ",", "we", "investigate", "the", "impact", "of", "the", "number", "of", "input", "frames", "on", "complexity", "and", "accuracy", "without", "motion", "compensation", ".", "We", "compare", "single", "frame", "models", "(", "SF", ")", "against", "early", "fusion", "spatio", "-", "temporal", "models", "using", "3", ",", "5", "and", "7", "input", "frames", "(", "E3", ",", "E5", "and", "E7", ")", ".", "PSNR", "results", "on", "the", "CDVL", "dataset", "for", "networks", "of", "6", "to", "11", "layers", "are", "plotted", "in", "fig", ":", "single", "-", "vs", "-", "multi", "-", "frame", ".", "Exploiting", "spatio", "-", "temporal", "correlations", "provides", "a", "more", "accurate", "result", "relative", "to", "an", "independent", "processing", "of", "frames", ".", "The", "increase", "in", "complexity", "from", "early", "fusion", "is", "marginal", "because", "only", "the", "first", "layer", "contributes", "to", "an", "increase", "of", "operations", ".", "Although", "the", "accuracy", "of", "spatio", "-", "temporal", "models", "is", "relatively", "similar", ",", "we", "find", "that", "E7", "slightly", "underperforms", ".", "It", "is", "likely", "that", "temporal", "dependencies", "beyond", "5", "frames", "become", "too", "complex", "for", "networks", "to", "learn", "useful", "information", "and", "act", "as", "noise", "degrading", "their", "performance", ".", "Notice", "also", "that", ",", "whereas", "the", "performance", "increase", "from", "network", "depth", "is", "minimal", "after", "8", "layers", "for", "single", "frame", "networks", ",", "this", "increase", "is", "more", "consistent", "for", "spatio", "-", "temporal", "models", ".", "subsubsection", ":", "Early", "vs", "slow", "fusion", "Here", "we", "compare", "the", "different", "treatments", "of", "the", "temporal", "dimension", "discussed", "in", "ssec", ":", "st", "-", "networks", ".", "We", "assume", "networks", "with", "an", "input", "of", "frames", "and", "slow", "fusion", "models", "with", "filter", "temporal", "depths", "as", "in", "fig", ":", "st", "-", "networks", ".", "Using", "SF", ",", "E5", ",", "S5", ",", "and", "S5", "-", "SW", "to", "refer", "to", "single", "frame", "networks", "and", "5", "frame", "input", "networks", "using", "early", "fusion", ",", "slow", "fusion", ",", "and", "slow", "fusion", "with", "shared", "weights", ",", "we", "show", "in", "tab", ":", "early", "-", "vs", "-", "slow", "-", "fusion", "results", "for", "7", "and", "9", "layer", "networks", ".", "As", "seen", "previously", ",", "early", "fusion", "networks", "attain", "a", "higher", "accuracy", "at", "a", "marginal", "3", "%", "increase", "in", "operations", "relative", "to", "the", "single", "frame", "models", ",", "and", "as", "expected", ",", "slow", "fusion", "architectures", "provide", "efficiency", "advantages", ".", "Slow", "fusion", "is", "faster", "than", "early", "fusion", "because", "it", "uses", "fewer", "features", "in", "the", "initial", "layers", ".", "Referring", "to", "eq", ":", "operations", ",", "slow", "fusion", "uses", "in", "the", "first", "layers", "and", ",", "which", "results", "in", "fewer", "operations", "than", ",", "as", "used", "in", "early", "fusion", ".", "While", "the", "7", "layer", "network", "sees", "a", "considerable", "decrease", "in", "accuracy", "using", "slow", "fusion", "relative", "to", "early", "fusion", ",", "the", "9", "layer", "network", "can", "benefit", "from", "the", "same", "accuracy", "while", "reducing", "its", "complexity", "with", "slow", "fusion", "by", "about", "30", "%", ".", "This", "suggests", "that", "in", "shallow", "networks", "the", "best", "use", "of", "network", "resources", "is", "to", "utilise", "the", "full", "network", "capacity", "to", "jointly", "process", "all", "temporal", "information", "as", "done", "by", "early", "fusion", ",", "but", "that", "in", "deeper", "networks", "slowly", "fusing", "the", "temporal", "dimension", "is", "beneficial", ",", "which", "is", "in", "line", "with", "the", "results", "presented", "by", "for", "video", "classification", ".", "Additionally", ",", "weight", "sharing", "decreases", "accuracy", "because", "of", "the", "reduction", "in", "network", "parameters", ",", "but", "the", "reusability", "of", "network", "features", "means", "fewer", "operations", "are", "needed", "per", "frame", ".", "For", "instance", ",", "the", "7", "layer", "S5", "-", "SW", "network", "shows", "a", "reduction", "of", "almost", "30", "%", "of", "operations", "with", "a", "minimal", "decrease", "in", "accuracy", "relative", "to", "SF", ".", "Using", "7", "layers", "with", "E5", "nevertheless", "shows", "better", "performance", "and", "faster", "operation", "than", "S5", "-", "SW", "with", "9", "layers", ",", "and", "in", "all", "cases", "we", "found", "that", "early", "or", "slow", "fusion", "consistently", "outperformed", "slow", "fusion", "with", "shared", "weights", "in", "this", "performance", "and", "efficiency", "trade", "-", "off", ".", "Convolutions", "in", "spatio", "-", "temporal", "domain", "were", "shown", "in", "to", "work", "well", "for", "video", "action", "recognition", ",", "but", "with", "larger", "capacity", "and", "many", "more", "frames", "processed", "jointly", ".", "We", "speculate", "this", "could", "be", "the", "reason", "why", "the", "conclusions", "drawn", "from", "this", "high", "-", "level", "vision", "task", "do", "not", "extrapolate", "to", "the", "SR", "problem", ".", "subsection", ":", "Motion", "compensated", "video", "SR", "[", "b", "]", "0.32", "[", "b", "]", "0.32", "[", "b", "]", "0.32", "In", "this", "section", ",", "the", "proposed", "frame", "motion", "compensation", "is", "combined", "with", "an", "early", "fusion", "network", "of", "temporal", "depth", ".", "First", ",", "the", "motion", "compensation", "module", "is", "trained", "independently", "using", "eq", ":", "video", "-", "sr", "-", "memc", ",", "where", "the", "first", "term", "is", "ignored", "and", ",", ".", "This", "results", "in", "a", "network", "that", "will", "compensate", "the", "motion", "of", "three", "consecutive", "frames", "by", "estimating", "the", "flow", "maps", "of", "outer", "frames", "relative", "to", "the", "middle", "frame", ".", "An", "example", "of", "a", "flow", "map", "obtained", "for", "one", "frame", "is", "shown", "in", "fig", ":", "memc", ",", "where", "we", "also", "show", "the", "effect", "the", "motion", "compensation", "module", "has", "on", "three", "consecutive", "frames", ".", "The", "early", "fusion", "motion", "compensated", "SR", "network", "(", "E3", "-", "MC", ")", "is", "initialised", "with", "a", "compensation", "and", "a", "SR", "network", "pretrained", "separately", ",", "and", "the", "full", "model", "is", "then", "jointly", "optimised", "with", "eq", ":", "video", "-", "sr", "-", "memc", "(", ",", ")", ".", "Results", "for", "SR", "on", "CDVL", "are", "compared", "in", "tab", ":", "motion", "-", "compensated", "-", "video", "-", "sr", "against", "a", "single", "frame", "(", "SF", ")", "model", "and", "early", "fusion", "without", "motion", "compensation", "(", "E3", ")", ".", "E3", "-", "MC", "results", "in", "a", "PSNR", "that", "is", "sometimes", "almost", "twice", "the", "improvement", "of", "E3", "relative", "to", "SF", ",", "which", "we", "attribute", "to", "the", "fact", "that", "the", "network", "adapts", "the", "SR", "input", "to", "maximise", "temporal", "redundancy", ".", "In", "fig", ":", "mcsr", "-", "x3", "we", "show", "how", "this", "improvement", "is", "reflected", "in", "better", "structure", "preservation", ".", "subsection", ":", "Comparison", "to", "state", "-", "of", "-", "the", "-", "art", "We", "show", "in", "tab", ":", "set4", "the", "performance", "on", "Vid4", "for", "SRCNN", ",", "ESPCN", ",", "VSRnet", "and", "the", "proposed", "method", ",", "which", "we", "refer", "to", "as", "video", "ESPCN", "(", "VESPCN", ")", ".", "To", "demonstrate", "its", "benefits", "in", "efficiency", "and", "quality", "we", "evaluate", "two", "early", "fusion", "models", ":", "a", "5", "layer", "3", "frame", "network", "(", "5L", "-", "E3", ")", "and", "a", "9", "layer", "3", "frame", "network", "with", "motion", "compensation", "(", "9L", "-", "E3", "-", "MC", ")", ".", "The", "metrics", "compared", "are", "PSNR", ",", "SSIM", "and", "MOVIE", "indices", ".", "The", "MOVIE", "index", "was", "designed", "as", "a", "metric", "measuring", "video", "quality", "that", "correlates", "with", "human", "perception", "and", "incorporates", "a", "notion", "of", "temporal", "consistency", ".", "We", "also", "directly", "compare", "the", "number", "of", "operations", "per", "frame", "of", "all", "CNN", "-", "based", "approaches", "for", "upscaling", "a", "generic", "p", "frame", ".", "Reconstructions", "for", "SRCNN", ",", "ESPCN", "and", "VSRnet", "use", "models", "provided", "by", "the", "authors", ".", "SRCNN", ",", "ESPCN", "and", "VESPCN", "were", "tested", "on", "Theano", "and", "Lasagne", ",", "and", "for", "VSRnet", "we", "used", "available", "Caffe", "Matlab", "code", ".", "We", "crop", "spatial", "borders", "as", "well", "as", "initial", "and", "final", "frames", "on", "all", "reconstructions", "for", "fair", "comparison", "against", "VSRnet", ".", "subsubsection", ":", "Quality", "comparison", "An", "example", "of", "visual", "differences", "is", "shown", "in", "fig", ":", "set4", "-", "visualisation", "against", "the", "motion", "compensated", "network", ".", "From", "the", "close", "-", "up", "images", ",", "we", "see", "how", "the", "structural", "detail", "of", "the", "original", "video", "is", "better", "recovered", "by", "the", "proposed", "VESPCN", "method", ".", "This", "is", "reflected", "in", "tab", ":", "set4", ",", "where", "it", "surpasses", "any", "other", "method", "in", "PSNR", "and", "SSIM", "by", "a", "large", "margin", ".", "fig", ":", "set4", "-", "visualisation", "also", "shows", "temporal", "profiles", "on", "the", "row", "highlighted", "by", "a", "dashed", "line", "through", "25", "consecutive", "frames", ",", "demonstrating", "a", "better", "temporal", "coherence", "of", "the", "reconstruction", "proposed", ".", "The", "great", "temporal", "coherence", "of", "VESPCN", "also", "explains", "the", "significant", "reduction", "in", "the", "MOVIE", "index", ".", "subsubsection", ":", "Efficiency", "comparison", "The", "complexity", "of", "methods", "in", "tab", ":", "set4", "is", "determined", "by", "network", "and", "input", "image", "sizes", ".", "SRCNN", "and", "VSRnet", "upsample", "LR", "images", "before", "attempting", "to", "super", "-", "resolve", "them", ",", "which", "considerably", "increases", "the", "required", "number", "of", "operations", ".", "VSRnet", "is", "particularly", "expensive", "because", "it", "processes", "input", "frames", "in", "and", "feature", "layers", ",", "whereas", "sub", "-", "pixel", "convolution", "greatly", "reduces", "the", "number", "of", "operations", "required", "in", "ESPCN", "and", "VESPCN", ".", "As", "a", "reference", ",", "ESPCN", "runs", "at", "ms", "per", "frame", "on", "a", "K2", "GPU", ".", "The", "enhanced", "capabilities", "of", "spatio", "-", "temporal", "networks", "allow", "to", "reduce", "the", "network", "operations", "of", "VESPCN", "relative", "to", "ESPCN", "while", "still", "matching", "its", "accuracy", ".", "As", "an", "example", "we", "show", "VESPCN", "with", "5L", "-", "E3", ",", "which", "reduces", "the", "number", "of", "operations", "by", "about", "20", "%", "relative", "to", "ESPCN", "while", "maintaining", "a", "similar", "performance", "in", "all", "evaluated", "quality", "metrics", ".", "The", "operations", "for", "motion", "compensation", "in", "VESPCN", "with", "9L", "-", "E3", "-", "MC", ",", "included", "in", "tab", ":", "set4", "results", ",", "amount", "to", "and", "GOps", "for", "and", "upscaling", ",", "applied", "twice", "for", "each", "input", "frame", "requiring", "motion", "compensation", ".", "This", "makes", "the", "proposed", "motion", "compensated", "video", "SR", "very", "efficient", "relative", "to", "other", "approaches", ".", "For", "example", ",", "motion", "compensation", "in", "VSRnet", "is", "said", "to", "require", "55", "seconds", "per", "frame", "and", "is", "the", "computational", "bottleneck", ".", "This", "is", "not", "accounted", "for", "in", "tab", ":", "set4", "but", "is", "slower", "than", "VESPCN", "with", "9L", "-", "E3", "-", "MC", ",", "which", "can", "run", "in", "the", "order", "of", "seconds", ".", "The", "optical", "flow", "method", "in", "VSRnet", "was", "originally", "shown", "to", "run", "at", "ms", "on", "GPU", "for", "each", "frame", "of", "dimensions", ",", "but", "this", "is", "still", "considerably", "slower", "than", "the", "proposed", "solution", "considering", "motion", "compensation", "is", "required", "for", "more", "than", "a", "single", "frame", "of", "HD", "dimensions", ".", "section", ":", "Conclusion", "In", "this", "paper", "we", "combine", "the", "efficiency", "advantages", "of", "sub", "-", "pixel", "convolutions", "with", "temporal", "fusion", "strategies", "to", "present", "real", "-", "time", "spatio", "-", "temporal", "models", "for", "video", "SR", ".", "The", "spatio", "-", "temporal", "models", "used", "are", "shown", "to", "facilitate", "an", "improvement", "in", "reconstruction", "accuracy", "and", "temporal", "consistency", "or", "reduce", "computational", "complexity", "relative", "to", "independent", "single", "frame", "processing", ".", "The", "models", "investigated", "are", "extended", "with", "a", "motion", "compensation", "mechanism", "based", "on", "spatial", "transformer", "networks", "that", "is", "efficient", "and", "jointly", "trainable", "for", "video", "SR", ".", "Results", "obtained", "with", "approaches", "that", "incorporate", "explicit", "motion", "compensation", "are", "demonstrated", "to", "be", "superior", "in", "terms", "of", "PSNR", "and", "temporal", "consistency", "compared", "to", "spatio", "-", "temporal", "models", "alone", ",", "and", "outperform", "the", "current", "state", "of", "the", "art", "in", "video", "SR", ".", "bibliography", ":", "References"], "entity": [[2, 9, ["Real", "-", "Time", "Video", "Super", "-", "Resolution"], "Task"], [10, 14, ["Spatio", "-", "Temporal", "Networks"], "Method"], [15, 17, ["Motion", "Compensation"], "Method"], [17, 20, ["Convolutional", "neural", "networks"], "Method"], [23, 31, ["image", "super", "-", "resolution", "in", "real", "-", "time"], "Task"], [42, 46, ["video", "super", "-", "resolution"], "Task"], [61, 69, ["spatio", "-", "temporal", "sub", "-", "pixel", "convolution", "networks"], "Method"], [76, 78, ["reconstruction", "accuracy"], "Metric"], [80, 84, ["real", "-", "time", "speed"], "Metric"], [92, 94, ["early", "fusion"], "Method"], [95, 97, ["slow", "fusion"], "Method"], [98, 100, ["3D", "convolutions"], "Method"], [115, 118, ["joint", "motion", "compensation"], "Method"], [139, 145, ["multi", "-", "resolution", "spatial", "transformer", "module"], "Method"], [159, 160, ["accuracy"], "Metric"], [175, 179, ["single", "-", "frame", "models"], "Method"], [180, 184, ["spatio", "-", "temporal", "networks"], "Method"], [188, 190, ["computational", "cost"], "Metric"], [197, 198, ["quality"], "Metric"], [206, 208, ["computational", "cost"], "Metric"], [231, 232, ["accuracy"], "Metric"], [233, 234, ["efficiency"], "Metric"], [267, 269, ["signal", "processing"], "Task"], [270, 271, ["SR"], "Method"], [275, 277, ["HR", "image"], "Task"], [278, 279, ["video"], "Task"], [290, 292, ["medical", "imaging"], "Task"], [293, 295, ["satellite", "imaging"], "Task"], [299, 301, ["facilitating", "tasks"], "Task"], [303, 305, ["face", "recognition"], "Task"], [307, 311, ["reconstruction", "of", "HR", "data"], "Task"], [319, 323, ["ill", "-", "posed", "problem"], "Task"], [348, 351, ["single", "image", "SR"], "Task"], [391, 396, ["multi", "-", "scale", "patch", "recurrence"], "Method"], [398, 402, ["multi", "-", "image", "SR"], "Task"], [432, 434, ["downscaling", "process"], "Method"], [466, 467, ["accuracy"], "Metric"], [470, 471, ["efficiency"], "Metric"], [476, 479, ["Video", "SR", "methods"], "Method"], [485, 488, ["image", "SR", "techniques"], "Method"], [489, 492, ["Kernel", "regression", "methods"], "Method"], [508, 511, ["Dictionary", "learning", "approaches"], "Method"], [518, 524, ["sparse", "linear", "combination", "of", "dictionary", "atoms"], "Method"], [542, 547, ["example", "-", "based", "patch", "recurrence"], "Method"], [592, 594, ["image", "SR"], "Task"], [640, 644, ["Estimating", "and", "compensating", "motion"], "Task"], [695, 703, ["off", "-", "the", "-", "shelf", "optical", "flow", "algorithms"], "Method"], [707, 709, ["frame", "preprocessing"], "Method"], [716, 718, ["motion", "compensation"], "Task"], [725, 727, ["SR", "task"], "Task"], [732, 734, ["Bayesian", "approach"], "Method"], [741, 743, ["estimating", "motion"], "Task"], [751, 753, ["downscaling", "process"], "Task"], [757, 760, ["neural", "network", "techniques"], "Method"], [780, 781, ["SR"], "Task"], [790, 792, ["CNN", "architecture"], "Method"], [793, 796, ["single", "image", "SR"], "Task"], [811, 812, ["VSRnet"], "Method"], [830, 836, ["TV", "-", "based", "optical", "flow", "algorithm"], "Method"], [839, 840, ["accuracy"], "Metric"], [841, 844, ["Joint", "motion", "compensation"], "Task"], [845, 846, ["SR"], "Task"], [847, 849, ["neural", "networks"], "Method"], [854, 857, ["recurrent", "bidirectional", "networks"], "Method"], [862, 865, ["CNN", "based", "approaches"], "Method"], [873, 875, ["bicubic", "interpolation"], "Method"], [880, 882, ["SR", "problem"], "Task"], [890, 891, ["interpolation"], "Method"], [894, 896, ["computational", "burden"], "Metric"], [897, 899, ["CNN", "processing"], "Task"], [911, 917, ["efficient", "sub", "-", "pixel", "convolution", "network"], "Method"], [918, 920, ["ESPCN", ")"], "Method"], [923, 925, ["upscaling", "operation"], "Method"], [941, 942, ["runtime"], "Metric"], [949, 954, ["real", "-", "time", "video", "SR"], "Task"], [960, 963, ["single", "frame", "model"], "Method"], [975, 977, ["transposed", "convolutions"], "Method"], [998, 1001, ["HD", "video", "SR"], "Task"], [1013, 1016, ["real", "-", "time"], "Task"], [1022, 1023, ["ESPCN"], "Method"], [1024, 1028, ["sub", "-", "pixel", "convolution"], "Method"], [1059, 1060, ["VSRnet"], "Method"], [1068, 1070, ["reconstruction", "quality"], "Metric"], [1080, 1084, ["preprocessing", "of", "LR", "images"], "Task"], [1085, 1087, ["bicubic", "upscaling"], "Method"], [1093, 1096, ["motion", "compensation", "mechanism"], "Method"], [1097, 1098, ["runtime"], "Metric"], [1112, 1115, ["Spatial", "transformer", "networks"], "Method"], [1123, 1125, ["spatial", "mapping"], "Task"], [1131, 1133, ["differentiable", "networks"], "Method"], [1154, 1157, ["spatial", "transformer", "networks"], "Method"], [1162, 1164, ["image", "classification"], "Task"], [1181, 1183, ["spatial", "transformers"], "Method"], [1189, 1191, ["unsupervised", "training"], "Method"], [1210, 1212, ["view", "synthesis"], "Task"], [1213, 1215, ["rigid", "transformations"], "Task"], [1225, 1229, ["sub", "-", "pixel", "convolution"], "Method"], [1233, 1237, ["spatio", "-", "temporal", "networks"], "Method"], [1238, 1240, ["motion", "compensation"], "Method"], [1259, 1261, ["early", "fusion"], "Method"], [1262, 1264, ["slow", "fusion"], "Method"], [1265, 1267, ["3D", "convolutions"], "Method"], [1275, 1276, ["classification"], "Task"], [1286, 1289, ["motion", "compensation", "scheme"], "Method"], [1291, 1293, ["spatial", "transformers"], "Method"], [1298, 1302, ["spatio", "-", "temporal", "models"], "Method"], [1313, 1315, ["motion", "compensation"], "Task"], [1350, 1354, ["real", "-", "time", "approach"], "Method"], [1359, 1363, ["sub", "-", "pixel", "convolution"], "Method"], [1364, 1368, ["spatio", "-", "temporal", "networks"], "Method"], [1370, 1371, ["accuracy"], "Metric"], [1372, 1374, ["temporal", "consistency"], "Metric"], [1376, 1378, ["early", "fusion"], "Method"], [1379, 1381, ["slow", "fusion"], "Method"], [1382, 1384, ["3D", "convolutions"], "Method"], [1388, 1393, ["discovering", "spatio", "-", "temporal", "correlations"], "Task"], [1399, 1405, ["dense", "inter", "-", "frame", "motion", "compensation"], "Task"], [1408, 1414, ["multi", "-", "scale", "spatial", "transformer", "network"], "Method"], [1418, 1421, ["motion", "compensation", "technique"], "Method"], [1422, 1426, ["spatio", "-", "temporal", "models"], "Method"], [1482, 1484, ["training", "networks"], "Method"], [1496, 1498, ["image", "processing"], "Task"], [1521, 1526, ["Sub", "-", "pixel", "convolution", "SR"], "Method"], [1539, 1543, ["low", "-", "pass", "filtering"], "Method"], [1544, 1545, ["downscaling"], "Method"], [1553, 1558, ["CNN", "super", "-", "resolved", "solution"], "Method"], [1578, 1580, ["convolutional", "network"], "Method"], [1585, 1588, ["concatenation", "of", "layers"], "Method"], [1646, 1648, ["Model", "parameters"], "Method"], [1665, 1666, ["MSE"], "Metric"], [1670, 1672, ["bicubic", "upsampling"], "Method"], [1688, 1692, ["sub", "-", "pixel", "convolution"], "Method"], [1703, 1705, ["output", "filters"], "Method"], [1727, 1729, ["upscaling", "operation"], "Method"], [1737, 1739, ["bicubic", "upsampling"], "Method"], [1753, 1755, ["convolutional", "processing"], "Method"], [1769, 1773, ["Spatio", "-", "temporal", "networks"], "Task"], [1773, 1777, ["Spatio", "-", "temporal", "networks"], "Method"], [1869, 1872, ["image", "-", "sr"], "Task"], [1876, 1878, ["weighting", "filters"], "Method"], [1922, 1924, ["computational", "efficiency"], "Metric"], [1934, 1935, ["reconstruction"], "Task"], [1944, 1946, ["Early", "fusion"], "Method"], [1954, 1955, ["CNN"], "Method"], [1995, 1998, ["image", "SR", "network"], "Method"], [2004, 2006, ["early", "fusion"], "Task"], [2011, 2014, ["early", "-", "fusion"], "Task"], [2043, 2045, ["action", "recognition"], "Task"], [2055, 2056, ["VSRnet"], "Method"], [2059, 2060, ["VSRnet"], "Method"], [2061, 2063, ["bicubic", "upsampling"], "Method"], [2066, 2070, ["sub", "-", "pixel", "convolution"], "Method"], [2083, 2085, ["Slow", "fusion"], "Method"], [2153, 2155, ["slow", "fusion"], "Method"], [2161, 2163, ["early", "fusion"], "Method"], [2170, 2173, ["slow", "-", "fusion"], "Method"], [2176, 2179, ["slow", "fusion", "network"], "Method"], [2182, 2185, ["rate", "of", "fusion"], "Metric"], [2217, 2219, ["early", "fusion"], "Method"], [2224, 2226, ["slow", "fusion"], "Method"], [2229, 2231, ["3D", "convolutions"], "Method"], [2234, 2236, ["slow", "fusion"], "Method"], [2256, 2260, ["online", "processing", "of", "frames"], "Task"], [2290, 2293, ["slow", "-", "fusion"], "Method"], [2330, 2332, ["3D", "convolutions"], "Method"], [2360, 2363, ["3D", "convolution", "perspective"], "Task"], [2383, 2386, ["slow", "-", "fusion"], "Method"], [2389, 2393, ["Spatial", "transformer", "motion", "compensation"], "Task"], [2400, 2403, ["spatial", "transformer", "network"], "Method"], [2412, 2414, ["SR", "network"], "Method"], [2436, 2438, ["motion", "compensation"], "Task"], [2450, 2452, ["compensation", "module"], "Method"], [2454, 2456, ["SR", "network"], "Method"], [2469, 2471, ["motion", "compensation"], "Task"], [2515, 2518, ["optical", "flow", "representation"], "Method"], [2562, 2564, ["bilinear", "interpolation"], "Method"], [2572, 2577, ["thin", "-", "plate", "spline", "interpolation"], "Method"], [2581, 2583, ["Optical", "flow"], "Task"], [2618, 2622, ["multi", "-", "scale", "design"], "Method"], [2643, 2646, ["spatial", "transformer", "techniques"], "Method"], [2657, 2658, ["transformer"], "Method"], [2659, 2662, ["flow", "estimation", "modules"], "Method"], [2667, 2668, ["transformer"], "Method"], [2672, 2677, ["coarse", "estimate", "of", "the", "flow"], "Task"], [2691, 2693, ["strided", "convolutions"], "Method"], [2700, 2704, ["sub", "-", "pixel", "convolution"], "Method"], [2733, 2737, ["fine", "flow", "estimation", "module"], "Method"], [2742, 2744, ["strided", "convolution"], "Method"], [2745, 2746, ["stride"], "Method"], [2749, 2751, ["upscaling", "stage"], "Method"], [2778, 2779, ["tanh"], "Method"], [2808, 2810, ["spatial", "transformer"], "Method"], [2812, 2814, ["motion", "compensation"], "Task"], [2821, 2822, ["MSE"], "Metric"], [2834, 2837, ["optical", "flow", "methods"], "Method"], [2898, 2901, ["spatial", "transformer", "module"], "Method"], [2906, 2909, ["motion", "compensation", "mechanisms"], "Method"], [2917, 2919, ["SR", "network"], "Method"], [2921, 2924, ["joint", "motion", "compensation"], "Task"], [2955, 2957, ["spatial", "transformer"], "Method"], [2958, 2960, ["SR", "modules"], "Method"], [2984, 2986, ["composite", "loss"], "Metric"], [2988, 2989, ["accuracy"], "Metric"], [2991, 2992, ["reconstruction"], "Task"], [2995, 3000, ["image", "-", "sr", "-", "objective"], "Task"], [3002, 3006, ["fidelity", "of", "motion", "compensation"], "Task"], [3009, 3014, ["motion", "-", "compensation", "-", "objective"], "Task"], [3028, 3032, ["spatio", "-", "temporal", "networks"], "Method"], [3044, 3046, ["motion", "compensation"], "Task"], [3065, 3066, ["compression"], "Method"], [3098, 3101, ["network", "resource", "allocation"], "Task"], [3131, 3132, ["complexity"], "Metric"], [3133, 3136, ["slow", "fusion", "networks"], "Method"], [3137, 3139, ["weight", "sharing"], "Method"], [3167, 3169, ["VSRnet", "variants"], "Task"], [3236, 3237, ["validation"], "Task"], [3241, 3243, ["network", "architecture"], "Method"], [3258, 3261, ["single", "frame", "networks"], "Method"], [3278, 3282, ["spatio", "-", "temporal", "networks"], "Method"], [3327, 3329, ["SR", "methods"], "Method"], [3345, 3346, ["Vid4"], "Material"], [3349, 3351, ["Network", "training"], "Method"], [3354, 3356, ["SR", "models"], "Method"], [3379, 3382, ["rectified", "linear", "units"], "Method"], [3401, 3403, ["orthogonal", "initialisation"], "Method"], [3444, 3446, ["stochastic", "optimisation"], "Task"], [3449, 3450, ["Adam"], "Method"], [3452, 3454, ["learning", "rate"], "Metric"], [3527, 3530, ["slow", "fusion", "networks"], "Method"], [3544, 3547, ["slow", "-", "fusion"], "Method"], [3559, 3561, ["layer", "network"], "Method"], [3562, 3563, ["SR"], "Task"], [3578, 3583, ["Spatio", "-", "temporal", "video", "SR"], "Task"], [3587, 3591, ["multi", "frame", "early", "fusion"], "Method"], [3604, 3605, ["complexity"], "Metric"], [3606, 3607, ["accuracy"], "Metric"], [3608, 3610, ["motion", "compensation"], "Task"], [3613, 3616, ["single", "frame", "models"], "Method"], [3617, 3618, ["SF"], "Method"], [3620, 3626, ["early", "fusion", "spatio", "-", "temporal", "models"], "Method"], [3642, 3643, ["PSNR"], "Metric"], [3689, 3690, ["complexity"], "Metric"], [3691, 3693, ["early", "fusion"], "Method"], [3705, 3706, ["operations"], "Metric"], [3709, 3710, ["accuracy"], "Metric"], [3711, 3715, ["spatio", "-", "temporal", "models"], "Method"], [3769, 3772, ["single", "frame", "networks"], "Method"], [3779, 3783, ["spatio", "-", "temporal", "models"], "Method"], [3802, 3807, ["ssec", ":", "st", "-", "networks"], "Method"], [3817, 3820, ["slow", "fusion", "models"], "Method"], [3828, 3831, ["st", "-", "networks"], "Method"], [3846, 3849, ["single", "frame", "networks"], "Method"], [3851, 3854, ["frame", "input", "networks"], "Method"], [3855, 3857, ["early", "fusion"], "Method"], [3858, 3860, ["slow", "fusion"], "Method"], [3862, 3864, ["slow", "fusion"], "Method"], [3877, 3880, ["slow", "-", "fusion"], "Method"], [3885, 3887, ["layer", "networks"], "Method"], [3893, 3895, ["fusion", "networks"], "Method"], [3898, 3899, ["accuracy"], "Metric"], [3906, 3907, ["operations"], "Metric"], [3910, 3913, ["single", "frame", "models"], "Method"], [3918, 3921, ["slow", "fusion", "architectures"], "Method"], [3925, 3927, ["Slow", "fusion"], "Method"], [3930, 3932, ["early", "fusion"], "Method"], [3948, 3950, ["slow", "fusion"], "Method"], [3967, 3969, ["early", "fusion"], "Method"], [3972, 3975, ["7", "layer", "network"], "Method"], [3980, 3981, ["accuracy"], "Metric"], [3982, 3984, ["slow", "fusion"], "Method"], [3986, 3988, ["early", "fusion"], "Method"], [3990, 3993, ["9", "layer", "network"], "Method"], [3998, 3999, ["accuracy"], "Metric"], [4002, 4003, ["complexity"], "Metric"], [4004, 4006, ["slow", "fusion"], "Method"], [4015, 4017, ["shallow", "networks"], "Method"], [4027, 4030, ["full", "network", "capacity"], "Method"], [4039, 4041, ["early", "fusion"], "Method"], [4045, 4047, ["deeper", "networks"], "Method"], [4070, 4072, ["weight", "sharing"], "Method"], [4073, 4074, ["accuracy"], "Metric"], [4100, 4106, ["7", "layer", "S5", "-", "SW", "network"], "Method"], [4120, 4121, ["accuracy"], "Metric"], [4123, 4124, ["SF"], "Method"], [4138, 4141, ["S5", "-", "SW"], "Method"], [4154, 4156, ["slow", "fusion"], "Method"], [4158, 4160, ["slow", "fusion"], "Method"], [4167, 4168, ["efficiency"], "Metric"], [4172, 4173, ["Convolutions"], "Method"], [4213, 4218, ["high", "-", "level", "vision", "task"], "Task"], [4223, 4225, ["SR", "problem"], "Task"], [4250, 4253, ["frame", "motion", "compensation"], "Method"], [4257, 4263, ["early", "fusion", "network", "of", "temporal", "depth"], "Method"], [4267, 4270, ["motion", "compensation", "module"], "Method"], [4323, 4325, ["flow", "map"], "Task"], [4343, 4346, ["motion", "compensation", "module"], "Method"], [4353, 4359, ["early", "fusion", "motion", "compensated", "SR", "network"], "Method"], [4360, 4363, ["E3", "-", "MC"], "Method"], [4368, 4369, ["compensation"], "Method"], [4371, 4373, ["SR", "network"], "Method"], [4378, 4380, ["full", "model"], "Method"], [4398, 4399, ["SR"], "Task"], [4400, 4401, ["CDVL"], "Task"], [4418, 4421, ["SF", ")", "model"], "Method"], [4422, 4424, ["early", "fusion"], "Method"], [4425, 4427, ["motion", "compensation"], "Method"], [4431, 4434, ["E3", "-", "MC"], "Method"], [4437, 4438, ["PSNR"], "Metric"], [4446, 4447, ["E3"], "Metric"], [4484, 4486, ["structure", "preservation"], "Task"], [4507, 4508, ["Vid4"], "Material"], [4509, 4510, ["SRCNN"], "Method"], [4511, 4512, ["ESPCN"], "Method"], [4513, 4514, ["VSRnet"], "Method"], [4524, 4526, ["video", "ESPCN"], "Method"], [4527, 4528, ["VESPCN"], "Method"], [4537, 4538, ["quality"], "Metric"], [4541, 4544, ["early", "fusion", "models"], "Method"], [4546, 4551, ["5", "layer", "3", "frame", "network"], "Method"], [4554, 4555, ["E3"], "Method"], [4558, 4563, ["9", "layer", "3", "frame", "network"], "Method"], [4564, 4566, ["motion", "compensation"], "Method"], [4578, 4579, ["PSNR"], "Metric"], [4580, 4581, ["SSIM"], "Metric"], [4582, 4583, ["MOVIE"], "Metric"], [4586, 4588, ["MOVIE", "index"], "Metric"], [4621, 4625, ["CNN", "-", "based", "approaches"], "Method"], [4632, 4633, ["Reconstructions"], "Method"], [4634, 4635, ["SRCNN"], "Method"], [4636, 4637, ["ESPCN"], "Method"], [4638, 4639, ["VSRnet"], "Method"], [4646, 4647, ["SRCNN"], "Method"], [4648, 4649, ["ESPCN"], "Method"], [4650, 4651, ["VESPCN"], "Method"], [4660, 4661, ["VSRnet"], "Method"], [4686, 4687, ["VSRnet"], "Method"], [4690, 4692, ["Quality", "comparison"], "Metric"], [4707, 4710, ["motion", "compensated", "network"], "Method"], [4734, 4736, ["VESPCN", "method"], "Method"], [4752, 4753, ["PSNR"], "Metric"], [4754, 4755, ["SSIM"], "Metric"], [4789, 4790, ["reconstruction"], "Method"], [4797, 4798, ["VESPCN"], "Method"], [4805, 4807, ["MOVIE", "index"], "Metric"], [4810, 4812, ["Efficiency", "comparison"], "Metric"], [4813, 4814, ["complexity"], "Metric"], [4823, 4824, ["network"], "Method"], [4829, 4830, ["SRCNN"], "Method"], [4831, 4832, ["VSRnet"], "Method"], [4852, 4853, ["VSRnet"], "Method"], [4867, 4871, ["sub", "-", "pixel", "convolution"], "Method"], [4879, 4880, ["ESPCN"], "Method"], [4881, 4882, ["VESPCN"], "Method"], [4887, 4888, ["ESPCN"], "Method"], [4902, 4906, ["spatio", "-", "temporal", "networks"], "Method"], [4913, 4914, ["VESPCN"], "Method"], [4916, 4917, ["ESPCN"], "Method"], [4921, 4922, ["accuracy"], "Metric"], [4928, 4929, ["VESPCN"], "Method"], [4946, 4947, ["ESPCN"], "Method"], [4955, 4957, ["quality", "metrics"], "Metric"], [4961, 4963, ["motion", "compensation"], "Task"], [4964, 4965, ["VESPCN"], "Method"], [4994, 4996, ["motion", "compensation"], "Task"], [5015, 5017, ["motion", "compensation"], "Task"], [5018, 5019, ["VSRnet"], "Task"], [5030, 5032, ["computational", "bottleneck"], "Metric"], [5046, 5047, ["VESPCN"], "Method"], [5064, 5067, ["optical", "flow", "method"], "Method"], [5068, 5069, ["VSRnet"], "Method"], [5095, 5097, ["motion", "compensation"], "Task"], [5121, 5125, ["sub", "-", "pixel", "convolutions"], "Method"], [5126, 5129, ["temporal", "fusion", "strategies"], "Method"], [5131, 5138, ["real", "-", "time", "spatio", "-", "temporal", "models"], "Method"], [5143, 5147, ["spatio", "-", "temporal", "models"], "Method"], [5155, 5157, ["reconstruction", "accuracy"], "Metric"], [5158, 5160, ["temporal", "consistency"], "Metric"], [5162, 5164, ["computational", "complexity"], "Metric"], [5167, 5170, ["single", "frame", "processing"], "Method"], [5178, 5181, ["motion", "compensation", "mechanism"], "Method"], [5183, 5186, ["spatial", "transformer", "networks"], "Method"], [5202, 5205, ["explicit", "motion", "compensation"], "Task"], [5213, 5214, ["PSNR"], "Metric"], [5215, 5217, ["temporal", "consistency"], "Metric"], [5219, 5223, ["spatio", "-", "temporal", "models"], "Method"], [107, 108, ["video"], "Task"], [119, 120, ["video"], "Task"], [259, 260, ["video"], "Task"], [556, 557, ["video"], "Task"], [627, 628, ["video"], "Task"], [665, 666, ["video"], "Task"], [807, 808, ["video"], "Task"], [1201, 1202, ["video"], "Task"], [1246, 1247, ["video"], "Task"], [1310, 1311, ["video"], "Task"], [1355, 1356, ["video"], "Task"], [1439, 1440, ["video"], "Task"], [1457, 1458, ["ESPCN"], "Method"], [2040, 2041, ["video"], "Task"], [2164, 2165, ["video"], "Task"], [2349, 2350, ["video"], "Task"], [2925, 2926, ["video"], "Task"], [3033, 3034, ["video"], "Task"], [3058, 3059, ["video"], "Task"], [3224, 3225, ["video"], "Task"], [4065, 4066, ["video"], "Task"], [4185, 4186, ["video"], "Task"], [4230, 4231, ["video"], "Task"], [4276, 4277, ["video"], "Task"], [4387, 4388, ["video"], "Task"], [4410, 4411, ["video"], "Task"], [4594, 4595, ["video"], "Task"], [4606, 4608, ["temporal", "consistency"], "Metric"], [4727, 4728, ["video"], "Task"], [5003, 5004, ["video"], "Task"], [5139, 5140, ["video"], "Task"], [5193, 5194, ["video"], "Task"], [5234, 5235, ["video"], "Task"]]} {"doc_id": "0834e74304b547c9354b6d7da6fa78ef47a48fa8", "text": ["document", ":", "LINE", ":", "Large", "-", "scale", "Information", "Network", "Embedding", "This", "paper", "studies", "the", "problem", "of", "embedding", "very", "large", "information", "networks", "into", "low", "-", "dimensional", "vector", "spaces", ",", "which", "is", "useful", "in", "many", "tasks", "such", "as", "visualization", ",", "node", "classification", ",", "and", "link", "prediction", ".", "Most", "existing", "graph", "embedding", "methods", "do", "not", "scale", "for", "real", "world", "information", "networks", "which", "usually", "contain", "millions", "of", "nodes", ".", "In", "this", "paper", ",", "we", "propose", "a", "novel", "network", "embedding", "method", "called", "the", "“", "LINE", ",", "”", "which", "is", "suitable", "for", "arbitrary", "types", "of", "information", "networks", ":", "undirected", ",", "directed", ",", "and", "/", "or", "weighted", ".", "The", "method", "optimizes", "a", "carefully", "designed", "objective", "function", "that", "preserves", "both", "the", "local", "and", "global", "network", "structures", ".", "An", "edge", "-", "sampling", "algorithm", "is", "proposed", "that", "addresses", "the", "limitation", "of", "the", "classical", "stochastic", "gradient", "descent", "and", "improves", "both", "the", "effectiveness", "and", "the", "efficiency", "of", "the", "inference", ".", "Empirical", "experiments", "prove", "the", "effectiveness", "of", "the", "LINE", "on", "a", "variety", "of", "real", "-", "world", "information", "networks", ",", "including", "language", "networks", ",", "social", "networks", ",", "and", "citation", "networks", ".", "The", "algorithm", "is", "very", "efficient", ",", "which", "is", "able", "to", "learn", "the", "embedding", "of", "a", "network", "with", "millions", "of", "vertices", "and", "billions", "of", "edges", "in", "a", "few", "hours", "on", "a", "typical", "single", "machine", ".", "The", "source", "code", "of", "the", "LINE", "is", "available", "online", ".", "CopyrightisheldbytheInternationalWorldWideWebConferenceCommittee", "(", "IW3C2", ")", ".IW3C2reservestherighttoprovideahyperlinktotheauthor’ssiteiftheMaterialisusedinelectronicmedia", ".", "WWW2015", ",", "May18–22", ",", "2015", ",", "Florence", ",", "Italy", ".", "ACM", "978", "-", "1", "-", "4503", "-", "3469", "-", "3", "/", "15", "/", "05", ".", "http:", "//", "dx.doi.org", "/", "10.1145", "/", "2736277.2741093", "I.2.6Artificial", "IntelligenceLearning", "Algorithms", ",", "Experimentation", "section", ":", "Introduction", "Information", "networks", "are", "ubiquitous", "in", "the", "real", "world", "with", "examples", "such", "as", "airline", "networks", ",", "publication", "networks", ",", "social", "and", "communication", "networks", ",", "and", "the", "World", "Wide", "Web", ".", "The", "size", "of", "these", "information", "networks", "ranges", "from", "hundreds", "of", "nodes", "to", "millions", "and", "billions", "of", "nodes", ".", "Analyzing", "large", "information", "networks", "has", "been", "attracting", "increasing", "attention", "in", "both", "academia", "and", "industry", ".", "This", "paper", "studies", "the", "problem", "of", "embedding", "information", "networks", "into", "low", "-", "dimensional", "spaces", ",", "in", "which", "every", "vertex", "is", "represented", "as", "a", "low", "-", "dimensional", "vector", ".", "Such", "a", "low", "-", "dimensional", "embedding", "is", "very", "useful", "in", "a", "variety", "of", "applications", "such", "as", "visualization", ",", "node", "classification", ",", "link", "prediction", ",", "and", "recommendation", ".", "Various", "methods", "of", "graph", "embedding", "have", "been", "proposed", "in", "the", "machine", "learning", "literature", "(", "e.g.", ",", ")", ".", "They", "generally", "perform", "well", "on", "smaller", "networks", ".", "The", "problem", "becomes", "much", "more", "challenging", "when", "a", "real", "world", "information", "network", "is", "concerned", ",", "which", "typically", "contains", "millions", "of", "nodes", "and", "billions", "of", "edges", ".", "For", "example", ",", "the", "Twitter", "followee", "-", "follower", "network", "contains", "175", "million", "active", "users", "and", "around", "twenty", "billion", "edges", "in", "2012", ".", "Most", "existing", "graph", "embedding", "algorithms", "do", "not", "scale", "for", "networks", "of", "this", "size", ".", "For", "example", ",", "the", "time", "complexity", "of", "classical", "graph", "embedding", "algorithms", "such", "as", "MDS", ",", "IsoMap", ",", "Laplacian", "eigenmap", "are", "at", "least", "quadratic", "to", "the", "number", "of", "vertices", ",", "which", "is", "too", "expensive", "for", "networks", "with", "millions", "of", "nodes", ".", "Although", "a", "few", "very", "recent", "studies", "approach", "the", "embedding", "of", "large", "-", "scale", "networks", ",", "these", "methods", "either", "use", "an", "indirect", "approach", "that", "is", "not", "designed", "for", "networks", "(", "e.g.", ",", ")", "or", "lack", "a", "clear", "objective", "function", "tailored", "for", "network", "embedding", "(", "e.g.", ",", ")", ".", "We", "anticipate", "that", "a", "new", "model", "with", "a", "carefully", "designed", "objective", "function", "that", "preserves", "properties", "of", "the", "graph", "and", "an", "efficient", "optimization", "technique", "should", "effectively", "find", "the", "embedding", "of", "millions", "of", "nodes", ".", "In", "this", "paper", ",", "we", "propose", "such", "a", "network", "embedding", "model", "called", "the", "“", "LINE", ",", "”", "which", "is", "able", "to", "scale", "to", "very", "large", ",", "arbitrary", "types", "of", "networks", ":", "undirected", ",", "directed", "and", "/", "or", "weighted", ".", "The", "model", "optimizes", "an", "objective", "which", "preserves", "both", "the", "local", "and", "global", "network", "structures", ".", "Naturally", ",", "the", "local", "structures", "are", "represented", "by", "the", "observed", "links", "in", "the", "networks", ",", "which", "capture", "the", "first", "-", "order", "proximity", "between", "the", "vertices", ".", "Most", "existing", "graph", "embedding", "algorithms", "are", "designed", "to", "preserve", "this", "first", "-", "order", "proximity", ",", "e.g.", ",", "IsoMap", "and", "Laplacian", "eigenmap", ",", "even", "if", "they", "do", "not", "scale", ".", "We", "observe", "that", "in", "a", "real", "-", "world", "network", "many", "(", "if", "not", "the", "majority", "of", ")", "legitimate", "links", "are", "actually", "not", "observed", ".", "In", "other", "words", ",", "the", "observed", "first", "-", "order", "proximity", "in", "the", "real", "world", "data", "is", "not", "sufficient", "for", "preserving", "the", "global", "network", "structures", ".", "As", "a", "complement", ",", "we", "explore", "the", "second", "-", "order", "proximity", "between", "the", "vertices", ",", "which", "is", "not", "determined", "through", "the", "observed", "tie", "strength", "but", "through", "the", "shared", "neighborhood", "structures", "of", "the", "vertices", ".", "The", "general", "notion", "of", "the", "second", "-", "order", "proximity", "can", "be", "interpreted", "as", "nodes", "with", "shared", "neighbors", "being", "likely", "to", "be", "similar", ".", "Such", "an", "intuition", "can", "be", "found", "in", "the", "theories", "of", "sociology", "and", "linguistics", ".", "For", "example", ",", "“", "the", "degree", "of", "overlap", "of", "two", "people", "’s", "friendship", "networks", "correlates", "with", "the", "strength", "of", "ties", "between", "them", ",", "”", "in", "a", "social", "network", ";", "and", "“", "You", "shall", "know", "a", "word", "by", "the", "company", "it", "keeps", "”", "(", "Firth", ",", "J.", "R.", "1957:11", ")", "in", "text", "corpora", ".", "Indeed", ",", "people", "who", "share", "many", "common", "friends", "are", "likely", "to", "share", "the", "same", "interest", "and", "become", "friends", ",", "and", "words", "that", "are", "used", "together", "with", "many", "similar", "words", "are", "likely", "to", "have", "similar", "meanings", ".", "Fig", ".", "[", "reference", "]", "presents", "an", "illustrative", "example", ".", "As", "the", "weight", "of", "the", "edge", "between", "vertex", "6", "and", "7", "is", "large", ",", "i.e.", ",", "6", "and", "7", "have", "a", "high", "first", "-", "order", "proximity", ",", "they", "should", "be", "represented", "closely", "to", "each", "other", "in", "the", "embedded", "space", ".", "On", "the", "other", "hand", ",", "though", "there", "is", "no", "link", "between", "vertex", "5", "and", "6", ",", "they", "share", "many", "common", "neighbors", ",", "i.e.", ",", "they", "have", "a", "high", "second", "-", "order", "proximity", "and", "therefore", "should", "also", "be", "represented", "closely", "to", "each", "other", ".", "We", "expect", "that", "the", "consideration", "of", "the", "second", "-", "order", "proximity", "effectively", "complements", "the", "sparsity", "of", "the", "first", "-", "order", "proximity", "and", "better", "preserves", "the", "global", "structure", "of", "the", "network", ".", "In", "this", "paper", ",", "we", "will", "present", "carefully", "designed", "objectives", "that", "preserve", "the", "first", "-", "order", "and", "the", "second", "-", "order", "proximities", ".", "Even", "if", "a", "sound", "objective", "is", "found", ",", "optimizing", "it", "for", "a", "very", "large", "network", "is", "challenging", ".", "One", "approach", "that", "attracts", "attention", "in", "recent", "years", "is", "using", "the", "stochastic", "gradient", "descent", "for", "the", "optimization", ".", "However", ",", "we", "show", "that", "directly", "deploying", "the", "stochastic", "gradient", "descent", "is", "problematic", "for", "real", "world", "information", "networks", ".", "This", "is", "because", "in", "many", "networks", ",", "edges", "are", "weighted", "and", "the", "weights", "usually", "present", "a", "high", "variance", ".", "Consider", "a", "word", "co", "-", "occurrence", "network", ",", "in", "which", "the", "weights", "(", "co", "-", "occurrences", ")", "of", "word", "pairs", "may", "range", "from", "one", "to", "hundreds", "of", "thousands", ".", "These", "weights", "of", "the", "edges", "will", "be", "multiplied", "into", "the", "gradients", ",", "resulting", "in", "the", "explosion", "of", "the", "gradients", "and", "thus", "compromise", "the", "performance", ".", "To", "address", "this", ",", "we", "propose", "a", "novel", "edge", "-", "sampling", "method", ",", "which", "improves", "both", "the", "effectiveness", "and", "efficiency", "of", "the", "inference", ".", "We", "sample", "the", "edges", "with", "the", "probabilities", "proportional", "to", "their", "weights", ",", "and", "then", "treat", "the", "sampled", "edges", "as", "binary", "edges", "for", "model", "updating", ".", "With", "this", "sampling", "process", ",", "the", "objective", "function", "remains", "the", "same", "and", "the", "weights", "of", "the", "edges", "no", "longer", "affect", "the", "gradients", ".", "The", "LINE", "is", "very", "general", ",", "which", "works", "well", "for", "directed", "or", "undirected", ",", "weighted", "or", "unweighted", "graphs", ".", "We", "evaluate", "the", "performance", "of", "the", "LINE", "with", "various", "real", "-", "world", "information", "networks", ",", "including", "language", "networks", ",", "social", "networks", ",", "and", "citation", "networks", ".", "The", "effectiveness", "of", "the", "learned", "embeddings", "is", "evaluated", "within", "multiple", "data", "mining", "tasks", ",", "including", "word", "analogy", ",", "text", "classification", ",", "and", "node", "classification", ".", "The", "results", "suggest", "that", "the", "LINE", "model", "outperforms", "other", "competitive", "baselines", "in", "terms", "of", "both", "effectiveness", "and", "efficiency", ".", "It", "is", "able", "to", "learn", "the", "embedding", "of", "a", "network", "with", "millions", "of", "nodes", "and", "billions", "of", "edges", "in", "a", "few", "hours", "on", "a", "single", "machine", ".", "To", "summarize", ",", "we", "make", "the", "following", "contributions", ":", "We", "propose", "a", "novel", "network", "embedding", "model", "called", "the", "“", "LINE", ",", "”", "which", "suits", "arbitrary", "types", "of", "information", "networks", "and", "easily", "scales", "to", "millions", "of", "nodes", ".", "It", "has", "a", "carefully", "designed", "objective", "function", "that", "preserves", "both", "the", "first", "-", "order", "and", "second", "-", "order", "proximities", ".", "We", "propose", "an", "edge", "-", "sampling", "algorithm", "for", "optimizing", "the", "objective", ".", "The", "algorithm", "tackles", "the", "limitation", "of", "the", "classical", "stochastic", "gradient", "decent", "and", "improves", "the", "effectiveness", "and", "efficiency", "of", "the", "inference", ".", "We", "conduct", "extensive", "experiments", "on", "real", "-", "world", "information", "networks", ".", "Experimental", "results", "prove", "the", "effectiveness", "and", "efficiency", "of", "the", "proposed", "LINE", "model", ".", "Organization", ".", "The", "rest", "of", "this", "paper", "is", "organized", "as", "follows", ".", "Section", "[", "reference", "]", "summarizes", "the", "related", "work", ".", "Section", "[", "reference", "]", "formally", "defines", "the", "problem", "of", "large", "-", "scale", "information", "network", "embedding", ".", "Section", "[", "reference", "]", "introduces", "the", "LINE", "model", "in", "details", ".", "Section", "[", "reference", "]", "presents", "the", "experimental", "results", ".", "Finally", "we", "conclude", "in", "Section", "[", "reference", "]", ".", "section", ":", "Related", "Work", "Our", "work", "is", "related", "to", "classical", "methods", "of", "graph", "embedding", "or", "dimension", "reduction", "in", "general", ",", "such", "as", "multidimensional", "scaling", "(", "MDS", ")", ",", "IsoMap", ",", "LLE", "and", "Laplacian", "Eigenmap", ".", "These", "approaches", "typically", "first", "construct", "the", "affinity", "graph", "using", "the", "feature", "vectors", "of", "the", "data", "points", ",", "e.g.", ",", "the", "K", "-", "nearest", "neighbor", "graph", "of", "data", ",", "and", "then", "embed", "the", "affinity", "graph", "into", "a", "low", "dimensional", "space", ".", "However", ",", "these", "algorithms", "usually", "rely", "on", "solving", "the", "leading", "eigenvectors", "of", "the", "affinity", "matrices", ",", "the", "complexity", "of", "which", "is", "at", "least", "quadratic", "to", "the", "number", "of", "nodes", ",", "making", "them", "inefficient", "to", "handle", "large", "-", "scale", "networks", ".", "Among", "the", "most", "recent", "literature", "is", "a", "technique", "called", "graph", "factorization", ".", "It", "finds", "the", "low", "-", "dimensional", "embedding", "of", "a", "large", "graph", "through", "matrix", "factorization", ",", "which", "is", "optimized", "using", "stochastic", "gradient", "descent", ".", "This", "is", "possible", "because", "a", "graph", "can", "be", "represented", "as", "an", "affinity", "matrix", ".", "However", ",", "the", "objective", "of", "matrix", "factorization", "is", "not", "designed", "for", "networks", ",", "therefore", "does", "not", "necessarily", "preserve", "the", "global", "network", "structure", ".", "Intuitively", ",", "graph", "factorization", "expects", "nodes", "with", "higher", "first", "-", "order", "proximity", "are", "represented", "closely", ".", "Instead", ",", "the", "LINE", "model", "uses", "an", "objective", "that", "is", "particularly", "designed", "for", "networks", ",", "which", "preserves", "both", "the", "first", "-", "order", "and", "the", "second", "-", "order", "proximities", ".", "Practically", ",", "the", "graph", "factorization", "method", "only", "applies", "to", "undirected", "graphs", "while", "the", "proposed", "model", "is", "applicable", "for", "both", "undirected", "and", "directed", "graphs", ".", "The", "most", "recent", "work", "related", "with", "ours", "is", "DeepWalk", ",", "which", "deploys", "a", "truncated", "random", "walk", "for", "social", "network", "embedding", ".", "Although", "empirically", "effective", ",", "the", "DeepWalk", "does", "not", "provide", "a", "clear", "objective", "that", "articulates", "what", "network", "properties", "are", "preserved", ".", "Intuitively", ",", "DeepWalk", "expects", "nodes", "with", "higher", "second", "-", "order", "proximity", "yield", "similar", "low", "-", "dimensional", "representations", ",", "while", "the", "LINE", "preserves", "both", "first", "-", "order", "and", "second", "-", "order", "proximities", ".", "DeepWalk", "uses", "random", "walks", "to", "expand", "the", "neighborhood", "of", "a", "vertex", ",", "which", "is", "analogical", "to", "a", "depth", "-", "first", "search", ".", "We", "use", "a", "breadth", "-", "first", "search", "strategy", ",", "which", "is", "a", "more", "reasonable", "approach", "to", "the", "second", "-", "order", "proximity", ".", "Practically", ",", "DeepWalk", "only", "applies", "to", "unweighted", "networks", ",", "while", "our", "model", "is", "applicable", "for", "networks", "with", "both", "weighted", "and", "unweighted", "edges", ".", "In", "Section", "[", "reference", "]", ",", "we", "empirically", "compare", "the", "proposed", "model", "with", "these", "methods", "using", "various", "real", "world", "networks", ".", "section", ":", "Problem", "Definition", "We", "formally", "define", "the", "problem", "of", "large", "-", "scale", "information", "network", "embedding", "using", "first", "-", "order", "and", "second", "-", "order", "proximities", ".", "We", "first", "define", "an", "information", "network", "as", "follows", ":", "theorem", ":", "(", "Information", "Network", ")", "An", "information", "network", "is", "defined", "as", "=", "G", "(", "V", ",", "E", ")", ",", "where", "V", "is", "the", "set", "of", "vertices", ",", "each", "representing", "a", "data", "object", "and", "E", "is", "the", "set", "of", "edges", "between", "the", "vertices", ",", "each", "representing", "a", "relationship", "between", "two", "data", "objects", ".", "Each", "edge", "∈eE", "is", "an", "ordered", "pair", "=", "e", "(", "u", ",", "v", ")", "and", "is", "associated", "with", "a", "weight", ">", "w⁢uv0", ",", "which", "indicates", "the", "strength", "of", "the", "relation", ".", "If", "G", "is", "undirected", ",", "we", "have", "≡", "(", "u", ",", "v", ")(", "v", ",", "u", ")", "and", "≡w⁢uvw⁢vu", ";", "if", "G", "is", "directed", ",", "we", "have", "≢", "(", "u", ",", "v", ")(", "v", ",", "u", ")", "and", "≢w⁢uvw⁢vu", ".", "In", "practice", ",", "information", "networks", "can", "be", "either", "directed", "(", "e.g.", ",", "citation", "networks", ")", "or", "undirected", "(", "e.g.", ",", "social", "network", "of", "users", "in", "Facebook", ")", ".", "The", "weights", "of", "the", "edges", "can", "be", "either", "binary", "or", "take", "any", "real", "value", ".", "Note", "that", "while", "negative", "edge", "weights", "are", "possible", ",", "in", "this", "study", "we", "only", "consider", "non", "-", "negative", "weights", ".", "For", "example", ",", "in", "citation", "networks", "and", "social", "networks", ",", "takes", "binary", "values", ";", "in", "co", "-", "occurrence", "networks", "between", "different", "objects", ",", "can", "take", "any", "non", "-", "negative", "value", ".", "The", "weights", "of", "the", "edges", "in", "some", "networks", "may", "diverge", "as", "some", "objects", "co", "-", "occur", "many", "times", "while", "others", "may", "just", "co", "-", "occur", "a", "few", "times", ".", "Embedding", "an", "information", "network", "into", "a", "low", "-", "dimensional", "space", "is", "useful", "in", "a", "variety", "of", "applications", ".", "To", "conduct", "the", "embedding", ",", "the", "network", "structures", "must", "be", "preserved", ".", "The", "first", "intuition", "is", "that", "the", "local", "network", "structure", ",", "i.e.", ",", "the", "local", "pairwise", "proximity", "between", "the", "vertices", ",", "must", "be", "preserved", ".", "We", "define", "the", "local", "network", "structures", "as", "the", "first", "-", "order", "proximity", "between", "the", "vertices", ":", "theorem", ":", "(", "First", "-", "order", "Proximity", ")", "The", "first", "-", "order", "proximity", "in", "a", "network", "is", "the", "local", "pairwise", "proximity", "between", "two", "vertices", ".", "For", "each", "pair", "of", "vertices", "linked", "by", "an", "edge", "(", "u", ",", "v", ")", ",", "the", "weight", "on", "that", "edge", ",", "w⁢uv", ",", "indicates", "the", "first", "-", "order", "proximity", "between", "u", "and", "v.", "If", "no", "edge", "is", "observed", "between", "u", "and", "v", ",", "their", "first", "-", "order", "proximity", "is", "0", ".", "The", "first", "-", "order", "proximity", "usually", "implies", "the", "similarity", "of", "two", "nodes", "in", "a", "real", "-", "world", "network", ".", "For", "example", ",", "people", "who", "are", "friends", "with", "each", "other", "in", "a", "social", "network", "tend", "to", "share", "similar", "interests", ";", "pages", "linking", "to", "each", "other", "in", "World", "Wide", "Web", "tend", "to", "talk", "about", "similar", "topics", ".", "Because", "of", "this", "importance", ",", "many", "existing", "graph", "embedding", "algorithms", "such", "as", "IsoMap", ",", "LLE", ",", "Laplacian", "eigenmap", ",", "and", "graph", "factorization", "have", "the", "objective", "to", "preserve", "the", "first", "-", "order", "proximity", ".", "However", ",", "in", "a", "real", "world", "information", "network", ",", "the", "links", "observed", "are", "only", "a", "small", "proportion", ",", "with", "many", "others", "missing", ".", "A", "pair", "of", "nodes", "on", "a", "missing", "link", "has", "a", "zero", "first", "-", "order", "proximity", ",", "even", "though", "they", "are", "intrinsically", "very", "similar", "to", "each", "other", ".", "Therefore", ",", "first", "-", "order", "proximity", "alone", "is", "not", "sufficient", "for", "preserving", "the", "network", "structures", ",", "and", "it", "is", "important", "to", "seek", "an", "alternative", "notion", "of", "proximity", "that", "addresses", "the", "problem", "of", "sparsity", ".", "A", "natural", "intuition", "is", "that", "vertices", "that", "share", "similar", "neighbors", "tend", "to", "be", "similar", "to", "each", "other", ".", "For", "example", ",", "in", "social", "networks", ",", "people", "who", "share", "similar", "friends", "tend", "to", "have", "similar", "interests", "and", "thus", "become", "friends", ";", "in", "word", "co", "-", "occurrence", "networks", ",", "words", "that", "always", "co", "-", "occur", "with", "the", "same", "set", "of", "words", "tend", "to", "have", "similar", "meanings", ".", "We", "therefore", "define", "the", "second", "-", "order", "proximity", ",", "which", "complements", "the", "first", "-", "order", "proximity", "and", "preserves", "the", "network", "structure", ".", "theorem", ":", "(", "Second", "-", "order", "Proximity", ")", "The", "second", "-", "order", "proximity", "between", "a", "pair", "of", "vertices", "(", "u", ",", "v", ")", "in", "a", "network", "is", "the", "similarity", "between", "their", "neighborhood", "network", "structures", ".", "Mathematically", ",", "let", "=", "pu", "(", "wu", ",", "1", ",", "…", ",", "wu", ",", "|V|", ")", "denote", "the", "first", "-", "order", "proximity", "of", "u", "with", "all", "the", "other", "vertices", ",", "then", "the", "second", "-", "order", "proximity", "between", "u", "and", "v", "is", "determined", "by", "the", "similarity", "between", "pu", "and", "pv", ".", "If", "no", "vertex", "is", "linked", "from", "/", "to", "both", "u", "and", "v", ",", "the", "second", "-", "order", "proximity", "between", "u", "and", "v", "is", "0", ".", "We", "investigate", "both", "first", "-", "order", "and", "second", "-", "order", "proximity", "for", "network", "embedding", ",", "which", "is", "defined", "as", "follows", ".", "theorem", ":", "(", "Large", "-", "scale", "Information", "Network", "Embedding", ")", "Given", "a", "large", "network", "=", "G", "(", "V", ",", "E", ")", ",", "the", "problem", "of", "Large", "-", "scale", "Information", "Network", "Embedding", "aims", "to", "represent", "each", "vertex", "∈vV", "into", "a", "low", "-", "dimensional", "space", "Rd", ",", "i.e.", ",", "learning", "a", "function", ":", "fG→VRd", ",", "where", "≪d|V|.", "In", "the", "space", "Rd", ",", "both", "the", "first", "-", "order", "proximity", "and", "the", "second", "-", "order", "proximity", "between", "the", "vertices", "are", "preserved", ".", "Next", ",", "we", "introduce", "a", "large", "-", "scale", "network", "embedding", "model", "that", "preserves", "both", "first", "-", "and", "second", "-", "order", "proximities", ".", "section", ":", "LINE", ":", "Large", "-", "scale", "Information", "Network", "Embedding", "A", "desirable", "embedding", "model", "for", "real", "world", "information", "networks", "must", "satisfy", "several", "requirements", ":", "first", ",", "it", "must", "be", "able", "to", "preserve", "both", "the", "first", "-", "order", "proximity", "and", "the", "second", "-", "order", "proximity", "between", "the", "vertices", ";", "second", ",", "it", "must", "scale", "for", "very", "large", "networks", ",", "say", "millions", "of", "vertices", "and", "billions", "of", "edges", ";", "third", ",", "it", "can", "deal", "with", "networks", "with", "arbitrary", "types", "of", "edges", ":", "directed", ",", "undirected", "and", "/", "or", "weighted", ".", "In", "this", "section", ",", "we", "present", "a", "novel", "network", "embedding", "model", "called", "the", "“", "LINE", ",", "”", "which", "satisfies", "all", "the", "three", "requirements", ".", "subsection", ":", "Model", "Description", "We", "describe", "the", "LINE", "model", "to", "preserve", "the", "first", "-", "order", "proximity", "and", "second", "-", "order", "proximity", "separately", ",", "and", "then", "introduce", "a", "simple", "way", "to", "combine", "the", "two", "proximity", ".", "subsubsection", ":", "LINE", "with", "First", "-", "order", "Proximity", "The", "first", "-", "order", "proximity", "refers", "to", "the", "local", "pairwise", "proximity", "between", "the", "vertices", "in", "the", "network", ".", "To", "model", "the", "first", "-", "order", "proximity", ",", "for", "each", "undirected", "edge", ",", "we", "define", "the", "joint", "probability", "between", "vertex", "and", "as", "follows", ":", "where", "is", "the", "low", "-", "dimensional", "vector", "representation", "of", "vertex", ".", "Eqn", ".", "(", "[", "reference", "]", ")", "defines", "a", "distribution", "over", "the", "space", ",", "and", "its", "empirical", "probability", "can", "be", "defined", "as", ",", "where", ".", "To", "preserve", "the", "first", "-", "order", "proximity", ",", "a", "straightforward", "way", "is", "to", "minimize", "the", "following", "objective", "function", ":", "where", "is", "the", "distance", "between", "two", "distributions", ".", "We", "choose", "to", "minimize", "the", "KL", "-", "divergence", "of", "two", "probability", "distributions", ".", "Replacing", "with", "KL", "-", "divergence", "and", "omitting", "some", "constants", ",", "we", "have", ":", "Note", "that", "the", "first", "-", "order", "proximity", "is", "only", "applicable", "for", "undirected", "graphs", ",", "not", "for", "directed", "graphs", ".", "By", "finding", "the", "that", "minimize", "the", "objective", "in", "Eqn", ".", "(", "[", "reference", "]", ")", ",", "we", "can", "represent", "every", "vertex", "in", "the", "d", "-", "dimensional", "space", ".", "subsubsection", ":", "LINE", "with", "Second", "-", "order", "Proximity", "The", "second", "-", "order", "proximity", "is", "applicable", "for", "both", "directed", "and", "undirected", "graphs", ".", "Given", "a", "network", ",", "without", "loss", "of", "generality", ",", "we", "assume", "it", "is", "directed", "(", "an", "undirected", "edge", "can", "be", "considered", "as", "two", "directed", "edges", "with", "opposite", "directions", "and", "equal", "weights", ")", ".", "The", "second", "-", "order", "proximity", "assumes", "that", "vertices", "sharing", "many", "connections", "to", "other", "vertices", "are", "similar", "to", "each", "other", ".", "In", "this", "case", ",", "each", "vertex", "is", "also", "treated", "as", "a", "specific", "“", "context", "”", "and", "vertices", "with", "similar", "distributions", "over", "the", "“", "contexts", "”", "are", "assumed", "to", "be", "similar", ".", "Therefore", ",", "each", "vertex", "plays", "two", "roles", ":", "the", "vertex", "itself", "and", "a", "specific", "“", "context", "”", "of", "other", "vertices", ".", "We", "introduce", "two", "vectors", "and", ",", "where", "is", "the", "representation", "of", "when", "it", "is", "treated", "as", "a", "vertex", "while", "is", "the", "representation", "of", "when", "it", "is", "treated", "as", "a", "specific", "“", "context", "”", ".", "For", "each", "directed", "edge", ",", "we", "first", "define", "the", "probability", "of", "“", "context", "”", "generated", "by", "vertex", "as", ":", "where", "is", "the", "number", "of", "vertices", "or", "“", "contexts", ".", "”", "For", "each", "vertex", ",", "Eqn", ".", "(", "[", "reference", "]", ")", "actually", "defines", "a", "conditional", "distribution", "over", "the", "contexts", ",", "i.e.", ",", "the", "entire", "set", "of", "vertices", "in", "the", "network", ".", "As", "mentioned", "above", ",", "the", "second", "-", "order", "proximity", "assumes", "that", "vertices", "with", "similar", "distributions", "over", "the", "contexts", "are", "similar", "to", "each", "other", ".", "To", "preserve", "the", "second", "-", "order", "proximity", ",", "we", "should", "make", "the", "conditional", "distribution", "of", "the", "contexts", "specified", "by", "the", "low", "-", "dimensional", "representation", "be", "close", "to", "the", "empirical", "distribution", ".", "Therefore", ",", "we", "minimize", "the", "following", "objective", "function", ":", "where", "is", "the", "distance", "between", "two", "distributions", ".", "As", "the", "importance", "of", "the", "vertices", "in", "the", "network", "may", "be", "different", ",", "we", "introduce", "in", "the", "objective", "function", "to", "represent", "the", "prestige", "of", "vertex", "in", "the", "network", ",", "which", "can", "be", "measured", "by", "the", "degree", "or", "estimated", "through", "algorithms", "such", "as", "PageRank", ".", "The", "empirical", "distribution", "is", "defined", "as", ",", "where", "is", "the", "weight", "of", "the", "edge", "and", "is", "the", "out", "-", "degree", "of", "vertex", ",", "i.e.", ",", "where", "is", "the", "set", "of", "out", "-", "neighbors", "of", ".", "In", "this", "paper", ",", "for", "simplicity", "we", "set", "as", "the", "degree", "of", "vertex", ",", "i.e.", ",", ",", "and", "here", "we", "also", "adopt", "KL", "-", "divergence", "as", "the", "distance", "function", ".", "Replacing", "with", "KL", "-", "divergence", ",", "setting", "and", "omitting", "some", "constants", ",", "we", "have", ":", "By", "learning", "and", "that", "minimize", "this", "objective", ",", "we", "are", "able", "to", "represent", "every", "vertex", "with", "a", "d", "-", "dimensional", "vector", ".", "subsubsection", ":", "Combining", "first", "-", "order", "and", "second", "-", "order", "proximities", "To", "embed", "the", "networks", "by", "preserving", "both", "the", "first", "-", "order", "and", "second", "-", "order", "proximity", ",", "a", "simple", "and", "effective", "way", "we", "find", "in", "practice", "is", "to", "train", "the", "LINE", "model", "which", "preserves", "the", "first", "-", "order", "proximity", "and", "second", "-", "order", "proximity", "separately", "and", "then", "concatenate", "the", "embeddings", "trained", "by", "the", "two", "methods", "for", "each", "vertex", ".", "A", "more", "principled", "way", "to", "combine", "the", "two", "proximity", "is", "to", "jointly", "train", "the", "objective", "function", "(", "[", "reference", "]", ")", "and", "(", "[", "reference", "]", ")", ",", "which", "we", "leave", "as", "future", "work", ".", "subsection", ":", "Model", "Optimization", "Optimizing", "objective", "(", "[", "reference", "]", ")", "is", "computationally", "expensive", ",", "which", "requires", "the", "summation", "over", "the", "entire", "set", "of", "vertices", "when", "calculating", "the", "conditional", "probability", ".", "To", "address", "this", "problem", ",", "we", "adopt", "the", "approach", "of", "negative", "sampling", "proposed", "in", ",", "which", "samples", "multiple", "negative", "edges", "according", "to", "some", "noisy", "distribution", "for", "each", "edge", ".", "More", "specifically", ",", "it", "specifies", "the", "following", "objective", "function", "for", "each", "edge", ":", "where", "is", "the", "sigmoid", "function", ".", "The", "first", "term", "models", "the", "observed", "edges", ",", "the", "second", "term", "models", "the", "negative", "edges", "drawn", "from", "the", "noise", "distribution", "and", "is", "the", "number", "of", "negative", "edges", ".", "We", "set", "as", "proposed", "in", ",", "where", "is", "the", "out", "-", "degree", "of", "vertex", ".", "For", "the", "objective", "function", "(", "[", "reference", "]", ")", ",", "there", "exists", "a", "trivial", "solution", ":", ",", "for", "i=", "and", ".", "To", "avoid", "the", "trivial", "solution", ",", "we", "can", "still", "utilize", "the", "negative", "sampling", "approach", "(", "[", "reference", "]", ")", "by", "just", "changing", "to", ".", "We", "adopt", "the", "asynchronous", "stochastic", "gradient", "algorithm", "(", "ASGD", ")", "for", "optimizing", "Eqn", ".", "(", "[", "reference", "]", ")", ".", "In", "each", "step", ",", "the", "ASGD", "algorithm", "samples", "a", "mini", "-", "batch", "of", "edges", "and", "then", "updates", "the", "model", "parameters", ".", "If", "an", "edge", "is", "sampled", ",", "the", "gradient", "w.r.t", ".", "the", "embedding", "vector", "of", "vertex", "will", "be", "calculated", "as", ":", "Note", "that", "the", "gradient", "will", "be", "multiplied", "by", "the", "weight", "of", "the", "edge", ".", "This", "will", "become", "problematic", "when", "the", "weights", "of", "edges", "have", "a", "high", "variance", ".", "For", "example", ",", "in", "a", "word", "co", "-", "occurrence", "network", ",", "some", "words", "co", "-", "occur", "many", "times", "(", "e.g.", ",", "tens", "of", "thousands", ")", "while", "some", "words", "co", "-", "occur", "only", "a", "few", "times", ".", "In", "such", "networks", ",", "the", "scales", "of", "the", "gradients", "diverge", "and", "it", "is", "very", "hard", "to", "find", "a", "good", "learning", "rate", ".", "If", "we", "select", "a", "large", "learning", "rate", "according", "to", "the", "edges", "with", "small", "weights", ",", "the", "gradients", "on", "edges", "with", "large", "weights", "will", "explode", "while", "the", "gradients", "will", "become", "too", "small", "if", "we", "select", "the", "learning", "rate", "according", "to", "the", "edges", "with", "large", "weights", ".", "subsubsection", ":", "Optimization", "via", "Edge", "Sampling", "The", "intuition", "in", "solving", "the", "above", "problem", "is", "that", "if", "the", "weights", "of", "all", "the", "edges", "are", "equal", "(", "e.g.", ",", "network", "with", "binary", "edges", ")", ",", "then", "there", "will", "be", "no", "problem", "of", "choosing", "an", "appropriate", "learning", "rate", ".", "A", "simple", "treatment", "is", "thus", "to", "unfold", "a", "weighted", "edge", "into", "multiple", "binary", "edges", ",", "e.g.", ",", "an", "edge", "with", "weight", "is", "unfolded", "into", "binary", "edges", ".", "This", "will", "solve", "the", "problem", "but", "will", "significantly", "increase", "the", "memory", "requirement", ",", "especially", "when", "the", "weights", "of", "the", "edges", "are", "very", "large", ".", "To", "resolve", "this", ",", "one", "can", "sample", "from", "the", "original", "edges", "and", "treat", "the", "sampled", "edges", "as", "binary", "edges", ",", "with", "the", "sampling", "probabilities", "proportional", "to", "the", "original", "edge", "weights", ".", "With", "this", "edge", "-", "sampling", "treatment", ",", "the", "overall", "objective", "function", "remains", "the", "same", ".", "The", "problem", "boils", "down", "to", "how", "to", "sample", "the", "edges", "according", "to", "their", "weights", ".", "Let", "denote", "the", "sequence", "of", "the", "weights", "of", "the", "edges", ".", "One", "can", "simply", "calculate", "the", "sum", "of", "the", "weights", "first", ",", "and", "then", "to", "sample", "a", "random", "value", "within", "the", "range", "of", "to", "see", "which", "interval", "[", "the", "random", "value", "falls", "into", ".", "This", "approach", "takes", "time", "to", "draw", "a", "sample", ",", "which", "is", "costly", "when", "the", "number", "of", "edges", "is", "large", ".", "We", "use", "the", "alias", "table", "method", "to", "draw", "a", "sample", "according", "to", "the", "weights", "of", "the", "edges", ",", "which", "takes", "only", "time", "when", "repeatedly", "drawing", "samples", "from", "the", "same", "discrete", "distribution", ".", "Sampling", "an", "edge", "from", "the", "alias", "table", "takes", "constant", "time", ",", ",", "and", "optimization", "with", "negative", "sampling", "takes", "time", ",", "where", "is", "the", "number", "of", "negative", "samples", ".", "Therefore", ",", "overall", "each", "step", "takes", "time", ".", "In", "practice", ",", "we", "find", "that", "the", "number", "of", "steps", "used", "for", "optimization", "is", "usually", "proportional", "to", "the", "number", "of", "edges", ".", "Therefore", ",", "the", "overall", "time", "complexity", "of", "the", "LINE", "is", ",", "which", "is", "linear", "to", "the", "number", "of", "edges", ",", "and", "does", "not", "depend", "on", "the", "number", "of", "vertices", ".", "The", "edge", "sampling", "treatment", "improves", "the", "effectiveness", "of", "the", "stochastic", "gradient", "descent", "without", "compromising", "the", "efficiency", ".", "subsection", ":", "Discussion", "We", "discuss", "several", "practical", "issues", "of", "the", "LINE", "model", ".", "Low", "degree", "vertices", ".", "One", "practical", "issue", "is", "how", "to", "accurately", "embed", "vertices", "with", "small", "degrees", ".", "As", "the", "number", "of", "neighbors", "of", "such", "a", "node", "is", "very", "small", ",", "it", "is", "very", "hard", "to", "accurately", "infer", "its", "representation", ",", "especially", "with", "the", "second", "-", "order", "proximity", "based", "methods", "which", "heavily", "rely", "on", "the", "number", "of", "“", "contexts", ".", "”", "An", "intuitive", "solution", "to", "this", "is", "expanding", "the", "neighbors", "of", "those", "vertices", "by", "adding", "higher", "order", "neighbors", ",", "such", "as", "neighbors", "of", "neighbors", ".", "In", "this", "paper", ",", "we", "only", "consider", "adding", "second", "-", "order", "neighbors", ",", "i.e.", ",", "neighbors", "of", "neighbors", ",", "to", "each", "vertex", ".", "The", "weight", "between", "vertex", "and", "its", "second", "-", "order", "neighbor", "is", "measured", "as", "In", "practice", ",", "one", "can", "only", "add", "a", "subset", "of", "vertices", "which", "have", "the", "largest", "proximity", "with", "the", "low", "degree", "vertex", ".", "New", "vertices", ".", "Another", "practical", "issue", "is", "how", "to", "find", "the", "representation", "of", "newly", "arrived", "vertices", ".", "For", "a", "new", "vertex", ",", "if", "its", "connections", "to", "the", "existing", "vertices", "are", "known", ",", "we", "can", "obtain", "the", "empirical", "distribution", "and", "over", "existing", "vertices", ".", "To", "obtain", "the", "embedding", "of", "the", "new", "vertex", ",", "according", "to", "the", "objective", "function", "Eqn", ".", "(", "[", "reference", "]", ")", "or", "Eqn", ".", "(", "[", "reference", "]", ")", ",", "a", "straightforward", "way", "is", "to", "minimize", "either", "one", "of", "the", "following", "objective", "functions", "by", "updating", "the", "embedding", "of", "the", "new", "vertex", "and", "keeping", "the", "embeddings", "of", "existing", "vertices", ".", "If", "no", "connections", "between", "the", "new", "vertex", "and", "existing", "vertices", "are", "observed", ",", "we", "must", "resort", "to", "other", "information", ",", "such", "as", "the", "textual", "information", "of", "the", "vertices", ",", "and", "we", "leave", "it", "as", "our", "future", "work", ".", "section", ":", "Experiments", "We", "empirically", "evaluated", "the", "effectiveness", "and", "efficiency", "of", "the", "LINE", ".", "We", "applied", "the", "method", "to", "several", "large", "-", "scale", "real", "-", "world", "networks", "of", "different", "types", ",", "including", "a", "language", "network", ",", "two", "social", "networks", ",", "and", "two", "citation", "networks", ".", "subsection", ":", "Experiment", "Setup", "paragraph", ":", "Data", "Sets", "(", "1", ")", "Language", "network", ".", "We", "constructed", "a", "word", "co", "-", "occurrence", "network", "from", "the", "entire", "set", "of", "English", "Wikipedia", "pages", ".", "Words", "within", "every", "5", "-", "word", "sliding", "window", "are", "considered", "to", "be", "co", "-", "occurring", "with", "each", "other", ".", "Words", "with", "frequency", "smaller", "than", "5", "are", "filtered", "out", ".", "(", "2", ")", "Social", "networks", ".", "We", "use", "two", "social", "networks", ":", "Flickr", "and", "YoutubeAvailable", "at", "http:", "//", "socialnetworks.mpi", "-", "sws.org", "/", "data", "-", "imc2007.html", ".", "The", "Flickr", "network", "is", "denser", "than", "the", "Youtube", "network", "(", "the", "same", "network", "as", "used", "in", "DeepWalk", ")", ".", "(", "3", ")", "Citation", "Networks", ".", "Two", "types", "of", "citation", "networks", "are", "used", ":", "an", "author", "citation", "network", "and", "a", "paper", "citation", "network", ".", "We", "use", "the", "DBLP", "data", "set", "to", "construct", "the", "citation", "networks", "between", "authors", "and", "between", "papers", ".", "The", "author", "citation", "network", "records", "the", "number", "of", "papers", "written", "by", "one", "author", "and", "cited", "by", "another", "author", ".", "The", "detailed", "statistics", "of", "these", "networks", "are", "summarized", "into", "Table", "[", "reference", "]", ".", "They", "represent", "a", "variety", "of", "information", "networks", ":", "directed", "and", "undirected", ",", "binary", "and", "weighted", ".", "Each", "network", "contains", "at", "least", "half", "a", "million", "nodes", "and", "millions", "of", "edges", ",", "with", "the", "largest", "network", "containing", "around", "two", "million", "nodes", "and", "a", "billion", "edges", ".", "paragraph", ":", "Compared", "Algorithms", "We", "compare", "the", "LINE", "model", "with", "several", "existing", "graph", "embedding", "methods", "that", "are", "able", "to", "scale", "up", "to", "very", "large", "networks", ".", "We", "do", "not", "compare", "with", "some", "classical", "graph", "embedding", "algorithms", "such", "as", "MDS", ",", "IsoMap", ",", "and", "Laplacian", "eigenmap", ",", "as", "they", "can", "not", "handle", "networks", "of", "this", "scale", ".", "Graph", "factorization", "(", "GF", ")", ".", "We", "compare", "with", "the", "matrix", "factorization", "techniques", "for", "graph", "factorization", ".", "An", "information", "network", "can", "be", "represented", "as", "an", "affinity", "matrix", ",", "and", "is", "able", "to", "represent", "each", "vertex", "with", "a", "low", "-", "dimensional", "vector", "through", "matrix", "factorization", ".", "Graph", "factorization", "is", "optimized", "through", "stochastic", "gradient", "descent", "and", "is", "able", "to", "handle", "large", "networks", ".", "It", "only", "applies", "to", "undirected", "networks", ".", "DeepWalk", ".", "DeepWalk", "is", "an", "approach", "recently", "proposed", "for", "social", "network", "embedding", ",", "which", "is", "only", "applicable", "for", "networks", "with", "binary", "edges", ".", "For", "each", "vertex", ",", "truncated", "random", "walks", "starting", "from", "the", "vertex", "are", "used", "to", "obtain", "the", "contextual", "information", ",", "and", "therefore", "only", "second", "-", "order", "proximity", "is", "utilized", ".", "LINE", "-", "SGD", ".", "This", "is", "the", "LINE", "model", "introduced", "in", "Section", "[", "reference", "]", "that", "optimizes", "the", "objective", "Eqn", ".", "(", "[", "reference", "]", ")", "or", "Eqn", ".", "(", "[", "reference", "]", ")", "directly", "with", "stochastic", "gradient", "descent", ".", "With", "this", "approach", ",", "the", "weights", "of", "the", "edges", "are", "directly", "multiplied", "into", "the", "gradients", "when", "the", "edges", "are", "sampled", "for", "model", "updating", ".", "There", "are", "two", "variants", "of", "this", "approach", ":", "LINE", "-", "SGD", "(", "1st", ")", "and", "LINE", "-", "SGD", "(", "2nd", ")", ",", "which", "use", "first", "-", "and", "second", "-", "order", "proximity", "respectively", ".", "LINE", ".", "This", "is", "the", "LINE", "model", "optimized", "through", "the", "edge", "-", "sampling", "treatment", "introduced", "in", "Section", "[", "reference", "]", ".", "In", "each", "stochastic", "gradient", "step", ",", "an", "edge", "is", "sampled", "with", "the", "probability", "proportional", "to", "its", "weight", "and", "then", "treated", "as", "binary", "for", "model", "updating", ".", "There", "are", "also", "two", "variants", ":", "LINE", "(", "1st", ")", "and", "LINE", "(", "2nd", ")", ".", "Like", "the", "graph", "factorization", ",", "both", "LINE", "(", "1st", ")", "and", "LINE", "-", "SGD", "(", "1st", ")", "only", "apply", "to", "undirected", "graphs", ".", "LINE", "(", "2nd", ")", "and", "LINE", "-", "SGD", "(", "2nd", ")", "apply", "to", "both", "undirected", "and", "directed", "graphs", ".", "LINE", "(", "1st", "+", "2nd", ")", ":", "To", "utilize", "both", "first", "-", "order", "and", "second", "-", "order", "proximity", ",", "a", "simple", "and", "effective", "way", "is", "to", "concatenate", "the", "vector", "representations", "learned", "by", "LINE", "(", "1st", ")", "and", "LINE", "(", "2nd", ")", "into", "a", "longer", "vector", ".", "After", "concatenation", ",", "the", "dimensions", "should", "be", "re", "-", "weighted", "to", "balance", "the", "two", "representations", ".", "In", "a", "supervised", "learning", "task", ",", "the", "weighting", "of", "dimensions", "can", "be", "automatically", "found", "based", "on", "the", "training", "data", ".", "In", "an", "unsupervised", "task", ",", "however", ",", "it", "is", "more", "difficult", "to", "set", "the", "weights", ".", "Therefore", "we", "only", "apply", "LINE", "(", "1st", "+", "2nd", ")", "to", "the", "scenario", "of", "supervised", "tasks", ".", "paragraph", ":", "Parameter", "Settings", "The", "mini", "-", "batch", "size", "of", "the", "stochastic", "gradient", "descent", "is", "set", "as", "1", "for", "all", "the", "methods", ".", "Similar", "to", ",", "the", "learning", "rate", "is", "set", "with", "the", "starting", "value", "and", ",", "where", "is", "the", "total", "number", "of", "mini", "-", "batches", "or", "edge", "samples", ".", "For", "fair", "comparisons", ",", "the", "dimensionality", "of", "the", "embeddings", "of", "the", "language", "network", "is", "set", "to", "200", ",", "as", "used", "in", "word", "embedding", ".", "For", "other", "networks", ",", "the", "dimension", "is", "set", "as", "128", "by", "default", ",", "as", "used", "in", ".", "Other", "default", "settings", "include", ":", "the", "number", "of", "negative", "samples", "for", "LINE", "and", "LINE", "-", "SGD", ";", "the", "total", "number", "of", "samples", "billion", "for", "LINE", "(", "1st", ")", "and", "LINE", "(", "2nd", ")", ",", "billion", "for", "GF", ";", "window", "size", ",", "walk", "length", ",", "walks", "per", "vertex", "for", "DeepWalk", ".", "All", "the", "embedding", "vectors", "are", "finally", "normalized", "by", "setting", ".", "subsection", ":", "Quantitative", "Results", "subsubsection", ":", "Language", "Network", "We", "start", "with", "the", "results", "on", "the", "language", "network", ",", "which", "contains", "two", "million", "nodes", "and", "a", "billion", "edges", ".", "Two", "applications", "are", "used", "to", "evaluate", "the", "effectiveness", "of", "the", "learned", "embeddings", ":", "word", "analogy", "and", "document", "classification", ".", "Word", "Analogy", ".", "This", "task", "is", "introduced", "by", "Mikolov", "et", "al", ".", ".", "Given", "a", "word", "pair", "and", "a", "word", ",", "the", "task", "aims", "to", "find", "a", "word", ",", "such", "that", "the", "relation", "between", "and", "is", "similar", "to", "the", "relation", "between", "and", ",", "or", "denoted", "as", ":", ".", "For", "instance", ",", "given", "a", "word", "pair", "(", "“", "China”", ",", "“Beijing", "”", ")", "and", "a", "word", "“", "France", ",", "”", "the", "right", "answer", "should", "be", "“", "Paris", "”", "because", "“", "Beijing", "”", "is", "the", "capital", "of", "“", "China", "”", "just", "as", "“", "Paris", "”", "is", "the", "capital", "of", "“", "France", ".", "”", "Given", "the", "word", "embeddings", ",", "this", "task", "is", "solved", "by", "finding", "the", "word", "whose", "embedding", "is", "closest", "to", "the", "vector", "in", "terms", "of", "cosine", "proximity", ",", "i.e.", ",", ".", "Two", "categories", "of", "word", "analogy", "are", "used", "in", "this", "task", ":", "semantic", "and", "syntactic", ".", "Significantly", "outperforms", "GF", "at", "the", ":", "*", "*", "0.01", "and", "*", "0.05", "level", ",", "paired", "t", "-", "test", ".", "Table", "[", "reference", "]", "reports", "the", "results", "of", "word", "analogy", "using", "the", "embeddings", "of", "words", "learned", "on", "the", "Wikipedia", "corpora", "(", "SkipGram", ")", "or", "the", "Wikipedia", "word", "network", "(", "all", "other", "methods", ")", ".", "For", "graph", "factorization", ",", "the", "weight", "between", "each", "pair", "of", "words", "is", "defined", "as", "the", "logarithm", "of", "the", "number", "of", "co", "-", "occurrences", ",", "which", "leads", "to", "better", "performance", "than", "the", "original", "value", "of", "co", "-", "occurrences", ".", "For", "DeepWalk", ",", "different", "cutoff", "thresholds", "are", "tried", "to", "convert", "the", "language", "network", "into", "a", "binary", "network", ",", "and", "the", "best", "performance", "is", "achieved", "when", "all", "the", "edges", "are", "kept", "in", "the", "network", ".", "We", "also", "compare", "with", "the", "state", "-", "of", "-", "the", "-", "art", "word", "embedding", "model", "SkipGram", ",", "which", "learns", "the", "word", "embeddings", "directly", "from", "the", "original", "Wikipedia", "pages", "and", "is", "also", "implicitly", "a", "matrix", "factorization", "approach", ".", "The", "window", "size", "is", "set", "as", "5", ",", "the", "same", "as", "used", "for", "constructing", "the", "language", "network", ".", "We", "can", "see", "that", "LINE", "(", "2nd", ")", "outperforms", "all", "other", "methods", ",", "including", "the", "graph", "embedding", "methods", "and", "the", "SkipGram", ".", "This", "indicates", "that", "the", "second", "-", "order", "proximity", "better", "captures", "the", "word", "semantics", "compared", "to", "the", "first", "-", "order", "proximity", ".", "This", "is", "not", "surprising", ",", "as", "a", "high", "second", "-", "order", "proximity", "implies", "that", "two", "words", "can", "be", "replaced", "in", "the", "same", "context", ",", "which", "is", "a", "stronger", "indicator", "of", "similar", "semantics", "than", "first", "-", "order", "co", "-", "occurrences", ".", "It", "is", "intriguing", "that", "the", "LINE", "(", "2nd", ")", "outperforms", "the", "state", "-", "of", "-", "the", "-", "art", "word", "embedding", "model", "trained", "on", "the", "original", "corpus", ".", "The", "reason", "may", "be", "that", "a", "language", "network", "better", "captures", "the", "global", "structure", "of", "word", "co", "-", "occurrences", "than", "the", "original", "word", "sequences", ".", "Among", "other", "methods", ",", "both", "graph", "factorization", "and", "LINE", "(", "1st", ")", "significantly", "outperform", "DeepWalk", "even", "if", "DeepWalk", "explores", "second", "-", "order", "proximity", ".", "This", "is", "because", "DeepWalk", "has", "to", "ignore", "the", "weights", "(", "i.e.", ",", "co", "-", "occurrences", ")", "of", "the", "edges", ",", "which", "is", "very", "important", "in", "a", "language", "network", ".", "The", "performance", "by", "the", "LINE", "models", "directly", "optimized", "with", "SGD", "is", "much", "worse", ",", "because", "the", "weights", "of", "the", "edges", "in", "the", "language", "network", "diverge", ",", "which", "range", "from", "a", "single", "digit", "to", "tens", "of", "thousands", ",", "making", "the", "learning", "process", "suffer", ".", "The", "LINE", "optimized", "by", "the", "edge", "-", "sampling", "treatment", "effectively", "addresses", "this", "problem", ",", "and", "performs", "very", "well", "using", "either", "first", "-", "order", "or", "second", "-", "order", "proximity", ".", "All", "the", "models", "are", "run", "on", "a", "single", "machine", "with", "1", "T", "memory", ",", "40", "CPU", "cores", "at", "2.0GHZ", "using", "16", "threads", ".", "Both", "the", "LINE", "(", "1st", ")", "and", "LINE", "(", "2nd", ")", "are", "quite", "efficient", ",", "which", "take", "less", "than", "3", "hours", "to", "process", "such", "a", "network", "with", "2", "million", "nodes", "and", "a", "billion", "edges", ".", "Both", "are", "at", "least", "10", "%", "faster", "than", "graph", "factorization", ",", "and", "much", "more", "efficient", "than", "DeepWalk", "(", "five", "times", "slower", ")", ".", "The", "reason", "that", "LINE", "-", "SGDs", "are", "slightly", "slower", "is", "that", "a", "threshold", "-", "cutting", "technique", "has", "to", "be", "applied", "to", "prevent", "the", "gradients", "from", "exploding", ".", "Document", "Classification", ".", "Another", "way", "to", "evaluate", "the", "quality", "of", "the", "word", "embeddings", "is", "to", "use", "the", "word", "vectors", "to", "compute", "document", "representation", ",", "which", "can", "be", "evaluated", "with", "document", "classification", "tasks", ".", "To", "obtain", "document", "vectors", ",", "we", "choose", "a", "very", "simple", "approach", ",", "taking", "the", "average", "of", "the", "word", "vector", "representations", "in", "that", "document", ".", "This", "is", "because", "we", "aim", "to", "compare", "the", "word", "embeddings", "with", "different", "approaches", "instead", "of", "finding", "the", "best", "method", "for", "document", "embeddings", ".", "The", "readers", "can", "find", "advanced", "document", "embedding", "approaches", "in", ".", "We", "download", "the", "abstracts", "of", "Wikipedia", "pages", "from", "and", "the", "categories", "of", "these", "pages", "from", ".", "We", "choose", "7", "diverse", "categories", "for", "classification", "including", "“", "Arts", ",", "”", "“", "History", ",", "”", "“", "Human", ",", "”", "“", "Mathematics", ",", "”", "“", "Nature", ",", "”", "“", "Technology", ",", "”", "and", "“", "Sports", ".", "”", "For", "each", "category", ",", "we", "randomly", "select", "10", ",", "000", "articles", ",", "and", "articles", "belonging", "to", "multiple", "categories", "are", "discarded", ".", "We", "randomly", "sample", "different", "percentages", "of", "the", "labeled", "documents", "for", "training", "and", "use", "the", "rest", "for", "evaluation", ".", "All", "document", "vectors", "are", "used", "to", "train", "a", "one", "-", "vs", "-", "rest", "logistic", "regression", "classifier", "using", "the", "LibLinear", "package", ".", "We", "report", "the", "classification", "metrics", "Micro", "-", "F1", "and", "Macro", "-", "F1", ".", "The", "results", "are", "averaged", "over", "10", "different", "runs", "by", "sampling", "different", "training", "data", ".", "Table", "[", "reference", "]", "reports", "the", "results", "of", "Wikipedia", "page", "classification", ".", "Similar", "conclusion", "can", "be", "made", "as", "in", "the", "word", "analogy", "task", ".", "The", "graph", "factorization", "outperforms", "DeepWalk", "as", "DeepWalk", "ignores", "the", "weights", "of", "the", "edges", ".", "The", "LINE", "-", "SGDs", "perform", "worse", "due", "to", "the", "divergence", "of", "the", "weights", "of", "the", "edges", ".", "The", "LINE", "optimized", "by", "the", "edge", "-", "sampling", "treatment", "performs", "much", "better", "than", "directly", "deploying", "SGD", ".", "The", "LINE", "(", "2nd", ")", "outperforms", "LINE", "(", "1st", ")", "and", "is", "slightly", "better", "than", "the", "graph", "factorization", ".", "Note", "that", "with", "the", "supervised", "task", ",", "it", "is", "feasible", "to", "concatenate", "the", "embeddings", "learned", "with", "LINE", "(", "1st", ")", "and", "LINE", "(", "2nd", ")", ".", "As", "a", "result", ",", "the", "LINE", "(", "1st", "+", "2nd", ")", "method", "performs", "significantly", "better", "than", "all", "other", "methods", ".", "This", "indicates", "that", "the", "first", "-", "order", "and", "second", "-", "order", "proximities", "are", "complementary", "to", "each", "other", ".", "To", "provide", "the", "readers", "more", "insight", "about", "the", "first", "-", "order", "and", "second", "-", "order", "proximities", ",", "Table", "[", "reference", "]", "compares", "the", "most", "similar", "words", "to", "a", "given", "word", "using", "first", "-", "order", "and", "second", "-", "order", "proximity", ".", "We", "can", "see", "that", "by", "using", "the", "contextual", "proximity", ",", "the", "most", "similar", "words", "returned", "by", "the", "second", "-", "order", "proximity", "are", "all", "semantically", "related", "words", ".", "The", "most", "similar", "words", "returned", "by", "the", "first", "-", "order", "proximity", "are", "a", "mixture", "of", "syntactically", "and", "semantically", "related", "words", ".", "subsubsection", ":", "Social", "Network", "Significantly", "outperforms", "DeepWalk", "at", "the", ":", "*", "*", "0.01", "and", "*", "0.05", "level", ",", "paired", "t", "-", "test", ".", "Compared", "with", "the", "language", "networks", ",", "the", "social", "networks", "are", "much", "sparser", ",", "especially", "the", "Youtube", "network", ".", "We", "evaluate", "the", "vertex", "embeddings", "through", "a", "multi", "-", "label", "classification", "task", "that", "assigns", "every", "node", "into", "one", "or", "more", "communities", ".", "Different", "percentages", "of", "the", "vertices", "are", "randomly", "sampled", "for", "training", "and", "the", "rest", "are", "used", "for", "evaluation", ".", "The", "results", "are", "averaged", "over", "10", "different", "runs", ".", "Flickr", "Network", ".", "Let", "us", "first", "take", "a", "look", "at", "the", "results", "on", "the", "Flickr", "network", ".", "We", "choose", "the", "most", "popular", "5", "communities", "as", "the", "categories", "of", "the", "vertices", "for", "multi", "-", "label", "classification", ".", "Table", "[", "reference", "]", "reports", "the", "results", ".", "Again", ",", "LINE", "(", "1st", "+", "2nd", ")", "significantly", "outperforms", "all", "other", "methods", ".", "LINE", "(", "1st", ")", "is", "slightly", "better", "than", "LINE", "(", "2nd", ")", ",", "which", "is", "opposite", "to", "the", "results", "on", "the", "language", "network", ".", "The", "reasons", "are", "two", "fold", ":", "(", "1", ")", "first", "-", "order", "proximity", "is", "still", "more", "important", "than", "second", "-", "order", "proximity", "in", "social", "network", ",", "which", "indicates", "strong", "ties", ";", "(", "2", ")", "when", "the", "network", "is", "too", "sparse", "and", "the", "average", "number", "of", "neighbors", "of", "a", "node", "is", "too", "small", ",", "the", "second", "-", "order", "proximity", "may", "become", "inaccurate", ".", "We", "will", "further", "investigate", "this", "issue", "in", "Section", "[", "reference", "]", ".", "LINE", "(", "1st", ")", "outperforms", "graph", "factorization", ",", "indicating", "a", "better", "capability", "of", "modeling", "the", "first", "-", "order", "proximity", ".", "LINE", "(", "2nd", ")", "outperforms", "DeepWalk", ",", "indicating", "a", "better", "capability", "of", "modeling", "the", "second", "-", "order", "proximity", ".", "By", "concatenating", "the", "representations", "learned", "by", "LINE", "(", "1st", ")", "and", "LINE", "(", "2nd", ")", ",", "the", "performance", "further", "improves", ",", "confirming", "that", "the", "two", "proximities", "are", "complementary", "to", "each", "other", ".", "Significantly", "outperforms", "DeepWalk", "at", "the", ":", "*", "*", "0.01", "and", "*", "0.05", "level", ",", "paired", "t", "-", "test", ".", "Significantly", "outperforms", "DeepWalk", "at", "the", ":", "*", "*", "0.01", "and", "*", "0.05", "level", ",", "paired", "t", "-", "test", ".", "Significantly", "outperforms", "DeepWalk", "at", "the", ":", "*", "*", "0.01", "and", "*", "0.05", "level", ",", "paired", "t", "-", "test", ".", "Youtube", "Network", ".", "Table", "[", "reference", "]", "reports", "the", "results", "on", "Youtube", "network", ",", "which", "is", "extremely", "sparse", "and", "the", "average", "degree", "is", "as", "low", "as", "5", ".", "In", "most", "cases", "with", "different", "percentages", "of", "training", "data", ",", "LINE", "(", "1st", ")", "outperforms", "LINE", "(", "2nd", ")", ",", "consistent", "with", "the", "results", "on", "the", "Flickr", "network", ".", "Due", "to", "the", "extreme", "sparsity", ",", "the", "performance", "of", "LINE", "(", "2nd", ")", "is", "even", "inferior", "to", "DeepWalk", ".", "By", "combining", "the", "representations", "learned", "by", "the", "LINE", "with", "both", "the", "first", "-", "and", "second", "-", "order", "proximity", ",", "the", "performance", "of", "LINE", "outperforms", "DeepWalk", "with", "either", "128", "or", "256", "dimension", ",", "showing", "that", "the", "two", "proximities", "are", "complementary", "to", "each", "other", "and", "able", "to", "address", "the", "problem", "of", "network", "sparsity", ".", "It", "is", "interesting", "to", "observe", "how", "DeepWalk", "tackles", "the", "network", "sparsity", "through", "truncated", "random", "walks", ",", "which", "enrich", "the", "neighbors", "or", "contexts", "of", "each", "vertex", ".", "The", "random", "walk", "approach", "acts", "like", "a", "depth", "-", "first", "search", ".", "Such", "an", "approach", "may", "quickly", "alleviate", "the", "sparsity", "of", "the", "neighborhood", "of", "nodes", "by", "bringing", "in", "indirect", "neighbors", ",", "but", "it", "may", "also", "introduce", "nodes", "that", "are", "long", "range", "away", ".", "A", "more", "reasonable", "way", "is", "to", "expand", "the", "neighborhood", "of", "each", "vertex", "using", "a", "breadth", "-", "first", "search", "strategy", ",", "i.e.", ",", "recursively", "adding", "neighbors", "of", "neighbors", ".", "To", "verify", "this", ",", "we", "expand", "the", "neighborhood", "of", "the", "vertices", "whose", "degree", "are", "less", "than", "1", ",", "000", "by", "adding", "the", "neighbors", "of", "neighbors", "until", "the", "size", "of", "the", "extended", "neighborhood", "reaches", "1", ",", "000", "nodes", ".", "We", "find", "that", "adding", "more", "than", "1", ",", "000", "vertices", "does", "not", "further", "increase", "the", "performance", ".", "The", "results", "in", "the", "brackets", "in", "Table", "[", "reference", "]", "are", "obtained", "on", "this", "reconstructed", "network", ".", "The", "performance", "of", "GF", ",", "LINE", "(", "1st", ")", "and", "LINE", "(", "2nd", ")", "all", "improves", ",", "especially", "LINE", "(", "2nd", ")", ".", "In", "the", "reconstructed", "network", ",", "the", "LINE", "(", "2nd", ")", "outperforms", "DeepWalk", "in", "most", "cases", ".", "We", "can", "also", "see", "that", "the", "performance", "of", "LINE", "(", "1st", "+", "2nd", ")", "on", "the", "reconstructed", "network", "does", "not", "improve", "too", "much", "compared", "with", "those", "on", "the", "original", "network", ".", "This", "implies", "that", "the", "combination", "of", "first", "-", "order", "and", "second", "-", "order", "proximity", "on", "the", "original", "network", "has", "already", "captured", "most", "information", "and", "LINE", "(", "1st", "+", "2nd", ")", "approach", "is", "a", "quite", "effective", "and", "efficient", "way", "for", "network", "embedding", ",", "suitable", "for", "both", "dense", "and", "sparse", "networks", ".", "subsubsection", ":", "Citation", "Network", "We", "present", "the", "results", "on", "two", "citation", "networks", ",", "both", "of", "which", "are", "directed", "networks", ".", "Both", "the", "GF", "and", "LINE", "methods", ",", "which", "use", "first", "-", "order", "proximity", ",", "are", "not", "applicable", "for", "directed", "networks", ",", "and", "hence", "we", "only", "compare", "DeepWalk", "and", "LINE", "(", "2nd", ")", ".", "We", "also", "evaluate", "the", "vertex", "embeddings", "through", "a", "multi", "-", "label", "classification", "task", ".", "We", "choose", "7", "popular", "conferences", "including", "AAAI", ",", "CIKM", ",", "ICML", ",", "KDD", ",", "NIPS", ",", "SIGIR", ",", "and", "WWW", "as", "the", "classification", "categories", ".", "Authors", "publishing", "in", "the", "conferences", "or", "papers", "published", "in", "the", "conferences", "are", "assumed", "to", "belong", "to", "the", "categories", "corresponding", "to", "the", "conferences", ".", "DBLP", "(", "AuthorCitation", ")", "Network", ".", "Table", "[", "reference", "]", "reports", "the", "results", "on", "the", "DBLP", "(", "AuthorCitation", ")", "network", ".", "As", "this", "network", "is", "also", "very", "sparse", ",", "DeepWalk", "outperforms", "LINE", "(", "2nd", ")", ".", "However", ",", "by", "reconstructing", "the", "network", "through", "recursively", "adding", "neighbors", "of", "neighbors", "for", "vertices", "with", "small", "degrees", "(", "smaller", "than", "500", ")", ",", "the", "performance", "of", "LINE", "(", "2nd", ")", "significantly", "increases", "and", "outperforms", "DeepWalk", ".", "The", "LINE", "model", "directly", "optimized", "by", "stochastic", "gradient", "descent", ",", "LINE", "(", "2nd", ")", ",", "does", "not", "perform", "well", "as", "expected", ".", "DBLP", "(", "PaperCitation", ")", "Network", ".", "Table", "[", "reference", "]", "reports", "the", "results", "on", "the", "DBLP", "(", "PaperCitation", ")", "network", ".", "The", "LINE", "(", "2nd", ")", "significantly", "outperforms", "DeepWalk", ".", "This", "is", "because", "the", "random", "walk", "on", "the", "paper", "citation", "network", "can", "only", "reach", "papers", "along", "the", "citing", "path", "(", "i.e.", ",", "older", "papers", ")", "and", "can", "not", "reach", "other", "references", ".", "Instead", ",", "the", "LINE", "(", "2nd", ")", "represents", "each", "paper", "with", "its", "references", ",", "which", "is", "obviously", "more", "reasonable", ".", "The", "performance", "of", "LINE", "(", "2nd", ")", "is", "further", "improved", "when", "the", "network", "is", "reconstructed", "by", "enriching", "the", "neighbors", "of", "vertices", "with", "small", "degrees", "(", "smaller", "than", "200", ")", ".", "subsection", ":", "Network", "Layouts", "An", "important", "application", "of", "network", "embedding", "is", "to", "create", "meaningful", "visualizations", "that", "layout", "a", "network", "on", "a", "two", "dimensional", "space", ".", "We", "visualize", "a", "co", "-", "author", "network", "extracted", "from", "the", "DBLP", "data", ".", "We", "select", "some", "conferences", "from", "three", "different", "research", "fields", ":", "WWW", ",", "KDD", "from", "“", "data", "mining", ",", "”", "NIPS", ",", "ICML", "from", "“", "machine", "learning", ",", "”", "and", "CVPR", ",", "ICCV", "from", "“", "computer", "vision", ".", "”", "The", "co", "-", "author", "network", "is", "built", "from", "the", "papers", "published", "in", "these", "conferences", ".", "Authors", "with", "degree", "less", "than", "3", "are", "filtered", "out", ",", "and", "finally", "the", "network", "contains", "18", ",", "561", "authors", "and", "207", ",", "074", "edges", ".", "Laying", "out", "this", "co", "-", "author", "network", "is", "very", "challenging", "as", "the", "three", "research", "fields", "are", "very", "close", "to", "each", "other", ".", "We", "first", "map", "the", "co", "-", "author", "network", "into", "a", "low", "-", "dimensional", "space", "with", "different", "embedding", "approaches", "and", "then", "further", "map", "the", "low", "-", "dimensional", "vectors", "of", "the", "vertices", "to", "a", "2", "-", "D", "space", "with", "the", "t", "-", "SNE", "package", ".", "Fig", ".", "[", "reference", "]", "compares", "the", "visualization", "results", "with", "different", "embedding", "approaches", ".", "The", "visualization", "using", "graph", "factorization", "is", "not", "very", "meaningful", ",", "in", "which", "the", "authors", "belonging", "to", "the", "same", "communities", "are", "not", "clustered", "together", ".", "The", "result", "of", "DeepWalk", "is", "much", "better", ".", "However", ",", "many", "authors", "belonging", "to", "different", "communities", "are", "clustered", "tightly", "into", "the", "center", "area", ",", "most", "of", "which", "are", "high", "degree", "vertices", ".", "This", "is", "because", "DeepWalk", "uses", "a", "random", "walk", "based", "approach", "to", "enrich", "the", "neighbors", "of", "the", "vertices", ",", "which", "brings", "in", "a", "lot", "of", "noise", "due", "to", "the", "randomness", ",", "especially", "for", "vertices", "with", "higher", "degrees", ".", "The", "LINE", "(", "2nd", ")", "performs", "quite", "well", "and", "generates", "meaningful", "layout", "of", "the", "network", "(", "nodes", "with", "same", "colors", "are", "distributed", "closer", ")", ".", "subsection", ":", "Performance", "w.r.t", ".", "Network", "Sparsity", "In", "this", "subsection", ",", "we", "formally", "analyze", "the", "performance", "of", "the", "above", "models", "w.r.t", ".", "the", "sparsity", "of", "networks", ".", "We", "use", "the", "social", "networks", "as", "examples", ".", "We", "first", "investigate", "how", "the", "sparsity", "of", "the", "networks", "affects", "the", "LINE", "(", "1st", ")", "and", "LINE", "(", "2nd", ")", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "results", "w.r.t", ".", "the", "percentage", "of", "links", "on", "the", "Flickr", "network", ".", "We", "choose", "Flickr", "network", "as", "it", "is", "much", "denser", "than", "the", "Youtube", "network", ".", "We", "randomly", "select", "different", "percentages", "of", "links", "from", "the", "original", "network", "to", "construct", "networks", "with", "different", "levels", "of", "sparsity", ".", "We", "can", "see", "that", "in", "the", "beginning", ",", "when", "the", "network", "is", "very", "sparse", ",", "the", "LINE", "(", "1st", ")", "outperforms", "LINE", "(", "2nd", ")", ".", "As", "we", "gradually", "increase", "the", "percentage", "of", "links", ",", "the", "LINE", "(", "2nd", ")", "begins", "to", "outperform", "the", "LINE", "(", "1st", ")", ".", "This", "shows", "that", "the", "second", "-", "order", "proximity", "suffers", "when", "the", "network", "is", "extremely", "sparse", ",", "and", "it", "outperforms", "first", "-", "order", "proximity", "when", "there", "are", "sufficient", "nodes", "in", "the", "neighborhood", "of", "a", "node", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "performance", "w.r.t", ".", "the", "degrees", "of", "the", "vertices", "on", "both", "the", "original", "and", "reconstructed", "Youtube", "networks", ".", "We", "categorize", "the", "vertices", "into", "different", "groups", "according", "to", "their", "degrees", "including", ",", "and", "then", "evaluate", "the", "performance", "of", "vertices", "in", "different", "groups", ".", "Overall", ",", "the", "performance", "of", "different", "models", "increases", "when", "the", "degrees", "of", "the", "vertices", "increase", ".", "In", "the", "original", "network", ",", "the", "LINE", "(", "2nd", ")", "outperforms", "LINE", "(", "1st", ")", "except", "for", "the", "first", "group", ",", "which", "confirms", "that", "the", "second", "-", "order", "proximity", "does", "not", "work", "well", "for", "nodes", "with", "a", "low", "degree", ".", "In", "the", "reconstructed", "dense", "network", ",", "the", "performance", "of", "the", "LINE", "(", "1st", ")", "or", "LINE", "(", "2nd", ")", "improves", ",", "especially", "the", "LINE", "(", "2nd", ")", "that", "preserves", "the", "second", "-", "order", "proximity", ".", "We", "can", "also", "see", "that", "the", "LINE", "(", "2nd", ")", "model", "on", "the", "reconstructed", "network", "outperforms", "DeepWalk", "in", "all", "the", "groups", ".", "subsection", ":", "Parameter", "Sensitivity", "Next", ",", "we", "investigate", "the", "performance", "w.r.t", ".", "the", "parameter", "dimension", "and", "the", "converging", "performance", "of", "different", "models", "w.r.t", "the", "number", "of", "samples", "on", "the", "reconstructed", "Youtube", "network", ".", "Fig", ".", "[", "reference", "]", "reports", "the", "performance", "of", "the", "LINE", "model", "w.r.t", ".", "the", "dimension", ".", "We", "can", "see", "that", "the", "performance", "of", "the", "LINE", "(", "1st", ")", "or", "LINE", "(", "2nd", ")", "drops", "when", "the", "dimension", "becomes", "too", "large", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "results", "of", "the", "LINE", "and", "DeepWalk", "w.r.t", ".", "the", "number", "of", "samples", "during", "the", "optimization", ".", "The", "LINE", "(", "2nd", ")", "consistently", "outperforms", "LINE", "(", "1st", ")", "and", "DeepWalk", ",", "and", "both", "the", "LINE", "(", "1st", ")", "and", "LINE", "(", "2nd", ")", "converge", "much", "faster", "than", "DeepWalk", ".", "subsection", ":", "Scalability", "Finally", ",", "we", "investigate", "the", "scalability", "of", "the", "LINE", "model", "optimized", "by", "the", "edge", "-", "sampling", "treatment", "and", "asynchronous", "stochastic", "gradient", "descent", ",", "which", "deploys", "multiple", "threads", "for", "optimization", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "speed", "up", "w.r.t", ".", "the", "number", "of", "threads", "on", "the", "Youtube", "data", "set", ".", "The", "speed", "up", "is", "quite", "close", "to", "linear", ".", "Fig", ".", "[", "reference", "]", "shows", "that", "the", "classification", "performance", "remains", "stable", "when", "using", "multiple", "threads", "for", "model", "updating", ".", "The", "two", "figures", "together", "show", "that", "the", "inference", "algorithm", "of", "the", "LINE", "model", "is", "quite", "scalable", ".", "section", ":", "Conclusion", "This", "paper", "presented", "a", "novel", "network", "embedding", "model", "called", "the", "“", "LINE", ",", "”", "which", "can", "easily", "scale", "up", "to", "networks", "with", "millions", "of", "vertices", "and", "billions", "of", "edges", ".", "It", "has", "carefully", "designed", "objective", "functions", "that", "preserve", "both", "the", "first", "-", "order", "and", "second", "-", "order", "proximities", ",", "which", "are", "complementary", "to", "each", "other", ".", "An", "efficient", "and", "effective", "edge", "-", "sampling", "method", "is", "proposed", "for", "model", "inference", ",", "which", "solved", "the", "limitation", "of", "stochastic", "gradient", "descent", "on", "weighted", "edges", "without", "compromising", "the", "efficiency", ".", "Experimental", "results", "on", "various", "real", "-", "world", "networks", "prove", "the", "efficiency", "and", "effectiveness", "of", "LINE", ".", "In", "the", "future", ",", "we", "plan", "to", "investigate", "higher", "-", "order", "proximity", "beyond", "the", "first", "-", "order", "and", "second", "-", "order", "proximities", "in", "the", "network", ".", "Besides", ",", "we", "also", "plan", "to", "investigate", "the", "embedding", "of", "heterogeneous", "information", "networks", ",", "e.g.", ",", "vertices", "with", "multiple", "types", ".", "section", ":", "Acknowledgments", "The", "authors", "thank", "the", "three", "anonymous", "reviewers", "for", "the", "helpful", "comments", ".", "The", "co", "-", "author", "Ming", "Zhang", "is", "supported", "by", "the", "National", "Natural", "Science", "Foundation", "of", "China", "(", "NSFC", "Grant", "No", ".", "61472006", ")", ";", "Qiaozhu", "Mei", "is", "supported", "by", "the", "National", "Science", "Foundation", "under", "grant", "numbers", "IIS", "-", "1054199", "and", "CCF", "-", "1048168", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["LINE"], "Method"], [4, 10, ["Large", "-", "scale", "Information", "Network", "Embedding"], "Method"], [16, 21, ["embedding", "very", "large", "information", "networks"], "Task"], [22, 27, ["low", "-", "dimensional", "vector", "spaces"], "Task"], [36, 37, ["visualization"], "Task"], [38, 40, ["node", "classification"], "Task"], [42, 44, ["link", "prediction"], "Task"], [47, 50, ["graph", "embedding", "methods"], "Method"], [54, 58, ["real", "world", "information", "networks"], "Task"], [73, 76, ["network", "embedding", "method"], "Method"], [79, 80, ["LINE"], "Method"], [89, 91, ["information", "networks"], "Task"], [107, 109, ["objective", "function"], "Method"], [120, 124, ["edge", "-", "sampling", "algorithm"], "Method"], [133, 136, ["stochastic", "gradient", "descent"], "Method"], [146, 147, ["inference"], "Task"], [155, 156, ["LINE"], "Method"], [160, 165, ["real", "-", "world", "information", "networks"], "Task"], [189, 190, ["embedding"], "Task"], [216, 217, ["LINE"], "Method"], [267, 269, ["Information", "networks"], "Task"], [279, 281, ["airline", "networks"], "Task"], [282, 284, ["publication", "networks"], "Task"], [285, 289, ["social", "and", "communication", "networks"], "Task"], [292, 295, ["World", "Wide", "Web"], "Task"], [300, 302, ["information", "networks"], "Task"], [314, 318, ["Analyzing", "large", "information", "networks"], "Task"], [335, 338, ["embedding", "information", "networks"], "Task"], [339, 343, ["low", "-", "dimensional", "spaces"], "Task"], [359, 363, ["low", "-", "dimensional", "embedding"], "Method"], [373, 374, ["visualization"], "Task"], [375, 377, ["node", "classification"], "Task"], [378, 380, ["link", "prediction"], "Task"], [382, 383, ["recommendation"], "Task"], [387, 389, ["graph", "embedding"], "Method"], [418, 422, ["real", "world", "information", "network"], "Task"], [440, 445, ["Twitter", "followee", "-", "follower", "network"], "Task"], [460, 463, ["graph", "embedding", "algorithms"], "Method"], [476, 478, ["time", "complexity"], "Metric"], [480, 483, ["graph", "embedding", "algorithms"], "Method"], [485, 486, ["MDS"], "Method"], [487, 488, ["IsoMap"], "Method"], [489, 491, ["Laplacian", "eigenmap"], "Method"], [520, 526, ["embedding", "of", "large", "-", "scale", "networks"], "Task"], [532, 534, ["indirect", "approach"], "Method"], [552, 554, ["network", "embedding"], "Task"], [569, 571, ["objective", "function"], "Metric"], [580, 582, ["optimization", "technique"], "Method"], [586, 591, ["embedding", "of", "millions", "of", "nodes"], "Task"], [600, 603, ["network", "embedding", "model"], "Method"], [606, 607, ["LINE"], "Method"], [674, 677, ["graph", "embedding", "algorithms"], "Method"], [817, 818, ["sociology"], "Task"], [819, 820, ["linguistics"], "Task"], [847, 849, ["social", "network"], "Task"], [1086, 1089, ["stochastic", "gradient", "descent"], "Method"], [1091, 1092, ["optimization"], "Task"], [1101, 1104, ["stochastic", "gradient", "descent"], "Method"], [1107, 1111, ["real", "world", "information", "networks"], "Task"], [1133, 1138, ["word", "co", "-", "occurrence", "network"], "Method"], [1193, 1197, ["edge", "-", "sampling", "method"], "Method"], [1207, 1208, ["inference"], "Task"], [1231, 1233, ["model", "updating"], "Task"], [1236, 1238, ["sampling", "process"], "Method"], [1258, 1259, ["LINE"], "Method"], [1267, 1268, ["directed"], "Task"], [1269, 1270, ["undirected"], "Task"], [1271, 1272, ["weighted"], "Task"], [1273, 1275, ["unweighted", "graphs"], "Task"], [1282, 1283, ["LINE"], "Method"], [1285, 1290, ["real", "-", "world", "information", "networks"], "Task"], [1312, 1315, ["data", "mining", "tasks"], "Task"], [1317, 1319, ["word", "analogy"], "Task"], [1320, 1322, ["text", "classification"], "Task"], [1324, 1326, ["node", "classification"], "Task"], [1342, 1343, ["effectiveness"], "Metric"], [1344, 1345, ["efficiency"], "Metric"], [1386, 1389, ["network", "embedding", "model"], "Method"], [1392, 1393, ["LINE"], "Method"], [1400, 1402, ["information", "networks"], "Task"], [1433, 1437, ["edge", "-", "sampling", "algorithm"], "Method"], [1440, 1441, ["objective"], "Task"], [1450, 1453, ["stochastic", "gradient", "decent"], "Method"], [1461, 1462, ["inference"], "Task"], [1468, 1473, ["real", "-", "world", "information", "networks"], "Task"], [1517, 1523, ["large", "-", "scale", "information", "network", "embedding"], "Task"], [1565, 1567, ["graph", "embedding"], "Task"], [1568, 1570, ["dimension", "reduction"], "Task"], [1575, 1577, ["multidimensional", "scaling"], "Method"], [1578, 1579, ["MDS"], "Method"], [1581, 1582, ["IsoMap"], "Method"], [1583, 1584, ["LLE"], "Method"], [1585, 1587, ["Laplacian", "Eigenmap"], "Method"], [1594, 1596, ["affinity", "graph"], "Method"], [1608, 1615, ["K", "-", "nearest", "neighbor", "graph", "of", "data"], "Method"], [1620, 1622, ["affinity", "graph"], "Method"], [1645, 1646, ["complexity"], "Metric"], [1663, 1667, ["large", "-", "scale", "networks"], "Task"], [1677, 1679, ["graph", "factorization"], "Method"], [1683, 1691, ["low", "-", "dimensional", "embedding", "of", "a", "large", "graph"], "Task"], [1692, 1694, ["matrix", "factorization"], "Method"], [1699, 1702, ["stochastic", "gradient", "descent"], "Method"], [1722, 1724, ["matrix", "factorization"], "Method"], [1728, 1729, ["networks"], "Task"], [1742, 1744, ["graph", "factorization"], "Method"], [1769, 1770, ["networks"], "Task"], [1788, 1791, ["graph", "factorization", "method"], "Method"], [1804, 1808, ["undirected", "and", "directed", "graphs"], "Task"], [1817, 1818, ["DeepWalk"], "Method"], [1822, 1825, ["truncated", "random", "walk"], "Method"], [1826, 1829, ["social", "network", "embedding"], "Task"], [1835, 1836, ["DeepWalk"], "Method"], [1870, 1871, ["LINE"], "Method"], [1882, 1883, ["DeepWalk"], "Method"], [1884, 1886, ["random", "walks"], "Method"], [1899, 1903, ["depth", "-", "first", "search"], "Method"], [1907, 1912, ["breadth", "-", "first", "search", "strategy"], "Method"], [1928, 1929, ["DeepWalk"], "Method"], [1932, 1934, ["unweighted", "networks"], "Task"], [1944, 1945, ["weighted"], "Task"], [1946, 1948, ["unweighted", "edges"], "Task"], [1980, 1986, ["large", "-", "scale", "information", "network", "embedding"], "Task"], [2000, 2002, ["information", "network"], "Task"], [2008, 2010, ["Information", "Network"], "Task"], [2012, 2014, ["information", "network"], "Task"], [2132, 2134, ["information", "networks"], "Task"], [2196, 2198, ["citation", "networks"], "Task"], [2199, 2201, ["social", "networks"], "Task"], [2254, 2256, ["information", "network"], "Task"], [2273, 2274, ["embedding"], "Task"], [2460, 2463, ["graph", "embedding", "algorithms"], "Method"], [2465, 2466, ["IsoMap"], "Method"], [2467, 2468, ["LLE"], "Method"], [2469, 2471, ["Laplacian", "eigenmap"], "Method"], [2473, 2475, ["graph", "factorization"], "Method"], [2490, 2494, ["real", "world", "information", "network"], "Task"], [2538, 2542, ["first", "-", "order", "proximity"], "Method"], [2568, 2569, ["sparsity"], "Task"], [2592, 2594, ["social", "networks"], "Task"], [2774, 2778, ["second", "-", "order", "proximity"], "Task"], [2779, 2781, ["network", "embedding"], "Task"], [2791, 2797, ["Large", "-", "scale", "Information", "Network", "Embedding"], "Method"], [2813, 2819, ["Large", "-", "scale", "Information", "Network", "Embedding"], "Method"], [2839, 2840, ["fG→VRd"], "Method"], [2871, 2877, ["large", "-", "scale", "network", "embedding", "model"], "Method"], [2890, 2891, ["LINE"], "Method"], [2892, 2898, ["Large", "-", "scale", "Information", "Network", "Embedding"], "Method"], [2900, 2902, ["embedding", "model"], "Method"], [2903, 2907, ["real", "world", "information", "networks"], "Task"], [2984, 2987, ["network", "embedding", "model"], "Method"], [2990, 2991, ["LINE"], "Method"], [3037, 3038, ["LINE"], "Method"], [3039, 3043, ["First", "-", "order", "Proximity"], "Method"], [3088, 3095, ["low", "-", "dimensional", "vector", "representation", "of", "vertex"], "Method"], [3225, 3229, ["Second", "-", "order", "Proximity"], "Method"], [3238, 3242, ["directed", "and", "undirected", "graphs"], "Task"], [3487, 3491, ["low", "-", "dimensional", "representation"], "Method"], [3504, 3506, ["objective", "function"], "Metric"], [3557, 3558, ["PageRank"], "Method"], [3664, 3672, ["first", "-", "order", "and", "second", "-", "order", "proximities"], "Method"], [3768, 3770, ["Model", "Optimization"], "Task"], [3770, 3772, ["Optimizing", "objective"], "Task"], [3807, 3809, ["negative", "sampling"], "Method"], [3890, 3892, ["objective", "function"], "Metric"], [3920, 3923, ["negative", "sampling", "approach"], "Method"], [3936, 3940, ["asynchronous", "stochastic", "gradient", "algorithm"], "Method"], [3941, 3942, ["ASGD"], "Method"], [3944, 3946, ["optimizing", "Eqn"], "Task"], [3958, 3960, ["ASGD", "algorithm"], "Method"], [4027, 4032, ["word", "co", "-", "occurrence", "network"], "Method"], [4077, 4079, ["learning", "rate"], "Metric"], [4085, 4087, ["learning", "rate"], "Metric"], [4127, 4128, ["Optimization"], "Task"], [4129, 4131, ["Edge", "Sampling"], "Method"], [4208, 4210, ["memory", "requirement"], "Metric"], [4255, 4259, ["edge", "-", "sampling", "treatment"], "Method"], [4262, 4264, ["objective", "function"], "Metric"], [4350, 4353, ["alias", "table", "method"], "Method"], [4392, 4393, ["optimization"], "Task"], [4394, 4396, ["negative", "sampling"], "Task"], [4427, 4428, ["optimization"], "Task"], [4441, 4443, ["time", "complexity"], "Metric"], [4445, 4446, ["LINE"], "Method"], [4468, 4471, ["edge", "sampling", "treatment"], "Method"], [4476, 4479, ["stochastic", "gradient", "descent"], "Method"], [4540, 4546, ["second", "-", "order", "proximity", "based", "methods"], "Method"], [4791, 4792, ["LINE"], "Method"], [4799, 4806, ["large", "-", "scale", "real", "-", "world", "networks"], "Task"], [4835, 4837, ["Language", "network"], "Method"], [4841, 4846, ["word", "co", "-", "occurrence", "network"], "Method"], [4851, 4854, ["English", "Wikipedia", "pages"], "Material"], [4896, 4897, ["Flickr"], "Method"], [4898, 4899, ["YoutubeAvailable"], "Method"], [4911, 4913, ["Flickr", "network"], "Method"], [4917, 4919, ["Youtube", "network"], "Method"], [4938, 4940, ["citation", "networks"], "Method"], [4944, 4947, ["author", "citation", "network"], "Method"], [4949, 4952, ["paper", "citation", "network"], "Method"], [4971, 4974, ["author", "citation", "network"], "Method"], [5059, 5062, ["graph", "embedding", "methods"], "Method"], [5080, 5083, ["graph", "embedding", "algorithms"], "Method"], [5085, 5086, ["MDS"], "Method"], [5087, 5088, ["IsoMap"], "Method"], [5090, 5092, ["Laplacian", "eigenmap"], "Method"], [5103, 5105, ["Graph", "factorization"], "Method"], [5106, 5107, ["GF"], "Method"], [5113, 5116, ["matrix", "factorization", "techniques"], "Method"], [5117, 5119, ["graph", "factorization"], "Task"], [5121, 5123, ["information", "network"], "Task"], [5128, 5130, ["affinity", "matrix"], "Method"], [5145, 5147, ["matrix", "factorization"], "Method"], [5148, 5150, ["Graph", "factorization"], "Method"], [5153, 5156, ["stochastic", "gradient", "descent"], "Method"], [5168, 5170, ["undirected", "networks"], "Task"], [5171, 5172, ["DeepWalk"], "Method"], [5173, 5174, ["DeepWalk"], "Method"], [5180, 5183, ["social", "network", "embedding"], "Task"], [5241, 5243, ["objective", "Eqn"], "Metric"], [5259, 5262, ["stochastic", "gradient", "descent"], "Method"], [5284, 5286, ["model", "updating"], "Task"], [5320, 5321, ["LINE"], "Method"], [5330, 5334, ["edge", "-", "sampling", "treatment"], "Method"], [5364, 5366, ["model", "updating"], "Task"], [5373, 5377, ["LINE", "(", "1st", ")"], "Method"], [5378, 5382, ["LINE", "(", "2nd", ")"], "Method"], [5385, 5387, ["graph", "factorization"], "Method"], [5389, 5393, ["LINE", "(", "1st", ")"], "Method"], [5403, 5405, ["undirected", "graphs"], "Method"], [5406, 5410, ["LINE", "(", "2nd", ")"], "Method"], [5420, 5421, ["undirected"], "Method"], [5422, 5424, ["directed", "graphs"], "Method"], [5425, 5426, ["LINE"], "Method"], [5427, 5430, ["1st", "+", "2nd"], "Method"], [5453, 5455, ["vector", "representations"], "Method"], [5457, 5461, ["LINE", "(", "1st", ")"], "Method"], [5462, 5466, ["LINE", "(", "2nd", ")"], "Method"], [5489, 5492, ["supervised", "learning", "task"], "Task"], [5509, 5511, ["unsupervised", "task"], "Task"], [5527, 5528, ["LINE"], "Method"], [5537, 5539, ["supervised", "tasks"], "Task"], [5551, 5554, ["stochastic", "gradient", "descent"], "Method"], [5567, 5569, ["learning", "rate"], "Metric"], [5595, 5596, ["dimensionality"], "Metric"], [5601, 5603, ["language", "network"], "Method"], [5611, 5613, ["word", "embedding"], "Task"], [5642, 5643, ["LINE"], "Method"], [5655, 5659, ["LINE", "(", "1st", ")"], "Method"], [5660, 5664, ["LINE", "(", "2nd", ")"], "Method"], [5667, 5668, ["GF"], "Method"], [5679, 5680, ["DeepWalk"], "Method"], [5697, 5699, ["Language", "Network"], "Method"], [5706, 5708, ["language", "network"], "Task"], [5730, 5731, ["embeddings"], "Task"], [5732, 5734, ["word", "analogy"], "Task"], [5735, 5737, ["document", "classification"], "Task"], [5738, 5740, ["Word", "Analogy"], "Task"], [5871, 5873, ["word", "analogy"], "Task"], [5879, 5882, ["semantic", "and", "syntactic"], "Task"], [5885, 5886, ["GF"], "Method"], [5910, 5912, ["word", "analogy"], "Task"], [5920, 5922, ["Wikipedia", "corpora"], "Material"], [5923, 5924, ["SkipGram"], "Material"], [5927, 5930, ["Wikipedia", "word", "network"], "Material"], [5937, 5939, ["graph", "factorization"], "Task"], [5975, 5976, ["DeepWalk"], "Method"], [6034, 6036, ["Wikipedia", "pages"], "Material"], [6041, 6044, ["matrix", "factorization", "approach"], "Method"], [6060, 6062, ["language", "network"], "Method"], [6067, 6071, ["LINE", "(", "2nd", ")"], "Method"], [6078, 6081, ["graph", "embedding", "methods"], "Method"], [6083, 6084, ["SkipGram"], "Material"], [6151, 6155, ["LINE", "(", "2nd", ")"], "Method"], [6164, 6167, ["word", "embedding", "model"], "Method"], [6179, 6181, ["language", "network"], "Method"], [6202, 6204, ["graph", "factorization"], "Method"], [6205, 6209, ["LINE", "(", "1st", ")"], "Method"], [6211, 6212, ["DeepWalk"], "Method"], [6214, 6215, ["DeepWalk"], "Method"], [6224, 6225, ["DeepWalk"], "Method"], [6247, 6249, ["language", "network"], "Task"], [6259, 6260, ["SGD"], "Method"], [6272, 6274, ["language", "network"], "Method"], [6289, 6291, ["learning", "process"], "Method"], [6294, 6295, ["LINE"], "Method"], [6298, 6302, ["edge", "-", "sampling", "treatment"], "Method"], [6313, 6316, ["first", "-", "order"], "Method"], [6347, 6351, ["LINE", "(", "1st", ")"], "Method"], [6352, 6356, ["LINE", "(", "2nd", ")"], "Method"], [6388, 6390, ["graph", "factorization"], "Method"], [6396, 6397, ["DeepWalk"], "Method"], [6415, 6419, ["threshold", "-", "cutting", "technique"], "Method"], [6430, 6432, ["Document", "Classification"], "Task"], [6451, 6453, ["document", "representation"], "Task"], [6459, 6462, ["document", "classification", "tasks"], "Task"], [6480, 6483, ["word", "vector", "representations"], "Method"], [6495, 6497, ["word", "embeddings"], "Method"], [6507, 6509, ["document", "embeddings"], "Task"], [6515, 6518, ["document", "embedding", "approaches"], "Method"], [6525, 6527, ["Wikipedia", "pages"], "Material"], [6542, 6543, ["classification"], "Task"], [6610, 6611, ["evaluation"], "Task"], [6620, 6628, ["one", "-", "vs", "-", "rest", "logistic", "regression", "classifier"], "Method"], [6630, 6632, ["LibLinear", "package"], "Method"], [6636, 6641, ["classification", "metrics", "Micro", "-", "F1"], "Metric"], [6642, 6645, ["Macro", "-", "F1"], "Metric"], [6668, 6671, ["Wikipedia", "page", "classification"], "Task"], [6680, 6683, ["word", "analogy", "task"], "Task"], [6685, 6687, ["graph", "factorization"], "Method"], [6688, 6689, ["DeepWalk"], "Method"], [6690, 6691, ["DeepWalk"], "Method"], [6716, 6717, ["LINE"], "Method"], [6720, 6724, ["edge", "-", "sampling", "treatment"], "Method"], [6730, 6731, ["SGD"], "Method"], [6733, 6737, ["LINE", "(", "2nd", ")"], "Method"], [6738, 6742, ["LINE", "(", "1st", ")"], "Method"], [6748, 6750, ["graph", "factorization"], "Method"], [6755, 6757, ["supervised", "task"], "Task"], [6767, 6771, ["LINE", "(", "1st", ")"], "Method"], [6772, 6776, ["LINE", "(", "2nd", ")"], "Method"], [6889, 6893, ["first", "-", "order", "proximity"], "Method"], [6905, 6907, ["Social", "Network"], "Method"], [6909, 6910, ["DeepWalk"], "Method"], [6929, 6931, ["language", "networks"], "Method"], [6933, 6935, ["social", "networks"], "Method"], [6941, 6943, ["Youtube", "network"], "Method"], [6951, 6956, ["multi", "-", "label", "classification", "task"], "Task"], [6993, 6995, ["Flickr", "Network"], "Method"], [7007, 7009, ["Flickr", "network"], "Method"], [7024, 7028, ["multi", "-", "label", "classification"], "Task"], [7051, 7055, ["LINE", "(", "1st", ")"], "Method"], [7059, 7063, ["LINE", "(", "2nd", ")"], "Method"], [7072, 7074, ["language", "network"], "Method"], [7098, 7100, ["social", "network"], "Method"], [7149, 7153, ["LINE", "(", "1st", ")"], "Method"], [7154, 7156, ["graph", "factorization"], "Method"], [7169, 7173, ["LINE", "(", "2nd", ")"], "Method"], [7174, 7175, ["DeepWalk"], "Method"], [7194, 7198, ["LINE", "(", "1st", ")"], "Method"], [7199, 7203, ["LINE", "(", "2nd", ")"], "Method"], [7222, 7223, ["DeepWalk"], "Method"], [7241, 7242, ["DeepWalk"], "Method"], [7260, 7261, ["DeepWalk"], "Method"], [7277, 7279, ["Youtube", "Network"], "Method"], [7288, 7290, ["Youtube", "network"], "Method"], [7315, 7319, ["LINE", "(", "1st", ")"], "Method"], [7320, 7324, ["LINE", "(", "2nd", ")"], "Method"], [7331, 7333, ["Flickr", "network"], "Method"], [7343, 7347, ["LINE", "(", "2nd", ")"], "Method"], [7351, 7352, ["DeepWalk"], "Method"], [7360, 7361, ["LINE"], "Method"], [7375, 7376, ["LINE"], "Method"], [7377, 7378, ["DeepWalk"], "Method"], [7402, 7404, ["network", "sparsity"], "Task"], [7411, 7412, ["DeepWalk"], "Method"], [7432, 7435, ["random", "walk", "approach"], "Method"], [7438, 7442, ["depth", "-", "first", "search"], "Method"], [7488, 7493, ["breadth", "-", "first", "search", "strategy"], "Method"], [7571, 7573, ["reconstructed", "network"], "Method"], [7577, 7578, ["GF"], "Method"], [7579, 7583, ["LINE", "(", "1st", ")"], "Method"], [7584, 7588, ["LINE", "(", "2nd", ")"], "Method"], [7592, 7596, ["LINE", "(", "2nd", ")"], "Method"], [7599, 7601, ["reconstructed", "network"], "Method"], [7603, 7607, ["LINE", "(", "2nd", ")"], "Method"], [7608, 7609, ["DeepWalk"], "Method"], [7629, 7631, ["reconstructed", "network"], "Method"], [7683, 7685, ["network", "embedding"], "Task"], [7689, 7693, ["dense", "and", "sparse", "networks"], "Task"], [7696, 7698, ["Citation", "Network"], "Method"], [7704, 7706, ["citation", "networks"], "Method"], [7711, 7713, ["directed", "networks"], "Method"], [7732, 7734, ["directed", "networks"], "Task"], [7740, 7741, ["DeepWalk"], "Method"], [7742, 7746, ["LINE", "(", "2nd", ")"], "Method"], [7755, 7760, ["multi", "-", "label", "classification", "task"], "Task"], [7767, 7768, ["AAAI"], "Method"], [7769, 7770, ["CIKM"], "Method"], [7771, 7772, ["ICML"], "Method"], [7773, 7774, ["KDD"], "Method"], [7775, 7776, ["NIPS"], "Method"], [7777, 7778, ["SIGIR"], "Method"], [7780, 7781, ["WWW"], "Method"], [7783, 7785, ["classification", "categories"], "Task"], [7809, 7814, ["DBLP", "(", "AuthorCitation", ")", "Network"], "Method"], [7824, 7829, ["DBLP", "(", "AuthorCitation", ")", "network"], "Method"], [7838, 7839, ["DeepWalk"], "Method"], [7871, 7875, ["LINE", "(", "2nd", ")"], "Method"], [7879, 7880, ["DeepWalk"], "Method"], [7887, 7890, ["stochastic", "gradient", "descent"], "Method"], [7891, 7895, ["LINE", "(", "2nd", ")"], "Method"], [7903, 7908, ["DBLP", "(", "PaperCitation", ")", "Network"], "Method"], [7918, 7923, ["DBLP", "(", "PaperCitation", ")", "network"], "Method"], [7925, 7929, ["LINE", "(", "2nd", ")"], "Method"], [7931, 7932, ["DeepWalk"], "Method"], [7941, 7944, ["paper", "citation", "network"], "Method"], [7968, 7972, ["LINE", "(", "2nd", ")"], "Method"], [7988, 7992, ["LINE", "(", "2nd", ")"], "Method"], [8017, 8019, ["Network", "Layouts"], "Method"], [8023, 8025, ["network", "embedding"], "Task"], [8028, 8030, ["meaningful", "visualizations"], "Task"], [8043, 8047, ["co", "-", "author", "network"], "Method"], [8063, 8064, ["WWW"], "Method"], [8065, 8066, ["KDD"], "Method"], [8068, 8070, ["data", "mining"], "Task"], [8072, 8073, ["NIPS"], "Method"], [8074, 8075, ["ICML"], "Method"], [8077, 8079, ["machine", "learning"], "Task"], [8082, 8083, ["CVPR"], "Method"], [8084, 8085, ["ICCV"], "Method"], [8087, 8089, ["computer", "vision"], "Task"], [8092, 8096, ["co", "-", "author", "network"], "Method"], [8134, 8138, ["co", "-", "author", "network"], "Method"], [8157, 8161, ["co", "-", "author", "network"], "Method"], [8169, 8171, ["embedding", "approaches"], "Method"], [8191, 8195, ["t", "-", "SNE", "package"], "Method"], [8203, 8204, ["visualization"], "Task"], [8207, 8209, ["embedding", "approaches"], "Method"], [8211, 8212, ["visualization"], "Task"], [8213, 8215, ["graph", "factorization"], "Method"], [8237, 8238, ["DeepWalk"], "Method"], [8269, 8270, ["DeepWalk"], "Method"], [8272, 8276, ["random", "walk", "based", "approach"], "Method"], [8351, 8354, ["sparsity", "of", "networks"], "Task"], [8358, 8360, ["social", "networks"], "Method"], [8374, 8378, ["LINE", "(", "1st", ")"], "Method"], [8379, 8383, ["LINE", "(", "2nd", ")"], "Method"], [8400, 8402, ["Flickr", "network"], "Method"], [8405, 8407, ["Flickr", "network"], "Method"], [8414, 8416, ["Youtube", "network"], "Method"], [8453, 8457, ["LINE", "(", "1st", ")"], "Method"], [8458, 8462, ["LINE", "(", "2nd", ")"], "Method"], [8473, 8477, ["LINE", "(", "2nd", ")"], "Method"], [8481, 8485, ["LINE", "(", "1st", ")"], "Method"], [8490, 8494, ["second", "-", "order", "proximity"], "Method"], [8505, 8509, ["first", "-", "order", "proximity"], "Method"], [8541, 8544, ["reconstructed", "Youtube", "networks"], "Method"], [8591, 8595, ["LINE", "(", "2nd", ")"], "Method"], [8596, 8600, ["LINE", "(", "1st", ")"], "Method"], [8627, 8630, ["reconstructed", "dense", "network"], "Task"], [8635, 8639, ["LINE", "(", "1st", ")"], "Method"], [8640, 8644, ["LINE", "(", "2nd", ")"], "Method"], [8648, 8652, ["LINE", "(", "2nd", ")"], "Method"], [8673, 8675, ["reconstructed", "network"], "Method"], [8676, 8677, ["DeepWalk"], "Method"], [8684, 8686, ["Parameter", "Sensitivity"], "Metric"], [8711, 8714, ["reconstructed", "Youtube", "network"], "Method"], [8740, 8744, ["LINE", "(", "1st", ")"], "Method"], [8745, 8749, ["LINE", "(", "2nd", ")"], "Method"], [8767, 8768, ["LINE"], "Method"], [8769, 8770, ["DeepWalk"], "Method"], [8778, 8779, ["optimization"], "Task"], [8781, 8785, ["LINE", "(", "2nd", ")"], "Method"], [8787, 8791, ["LINE", "(", "1st", ")"], "Method"], [8792, 8793, ["DeepWalk"], "Method"], [8797, 8801, ["LINE", "(", "1st", ")"], "Method"], [8810, 8811, ["DeepWalk"], "Method"], [8828, 8832, ["edge", "-", "sampling", "treatment"], "Method"], [8833, 8837, ["asynchronous", "stochastic", "gradient", "descent"], "Method"], [8843, 8844, ["optimization"], "Task"], [8862, 8865, ["Youtube", "data", "set"], "Method"], [8883, 8884, ["classification"], "Task"], [8892, 8894, ["model", "updating"], "Task"], [8902, 8904, ["inference", "algorithm"], "Method"], [8920, 8923, ["network", "embedding", "model"], "Method"], [8926, 8927, ["LINE"], "Method"], [8975, 8979, ["edge", "-", "sampling", "method"], "Method"], [8982, 8984, ["model", "inference"], "Task"], [8990, 8993, ["stochastic", "gradient", "descent"], "Method"], [9005, 9009, ["real", "-", "world", "networks"], "Task"], [9015, 9016, ["LINE"], "Method"], [9051, 9056, ["embedding", "of", "heterogeneous", "information", "networks"], "Task"], [691, 693, ["Laplacian", "eigenmap"], "Method"], [1332, 1333, ["LINE"], "Method"], [1484, 1485, ["LINE"], "Method"], [1530, 1531, ["LINE"], "Method"], [1759, 1760, ["LINE"], "Method"], [3007, 3008, ["LINE"], "Method"], [3223, 3224, ["LINE"], "Method"], [3702, 3703, ["LINE"], "Method"], [4494, 4495, ["LINE"], "Method"], [5054, 5055, ["LINE"], "Method"], [5223, 5224, ["LINE"], "Method"], [5230, 5231, ["LINE"], "Method"], [5295, 5296, ["LINE"], "Method"], [5302, 5303, ["LINE"], "Method"], [5325, 5326, ["LINE"], "Method"], [5394, 5395, ["LINE"], "Method"], [5411, 5412, ["LINE"], "Method"], [5529, 5532, ["1st", "+", "2nd"], "Method"], [5644, 5645, ["LINE"], "Method"], [6023, 6024, ["SkipGram"], "Material"], [6254, 6255, ["LINE"], "Method"], [6406, 6407, ["LINE"], "Method"], [6699, 6700, ["LINE"], "Method"], [6782, 6783, ["LINE"], "Method"], [6784, 6787, ["1st", "+", "2nd"], "Method"], [7039, 7040, ["LINE"], "Method"], [7041, 7044, ["1st", "+", "2nd"], "Method"], [7621, 7622, ["LINE"], "Method"], [7623, 7626, ["1st", "+", "2nd"], "Method"], [7668, 7669, ["LINE"], "Method"], [7670, 7673, ["1st", "+", "2nd"], "Method"], [7718, 7719, ["LINE"], "Method"], [7840, 7841, ["LINE"], "Method"], [7882, 7883, ["LINE"], "Method"], [8304, 8305, ["LINE"], "Method"], [8666, 8667, ["LINE"], "Method"], [8725, 8726, ["LINE"], "Method"], [8802, 8803, ["LINE"], "Method"], [8823, 8824, ["LINE"], "Method"], [8906, 8907, ["LINE"], "Method"]]} {"doc_id": "0a3381f0432c5cfe491c718349d7a44e5814592c", "text": ["document", ":", "Compositional", "Sequence", "Labeling", "Models", "for", "Error", "Detection", "in", "Learner", "Writing", "In", "this", "paper", ",", "we", "present", "the", "first", "experiments", "using", "neural", "network", "models", "for", "the", "task", "of", "error", "detection", "in", "learner", "writing", ".", "We", "perform", "a", "systematic", "comparison", "of", "alternative", "compositional", "architectures", "and", "propose", "a", "framework", "for", "error", "detection", "based", "on", "bidirectional", "LSTMs", ".", "Experiments", "on", "the", "CoNLL", "-", "14", "shared", "task", "dataset", "show", "the", "model", "is", "able", "to", "outperform", "other", "participants", "on", "detecting", "errors", "in", "learner", "writing", ".", "Finally", ",", "the", "model", "is", "integrated", "with", "a", "publicly", "deployed", "self", "-", "assessment", "system", ",", "leading", "to", "performance", "comparable", "to", "human", "annotators", ".", "section", ":", "Introduction", "Automated", "systems", "for", "detecting", "errors", "in", "learner", "writing", "are", "valuable", "tools", "for", "second", "language", "learning", "and", "assessment", ".", "Most", "work", "in", "recent", "years", "has", "focussed", "on", "error", "correction", ",", "with", "error", "detection", "performance", "measured", "as", "a", "byproduct", "of", "the", "correction", "output", ".", "However", ",", "this", "assumes", "that", "systems", "are", "able", "to", "propose", "a", "correction", "for", "every", "detected", "error", ",", "and", "accurate", "systems", "for", "correction", "might", "not", "be", "optimal", "for", "detection", ".", "While", "closed", "-", "class", "errors", "such", "as", "incorrect", "prepositions", "and", "determiners", "can", "be", "modeled", "with", "a", "supervised", "classification", "approach", ",", "content", "-", "content", "word", "errors", "are", "the", "3rd", "most", "frequent", "error", "type", "and", "pose", "a", "serious", "challenge", "to", "error", "correction", "frameworks", ".", "Evaluation", "of", "error", "correction", "is", "also", "highly", "subjective", "and", "human", "annotators", "have", "rather", "low", "agreement", "on", "gold", "-", "standard", "corrections", ".", "Therefore", ",", "we", "treat", "error", "detection", "in", "learner", "writing", "as", "an", "independent", "task", "and", "propose", "a", "system", "for", "labeling", "each", "token", "as", "being", "correct", "or", "incorrect", "in", "context", ".", "Common", "approaches", "to", "similar", "sequence", "labeling", "tasks", "involve", "learning", "weights", "or", "probabilities", "for", "context", "n", "-", "grams", "of", "varying", "sizes", ",", "or", "relying", "on", "previously", "extracted", "high", "-", "confidence", "context", "patterns", ".", "Both", "of", "these", "methods", "can", "suffer", "from", "data", "sparsity", ",", "as", "they", "treat", "words", "as", "independent", "units", "and", "miss", "out", "on", "potentially", "related", "patterns", ".", "In", "addition", ",", "they", "need", "to", "specify", "a", "fixed", "context", "size", "and", "are", "therefore", "often", "limited", "to", "using", "a", "small", "window", "near", "the", "target", ".", "Neural", "network", "models", "aim", "to", "address", "these", "weaknesses", "and", "have", "achieved", "success", "in", "various", "NLP", "tasks", "such", "as", "language", "modeling", "and", "speech", "recognition", ".", "Recent", "developments", "in", "machine", "translation", "have", "also", "shown", "that", "text", "of", "varying", "length", "can", "be", "represented", "as", "a", "fixed", "-", "size", "vector", "using", "convolutional", "networks", "or", "recurrent", "neural", "networks", ".", "In", "this", "paper", ",", "we", "present", "the", "first", "experiments", "using", "neural", "network", "models", "for", "the", "task", "of", "error", "detection", "in", "learner", "writing", ".", "We", "perform", "a", "systematic", "comparison", "of", "alternative", "compositional", "structures", "for", "constructing", "informative", "context", "representations", ".", "Based", "on", "the", "findings", ",", "we", "propose", "a", "novel", "framework", "for", "performing", "error", "detection", "in", "learner", "writing", ",", "which", "achieves", "state", "-", "of", "-", "the", "-", "art", "results", "on", "two", "datasets", "of", "error", "-", "annotated", "learner", "essays", ".", "The", "sequence", "labeling", "model", "creates", "a", "single", "variable", "-", "size", "network", "over", "the", "whole", "sentence", ",", "conditions", "each", "label", "on", "all", "the", "words", ",", "and", "predicts", "all", "labels", "together", ".", "The", "effects", "of", "different", "datasets", "on", "the", "overall", "performance", "are", "investigated", "by", "incrementally", "providing", "additional", "training", "data", "to", "the", "model", ".", "Finally", ",", "we", "integrate", "the", "error", "detection", "framework", "with", "a", "publicly", "deployed", "self", "-", "assessment", "system", ",", "leading", "to", "performance", "comparable", "to", "human", "annotators", ".", "section", ":", "Background", "and", "Related", "Work", "The", "field", "of", "automatically", "detecting", "errors", "in", "learner", "text", "has", "a", "long", "and", "rich", "history", ".", "Most", "work", "has", "focussed", "on", "tackling", "specific", "types", "of", "errors", ",", "such", "as", "usage", "of", "incorrect", "prepositions", ",", "articles", ",", "verb", "forms", ",", "and", "adjective", "-", "noun", "pairs", ".", "However", ",", "there", "has", "been", "limited", "work", "on", "more", "general", "error", "detection", "systems", "that", "could", "handle", "all", "types", "of", "errors", "in", "learner", "text", ".", "Chodorow1998", "proposed", "a", "method", "based", "on", "mutual", "information", "and", "the", "chi", "-", "square", "statistic", "to", "detect", "sequences", "of", "part", "-", "of", "-", "speech", "tags", "and", "function", "words", "that", "are", "likely", "to", "be", "ungrammatical", "in", "English", ".", "Gamon2011", "used", "Maximum", "Entropy", "Markov", "Models", "with", "a", "range", "of", "features", ",", "such", "as", "POS", "tags", ",", "string", "features", ",", "and", "outputs", "from", "a", "constituency", "parser", ".", "The", "pilot", "Helping", "Our", "Own", "shared", "task", "also", "evaluated", "grammatical", "error", "detection", "of", "a", "number", "of", "different", "error", "types", ",", "though", "most", "systems", "were", "error", "-", "type", "specific", "and", "the", "best", "approach", "was", "heavily", "skewed", "towards", "article", "and", "preposition", "errors", ".", "We", "extend", "this", "line", "of", "research", ",", "working", "towards", "general", "error", "detection", "systems", ",", "and", "investigate", "the", "use", "of", "neural", "compositional", "models", "on", "this", "task", ".", "The", "related", "area", "of", "grammatical", "error", "correction", "has", "also", "gained", "considerable", "momentum", "in", "the", "past", "years", ",", "with", "four", "recent", "shared", "tasks", "highlighting", "several", "emerging", "directions", ".", "The", "current", "state", "-", "of", "-", "the", "-", "art", "approaches", "can", "broadly", "be", "separated", "into", "two", "categories", ":", "Phrase", "-", "based", "statistical", "machine", "translation", "techniques", ",", "essentially", "translating", "the", "incorrect", "source", "text", "into", "the", "corrected", "version", "Averaged", "Perceptrons", "and", "Naive", "Bayes", "classifiers", "making", "use", "of", "native", "-", "language", "error", "correction", "priors", ".", "Error", "correction", "systems", "require", "very", "specialised", "models", ",", "as", "they", "need", "to", "generate", "an", "improved", "version", "of", "the", "input", "text", ",", "whereas", "a", "wider", "range", "of", "tagging", "and", "classification", "models", "can", "be", "deployed", "on", "error", "detection", ".", "In", "addition", ",", "automated", "writing", "feedback", "systems", "that", "indicate", "the", "presence", "and", "location", "of", "errors", "may", "be", "better", "from", "a", "pedagogic", "point", "of", "view", ",", "rather", "than", "providing", "a", "panacea", "and", "correcting", "all", "errors", "in", "learner", "text", ".", "In", "Section", "[", "reference", "]", "we", "evaluate", "a", "neural", "sequence", "tagging", "model", "on", "the", "latest", "shared", "task", "test", "data", ",", "and", "compare", "it", "to", "the", "top", "participating", "systems", "on", "the", "task", "of", "error", "detection", ".", "section", ":", "Sequence", "Labeling", "Architectures", "We", "construct", "a", "neural", "network", "sequence", "labeling", "framework", "for", "the", "task", "of", "error", "detection", "in", "learner", "writing", ".", "The", "model", "receives", "only", "a", "series", "of", "tokens", "as", "input", ",", "and", "outputs", "the", "probability", "of", "each", "token", "in", "the", "sentence", "being", "correct", "or", "incorrect", "in", "a", "given", "context", ".", "The", "architectures", "start", "with", "the", "vector", "representations", "of", "individual", "words", ",", ",", "where", "is", "the", "length", "of", "the", "sentence", ".", "Different", "composition", "functions", "are", "then", "used", "to", "calculate", "a", "hidden", "vector", "representation", "of", "each", "token", "in", "context", ",", ".", "These", "representations", "are", "passed", "through", "a", "softmax", "layer", ",", "producing", "a", "probability", "distribution", "over", "the", "possible", "labels", "for", "every", "token", "in", "context", ":", "where", "is", "the", "weight", "matrix", "between", "the", "hidden", "vector", "and", "the", "output", "layer", ".", "We", "investigate", "six", "alternative", "neural", "network", "architectures", "for", "the", "task", "of", "error", "detection", ":", "convolutional", ",", "bidirectional", "recurrent", ",", "bidirectional", "LSTM", ",", "and", "multi", "-", "layer", "variants", "of", "each", "of", "them", ".", "In", "the", "convolutional", "neural", "network", "(", "CNN", ",", "Figure", "[", "reference", "]", "a", ")", "for", "token", "labeling", ",", "the", "hidden", "vector", "is", "calculated", "based", "on", "a", "fixed", "-", "size", "context", "window", ".", "The", "convolution", "acts", "as", "a", "feedforward", "network", ",", "using", "surrounding", "context", "words", "as", "input", ",", "and", "therefore", "it", "will", "learn", "to", "detect", "the", "presence", "of", "different", "types", "of", "n", "-", "grams", ".", "The", "assumption", "behind", "the", "convolutional", "architecture", "is", "that", "memorising", "erroneous", "token", "sequences", "from", "the", "training", "data", "is", "sufficient", "for", "performing", "error", "detection", ".", "The", "convolution", "uses", "tokens", "on", "either", "side", "of", "the", "target", "token", ",", "and", "the", "vectors", "for", "these", "tokens", "are", "concatenated", ",", "preserving", "the", "ordering", ":", "where", "is", "used", "as", "notation", "for", "vector", "concatenation", "of", "and", ".", "The", "combined", "vector", "is", "then", "passed", "through", "a", "non", "-", "linear", "layer", "to", "produce", "the", "hidden", "representation", ":", "The", "deep", "convolutional", "network", "(", "Figure", "[", "reference", "]", "b", ")", "adds", "an", "extra", "convolutional", "layer", "to", "the", "architecture", ",", "using", "the", "first", "layer", "as", "input", ".", "It", "creates", "convolutions", "of", "convolutions", ",", "thereby", "capturing", "more", "complex", "higher", "-", "order", "features", "from", "the", "dataset", ".", "In", "a", "recurrent", "neural", "network", "(", "RNN", ")", ",", "each", "hidden", "representation", "is", "calculated", "based", "on", "the", "current", "token", "embedding", "and", "the", "hidden", "vector", "at", "the", "previous", "time", "step", ":", "where", "is", "a", "nonlinear", "function", ",", "such", "as", "the", "sigmoid", "function", ".", "Instead", "of", "a", "fixed", "context", "window", ",", "information", "is", "passed", "through", "the", "sentence", "using", "a", "recursive", "function", "and", "the", "network", "is", "able", "to", "learn", "which", "patterns", "to", "disregard", "or", "pass", "forward", ".", "This", "recurrent", "network", "structure", "is", "referred", "to", "as", "an", "Elman", "-", "type", "network", ",", "after", "Elman1990", ".", "The", "bidirectional", "RNN", "(", "Figure", "[", "reference", "]", "c", ")", "consists", "of", "two", "recurrent", "components", ",", "moving", "in", "opposite", "directions", "through", "the", "sentence", ".", "While", "the", "unidirectional", "version", "takes", "into", "account", "only", "context", "on", "the", "left", "of", "the", "target", "token", ",", "the", "bidirectional", "version", "recursively", "builds", "separate", "context", "representations", "from", "either", "side", "of", "the", "target", "token", ".", "The", "left", "and", "right", "context", "are", "then", "concatenated", "and", "used", "as", "the", "hidden", "representation", ":", "Recurrent", "networks", "have", "been", "shown", "to", "perform", "well", "on", "the", "task", "of", "language", "modeling", ",", "where", "they", "learn", "an", "incremental", "composition", "function", "for", "predicting", "the", "next", "token", "in", "the", "sequence", ".", "However", ",", "while", "language", "models", "can", "estimate", "the", "probability", "of", "each", "token", ",", "they", "are", "unable", "to", "differentiate", "between", "infrequent", "and", "incorrect", "token", "sequences", ".", "For", "error", "detection", ",", "the", "composition", "function", "needs", "to", "learn", "to", "identify", "semantic", "anomalies", "or", "ungrammatical", "combinations", ",", "independent", "of", "their", "frequency", ".", "The", "bidirectional", "model", "provides", "extra", "information", ",", "as", "it", "allows", "the", "network", "to", "use", "context", "on", "both", "sides", "of", "the", "target", "token", ".", "Irsoy2014a", "created", "an", "extension", "of", "this", "architecture", "by", "connecting", "together", "multiple", "layers", "of", "bidirectional", "Elman", "-", "type", "recurrent", "network", "modules", ".", "This", "deep", "bidirectional", "RNN", "(", "Figure", "[", "reference", "]", "d", ")", "calculates", "a", "context", "-", "dependent", "representation", "for", "each", "token", "using", "a", "bidirectional", "RNN", ",", "and", "then", "uses", "this", "as", "input", "to", "another", "bidirectional", "RNN", ".", "The", "multi", "-", "layer", "structure", "allows", "the", "model", "to", "learn", "more", "complex", "higher", "-", "level", "features", "and", "effectively", "perform", "multiple", "recurrent", "passes", "through", "the", "sentence", ".", "The", "long", "-", "short", "term", "memory", "(", "LSTM", ")", "is", "an", "advanced", "alternative", "to", "the", "Elman", "-", "type", "networks", "that", "has", "recently", "become", "increasingly", "popular", ".", "It", "uses", "two", "separate", "hidden", "vectors", "to", "pass", "information", "between", "different", "time", "steps", ",", "and", "includes", "gating", "mechanisms", "for", "modulating", "its", "own", "output", ".", "LSTMs", "have", "been", "successfully", "applied", "to", "various", "tasks", ",", "such", "as", "speech", "recognition", ",", "machine", "translation", ",", "and", "natural", "language", "generation", ".", "Two", "sets", "of", "gating", "values", "(", "referred", "to", "as", "the", "input", "and", "forget", "gates", ")", "are", "first", "calculated", "based", "on", "the", "previous", "states", "of", "the", "network", ":", "where", "is", "the", "current", "input", ",", "is", "the", "previous", "hidden", "state", ",", "and", "are", "biases", ",", "is", "the", "previous", "internal", "state", "(", "referred", "to", "as", "the", "cell", ")", ",", "and", "is", "the", "logistic", "function", ".", "The", "new", "internal", "state", "is", "calculated", "based", "on", "the", "current", "input", "and", "the", "previous", "hidden", "state", ",", "and", "then", "interpolated", "with", "the", "previous", "internal", "state", "using", "and", "as", "weights", ":", "where", "is", "element", "-", "wise", "multiplication", ".", "Finally", ",", "the", "hidden", "state", "is", "calculated", "by", "passing", "the", "internal", "state", "through", "a", "nonlinearity", ",", "and", "weighting", "it", "with", ".", "The", "values", "of", "are", "conditioned", "on", "the", "new", "internal", "state", "(", ")", ",", "as", "opposed", "to", "the", "previous", "one", "(", ")", ":", "Because", "of", "the", "linear", "combination", "in", "equation", "(", "[", "reference", "]", ")", ",", "the", "LSTM", "is", "less", "susceptible", "to", "vanishing", "gradients", "over", "time", ",", "thereby", "being", "able", "to", "make", "use", "of", "longer", "context", "when", "making", "predictions", ".", "In", "addition", ",", "the", "network", "learns", "to", "modulate", "itself", ",", "effectively", "using", "the", "gates", "to", "predict", "which", "operation", "is", "required", "at", "each", "time", "step", ",", "thereby", "incorporating", "higher", "-", "level", "features", ".", "In", "order", "to", "use", "this", "architecture", "for", "error", "detection", ",", "we", "create", "a", "bidirectional", "LSTM", ",", "making", "use", "of", "the", "advanced", "features", "of", "LSTM", "and", "incorporating", "context", "on", "both", "sides", "of", "the", "target", "token", ".", "In", "addition", ",", "we", "experiment", "with", "a", "deep", "bidirectional", "LSTM", ",", "which", "includes", "two", "consecutive", "layers", "of", "bidirectional", "LSTMs", ",", "modeling", "even", "more", "complex", "features", "and", "performing", "multiple", "passes", "through", "the", "sentence", ".", "For", "comparison", "with", "non", "-", "neural", "models", ",", "we", "also", "report", "results", "using", "CRFs", ",", "which", "are", "a", "popular", "choice", "for", "sequence", "labeling", "tasks", ".", "We", "trained", "the", "CRF", "++", "implementation", "on", "the", "same", "dataset", ",", "using", "as", "features", "unigrams", ",", "bigrams", "and", "trigrams", "in", "a", "7", "-", "word", "window", "surrouding", "the", "target", "word", "(", "3", "words", "before", "and", "after", ")", ".", "The", "predicted", "label", "is", "also", "conditioned", "on", "the", "previous", "label", "in", "the", "sequence", ".", "section", ":", "Experiments", "We", "evaluate", "the", "alternative", "network", "structures", "on", "the", "publicly", "released", "First", "Certificate", "in", "English", "dataset", "(", "FCE", "-", "public", ",", "Yannakoudakis2011", ")", ".", "The", "dataset", "contains", "short", "texts", ",", "written", "by", "learners", "of", "English", "as", "an", "additional", "language", "in", "response", "to", "exam", "prompts", "eliciting", "free", "-", "text", "answers", "and", "assessing", "mastery", "of", "the", "upper", "-", "intermediate", "proficiency", "level", ".", "The", "texts", "have", "been", "manually", "error", "-", "annotated", "using", "a", "taxonomy", "of", "77", "error", "types", ".", "We", "use", "the", "released", "test", "set", "for", "evaluation", ",", "containing", "2", ",", "720", "sentences", ",", "leaving", "30", ",", "953", "sentences", "for", "training", ".", "We", "further", "separate", "2", ",", "222", "sentences", "from", "the", "training", "set", "for", "development", "and", "hyper", "-", "parameter", "tuning", ".", "The", "dataset", "contains", "manually", "annotated", "error", "spans", "of", "various", "types", "of", "errors", ",", "together", "with", "their", "suggested", "corrections", ".", "We", "convert", "this", "to", "a", "token", "-", "level", "error", "detection", "task", "by", "labeling", "each", "token", "inside", "the", "error", "span", "as", "being", "incorrect", ".", "In", "order", "to", "capture", "errors", "involving", "missing", "words", ",", "the", "error", "label", "is", "assigned", "to", "the", "token", "immediately", "after", "the", "incorrect", "gap", "–", "this", "is", "motivated", "by", "the", "intuition", "that", "while", "this", "token", "is", "correct", "when", "considered", "in", "isolation", ",", "it", "is", "incorrect", "in", "the", "current", "context", ",", "as", "another", "token", "should", "have", "preceeded", "it", ".", "As", "the", "main", "evaluation", "measure", "for", "error", "detection", "we", "use", ",", "which", "was", "also", "the", "measure", "adopted", "in", "the", "CoNLL", "-", "14", "shared", "task", "on", "error", "correction", ".", "It", "combines", "both", "precision", "and", "recall", ",", "while", "assigning", "twice", "as", "much", "weight", "to", "precision", ",", "since", "accurate", "feedback", "is", "often", "more", "important", "than", "coverage", "in", "error", "detection", "applications", ".", "Following", "Chodorow2012", ",", "we", "also", "report", "raw", "counts", "for", "predicted", "and", "correct", "tokens", ".", "Related", "evaluation", "measures", ",", "such", "as", "the", "-", "scorer", "and", "the", "I", "-", "measure", ",", "require", "the", "system", "to", "propose", "a", "correction", "and", "are", "therefore", "not", "directly", "applicable", "on", "the", "task", "of", "error", "detection", ".", "During", "the", "experiments", ",", "the", "input", "text", "was", "lowercased", "and", "all", "tokens", "that", "occurred", "less", "than", "twice", "in", "the", "training", "data", "were", "represented", "as", "a", "single", "unk", "token", ".", "Word", "embeddings", "were", "set", "to", "size", "and", "initialised", "using", "the", "publicly", "released", "pretrained", "Word2Vec", "vectors", ".", "The", "convolutional", "networks", "use", "window", "size", "on", "either", "side", "of", "the", "target", "token", "and", "produce", "a", "300", "-", "dimensional", "context", "-", "dependent", "vector", ".", "The", "recurrent", "networks", "use", "hidden", "layers", "of", "size", "200", "in", "either", "direction", ".", "We", "also", "added", "an", "extra", "hidden", "layer", "of", "size", "between", "each", "of", "the", "composition", "functions", "and", "the", "output", "layer", "–", "this", "allows", "the", "network", "to", "learn", "a", "separate", "non", "-", "linear", "transformation", "and", "reduces", "the", "dimensionality", "of", "the", "compositional", "vectors", ".", "The", "parameters", "were", "optimised", "using", "gradient", "descent", "with", "initial", "learning", "rate", ",", "the", "ADAM", "algorithm", "for", "dynamically", "adapting", "the", "learning", "rate", ",", "and", "batch", "size", "of", "64", "sentences", ".", "on", "the", "development", "set", "was", "evaluated", "at", "each", "epoch", ",", "and", "the", "best", "model", "was", "used", "for", "final", "evaluations", ".", "section", ":", "Results", "Table", "[", "reference", "]", "contains", "results", "for", "experiments", "comparing", "different", "composition", "architectures", "on", "the", "task", "of", "error", "detection", ".", "The", "CRF", "has", "the", "lowest", "score", "compared", "to", "any", "of", "the", "neural", "models", ".", "It", "memorises", "frequent", "error", "sequences", "with", "high", "precision", ",", "but", "does", "not", "generalise", "sufficiently", ",", "resulting", "in", "low", "recall", ".", "The", "ability", "to", "condition", "on", "the", "previous", "label", "also", "does", "not", "provide", "much", "help", "on", "this", "task", "–", "there", "are", "only", "two", "possible", "labels", "and", "the", "errors", "are", "relatively", "sparse", ".", "The", "architecture", "using", "convolutional", "networks", "performs", "well", "and", "achieves", "the", "second", "-", "highest", "result", "on", "the", "test", "set", ".", "It", "is", "designed", "to", "detect", "error", "patterns", "from", "a", "fixed", "window", "of", "7", "words", ",", "which", "is", "large", "enough", "to", "not", "require", "the", "use", "of", "more", "advanced", "composition", "functions", ".", "In", "contrast", ",", "the", "performance", "of", "the", "bidirectional", "recurrent", "network", "(", "Bi", "-", "RNN", ")", "is", "somewhat", "lower", ",", "especially", "on", "the", "test", "set", ".", "In", "Elman", "-", "type", "recurrent", "networks", ",", "the", "context", "signal", "from", "distant", "words", "decreases", "fairly", "rapidly", "due", "to", "the", "sigmoid", "activation", "function", "and", "diminishing", "gradients", ".", "This", "is", "likely", "why", "the", "Bi", "-", "RNN", "achieves", "the", "highest", "precision", "of", "all", "systems", "–", "the", "predicted", "label", "is", "mostly", "influenced", "by", "the", "target", "token", "and", "its", "immediate", "neighbours", ",", "allowing", "the", "network", "to", "only", "detect", "short", "high", "-", "confidence", "error", "patterns", ".", "The", "convolutional", "network", ",", "which", "uses", "7", "context", "words", "with", "equal", "attention", ",", "is", "able", "to", "outperform", "the", "Bi", "-", "RNN", "despite", "the", "fixed", "-", "size", "context", "window", ".", "The", "best", "overall", "result", "and", "highest", "is", "achieved", "by", "the", "bidirectional", "LSTM", "composition", "model", "(", "Bi", "-", "LSTM", ")", ".", "This", "architecture", "makes", "use", "of", "the", "full", "sentence", "for", "building", "context", "vectors", "on", "both", "sides", "of", "the", "target", "token", ",", "but", "improves", "on", "Bi", "-", "RNN", "by", "utilising", "a", "more", "advanced", "composition", "function", ".", "Through", "the", "application", "of", "a", "linear", "update", "for", "the", "internal", "cell", "representation", ",", "the", "LSTM", "is", "able", "to", "capture", "dependencies", "over", "longer", "distances", ".", "In", "addition", ",", "the", "gating", "functions", "allow", "it", "to", "adaptively", "decide", "which", "information", "to", "include", "in", "the", "hidden", "representations", "or", "output", "for", "error", "detection", ".", "We", "found", "that", "using", "multiple", "layers", "of", "compositional", "functions", "in", "a", "deeper", "network", "gave", "comparable", "or", "slightly", "lower", "results", "for", "all", "the", "composition", "architectures", ".", "This", "is", "in", "contrast", "to", "Irsoy2014a", ",", "who", "experimented", "with", "Elman", "-", "type", "networks", "and", "found", "some", "improvements", "using", "multiple", "layers", "of", "Bi", "-", "RNNs", ".", "The", "differences", "can", "be", "explained", "by", "their", "task", "benefiting", "from", "alternative", "features", ":", "the", "evaluation", "was", "performed", "on", "opinion", "mining", "where", "most", "target", "sequences", "are", "longer", "phrases", "that", "need", "to", "be", "identified", "based", "on", "their", "semantics", ",", "whereas", "many", "errors", "in", "learner", "writing", "are", "short", "and", "can", "only", "be", "identified", "by", "a", "contextual", "mismatch", ".", "In", "addition", ",", "our", "networks", "contain", "an", "extra", "hidden", "layer", "before", "the", "output", ",", "which", "allows", "the", "models", "to", "learn", "higher", "-", "level", "representations", "without", "adding", "complexity", "through", "an", "extra", "compositional", "layer", ".", "section", ":", "Additional", "Training", "Data", "There", "are", "essentially", "infinitely", "many", "ways", "of", "committing", "errors", "in", "text", "and", "introducing", "additional", "training", "data", "should", "alleviate", "some", "of", "the", "problems", "with", "data", "sparsity", ".", "We", "experimented", "with", "incrementally", "adding", "different", "error", "-", "tagged", "corpora", "into", "the", "training", "set", "and", "measured", "the", "resulting", "performance", ".", "This", "allows", "us", "to", "provide", "some", "context", "to", "the", "results", "obtained", "by", "using", "each", "of", "the", "datasets", ",", "and", "gives", "us", "an", "estimate", "of", "how", "much", "annotated", "data", "is", "required", "for", "optimal", "performance", "on", "error", "detection", ".", "The", "datasets", "we", "consider", "are", "as", "follows", ":", "FCE", "-", "public", "–", "the", "publicly", "released", "subset", "of", "FCE", ",", "as", "described", "in", "Section", "[", "reference", "]", ".", "NUCLE", "–", "the", "NUS", "Corpus", "of", "Learner", "English", ",", "used", "as", "the", "main", "training", "set", "for", "CoNLL", "shared", "tasks", "on", "error", "correction", ".", "IELTS", "–", "a", "subset", "of", "the", "IELTS", "examination", "dataset", "extracted", "from", "the", "Cambridge", "Learner", "Corpus", "(", "CLC", ",", "Nicholls2003", ")", ",", "containing", "68", ",", "505", "sentences", "from", "all", "proficiency", "levels", ",", "also", "used", "by", "Felice2014", ".", "FCE", "–", "a", "larger", "selection", "of", "FCE", "texts", "from", "the", "CLC", ",", "containing", "323", ",", "192", "sentences", ".", "CPE", "–", "essays", "from", "the", "proficient", "examination", "level", "in", "the", "CLC", ",", "containing", "210", ",", "678", "sentences", ".", "CAE", "–", "essays", "from", "the", "advanced", "examination", "level", "in", "the", "CLC", ",", "containing", "219", ",", "953", "sentences", ".", "Table", "[", "reference", "]", "contains", "results", "obtained", "by", "incrementally", "adding", "training", "data", "to", "the", "Bi", "-", "LSTM", "model", ".", "We", "found", "that", "incorporating", "the", "NUCLE", "dataset", "does", "not", "improve", "performance", "over", "using", "only", "the", "FCE", "-", "public", "dataset", ",", "which", "is", "likely", "due", "to", "the", "two", "corpora", "containing", "texts", "with", "different", "domains", "and", "writing", "styles", ".", "The", "texts", "in", "FCE", "are", "written", "by", "young", "intermediate", "students", ",", "in", "response", "to", "prompts", "eliciting", "letters", ",", "emails", "and", "reviews", ",", "whereas", "NUCLE", "contains", "mostly", "argumentative", "essays", "written", "by", "advanced", "adult", "learners", ".", "The", "differences", "in", "the", "datasets", "offset", "the", "benefits", "from", "additional", "training", "data", ",", "and", "the", "performance", "remains", "roughly", "the", "same", ".", "In", "contrast", ",", "substantial", "improvements", "are", "obtained", "when", "introducing", "the", "IELTS", "and", "FCE", "datasets", ",", "with", "each", "of", "them", "increasing", "the", "score", "by", "roughly", ".", "The", "IELTS", "dataset", "contains", "essays", "from", "all", "proficiency", "levels", ",", "and", "FCE", "from", "mid", "-", "level", "English", "learners", ",", "which", "provides", "the", "model", "with", "a", "distribution", "of", "‘", "average", "’", "errors", "to", "learn", "from", ".", "Adding", "even", "more", "training", "data", "from", "high", "-", "proficiency", "essays", "in", "CPE", "and", "CAE", "only", "provides", "minor", "further", "improvements", ".", "Figure", "[", "reference", "]", "also", "shows", "on", "the", "FCE", "-", "public", "test", "set", "as", "a", "function", "of", "the", "total", "number", "of", "tokens", "in", "the", "training", "data", ".", "The", "optimal", "trade", "-", "off", "between", "performance", "and", "data", "size", "is", "obtained", "at", "around", "8", "million", "tokens", ",", "after", "introducing", "the", "FCE", "dataset", ".", "section", ":", "CoNLL", "-", "14", "Shared", "Task", "The", "CoNLL", "-", "14", "shared", "task", "focussed", "on", "automatically", "correcting", "errors", "in", "learner", "writing", ".", "The", "NUCLE", "dataset", "was", "provided", "as", "the", "main", "training", "dataset", ",", "but", "participants", "were", "allowed", "to", "include", "other", "annotated", "corpora", "and", "external", "resources", ".", "For", "evaluation", ",", "25", "students", "were", "recruited", "to", "each", "write", "two", "new", "essays", ",", "which", "were", "then", "annotated", "by", "two", "experts", ".", "We", "used", "the", "same", "methods", "from", "Section", "[", "reference", "]", "for", "converting", "the", "shared", "task", "annotation", "to", "a", "token", "-", "level", "labeling", "task", "in", "order", "to", "evaluate", "the", "models", "on", "error", "detection", ".", "In", "addition", ",", "the", "correction", "outputs", "of", "all", "the", "participating", "systems", "were", "made", "available", "online", ",", "therefore", "we", "are", "able", "to", "report", "their", "performance", "on", "this", "task", ".", "In", "order", "to", "convert", "their", "output", "to", "error", "detection", "labels", ",", "the", "corrected", "sentences", "were", "aligned", "with", "the", "original", "input", "using", "Levenshtein", "distance", ",", "and", "any", "changes", "proposed", "by", "the", "system", "resulted", "in", "the", "corresponding", "source", "words", "being", "labeled", "as", "errors", ".", "The", "results", "on", "the", "two", "annotations", "of", "the", "shared", "task", "test", "data", "can", "be", "seen", "in", "Table", "[", "reference", "]", ".", "We", "first", "evaluated", "each", "of", "the", "human", "annotators", "with", "respect", "to", "the", "other", ",", "in", "order", "to", "estimate", "the", "upper", "bound", "on", "this", "task", ".", "The", "average", "of", "roughly", "50", "%", "shows", "that", "the", "task", "is", "difficult", "and", "even", "human", "experts", "have", "a", "rather", "low", "agreement", ".", "It", "has", "been", "shown", "before", "that", "correcting", "grammatical", "errors", "is", "highly", "subjective", ",", "but", "these", "results", "indicate", "that", "trained", "annotators", "can", "disagree", "even", "on", "the", "number", "and", "location", "of", "errors", ".", "In", "the", "same", "table", ",", "we", "provide", "error", "detection", "results", "for", "the", "top", "3", "participants", "in", "the", "shared", "task", ":", "CAMB", ",", "CUUI", ",", "and", "AMU", ".", "They", "each", "preserve", "their", "relative", "ranking", "also", "in", "the", "error", "detection", "evaluation", ".", "The", "CAMB", "system", "has", "a", "lower", "precision", "but", "the", "highest", "recall", ",", "also", "resulting", "in", "the", "highest", ".", "CUUI", "and", "AMU", "are", "close", "in", "performance", ",", "with", "AMU", "having", "slightly", "higher", "precision", ".", "After", "the", "official", "shared", "task", ",", "Susanto2014", "published", "a", "system", "which", "combines", "several", "alternative", "models", "and", "outperforms", "the", "shared", "task", "participants", "when", "evaluated", "on", "error", "correction", ".", "However", ",", "on", "error", "detection", "it", "receives", "lower", "results", ",", "ranking", "3rd", "and", "4th", "when", "evaluated", "on", "(", "P1", "+", "P2", "+", "S1", "+", "S2", "in", "Table", "[", "reference", "]", ")", ".", "The", "system", "has", "detected", "a", "small", "number", "of", "errors", "with", "high", "precision", ",", "and", "does", "not", "reach", "the", "highest", ".", "Finally", ",", "we", "present", "results", "for", "the", "Bi", "-", "LSTM", "sequence", "labeling", "system", "for", "error", "detection", ".", "Using", "only", "FCE", "-", "public", "for", "training", ",", "the", "overall", "performance", "is", "rather", "low", "as", "the", "training", "set", "is", "very", "small", "and", "contains", "texts", "from", "a", "different", "domain", ".", "However", ",", "these", "results", "show", "that", "the", "model", "behaves", "as", "expected", "–", "since", "it", "has", "not", "seen", "similar", "language", "during", "training", ",", "it", "labels", "a", "very", "large", "portion", "of", "tokens", "as", "errors", ".", "This", "indicates", "that", "the", "network", "is", "trying", "to", "learn", "correct", "language", "constructions", "from", "the", "limited", "data", "and", "classifies", "unseen", "structures", "as", "errors", ",", "as", "opposed", "to", "simply", "memorising", "error", "sequences", "from", "the", "training", "data", ".", "When", "trained", "on", "all", "the", "datasets", "from", "Section", "[", "reference", "]", ",", "the", "model", "achieves", "the", "highest", "of", "all", "systems", "on", "both", "of", "the", "CoNLL", "-", "14", "shared", "task", "test", "annotations", ",", "with", "an", "absolute", "improvement", "of", "over", "the", "previous", "best", "result", ".", "It", "is", "worth", "noting", "that", "the", "full", "Bi", "-", "LSTM", "has", "been", "trained", "on", "more", "data", "than", "the", "other", "CoNLL", "contestants", ".", "However", ",", "as", "the", "shared", "task", "systems", "were", "not", "restricted", "to", "the", "NUCLE", "training", "set", ",", "all", "the", "submissions", "also", "used", "differing", "amounts", "of", "training", "data", "from", "various", "sources", ".", "In", "addition", ",", "the", "CoNLL", "systems", "are", "mostly", "combinations", "of", "many", "alternative", "models", ":", "the", "CAMB", "system", "is", "a", "hybrid", "of", "machine", "translation", ",", "a", "rule", "-", "based", "system", ",", "and", "a", "language", "model", "re", "-", "ranker", ";", "CUUI", "consists", "of", "different", "classifiers", "for", "each", "individual", "error", "type", ";", "and", "P1", "+", "P2", "+", "S1", "+", "S2", "is", "a", "combination", "of", "four", "different", "error", "correction", "systems", ".", "In", "contrast", ",", "the", "Bi", "-", "LSTM", "is", "a", "single", "model", "for", "detecting", "all", "error", "types", ",", "and", "therefore", "represents", "a", "more", "scalable", "data", "-", "driven", "approach", ".", "section", ":", "Essay", "Scoring", "In", "this", "section", ",", "we", "perform", "an", "extrinsic", "evaluation", "of", "the", "efficacy", "of", "the", "error", "detection", "system", "and", "examine", "the", "extent", "to", "which", "it", "generalises", "at", "higher", "levels", "of", "granularity", "on", "the", "task", "of", "automated", "essay", "scoring", ".", "More", "specifically", ",", "we", "replicate", "experiments", "using", "the", "text", "-", "level", "model", "described", "by", "Andersen2013", ",", "which", "is", "currently", "deployed", "in", "a", "self", "-", "assessment", "and", "tutoring", "system", "(", "SAT", ")", ",", "an", "online", "automated", "writing", "feedback", "tool", "actively", "used", "by", "language", "learners", ".", "The", "SAT", "system", "predicts", "an", "overall", "score", "for", "a", "given", "text", ",", "which", "provides", "a", "holistic", "assessment", "of", "linguistic", "competence", "and", "language", "proficiency", ".", "The", "authors", "trained", "a", "supervised", "ranking", "perceptron", "model", "on", "the", "FCE", "-", "public", "dataset", ",", "using", "features", "such", "as", "error", "-", "rate", "estimates", "from", "a", "language", "model", "and", "various", "lexical", "and", "grammatical", "properties", "of", "text", "(", "e.g.", ",", "word", "n", "-", "grams", ",", "part", "-", "of", "-", "speech", "n", "-", "grams", "and", "phrase", "-", "structure", "rules", ")", ".", "We", "replicate", "this", "experiment", "and", "add", "the", "average", "probability", "of", "each", "token", "in", "the", "essay", "being", "correct", ",", "according", "to", "the", "error", "detection", "model", ",", "as", "an", "additional", "feature", "for", "the", "scoring", "framework", ".", "The", "system", "was", "then", "retrained", "on", "FCE", "-", "public", "and", "evaluated", "on", "correctly", "predicting", "the", "assigned", "essay", "score", ".", "Table", "[", "reference", "]", "presents", "the", "experimental", "results", ".", "The", "human", "performance", "on", "the", "test", "set", "is", "calculated", "as", "the", "average", "inter", "-", "annotator", "correlation", "on", "the", "same", "data", ",", "and", "the", "existing", "SAT", "system", "has", "demonstrated", "levels", "of", "performance", "that", "are", "very", "close", "to", "that", "of", "human", "assessors", ".", "Nevertheless", ",", "the", "Bi", "-", "LSTM", "model", "trained", "only", "on", "FCE", "-", "public", "complements", "the", "existing", "features", ",", "and", "the", "combined", "model", "achieves", "an", "absolute", "improvement", "of", "around", "1", "%", "percent", ",", "corresponding", "to", "20", "-", "31", "%", "relative", "error", "reduction", "with", "respect", "to", "the", "human", "performance", ".", "Even", "though", "the", "Bi", "-", "LSTM", "is", "trained", "on", "the", "same", "dataset", "and", "the", "SAT", "system", "already", "includes", "various", "linguistic", "features", "for", "capturing", "errors", ",", "our", "error", "detection", "model", "manages", "to", "further", "improve", "its", "performance", ".", "When", "the", "Bi", "-", "LSTM", "is", "trained", "on", "all", "the", "available", "data", "from", "Section", "[", "reference", "]", ",", "the", "combination", "achieves", "further", "substantial", "improvements", ".", "The", "relative", "error", "reduction", "on", "Pearson", "’s", "correlation", "is", "64", "%", ",", "and", "the", "system", "actually", "outperforms", "human", "annotators", "on", "Spearman", "’s", "correlation", ".", "section", ":", "Conclusions", "In", "this", "paper", ",", "we", "presented", "the", "first", "experiments", "using", "neural", "network", "models", "for", "the", "task", "of", "error", "detection", "in", "learner", "writing", ".", "Six", "alternative", "compositional", "network", "architectures", "for", "modeling", "context", "were", "evaluated", ".", "Based", "on", "the", "findings", ",", "we", "propose", "a", "novel", "error", "detection", "framework", "using", "token", "-", "level", "embeddings", ",", "bidirectional", "LSTMs", "for", "context", "representation", ",", "and", "a", "multi", "-", "layer", "architecture", "for", "learning", "more", "complex", "features", ".", "This", "structure", "allows", "the", "model", "to", "classify", "each", "token", "as", "being", "correct", "or", "incorrect", ",", "using", "the", "full", "sentence", "as", "context", ".", "The", "self", "-", "modulation", "architecture", "of", "LSTMs", "was", "also", "shown", "to", "be", "beneficial", ",", "as", "it", "allows", "the", "network", "to", "learn", "more", "advanced", "composition", "rules", "and", "remember", "dependencies", "over", "longer", "distances", ".", "Substantial", "performance", "improvements", "were", "achieved", "by", "training", "the", "best", "model", "on", "additional", "datasets", ".", "We", "found", "that", "the", "largest", "benefit", "was", "obtained", "from", "training", "on", "8", "million", "tokens", "of", "text", "from", "learners", "with", "varying", "levels", "of", "language", "proficiency", ".", "In", "contrast", ",", "including", "even", "more", "data", "from", "higher", "-", "proficiency", "learners", "gave", "marginal", "further", "improvements", ".", "As", "part", "of", "future", "work", ",", "it", "would", "be", "beneficial", "to", "investigate", "the", "effect", "of", "automatically", "generated", "training", "data", "for", "error", "detection", "(", "e.g.", ",", "Rozovskaya2010", ")", ".", "We", "evaluated", "the", "performance", "of", "existing", "error", "correction", "systems", "from", "CoNLL", "-", "14", "on", "the", "task", "of", "error", "detection", ".", "The", "experiments", "showed", "that", "success", "on", "error", "correction", "does", "not", "necessarily", "mean", "success", "on", "error", "detection", ",", "as", "the", "current", "best", "correction", "system", "(", "P1", "+", "P2", "+", "S1", "+", "S2", ")", "is", "not", "the", "same", "as", "the", "best", "shared", "task", "detection", "system", "(", "CAMB", ")", ".", "In", "addition", ",", "the", "neural", "sequence", "tagging", "model", ",", "specialised", "for", "error", "detection", ",", "was", "able", "to", "outperform", "all", "other", "participating", "systems", ".", "Finally", ",", "we", "performed", "an", "extrinsic", "evaluation", "by", "incorporating", "probabilities", "from", "the", "error", "detection", "system", "as", "features", "in", "an", "essay", "scoring", "model", ".", "Even", "without", "any", "additional", "data", ",", "the", "combination", "further", "improved", "performance", "which", "is", "already", "close", "to", "the", "results", "from", "human", "annotators", ".", "In", "addition", ",", "when", "the", "error", "detection", "model", "was", "trained", "on", "a", "larger", "training", "set", ",", "the", "essay", "scorer", "was", "able", "to", "exceed", "human", "-", "level", "performance", ".", "section", ":", "Acknowledgments", "We", "would", "like", "to", "thank", "Prof", "Ted", "Briscoe", "and", "the", "reviewers", "for", "providing", "useful", "feedback", ".", "bibliography", ":", "References"], "entity": [[2, 6, ["Compositional", "Sequence", "Labeling", "Models"], "Method"], [7, 9, ["Error", "Detection"], "Task"], [10, 12, ["Learner", "Writing"], "Task"], [22, 25, ["neural", "network", "models"], "Method"], [29, 31, ["error", "detection"], "Task"], [32, 34, ["learner", "writing"], "Task"], [42, 44, ["compositional", "architectures"], "Method"], [49, 51, ["error", "detection"], "Task"], [59, 65, ["CoNLL", "-", "14", "shared", "task", "dataset"], "Material"], [75, 80, ["detecting", "errors", "in", "learner", "writing"], "Task"], [107, 109, ["Automated", "systems"], "Method"], [110, 112, ["detecting", "errors"], "Task"], [113, 115, ["learner", "writing"], "Task"], [119, 122, ["second", "language", "learning"], "Task"], [123, 124, ["assessment"], "Task"], [133, 135, ["error", "correction"], "Task"], [137, 139, ["error", "detection"], "Task"], [146, 148, ["correction", "output"], "Metric"], [170, 171, ["correction"], "Task"], [176, 177, ["detection"], "Task"], [194, 197, ["supervised", "classification", "approach"], "Method"], [198, 203, ["content", "-", "content", "word", "errors"], "Metric"], [216, 219, ["error", "correction", "frameworks"], "Method"], [220, 224, ["Evaluation", "of", "error", "correction"], "Task"], [245, 247, ["error", "detection"], "Task"], [248, 250, ["learner", "writing"], "Task"], [252, 254, ["independent", "task"], "Task"], [274, 277, ["sequence", "labeling", "tasks"], "Task"], [352, 355, ["Neural", "network", "models"], "Method"], [366, 368, ["NLP", "tasks"], "Task"], [370, 372, ["language", "modeling"], "Task"], [373, 375, ["speech", "recognition"], "Task"], [379, 381, ["machine", "translation"], "Task"], [394, 398, ["fixed", "-", "size", "vector"], "Method"], [399, 401, ["convolutional", "networks"], "Method"], [402, 405, ["recurrent", "neural", "networks"], "Method"], [416, 419, ["neural", "network", "models"], "Method"], [423, 425, ["error", "detection"], "Task"], [426, 428, ["learner", "writing"], "Task"], [436, 438, ["compositional", "structures"], "Method"], [440, 443, ["informative", "context", "representations"], "Task"], [456, 458, ["error", "detection"], "Task"], [459, 461, ["learner", "writing"], "Task"], [483, 486, ["sequence", "labeling", "model"], "Method"], [489, 493, ["variable", "-", "size", "network"], "Method"], [538, 541, ["error", "detection", "framework"], "Method"], [567, 570, ["automatically", "detecting", "errors"], "Task"], [571, 573, ["learner", "text"], "Task"], [619, 622, ["error", "detection", "systems"], "Method"], [643, 647, ["chi", "-", "square", "statistic"], "Method"], [671, 675, ["Maximum", "Entropy", "Markov", "Models"], "Method"], [693, 695, ["constituency", "parser"], "Method"], [705, 708, ["grammatical", "error", "detection"], "Task"], [747, 750, ["error", "detection", "systems"], "Task"], [756, 759, ["neural", "compositional", "models"], "Method"], [767, 770, ["grammatical", "error", "correction"], "Task"], [808, 815, ["Phrase", "-", "based", "statistical", "machine", "translation", "techniques"], "Method"], [827, 828, ["Perceptrons"], "Method"], [829, 832, ["Naive", "Bayes", "classifiers"], "Method"], [842, 845, ["Error", "correction", "systems"], "Method"], [868, 872, ["tagging", "and", "classification", "models"], "Method"], [876, 878, ["error", "detection"], "Task"], [882, 886, ["automated", "writing", "feedback", "systems"], "Task"], [925, 929, ["neural", "sequence", "tagging", "model"], "Method"], [949, 951, ["error", "detection"], "Task"], [954, 957, ["Sequence", "Labeling", "Architectures"], "Method"], [960, 965, ["neural", "network", "sequence", "labeling", "framework"], "Method"], [969, 971, ["error", "detection"], "Task"], [972, 974, ["learner", "writing"], "Task"], [1010, 1012, ["vector", "representations"], "Method"], [1026, 1028, ["composition", "functions"], "Method"], [1034, 1037, ["hidden", "vector", "representation"], "Method"], [1050, 1052, ["softmax", "layer"], "Method"], [1078, 1080, ["output", "layer"], "Method"], [1085, 1088, ["neural", "network", "architectures"], "Method"], [1092, 1094, ["error", "detection"], "Task"], [1095, 1096, ["convolutional"], "Method"], [1097, 1099, ["bidirectional", "recurrent"], "Method"], [1104, 1108, ["multi", "-", "layer", "variants"], "Method"], [1115, 1118, ["convolutional", "neural", "network"], "Method"], [1119, 1120, ["CNN"], "Method"], [1128, 1130, ["token", "labeling"], "Task"], [1146, 1147, ["convolution"], "Method"], [1150, 1152, ["feedforward", "network"], "Method"], [1181, 1183, ["convolutional", "architecture"], "Method"], [1197, 1199, ["error", "detection"], "Task"], [1201, 1202, ["convolution"], "Method"], [1244, 1248, ["non", "-", "linear", "layer"], "Method"], [1251, 1253, ["hidden", "representation"], "Method"], [1255, 1258, ["deep", "convolutional", "network"], "Method"], [1268, 1270, ["convolutional", "layer"], "Method"], [1283, 1286, ["convolutions", "of", "convolutions"], "Method"], [1301, 1304, ["recurrent", "neural", "network"], "Method"], [1305, 1307, ["RNN", ")"], "Method"], [1309, 1311, ["hidden", "representation"], "Method"], [1338, 1340, ["sigmoid", "function"], "Method"], [1356, 1358, ["recursive", "function"], "Method"], [1374, 1377, ["recurrent", "network", "structure"], "Method"], [1382, 1386, ["Elman", "-", "type", "network"], "Method"], [1391, 1393, ["bidirectional", "RNN"], "Method"], [1403, 1405, ["recurrent", "components"], "Method"], [1416, 1418, ["unidirectional", "version"], "Method"], [1432, 1434, ["bidirectional", "version"], "Method"], [1437, 1439, ["context", "representations"], "Method"], [1459, 1461, ["hidden", "representation"], "Method"], [1462, 1464, ["Recurrent", "networks"], "Method"], [1474, 1476, ["language", "modeling"], "Task"], [1481, 1484, ["incremental", "composition", "function"], "Method"], [1496, 1498, ["language", "models"], "Method"], [1519, 1521, ["error", "detection"], "Task"], [1523, 1525, ["composition", "function"], "Method"], [1542, 1544, ["bidirectional", "model"], "Method"], [1577, 1584, ["bidirectional", "Elman", "-", "type", "recurrent", "network", "modules"], "Method"], [1586, 1589, ["deep", "bidirectional", "RNN"], "Method"], [1598, 1602, ["context", "-", "dependent", "representation"], "Method"], [1607, 1609, ["bidirectional", "RNN"], "Method"], [1618, 1620, ["bidirectional", "RNN"], "Method"], [1648, 1653, ["long", "-", "short", "term", "memory"], "Method"], [1654, 1655, ["LSTM"], "Method"], [1662, 1666, ["Elman", "-", "type", "networks"], "Method"], [1689, 1691, ["gating", "mechanisms"], "Method"], [1697, 1698, ["LSTMs"], "Method"], [1708, 1710, ["speech", "recognition"], "Task"], [1711, 1713, ["machine", "translation"], "Task"], [1715, 1718, ["natural", "language", "generation"], "Task"], [1778, 1780, ["logistic", "function"], "Method"], [1864, 1866, ["linear", "combination"], "Method"], [1875, 1876, ["LSTM"], "Method"], [1896, 1897, ["predictions"], "Task"], [1937, 1939, ["error", "detection"], "Task"], [1953, 1954, ["LSTM"], "Method"], [1972, 1975, ["deep", "bidirectional", "LSTM"], "Method"], [2001, 2005, ["non", "-", "neural", "models"], "Method"], [2011, 2012, ["CRFs"], "Method"], [2019, 2022, ["sequence", "labeling", "tasks"], "Task"], [2026, 2029, ["CRF", "++", "implementation"], "Method"], [2081, 2083, ["network", "structures"], "Method"], [2087, 2092, ["First", "Certificate", "in", "English", "dataset"], "Material"], [2093, 2096, ["FCE", "-", "public"], "Material"], [2189, 2193, ["hyper", "-", "parameter", "tuning"], "Method"], [2218, 2224, ["token", "-", "level", "error", "detection", "task"], "Task"], [2295, 2297, ["evaluation", "measure"], "Metric"], [2298, 2300, ["error", "detection"], "Task"], [2311, 2316, ["CoNLL", "-", "14", "shared", "task"], "Material"], [2317, 2319, ["error", "correction"], "Task"], [2323, 2324, ["precision"], "Metric"], [2325, 2326, ["recall"], "Metric"], [2334, 2335, ["precision"], "Metric"], [2346, 2349, ["error", "detection", "applications"], "Task"], [2365, 2367, ["evaluation", "measures"], "Metric"], [2371, 2373, ["-", "scorer"], "Method"], [2375, 2378, ["I", "-", "measure"], "Metric"], [2385, 2386, ["correction"], "Task"], [2396, 2398, ["error", "detection"], "Task"], [2445, 2447, ["convolutional", "networks"], "Method"], [2469, 2471, ["recurrent", "networks"], "Method"], [2509, 2513, ["non", "-", "linear", "transformation"], "Method"], [2527, 2529, ["gradient", "descent"], "Method"], [2531, 2533, ["learning", "rate"], "Metric"], [2535, 2537, ["ADAM", "algorithm"], "Method"], [2541, 2543, ["learning", "rate"], "Metric"], [2584, 2586, ["composition", "architectures"], "Method"], [2590, 2592, ["error", "detection"], "Task"], [2594, 2595, ["CRF"], "Method"], [2604, 2606, ["neural", "models"], "Method"], [2614, 2615, ["precision"], "Metric"], [2625, 2626, ["recall"], "Metric"], [2661, 2663, ["convolutional", "networks"], "Method"], [2714, 2717, ["bidirectional", "recurrent", "network"], "Method"], [2718, 2721, ["Bi", "-", "RNN"], "Method"], [2733, 2738, ["Elman", "-", "type", "recurrent", "networks"], "Method"], [2763, 2766, ["Bi", "-", "RNN"], "Method"], [2769, 2770, ["precision"], "Metric"], [2803, 2805, ["convolutional", "network"], "Method"], [2820, 2823, ["Bi", "-", "RNN"], "Method"], [2841, 2845, ["bidirectional", "LSTM", "composition", "model"], "Method"], [2874, 2877, ["Bi", "-", "RNN"], "Method"], [2882, 2884, ["composition", "function"], "Method"], [2890, 2892, ["linear", "update"], "Method"], [2894, 2897, ["internal", "cell", "representation"], "Method"], [2899, 2900, ["LSTM"], "Method"], [2931, 2933, ["error", "detection"], "Task"], [2945, 2947, ["deeper", "network"], "Method"], [2956, 2958, ["composition", "architectures"], "Method"], [2969, 2973, ["Elman", "-", "type", "networks"], "Method"], [2978, 2984, ["multiple", "layers", "of", "Bi", "-", "RNNs"], "Method"], [3003, 3005, ["opinion", "mining"], "Task"], [3026, 3028, ["learner", "writing"], "Task"], [3070, 3072, ["compositional", "layer"], "Method"], [3158, 3160, ["error", "detection"], "Task"], [3169, 3172, ["FCE", "-", "public"], "Material"], [3178, 3179, ["FCE"], "Material"], [3204, 3207, ["CoNLL", "shared", "tasks"], "Task"], [3208, 3210, ["error", "correction"], "Task"], [3211, 3212, ["IELTS"], "Method"], [3247, 3248, ["FCE"], "Material"], [3253, 3255, ["FCE", "texts"], "Material"], [3257, 3258, ["CLC"], "Method"], [3265, 3266, ["CPE"], "Method"], [3275, 3276, ["CLC"], "Method"], [3283, 3284, ["CAE"], "Method"], [3293, 3294, ["CLC"], "Method"], [3360, 3361, ["FCE"], "Material"], [3424, 3426, ["FCE", "datasets"], "Material"], [3448, 3449, ["FCE"], "Material"], [3527, 3529, ["data", "size"], "Metric"], [3540, 3542, ["FCE", "dataset"], "Material"], [3545, 3550, ["CoNLL", "-", "14", "Shared", "Task"], "Material"], [3551, 3556, ["CoNLL", "-", "14", "shared", "task"], "Material"], [3558, 3561, ["automatically", "correcting", "errors"], "Task"], [3562, 3564, ["learner", "writing"], "Task"], [3624, 3627, ["shared", "task", "annotation"], "Task"], [3629, 3634, ["token", "-", "level", "labeling", "task"], "Task"], [3641, 3643, ["error", "detection"], "Task"], [3820, 3822, ["error", "detection"], "Task"], [3830, 3832, ["shared", "task"], "Task"], [3833, 3834, ["CAMB"], "Method"], [3835, 3836, ["CUUI"], "Method"], [3838, 3839, ["AMU"], "Method"], [3849, 3852, ["error", "detection", "evaluation"], "Task"], [3854, 3856, ["CAMB", "system"], "Method"], [3859, 3860, ["precision"], "Metric"], [3863, 3864, ["recall"], "Metric"], [3871, 3872, ["CUUI"], "Method"], [3873, 3874, ["AMU"], "Method"], [3880, 3881, ["AMU"], "Method"], [3884, 3885, ["precision"], "Metric"], [3888, 3891, ["official", "shared", "task"], "Task"], [3910, 3912, ["error", "correction"], "Task"], [3916, 3918, ["error", "detection"], "Task"], [3956, 3957, ["precision"], "Metric"], [3979, 3981, ["error", "detection"], "Task"], [3984, 3987, ["FCE", "-", "public"], "Material"], [3988, 3989, ["training"], "Task"], [4141, 4143, ["CoNLL", "contestants"], "Material"], [4148, 4151, ["shared", "task", "systems"], "Method"], [4178, 4180, ["CoNLL", "systems"], "Method"], [4189, 4191, ["CAMB", "system"], "Method"], [4195, 4197, ["machine", "translation"], "Method"], [4199, 4203, ["rule", "-", "based", "system"], "Method"], [4206, 4211, ["language", "model", "re", "-", "ranker"], "Method"], [4212, 4213, ["CUUI"], "Method"], [4216, 4217, ["classifiers"], "Method"], [4224, 4231, ["P1", "+", "P2", "+", "S1", "+", "S2"], "Method"], [4237, 4240, ["error", "correction", "systems"], "Method"], [4255, 4257, ["error", "types"], "Task"], [4264, 4268, ["data", "-", "driven", "approach"], "Method"], [4271, 4273, ["Essay", "Scoring"], "Metric"], [4287, 4290, ["error", "detection", "system"], "Method"], [4307, 4310, ["automated", "essay", "scoring"], "Task"], [4319, 4323, ["text", "-", "level", "model"], "Method"], [4333, 4336, ["self", "-", "assessment"], "Method"], [4337, 4339, ["tutoring", "system"], "Method"], [4340, 4341, ["SAT"], "Method"], [4344, 4349, ["online", "automated", "writing", "feedback", "tool"], "Task"], [4352, 4354, ["language", "learners"], "Task"], [4370, 4375, ["holistic", "assessment", "of", "linguistic", "competence"], "Task"], [4376, 4378, ["language", "proficiency"], "Task"], [4383, 4387, ["supervised", "ranking", "perceptron", "model"], "Method"], [4398, 4402, ["error", "-", "rate", "estimates"], "Metric"], [4404, 4406, ["language", "model"], "Method"], [4458, 4461, ["error", "detection", "model"], "Method"], [4468, 4470, ["scoring", "framework"], "Method"], [4487, 4489, ["essay", "score"], "Metric"], [4510, 4515, ["average", "inter", "-", "annotator", "correlation"], "Metric"], [4550, 4553, ["FCE", "-", "public"], "Material"], [4578, 4581, ["relative", "error", "reduction"], "Metric"], [4614, 4617, ["error", "detection", "model"], "Method"], [4650, 4653, ["relative", "error", "reduction"], "Metric"], [4654, 4657, ["Pearson", "’s", "correlation"], "Metric"], [4669, 4672, ["Spearman", "’s", "correlation"], "Metric"], [4686, 4689, ["neural", "network", "models"], "Method"], [4693, 4695, ["error", "detection"], "Task"], [4696, 4698, ["learner", "writing"], "Task"], [4701, 4704, ["compositional", "network", "architectures"], "Method"], [4705, 4707, ["modeling", "context"], "Task"], [4719, 4722, ["error", "detection", "framework"], "Method"], [4723, 4727, ["token", "-", "level", "embeddings"], "Method"], [4731, 4733, ["context", "representation"], "Method"], [4736, 4740, ["multi", "-", "layer", "architecture"], "Method"], [4769, 4773, ["self", "-", "modulation", "architecture"], "Method"], [4774, 4775, ["LSTMs"], "Method"], [4876, 4878, ["error", "detection"], "Task"], [4890, 4893, ["error", "correction", "systems"], "Method"], [4894, 4897, ["CoNLL", "-", "14"], "Material"], [4901, 4903, ["error", "detection"], "Task"], [4910, 4912, ["error", "correction"], "Task"], [4918, 4920, ["error", "detection"], "Task"], [4925, 4927, ["correction", "system"], "Method"], [4928, 4935, ["P1", "+", "P2", "+", "S1", "+", "S2"], "Method"], [4943, 4947, ["shared", "task", "detection", "system"], "Method"], [4948, 4949, ["CAMB"], "Method"], [4955, 4959, ["neural", "sequence", "tagging", "model"], "Method"], [4962, 4964, ["error", "detection"], "Task"], [4979, 4981, ["extrinsic", "evaluation"], "Task"], [4986, 4989, ["error", "detection", "system"], "Method"], [4993, 4996, ["essay", "scoring", "model"], "Method"], [5024, 5027, ["error", "detection", "model"], "Method"], [5036, 5038, ["essay", "scorer"], "Metric"], [54, 55, ["LSTMs"], "Method"], [91, 94, ["self", "-", "assessment"], "Method"], [545, 548, ["self", "-", "assessment"], "Method"], [1101, 1102, ["LSTM"], "Method"], [1944, 1945, ["LSTM"], "Method"], [1983, 1984, ["LSTMs"], "Method"], [2848, 2849, ["LSTM"], "Method"], [3317, 3318, ["LSTM"], "Method"], [3335, 3338, ["FCE", "-", "public"], "Material"], [3500, 3503, ["FCE", "-", "public"], "Material"], [3974, 3975, ["LSTM"], "Method"], [4103, 4108, ["CoNLL", "-", "14", "shared", "task"], "Material"], [4131, 4132, ["LSTM"], "Method"], [4247, 4248, ["LSTM"], "Method"], [4356, 4357, ["SAT"], "Method"], [4389, 4392, ["FCE", "-", "public"], "Material"], [4477, 4480, ["FCE", "-", "public"], "Material"], [4523, 4524, ["SAT"], "Method"], [4545, 4546, ["LSTM"], "Method"], [4593, 4594, ["LSTM"], "Method"], [4602, 4603, ["SAT"], "Method"], [4628, 4629, ["LSTM"], "Method"], [4729, 4730, ["LSTMs"], "Method"]]} {"doc_id": "0a6c36de8726b6feaab586046ddc1d1a008f44f9", "text": ["document", ":", "Filtered", "Channel", "Features", "for", "Pedestrian", "Detection", "This", "paper", "starts", "from", "the", "observation", "that", "multiple", "top", "performing", "pedestrian", "detectors", "can", "be", "modelled", "by", "using", "an", "intermediate", "layer", "filtering", "low", "-", "level", "features", "in", "combination", "with", "a", "boosted", "decision", "forest", ".", "Based", "on", "this", "observation", "we", "propose", "a", "unifying", "framework", "and", "experimentally", "explore", "different", "filter", "families", ".", "We", "report", "extensive", "results", "enabling", "a", "systematic", "analysis", ".", "Using", "filtered", "channel", "features", "we", "obtain", "top", "performance", "on", "the", "challenging", "Caltech", "and", "KITTI", "datasets", ",", "while", "using", "only", "HOG", "+", "LUV", "as", "low", "-", "level", "features", ".", "When", "adding", "optical", "flow", "features", "we", "further", "improve", "detection", "quality", "and", "report", "the", "best", "known", "results", "on", "the", "Caltech", "dataset", ",", "reaching", "93", "%", "recall", "at", "1", "FPPI", ".", "=", "-", "1", "section", ":", "Introduction", "Pedestrian", "detection", "is", "an", "active", "research", "area", ",", "with", "1000", "+", "papers", "published", "in", "the", "last", "decade", ",", "and", "well", "established", "benchmark", "datasets", ".", "It", "is", "considered", "a", "canonical", "case", "of", "object", "detection", ",", "and", "has", "served", "as", "playground", "to", "explore", "ideas", "that", "might", "be", "effective", "for", "generic", "object", "detection", ".", "Although", "many", "different", "ideas", "have", "been", "explored", ",", "and", "detection", "quality", "has", "been", "steadily", "improving", ",", "arguably", "it", "is", "still", "unclear", "what", "are", "the", "key", "ingredients", "for", "good", "pedestrian", "detection", ";", "e.g.", "it", "remains", "unclear", "how", "effective", "parts", ",", "components", ",", "and", "features", "learning", "are", "for", "this", "task", ".", "Current", "top", "performing", "pedestrian", "detection", "methods", "all", "point", "to", "an", "intermediate", "layer", "(", "such", "as", "max", "-", "pooling", "or", "filtering", ")", "between", "the", "low", "-", "level", "feature", "maps", "and", "the", "classification", "layer", ".", "In", "this", "paper", "we", "explore", "the", "simplest", "of", "such", "intermediary", ":", "a", "linear", "transformation", "implemented", "as", "convolution", "with", "a", "filter", "bank", ".", "We", "propose", "a", "framework", "for", "filtered", "channel", "features", "(", "see", "figure", "[", "reference", "]", ")", "that", "unifies", "multiple", "top", "performing", "methods", ",", "and", "that", "enables", "a", "systematic", "exploration", "of", "different", "filter", "banks", ".", "With", "our", "experiments", "we", "show", "that", ",", "with", "the", "proper", "filter", "bank", ",", "filtered", "channel", "features", "reach", "top", "detection", "quality", ".", "It", "has", "been", "shown", "that", "using", "extra", "information", "at", "test", "time", "(", "such", "as", "context", ",", "stereo", "images", ",", "optical", "flow", ",", "etc", ".", ")", "can", "boost", "detection", "quality", ".", "In", "this", "paper", "we", "focus", "on", "the", "‘", "‘", "core", "’", "’", "sliding", "window", "algorithm", "using", "solely", "HOG", "+", "LUV", "features", "(", "i.e.", "oriented", "gradient", "magnitude", "and", "colour", "features", ")", ".", "We", "consider", "context", "information", "and", "optical", "flow", "as", "add", "-", "ons", ",", "included", "in", "the", "experiments", "section", "for", "the", "sake", "of", "completeness", "and", "comparison", "with", "existing", "methods", ".", "Using", "only", "HOG", "+", "LUV", "features", "we", "already", "reach", "top", "performance", "on", "the", "challenging", "Caltech", "and", "KITTI", "datasets", ",", "matching", "results", "using", "optical", "flow", "and", "significantly", "more", "features", "(", "such", "as", "LBP", "and", "covariance", ")", ".", "subsection", ":", "Related", "work", "Recent", "survey", "papers", "discuss", "the", "diverse", "set", "of", "ideas", "explored", "for", "pedestrian", "detection", ".", "The", "most", "recent", "survey", "indicates", "that", "the", "classifier", "choice", "(", "e.g.", "linear", "/", "non", "-", "linear", "SVM", "versus", "decision", "forest", ")", "is", "not", "a", "clear", "differentiator", "regarding", "quality", ";", "rather", "the", "features", "used", "seem", "more", "important", ".", "Creativity", "regarding", "different", "types", "of", "features", "has", "not", "been", "lacking", ".", "HOG", ")", "The", "classic", "HOG", "descriptor", "is", "based", "on", "local", "image", "differences", "(", "plus", "pooling", "and", "normalization", "steps", ")", ",", "and", "has", "been", "used", "directly", ",", "as", "input", "for", "a", "deformable", "parts", "model", ",", "or", "as", "features", "to", "be", "boosted", ".", "The", "integral", "channel", "features", "detector", "uses", "a", "simpler", "HOG", "variant", "with", "sum", "pooling", "and", "no", "normalizations", ".", "Many", "extensions", "of", "HOG", "have", "been", "proposed", "(", "e.g.", ")", ".", "LBP", ")", "Instead", "of", "using", "the", "magnitude", "of", "local", "pixel", "differences", ",", "LBP", "uses", "the", "difference", "sign", "only", "as", "signal", ".", "Colour", ")", "Although", "the", "appearance", "of", "pedestrians", "is", "diverse", ",", "the", "background", "and", "skin", "areas", "do", "exhibit", "a", "colour", "bias", ".", "Colour", "has", "shown", "to", "be", "an", "effective", "feature", "for", "pedestrian", "detection", "and", "hence", "multiple", "colour", "spaces", "have", "been", "explored", "(", "both", "hand", "-", "crafted", "and", "learned", ")", ".", "Local", "structure", ")", "Instead", "of", "simple", "pixel", "values", ",", "some", "approaches", "try", "to", "encode", "a", "larger", "local", "structure", "based", "on", "colour", "similarities", "(", "soft", "-", "cue", ")", ",", "segmentation", "methods", "(", "hard", "-", "decision", ")", ",", "or", "by", "estimating", "local", "boundaries", ".", "Covariance", ")", "Another", "popular", "way", "to", "encode", "richer", "information", "is", "to", "compute", "the", "covariance", "amongst", "features", "(", "commonly", "colour", ",", "gradient", ",", "and", "oriented", "gradient", ")", ".", "Etc", ".", ")", "Other", "features", "include", "bag", "-", "of", "-", "words", "over", "colour", ",", "HOG", ",", "or", "LBP", "features", ";", "learning", "sparse", "dictionary", "encoders", ";", "and", "training", "features", "via", "a", "convolutional", "neural", "network", ".", "Additional", "features", "specific", "for", "stereo", "depth", "or", "optical", "flow", "have", "been", "proposed", ",", "however", "we", "consider", "these", "beyond", "the", "focus", "of", "this", "paper", ".", "For", "our", "flow", "experiments", "we", "will", "use", "difference", "of", "frames", "from", "weakly", "stabilized", "videos", "(", "SDt", ")", ".", "All", "the", "feature", "types", "listed", "above", "can", "be", "used", "in", "the", "integral", "channel", "features", "detector", "framework", ".", "This", "family", "of", "detectors", "is", "an", "extension", "of", "the", "old", "ideas", "from", "Viola", "&", "Jones", ".", "Sums", "of", "rectangular", "regions", "are", "used", "as", "input", "to", "decision", "trees", "trained", "via", "Adaboost", ".", "Both", "the", "regions", "to", "pool", "from", "and", "the", "thresholds", "in", "the", "decision", "trees", "are", "selected", "during", "training", ".", "The", "crucial", "difference", "from", "the", "pioneer", "work", "is", "that", "the", "sums", "are", "done", "over", "feature", "channels", "other", "than", "simple", "image", "luminance", ".", "Current", "top", "performing", "pedestrian", "detection", "methods", "(", "dominating", "INRIA", ",", "Caltech", "and", "KITTI", "datasets", ")", "are", "all", "extensions", "of", "the", "basic", "integral", "channel", "features", "detector", "(", "named", "ChnFtrs", "in", ",", "which", "uses", "only", "HOG", "+", "LUV", "features", ")", ".", "SquaresChnFtrs", ",", "InformedHaar", ",", "and", "LDCF", ",", "are", "discussed", "in", "detail", "in", "section", "[", "reference", "]", ".", "Katamari", "exploits", "context", "and", "optical", "flow", "for", "improved", "performance", ".", "SpatialPooling", "(+", ")", "adds", "max", "-", "pooling", "on", "top", "of", "sum", "-", "pooling", ",", "and", "uses", "additional", "features", "such", "as", "covariance", ",", "LBP", ",", "and", "optical", "flow", ".", "Similarly", ",", "Regionlets", "also", "uses", "extended", "features", "and", "max", "-", "pooling", ",", "together", "with", "stronger", "weak", "classifiers", "and", "training", "a", "cascade", "of", "classifiers", ".", "Out", "of", "these", ",", "Regionlets", "is", "the", "only", "method", "that", "has", "also", "shown", "good", "performance", "on", "general", "classes", "datasets", "such", "as", "Pascal", "VOC", "and", "ImageNet", ".", "In", "this", "paper", "we", "will", "show", "that", "vanilla", "HOG", "+", "LUV", "features", "have", "not", "yet", "saturated", ",", "and", "that", ",", "when", "properly", "used", ",", "they", "can", "reach", "top", "performance", "for", "pedestrian", "detection", ".", "subsection", ":", "Contributions", "We", "point", "out", "the", "link", "between", "ACF", ",", "(", "Squares", ")", "ChnFtrs", ",", "InformedHaar", ",", "and", "LDCF", ".", "See", "section", "[", "reference", "]", ".", "We", "provide", "extensive", "experiments", "to", "enable", "a", "systematic", "analysis", "of", "the", "filtered", "integral", "channels", ",", "covering", "aspects", "not", "explored", "by", "related", "work", ".", "We", "report", "the", "summary", "of", "trained", "models", "(", "corresponding", "days", "of", "single", "machine", "computation", ")", ".", "See", "sections", "[", "reference", "]", ",", "[", "reference", "]", "and", "[", "reference", "]", ".", "We", "show", "that", "top", "detection", "performance", "can", "be", "reached", "on", "Caltech", "and", "KITTI", "using", "HOG", "+", "LUV", "features", "only", ".", "We", "additionally", "report", "the", "best", "known", "results", "on", "Caltech", ".", "See", "section", "[", "reference", "]", ".", "section", ":", "Filtered", "channel", "features", "Before", "entering", "the", "experimental", "section", ",", "let", "us", "describe", "our", "general", "architecture", ".", "Methods", "such", "as", "ChnFtrs", ",", "SquaresChnFtrs", "and", "ACF", "all", "use", "the", "basic", "architecture", "depicted", "in", "figure", "[", "reference", "]", "top", "part", "(", "best", "viewed", "in", "colours", ")", ".", "The", "input", "image", "is", "transformed", "into", "a", "set", "of", "feature", "channels", "(", "also", "called", "feature", "maps", ")", ",", "the", "feature", "vector", "is", "constructed", "by", "sum", "-", "pooling", "over", "a", "(", "large", ")", "set", "of", "rectangular", "regions", ".", "This", "feature", "vector", "is", "fed", "into", "a", "decision", "forest", "learned", "via", "Adaboost", ".", "The", "split", "nodes", "in", "the", "trees", "are", "a", "simple", "comparison", "between", "a", "feature", "value", "and", "a", "learned", "threshold", ".", "Commonly", "only", "a", "subset", "of", "the", "feature", "vector", "is", "used", "by", "the", "learned", "decision", "forest", ".", "Adaboost", "serves", "both", "for", "feature", "selection", "and", "for", "learning", "the", "thresholds", "in", "the", "split", "nodes", ".", "A", "key", "observation", ",", "illustrated", "in", "figure", "[", "reference", "]", "(", "bottom", ")", ",", "is", "that", "such", "sum", "-", "pooling", "can", "be", "re", "-", "written", "as", "convolution", "with", "a", "filter", "bank", "(", "one", "filter", "per", "rectangular", "shape", ")", "followed", "by", "reading", "a", "single", "value", "of", "the", "convolution", "’s", "response", "map", ".", "This", "‘", "‘", "filter", "+", "pick", "’", "’", "view", "generalizes", "the", "integral", "channel", "features", "detectors", "by", "allowing", "to", "use", "any", "filter", "bank", "(", "instead", "of", "only", "rectangular", "shapes", ")", ".", "We", "name", "this", "generalization", "‘", "‘", "filtered", "channel", "features", "detectors", "’", "’", ".", "In", "our", "framework", ",", "ACF", "has", "a", "single", "filter", "in", "its", "bank", ",", "corresponding", "to", "a", "uniform", "pooling", "region", ".", "ChnFtrs", "was", "a", "very", "large", "(", "tens", "of", "thousands", ")", "filter", "bank", "comprised", "of", "random", "rectangular", "shapes", ".", "SquaresChnFtrs", ",", "on", "the", "other", "hand", ",", "was", "only", "filters", ",", "each", "with", "a", "square", "-", "shaped", "uniform", "pooling", "region", "of", "different", "sizes", ".", "See", "figure", "[", "reference", "]", "for", "an", "illustration", "of", "the", "SquaresChnFtrs", "filters", ",", "the", "upper", "-", "left", "filter", "corresponds", "to", "ACF", "’s", "one", ".", "The", "InformedHaar", "method", "can", "also", "be", "seen", "as", "a", "filtered", "channel", "features", "detector", ",", "where", "the", "filter", "bank", "(", "and", "read", "locations", ")", "are", "based", "on", "a", "human", "shape", "template", "(", "thus", "the", "‘", "‘", "informed", "’", "’", "naming", ")", ".", "LDCF", "is", "also", "a", "particular", "instance", "of", "this", "framework", ",", "where", "the", "filter", "bank", "consists", "of", "PCA", "bases", "of", "patches", "from", "the", "training", "dataset", ".", "In", "sections", "[", "reference", "]", "and", "[", "reference", "]", "we", "provide", "experiments", "revisiting", "some", "of", "the", "design", "decisions", "of", "these", "methods", ".", "Note", "that", "all", "the", "methods", "mentioned", "above", "(", "and", "in", "the", "majority", "of", "experiments", "below", ")", "use", "only", "HOG", "+", "LUV", "feature", "channels", "(", "10", "channels", "total", ")", ".", "Using", "linear", "filters", "and", "decision", "trees", "on", "top", "of", "these", "does", "not", "allow", "to", "reconstruct", "the", "decision", "functions", "obtained", "when", "using", "LBP", "or", "covariance", "features", "(", "used", "by", "SpatialPooling", "and", "Regionlets", ")", ".", "We", "thus", "consider", "the", "approach", "considered", "here", "orthogonal", "to", "adding", "such", "types", "of", "features", ".", "subsection", ":", "Evaluation", "protocol", "For", "our", "experiments", "we", "use", "the", "Caltech", "and", "KITTI", "datasets", ".", "The", "popular", "INRIA", "dataset", "is", "considered", "too", "small", "and", "too", "close", "to", "saturation", "to", "provide", "interesting", "results", ".", "All", "Caltech", "results", "are", "evaluated", "using", "the", "provided", "toolbox", ",", "and", "summarised", "by", "log", "-", "average", "miss", "-", "rate", "(", "MR", ",", "lower", "is", "better", ")", "in", "the", "range", "for", "the", "‘", "‘", "reasonable", "’", "’", "setup", ".", "KITTI", "results", "are", "evaluated", "via", "the", "online", "evaluation", "portal", ",", "and", "summarised", "as", "average", "precision", "(", "AP", ",", "higher", "is", "better", ")", "for", "the", "‘", "‘", "moderate", "’", "’", "setup", ".", "paragraph", ":", "Caltech10x", "The", "raw", "Caltech", "dataset", "consists", "of", "videos", "(", "acquired", "at", ")", "with", "every", "frame", "annotated", ".", "The", "standard", "training", "and", "evaluation", "considers", "one", "out", "of", "each", "frames", "(", "pedestrians", "over", "frames", "in", "training", ",", "pedestrians", "over", "frames", "in", "testing", ")", ".", "In", "our", "experiments", "of", "section", "[", "reference", "]", "we", "will", "also", "consider", "a", "increased", "training", "set", "where", "every", "rd", "frame", "is", "used", "(", "linear", "growth", "in", "pedestrians", "and", "images", ")", ".", "We", "name", "this", "extended", "training", "set", "‘", "‘", "Caltech10x", "’", "’", ".", "LDCF", "uses", "a", "similar", "extended", "set", "for", "training", "its", "model", "(", "every", "th", "frame", ")", ".", "paragraph", ":", "Flow", "Methods", "using", "optical", "flow", "do", "not", "only", "use", "additional", "neighbour", "frames", "during", "training", "(", "depending", "on", "the", "method", ")", ",", "but", "they", "also", "do", "so", "at", "test", "time", ".", "Because", "they", "have", "access", "to", "additional", "information", "at", "test", "time", ",", "we", "consider", "them", "as", "a", "separate", "group", "in", "our", "results", "section", ".", "paragraph", ":", "Validation", "set", "In", "order", "to", "explore", "the", "design", "space", "of", "our", "pedestrian", "detector", "we", "setup", "a", "Caltech", "validation", "set", "by", "splitting", "the", "six", "training", "videos", "into", "five", "for", "training", "and", "one", "for", "testing", "(", "one", "of", "the", "splits", "suggested", "in", ")", ".", "Most", "of", "our", "experiments", "use", "this", "validation", "setup", ".", "We", "also", "report", "(", "a", "posteriori", ")", "our", "key", "results", "on", "the", "standard", "test", "set", "for", "comparison", "to", "the", "state", "of", "the", "art", ".", "For", "the", "KITTI", "experiments", "we", "also", "validate", "some", "design", "choices", "(", "such", "as", "search", "range", "and", "number", "of", "scales", ")", "before", "submission", "on", "the", "evaluation", "server", ".", "There", "we", "use", "a", "validation", "setup", ".", "subsection", ":", "Baselines", "paragraph", ":", "ACF", "Our", "experiments", "are", "based", "on", "the", "open", "source", "release", "of", "ACF", ".", "Our", "first", "baseline", "is", "vanilla", "ACF", "re", "-", "trained", "on", "the", "standard", "Caltech", "set", "(", "not", "Caltech10x", ")", ".", "On", "the", "Caltech", "test", "set", "it", "obtains", "(", "on", "validation", "set", ")", ".", "Note", "that", "this", "baseline", "already", "improves", "over", "more", "than", "previously", "published", "methods", "on", "this", "dataset", ".", "There", "is", "also", "a", "large", "gap", "between", "ACF", "-", "Ours", "(", ")", "and", "the", "original", "number", "from", "ACF", "-", "Caltech", "(", ")", ".", "The", "improvement", "is", "mainly", "due", "to", "the", "change", "towards", "a", "larger", "model", "size", "(", "from", "to", ")", ".", "All", "parameter", "details", "are", "described", "in", "section", "[", "reference", "]", ",", "and", "kept", "identical", "across", "experiments", "unless", "explicitly", "stated", ".", "paragraph", ":", "InformedHaar", "Our", "second", "baseline", "is", "a", "re", "-", "implementation", "of", "InformedHaar", ".", "Here", "again", "we", "observe", "an", "important", "gain", "from", "using", "a", "larger", "model", "size", "(", "same", "change", "as", "for", "ACF", ")", ".", "While", "the", "original", "InformedHaar", "paper", "reports", ",", "InformedHaar", "-", "Ours", "reaches", "on", "the", "Caltech", "test", "set", "(", "on", "validation", "set", ")", ".", "For", "both", "our", "baselines", "we", "use", "exactly", "the", "same", "training", "set", "as", "the", "original", "papers", ".", "Note", "that", "the", "InformedHaar", "-", "Ours", "baseline", "(", ")", "is", "right", "away", "the", "best", "known", "result", "for", "a", "method", "trained", "on", "the", "standard", "Caltech", "training", "set", ".", "In", "section", "[", "reference", "]", "we", "will", "discuss", "our", "re", "-", "implementation", "of", "LDCF", ".", "subsection", ":", "Model", "parameters", "Unless", "otherwise", "specified", "we", "train", "all", "our", "models", "using", "the", "following", "parameters", ".", "Feature", "channels", "are", "HOG", "+", "LUV", "only", ".", "The", "final", "classifier", "includes", "level", "-", "2", "decision", "trees", "(", "L2", ",", "3", "stumps", "per", "tree", ")", ",", "trained", "via", "vanilla", "discrete", "Adaboost", ".", "Each", "tree", "is", "built", "by", "doing", "exhaustive", "greedy", "search", "for", "each", "node", "(", "no", "randomization", ")", ".", "The", "model", "has", "size", ",", "and", "is", "built", "via", "four", "rounds", "of", "hard", "negative", "mining", "(", "starting", "from", "a", "model", "with", "trees", ",", "and", "then", ",", ",", ",", "trees", ")", ".", "Each", "round", "adds", "additional", "negatives", "to", "the", "training", "set", ".", "The", "sliding", "window", "stride", "is", "(", "both", "during", "hard", "negative", "mining", "and", "at", "test", "time", ")", ".", "Compared", "to", "the", "default", "ACF", "parameters", ",", "we", "use", "a", "bigger", "model", ",", "more", "trees", ",", "more", "negative", "samples", ",", "and", "more", "boosting", "rounds", ".", "But", "we", "do", "use", "the", "same", "code", "-", "base", "and", "the", "same", "training", "set", ".", "Starting", "from", "section", "[", "reference", "]", "we", "will", "also", "consider", "results", "with", "the", "Caltech10x", "data", ",", "there", "we", "use", "level", "-", "4", "decision", "trees", "(", "L4", ")", ",", "and", "Realboost", "instead", "of", "discrete", "Adaboost", ".", "All", "other", "parameters", "are", "left", "unchanged", ".", "section", ":", "Filter", "bank", "families", "Given", "the", "general", "architecture", "and", "the", "baselines", "described", "in", "section", "[", "reference", "]", ",", "we", "now", "proceed", "to", "explore", "different", "types", "of", "filter", "banks", ".", "Some", "of", "them", "are", "designed", "using", "prior", "knowledge", "and", "they", "do", "not", "change", "when", "applied", "across", "datasets", ",", "others", "exploit", "data", "-", "driven", "techniques", "for", "learning", "their", "filters", ".", "Sections", "[", "reference", "]", "and", "[", "reference", "]", "will", "compare", "their", "detection", "quality", ".", "paragraph", ":", "InformedFilters", "Starting", "from", "the", "InformedHaar", "baseline", "we", "use", "the", "same", "‘", "‘", "informed", "’", "’", "filters", "but", "let", "free", "the", "positions", "where", "they", "are", "applied", "(", "instead", "of", "fixed", "in", "InformedHaar", ")", ";", "these", "are", "selected", "during", "the", "boosting", "learning", ".", "Our", "initial", "experiments", "show", "that", "removing", "the", "position", "constraint", "has", "a", "small", "(", "positive", ")", "effect", ".", "Additionally", "we", "observe", "that", "the", "original", "InformedHaar", "filters", "do", "not", "include", "simple", "square", "pooling", "regions", "(", "à", "la", "SquaresChnFtrs", ")", ",", "we", "thus", "add", "these", "too", ".", "We", "end", "up", "with", "filters", "in", "total", ",", "to", "be", "applied", "over", "each", "of", "the", "feature", "channels", ".", "This", "is", "equivalent", "to", "training", "decision", "trees", "over", "(", "non", "filtered", ")", "channel", "features", ".", "As", "illustrated", "in", "figure", "[", "reference", "]", "the", "InformedFilters", "have", "different", "sizes", ",", "from", "to", "cells", "(", ")", ",", "and", "each", "cell", "takes", "a", "value", "in", ".", "These", "filters", "are", "applied", "with", "a", "step", "size", "of", ".", "For", "a", "model", "of", "this", "results", "in", "features", "per", "channel", ",", "features", "in", "total", ".", "In", "practice", "considering", "border", "effects", "(", "large", "filters", "are", "not", "applied", "on", "the", "border", "of", "the", "model", "to", "avoid", "reading", "outside", "it", ")", "we", "end", "up", "with", "features", ".", "When", "training", "level", "-", "2", "decision", "trees", ",", "at", "most", "features", "will", "be", "used", ",", "that", "is", "of", "the", "total", ".", "In", "this", "scenario", "(", "and", "all", "others", "considered", "in", "this", "paper", ")", "Adaboost", "has", "a", "strong", "role", "of", "feature", "selection", ".", "paragraph", ":", "Checkerboards", "As", "seen", "in", "section", "[", "reference", "]", "InformedHaar", "is", "a", "strong", "baseline", ".", "It", "is", "however", "unclear", "how", "much", "the", "‘", "‘", "informed", "’", "’", "design", "of", "the", "filters", "is", "effective", "compared", "to", "other", "possible", "choices", ".", "Checkerboards", "is", "a", "naïve", "set", "of", "filters", "that", "covers", "the", "same", "sizes", "(", "in", "number", "of", "cells", ")", "as", "InformedHaar", "/", "InformedFilters", "and", "for", "each", "size", "defines", "(", "see", "figure", "[", "reference", "]", ")", ":", "a", "uniform", "square", ",", "all", "horizontal", "and", "vertical", "gradient", "detectors", "(", "values", ")", ",", "and", "all", "possible", "checkerboard", "patterns", ".", "These", "configurations", "are", "comparable", "to", "InformedFilters", "but", "do", "not", "use", "the", "human", "shape", "as", "prior", ".", "The", "total", "number", "of", "filters", "is", "a", "direct", "function", "of", "the", "maximum", "size", "selected", ".", "For", "up", "to", "cells", "we", "end", "up", "with", "filters", ",", "up", "to", "cells", "filters", ",", "up", "to", "cells", "filters", ",", "and", "up", "to", "cells", "filters", ".", "paragraph", ":", "RandomFilters", "Our", "next", "step", "towards", "removing", "a", "hand", "-", "crafted", "design", "is", "simply", "using", "random", "filters", "(", "see", "figure", "[", "reference", "]", ")", ".", "Given", "a", "desired", "number", "of", "filters", "and", "a", "maximum", "filter", "size", "(", "in", "cells", ")", ",", "we", "sample", "the", "filter", "size", "with", "uniform", "distribution", ",", "and", "set", "its", "cell", "values", "to", "with", "uniform", "probability", ".", "We", "also", "experimented", "with", "values", "and", "observed", "a", "(", "small", ")", "quality", "decrease", "compared", "to", "the", "binary", "option", ")", ".", "The", "design", "of", "the", "filters", "considered", "above", "completely", "ignores", "the", "available", "training", "data", ".", "In", "the", "following", ",", "we", "consider", "additional", "filters", "learned", "from", "data", ".", "paragraph", ":", "LDCF", "The", "work", "on", "PCANet", "showed", "that", "applying", "arbitrary", "non", "-", "linearities", "on", "top", "of", "PCA", "projections", "of", "image", "patches", "can", "be", "surprisingly", "effective", "for", "image", "classification", ".", "Following", "this", "intuition", "LDCF", "uses", "learned", "PCA", "eigenvectors", "as", "filters", "(", "see", "figure", "[", "reference", "]", ")", ".", "We", "present", "a", "re", "-", "implementation", "of", "based", "on", "ACF", "’s", "source", "code", ".", "We", "try", "to", "follow", "the", "original", "description", "as", "closely", "as", "possible", ".", "We", "use", "the", "same", "top", "filters", "of", ",", "selected", "per", "feature", "channel", "based", "on", "their", "eigenvalues", "(", "filters", "total", ")", ".", "We", "do", "change", "some", "parameters", "to", "be", "consistent", "amongst", "all", "experiments", ",", "see", "sections", "[", "reference", "]", "and", "[", "reference", "]", ".", "The", "main", "changes", "are", "the", "training", "set", "(", "we", "use", "Caltech10x", ",", "sampled", "every", "3", "frames", ",", "instead", "of", "every", "4", "frames", "in", ")", ",", "and", "the", "model", "size", "(", "pixels", "instead", "of", ")", ".", "As", "will", "be", "shown", "in", "section", "[", "reference", "]", ",", "our", "implementation", "(", "LDCF", "-", "Ours", ")", "clearly", "improves", "over", "the", "previously", "published", "numbers", ",", "showing", "the", "potential", "of", "the", "method", ".", "For", "comparison", "with", "PcaForeground", "we", "also", "consider", "training", "LDCF8", "where", "the", "top", "filters", "are", "selected", "per", "channel", "(", "filters", "total", ")", ".", "paragraph", ":", "PcaForeground", "In", "LDCF", "the", "filters", "are", "learned", "using", "all", "of", "the", "training", "data", "available", ".", "In", "practice", "this", "means", "that", "the", "learned", "filters", "will", "be", "dominated", "by", "background", "information", ",", "and", "will", "have", "minimal", "information", "about", "the", "pedestrians", ".", "Put", "differently", ",", "learning", "filters", "from", "all", "the", "data", "assumes", "that", "the", "decision", "boundary", "is", "defined", "by", "a", "single", "distribution", "(", "like", "in", "Linear", "Discriminant", "Analysis", ")", ",", "while", "we", "might", "want", "to", "define", "it", "based", "on", "the", "relation", "between", "the", "background", "distribution", "and", "the", "foreground", "distribution", "(", "like", "Fisher", "’s", "Discriminant", "Analysis", ")", ".", "In", "PcaForeground", "we", "train", "filters", "per", "feature", "channel", ",", "learned", "from", "background", "image", "patches", ",", "and", "learned", "from", "patches", "extracted", "over", "pedestrians", "(", "see", "figure", "[", "reference", "]", ")", ".", "Compared", "to", "LDCF8", "the", "obtained", "filters", "are", "similar", "but", "not", "identical", ",", "all", "other", "parameters", "are", "kept", "identical", ".", "Other", "than", "via", "PcaForeground", "/", "LDCF8", ",", "it", "is", "not", "clear", "how", "to", "further", "increase", "the", "number", "of", "filters", "used", "in", "LDCF", ".", "Past", "filters", "per", "channel", ",", "the", "eigenvalues", "decrease", "to", "negligible", "values", "and", "the", "eigenvectors", "become", "essentially", "random", "(", "similar", "to", "RandomFilters", ")", ".", "To", "keep", "the", "filtered", "channel", "features", "setup", "close", "to", "InformedHaar", ",", "the", "filters", "are", "applied", "with", "a", "step", "of", "However", ",", "to", "stay", "close", "to", "the", "original", "LDCF", ",", "the", "LDCF", "/", "PcaForeground", "filters", "are", "evaluated", "every", ".", "Although", "(", "for", "example", ")", "LDCF8", "uses", "only", "of", "the", "number", "of", "filters", "per", "channel", "compared", "to", "Checkerboards4x4", ",", "due", "to", "the", "step", "size", "increase", ",", "the", "obtained", "feature", "vector", "size", "is", ".", "section", ":", "How", "many", "filters", "?", "Given", "a", "fixed", "set", "of", "channel", "features", ",", "a", "larger", "filter", "bank", "provides", "a", "richer", "view", "over", "the", "data", "compared", "to", "a", "smaller", "one", ".", "With", "enough", "training", "data", "one", "would", "expect", "larger", "filter", "banks", "to", "perform", "best", ".", "We", "want", "thus", "to", "analyze", "the", "trade", "-", "off", "between", "number", "of", "filters", "and", "detection", "quality", ",", "as", "well", "as", "which", "filter", "bank", "family", "performs", "best", ".", "Figure", "[", "reference", "]", "presents", "the", "results", "of", "our", "initial", "experiments", "on", "the", "Caltech", "validation", "set", ".", "It", "shows", "detection", "quality", "versus", "number", "of", "filters", "per", "channel", ".", "This", "figure", "densely", "summarizes", "trained", "models", ".", "paragraph", ":", "InformedFilters", "The", "first", "aspect", "to", "notice", "is", "that", "there", "is", "a", "meaningful", "gap", "between", "InformedHaar", "-", "Ours", "and", "InformedFilters", "despite", "having", "a", "similar", "number", "of", "filters", "(", "versus", ")", ".", "This", "validates", "the", "importance", "of", "letting", "Adaboost", "choose", "the", "pooling", "locations", "instead", "of", "hand", "-", "crafting", "them", ".", "Keep", "in", "mind", "that", "InformedHaar", "-", "Ours", "is", "a", "top", "performing", "baseline", "(", "see", "§", "[", "reference", "]", ")", ".", "Secondly", ",", "we", "observe", "that", "(", "for", "the", "fixed", "training", "data", "available", ")", "filters", "is", "better", "than", ".", "Below", "filters", "the", "performance", "degrades", "for", "all", "methods", "(", "as", "expected", ")", ".", "To", "change", "the", "number", "of", "filters", "in", "InformedFilters", "we", "train", "a", "full", "model", "(", "filters", ")", ",", "pick", "the", "most", "frequently", "used", "filters", "(", "selected", "from", "node", "splitting", "in", "the", "decision", "forest", ")", ",", "and", "use", "these", "to", "train", "the", "desired", "reduced", "model", ".", "We", "can", "select", "the", "most", "frequent", "filters", "across", "channels", "or", "per", "channel", "(", "marked", "as", "Inf", ".", "FiltersPerChannel", ")", ".", "We", "observe", "that", "per", "channel", "selection", "is", "slightly", "worse", "than", "across", "channels", ",", "thus", "we", "stick", "to", "the", "latter", ".", "Using", "the", "most", "frequently", "used", "filters", "for", "selection", "is", "clearly", "a", "crude", "strategy", "since", "frequent", "usage", "does", "not", "guarantee", "discriminative", "power", ",", "and", "it", "ignores", "relation", "amongst", "filters", ".", "We", "find", "this", "strategy", "good", "enough", "to", "convey", "the", "main", "points", "of", "this", "work", ".", "paragraph", ":", "Checkerboards", "also", "reaches", "best", "results", "in", "the", "filters", "region", ".", "Here", "the", "number", "of", "filters", "is", "varied", "by", "changing", "the", "maximum", "filter", "size", "(", "in", "number", "of", "cells", ")", ".", "Regarding", "the", "lowest", "miss", "-", "rate", "there", "is", "no", "large", "gap", "between", "the", "‘", "‘", "informed", "’", "’", "filters", "and", "this", "naïve", "baseline", ".", "paragraph", ":", "RandomFilters", "The", "hexagonal", "dots", "and", "their", "deviation", "bars", "indicate", "the", "mean", ",", "maximum", "and", "minimum", "miss", "-", "rate", "obtained", "out", "of", "five", "random", "runs", ".", "When", "using", "a", "larger", "number", "of", "filters", "(", ")", "we", "observe", "a", "lower", "(", "better", ")", "mean", "but", "a", "larger", "variance", "compared", "to", "when", "using", "fewer", "filters", "(", ")", ".", "Here", "again", "the", "gap", "between", "the", "best", "random", "run", "and", "the", "best", "result", "of", "other", "methods", "is", "not", "large", ".", "Given", "a", "set", "of", "five", "models", ",", "we", "select", "the", "most", "frequently", "used", "filters", "and", "train", "new", "reduced", "models", ";", "these", "are", "shown", "in", "the", "RandomFilters", "line", ".", "Overall", "the", "random", "filters", "are", "surprisingly", "close", "to", "the", "other", "filter", "families", ".", "This", "indicates", "that", "expanding", "the", "feature", "channels", "via", "filtering", "is", "the", "key", "step", "for", "improving", "detection", "quality", ",", "while", "selecting", "the", "‘", "‘", "perfect", "’", "’", "filters", "is", "a", "secondary", "concern", ".", "paragraph", ":", "LDCF", "/", "PcaForeground", "In", "contrast", "to", "the", "other", "filter", "bank", "families", ",", "LDCF", "under", "-", "performs", "when", "increasing", "the", "number", "of", "filters", "(", "from", "to", ")", "while", "using", "the", "standard", "Caltech", "training", "set", "(", "consistent", "with", "the", "observations", "in", ")", ".", "PcaForeground", "improves", "marginally", "over", "LDCF8", ".", "paragraph", ":", "Takeaways", "From", "figure", "[", "reference", "]", "we", "observe", "two", "overall", "trends", ".", "First", ",", "the", "more", "filters", "the", "merrier", ",", "with", "filters", "as", "sweet", "spot", "for", "Caltech", "training", "data", ".", "Second", ",", "there", "is", "no", "flagrant", "difference", "between", "the", "different", "filter", "types", ".", "section", ":", "Additional", "training", "data", "One", "caveat", "of", "the", "previous", "experiments", "is", "that", "as", "we", "increase", "the", "number", "of", "filters", "used", ",", "so", "does", "the", "number", "of", "features", "Adaboost", "must", "pick", "from", ".", "Since", "we", "increased", "the", "model", "capacity", "(", "compared", "to", "ACF", "which", "uses", "a", "single", "filter", ")", ",", "we", "consider", "using", "the", "Caltech10x", "dataset", "(", "§", "[", "reference", "]", ")", "to", "verify", "that", "our", "models", "are", "not", "starving", "for", "data", ".", "Similar", "to", "the", "experiments", "in", ",", "we", "also", "reconsider", "the", "decision", "tree", "depth", ",", "since", "additional", "training", "data", "enables", "bigger", "models", ".", "Results", "for", "two", "representative", "methods", "are", "collected", "in", "table", "[", "reference", "]", ".", "First", "we", "observe", "that", "already", "with", "the", "original", "training", "data", ",", "deeper", "trees", "do", "provide", "significant", "improvement", "over", "level", "-", "2", "(", "which", "was", "selected", "when", "tuning", "over", "INRIA", "data", ")", ".", "Second", ",", "we", "notice", "that", "increasing", "the", "training", "data", "volume", "does", "provide", "the", "expected", "improvement", "only", "when", "the", "decision", "trees", "are", "deep", "enough", ".", "For", "our", "following", "experiments", "we", "choose", "to", "use", "level", "-", "4", "decision", "trees", "(", ")", "as", "a", "good", "balance", "between", "increased", "detection", "quality", "and", "reasonable", "training", "times", ".", "paragraph", ":", "Realboost", "Although", "previous", "papers", "on", "ChnFtrs", "detectors", "reported", "that", "different", "boosting", "variants", "all", "obtain", "equal", "results", "on", "this", "task", ",", "the", "recent", "indicated", "that", "Realboost", "has", "an", "edge", "over", "discrete", "Adaboost", "when", "additional", "training", "data", "is", "used", ".", "We", "observe", "the", "same", "behaviour", "in", "our", "Caltech10x", "setup", ".", "As", "summarized", "in", "table", "[", "reference", "]", "using", "filtered", "channels", ",", "deeper", "trees", ",", "additional", "training", "data", ",", "and", "Realboost", "does", "provide", "a", "significant", "detection", "quality", "boost", ".", "For", "the", "rest", "of", "the", "paper", "our", "models", "trained", "on", "Caltech10x", "all", "use", "level", "-", "4", "trees", "and", "RealBoost", ",", "instead", "of", "level", "-", "2", "and", "discrete", "Adaboost", "for", "the", "Caltech1x", "models", ".", "paragraph", ":", "Timing", "When", "using", "Caltech", "data", "ACF", "takes", "about", "one", "hour", "for", "training", "and", "one", "for", "testing", ".", "Checkerboards4x4", "takes", "about", "and", "hours", "respectively", ".", "When", "using", "Caltech10x", "the", "training", "times", "for", "these", "methods", "augment", "to", "2", "and", "29", "hours", ",", "respectively", ".", "The", "training", "time", "does", "not", "increase", "proportionally", "with", "the", "training", "data", "volume", "because", "the", "hard", "negative", "mining", "reads", "a", "variable", "amount", "of", "images", "to", "attain", "the", "desired", "quota", "of", "negative", "samples", ".", "This", "amount", "increases", "when", "a", "detector", "has", "less", "false", "positive", "mistakes", ".", "subsection", ":", "Validation", "set", "experiments", "Based", "on", "the", "results", "in", "table", "[", "reference", "]", "we", "proceed", "to", "evaluate", "on", "Caltech10x", "the", "most", "promising", "configurations", "(", "filter", "type", "and", "number", ")", "from", "section", "[", "reference", "]", ".", "The", "results", "over", "the", "Caltech", "validation", "set", "are", "collected", "in", "table", "[", "reference", "]", ".", "We", "observe", "a", "clear", "overall", "gain", "from", "increasing", "the", "training", "data", ".", "Interestingly", "with", "enough", "RandomFilters", "we", "can", "outperform", "the", "strong", "performance", "of", "LDCF", "-", "Ours", ".", "We", "also", "notice", "that", "the", "naïve", "Checkerboards", "outperforms", "the", "manual", "design", "of", "InformedFilters", ".", "section", ":", "Add", "-", "ons", "Before", "presenting", "the", "final", "test", "set", "results", "of", "our", "‘", "‘", "core", "’", "’", "method", "(", "section", "[", "reference", "]", ")", ",", "we", "also", "consider", "some", "possible", "‘", "‘", "add", "-", "ons", "’", "’", "based", "on", "the", "suggestions", "from", ".", "For", "the", "sake", "of", "evaluating", "complementarity", ",", "comparison", "with", "existing", "method", ",", "and", "reporting", "the", "best", "possible", "detection", "quality", ",", "we", "consider", "extending", "our", "detector", "with", "context", "and", "optical", "flow", "information", ".", "paragraph", ":", "Context", "Context", "is", "modelled", "via", "the", "2Ped", "re", "-", "scoring", "method", "of", ".", "It", "is", "a", "post", "-", "processing", "step", "that", "merges", "our", "detection", "scores", "with", "the", "results", "of", "a", "two", "person", "DPM", "trained", "on", "the", "INRIA", "dataset", "(", "with", "extended", "annotations", ")", ".", "In", "the", "authors", "reported", "an", "improvement", "of", "(", "percent", "points", ")", "on", "the", "Caltech", "set", ",", "across", "different", "methods", ".", "In", "an", "improvement", "of", "is", "reported", "over", "their", "strong", "detector", "(", "SquaresChnFtrs", "+", "DCT", "+", "SDt", ")", ".", "In", "our", "experiments", "however", "we", "obtain", "a", "gain", "inferior", "to", ".", "We", "have", "also", "investigated", "fusing", "the", "2Ped", "detection", "results", "via", "a", "different", ",", "more", "principled", ",", "fusion", "method", ".", "We", "observe", "consistent", "results", ":", "as", "the", "strength", "of", "the", "starting", "point", "increases", ",", "the", "gain", "from", "2Ped", "decreases", ".", "When", "reaching", "our", "Checkerboards", "results", ",", "all", "gains", "have", "evaporated", ".", "We", "believe", "that", "the", "2Ped", "approach", "is", "a", "promising", "one", ",", "but", "our", "experiments", "indicate", "that", "the", "used", "DPM", "template", "is", "simply", "too", "weak", "in", "comparison", "to", "our", "filtered", "channels", ".", "paragraph", ":", "Optical", "flow", "Optical", "flow", "is", "fed", "to", "our", "detector", "as", "an", "additional", "set", "of", "channels", "(", "not", "filtered", ")", ".", "We", "use", "the", "implementation", "from", "SDt", "which", "uses", "differences", "of", "weakly", "stabilized", "video", "frames", ".", "On", "Caltech", ",", "the", "authors", "of", "reported", "a", "gain", "over", "ACF", "(", ")", ",", "while", "reported", "a", "percent", "points", "improvement", "over", "their", "strong", "baseline", "(", "SquaresChnFtrs", "+", "DCT", "+", "2Ped", "⁢%27.4MR", ")", ".", "When", "using", "+", "SDt", "our", "results", "are", "directly", "comparable", "to", "Katamari", "and", "SpatialPooling", "+", "which", "both", "use", "optical", "flow", "too", ".", "Using", "our", "stronger", "Checkerboards", "results", "SDt", "provides", "a", "gain", ".", "Here", "again", "we", "observe", "an", "erosion", "as", "the", "starting", "point", "improves", "(", "for", "confirmation", ",", "reproduced", "the", "ACF", "+", "SDt", "results", ",", ")", ".", "We", "name", "our", "Checkerboards", "+", "SDt", "detector", "All", "-", "in", "-", "one", ".", "Our", "filtered", "channel", "features", "results", "are", "strong", "enough", "to", "erode", "existing", "context", "and", "flow", "features", ".", "Although", "these", "remain", "complementary", "cues", ",", "more", "sophisticated", "ways", "of", "extracting", "this", "information", "will", "be", "required", "to", "further", "progress", "in", "detection", "quality", ".", "It", "should", "be", "noted", "that", "despite", "our", "best", "efforts", "we", "could", "not", "reproduce", "the", "results", "from", "neither", "2Ped", "nor", "SDt", "on", "the", "KITTI", "dataset", "(", "in", "spite", "of", "its", "apparent", "similarity", "to", "Caltech", ")", ".", "Effective", "methods", "for", "context", "and", "optical", "flow", "across", "datasets", "have", "yet", "to", "be", "shown", ".", "Our", "main", "contribution", "remains", "on", "the", "core", "detector", "(", "only", "HOG", "+", "LUV", "features", "over", "local", "sliding", "window", "pixels", "in", "a", "single", "frame", ")", ".", "section", ":", "Test", "set", "results", "Having", "done", "our", "exploration", "of", "the", "parameters", "space", "on", "the", "validation", "set", ",", "we", "now", "evaluate", "the", "most", "promising", "methods", "on", "the", "Caltech", "and", "KITTI", "test", "sets", ".", "paragraph", ":", "Caltech", "test", "set", "Figures", "[", "reference", "]", "and", "[", "reference", "]", "present", "our", "key", "results", "on", "the", "Caltech", "test", "set", ".", "For", "proper", "comparison", ",", "only", "methods", "using", "the", "same", "training", "set", "should", "be", "compared", "(", "see", "for", "a", "similar", "table", "comparing", "previous", "methods", ")", ".", "We", "include", "for", "comparison", "the", "baselines", "mentioned", "in", "section", "[", "reference", "]", ",", "Roerei", "[", "]", "the", "best", "known", "method", "trained", "without", "any", "Caltech", "images", ",", "MT", "-", "DPM", "[", "]", "the", "best", "known", "method", "based", "on", "DPM", ",", "and", "SDN", "[", "]", "the", "best", "known", "method", "using", "convolutional", "neural", "networks", ".", "We", "also", "include", "the", "top", "performers", "Katamari", "and", "SpatialPooling", "+", ".", "We", "mark", "as", "‘", "‘", "’’", "both", "the", "Caltech10x", "training", "set", "and", "the", "one", "used", "in", "LDCF", "(", "see", "section", "[", "reference", "]", ")", ".", "paragraph", ":", "KITTI", "test", "set", "Figure", "[", "reference", "]", "presents", "the", "results", "on", "the", "KITTI", "test", "set", "(", "‘", "‘", "moderate", "’", "’", "setup", ")", ",", "together", "with", "all", "other", "reported", "methods", "using", "only", "monocular", "image", "content", "(", "no", "stereo", "or", "LIDAR", "data", ")", ".", "The", "KITTI", "evaluation", "server", "only", "recently", "has", "started", "receiving", "submissions", "(", "for", "this", "task", ",", "in", "the", "last", "year", ")", ",", "and", "thus", "is", "less", "prone", "to", "dataset", "over", "-", "fitting", ".", "We", "train", "our", "model", "on", "the", "KITTI", "training", "set", "using", "almost", "identical", "parameters", "as", "for", "Caltech", ".", "The", "only", "change", "is", "a", "subtle", "pre", "-", "processing", "step", "in", "the", "HOG", "+", "LUV", "computation", ".", "On", "KITTI", "the", "input", "image", "is", "smoothed", "(", "radius", ")", "before", "the", "feature", "channels", "are", "computed", ",", "while", "on", "Caltech", "we", "do", "not", ".", "This", "subtle", "change", "provided", "a", "(", "percent", "points", ")", "improvement", "on", "the", "KITTI", "validation", "set", ".", "subsection", ":", "Analysis", "With", "a", "(", "percent", "points", ")", "gap", "between", "ACF", "/", "InformedHaar", "and", "ACF", "/", "InformedHaar", "-", "Ours", "(", "see", "figure", "[", "reference", "]", ")", ",", "the", "results", "of", "our", "baselines", "show", "the", "importance", "of", "proper", "validation", "of", "training", "parameters", "(", "large", "enough", "model", "size", "and", "negative", "samples", ")", ".", "InformedHaar", "-", "Ours", "is", "the", "best", "reported", "result", "when", "training", "with", "Caltech1x", ".", "When", "considering", "methods", "trained", "on", "Caltech10x", ",", "we", "obtain", "a", "clear", "gap", "with", "the", "previous", "best", "results", "(", "LDCF", "Checkerboards", ")", ".", "Using", "our", "architecture", "and", "the", "adequate", "number", "of", "filters", "one", "can", "obtain", "strong", "results", "using", "only", "HOG", "+", "LUV", "features", ".", "The", "exact", "type", "of", "filters", "seems", "not", "critical", ",", "in", "our", "experiments", "Checkerboards4x3", "gets", "best", "performance", "given", "the", "available", "training", "data", ".", "RandomFilters", "reaches", "the", "same", "result", ",", "but", "requires", "training", "and", "merging", "multiple", "models", ".", "Our", "results", "cut", "by", "half", "miss", "-", "rate", "of", "the", "best", "known", "convnet", "for", "pedestrian", "detection", "(", "SDN", "[", "]", ")", ",", "which", "in", "principle", "could", "learn", "similar", "low", "-", "level", "features", "and", "their", "filtering", ".", "When", "adding", "optical", "flow", "we", "further", "push", "the", "state", "of", "the", "art", "and", "reach", ",", "a", "comfortable", "improvement", "over", "the", "previous", "best", "optical", "flow", "method", "(", "SpatialPooling", "+", ")", ".", "This", "is", "the", "best", "reported", "result", "on", "this", "challenging", "dataset", ".", "The", "results", "on", "the", "KITTI", "dataset", "confirm", "the", "strength", "of", "our", "approach", ",", "reaching", ",", "just", "below", "the", "best", "known", "result", "on", "this", "dataset", ".", "Competing", "methods", "(", "Regionlets", "[", "]", "and", "SpatialPooling", ")", "both", "use", "HOG", "together", "with", "additional", "LBP", "and", "covariance", "features", ".", "Adding", "these", "remains", "a", "possibility", "for", "our", "system", ".", "Note", "that", "our", "results", "also", "improve", "over", "methods", "using", "LIDAR", "+", "Image", ",", "such", "as", "Fusion", "-", "DPM", "(", ",", "not", "included", "in", "figure", "[", "reference", "]", "for", "clarity", ")", ".", "section", ":", "Conclusion", "Through", "this", "paper", "we", "have", "shown", "that", "the", "seemingly", "disconnected", "methods", "ACF", ",", "(", "Squares", ")", "ChnFtrs", ",", "InformedHaar", ",", "and", "LDCF", "can", "be", "all", "put", "under", "the", "filtered", "channel", "features", "detectors", "umbrella", ".", "We", "have", "systematically", "explored", "different", "filter", "banks", "for", "such", "architecture", "and", "shown", "that", "they", "provide", "means", "for", "important", "improvements", "for", "pedestrian", "detection", ".", "Our", "results", "indicate", "that", "HOG", "+", "LUV", "features", "have", "not", "yet", "saturated", ",", "and", "that", "competitive", "results", "(", "over", "Caltech", "and", "KITTI", "datasets", ")", "can", "be", "obtained", "using", "only", "them", ".", "When", "optical", "flow", "information", "is", "added", "we", "set", "the", "new", "state", "of", "art", "for", "the", "Caltech", "dataset", ",", "reaching", "(", "recall", "at", "false", "positive", "per", "image", ")", ".", "In", "future", "work", "we", "plan", "to", "explore", "how", "the", "insights", "of", "this", "work", "can", "be", "exploited", "into", "a", "more", "general", "detection", "architecture", "such", "as", "convolutional", "neural", "networks", ".", "paragraph", ":", "Acknowledgements", "We", "thank", "Jan", "Hosang", "for", "the", "help", "provided", "setting", "up", "some", "of", "the", "experiments", ".", "We", "also", "thank", "Seong", "Joon", "Oh", "and", "Sabrina", "Hoppe", "for", "their", "useful", "comments", ".", "bibliography", ":", "References", "appendix", ":", "Learned", "model", "In", "figures", "[", "reference", "]", "and", "[", "reference", "]", "we", "present", "some", "qualitative", "aspects", "of", "the", "final", "learned", "models", "Checkerboards4x3", "and", "RandomFilters", "(", "see", "results", "section", "of", "main", "paper", ")", ",", "not", "included", "in", "the", "main", "submission", "due", "to", "space", "limitations", ".", "In", "figure", "[", "reference", "]", "we", "compare", "the", "spatial", "distribution", "of", "our", "models", "versus", "a", "significantly", "weaker", "model", "(", "Roerei", ",", "trained", "on", "INRIA", ",", "see", "figure", "5", "of", "main", "paper", ")", ".", "We", "observe", "that", "our", "strong", "models", "focalize", "in", "similar", "areas", "than", "the", "weak", "Roerei", "model", ".", "This", "indicates", "that", "using", "filtered", "channels", "does", "not", "change", "which", "areas", "of", "the", "pedestrian", "are", "informative", ",", "but", "rather", "that", "at", "the", "same", "locations", "filtered", "channels", "are", "able", "to", "extract", "more", "discriminative", "information", ".", "In", "all", "three", "models", "we", "observe", "that", "diagonal", "oriented", "channels", "focus", "on", "left", "and", "right", "shoulders", ".", "The", "U", "colour", "channel", "is", "mainly", "used", "around", "the", "face", ",", "while", "L", "(", "luminance", ")", "and", "gradient", "magnitude", "(", ")", "channels", "are", "used", "all", "over", "the", "body", ".", "Overall", "head", ",", "feet", ",", "and", "upper", "torso", "areas", "provide", "most", "clues", "for", "detection", ".", "In", "figure", "[", "reference", "]", "we", "observe", "that", "the", "filters", "usage", "distribution", "is", "similar", "across", "different", "filter", "bank", "families", ".", "[", "Filters", "from", "Roerei", "(", "scale", ")", "model", ".", "Copied", "from", "]", "[", "Final", "Checkerboards4x3", "model", "]", "[", "Final", "RandomFilters", "model", "]", "[", "Filters", "used", "in", "our", "final", "Checkerboards4x3", "model", "]", "[", "Filters", "used", "in", "our", "final", "RandomFilters", "model", "]"], "entity": [[2, 5, ["Filtered", "Channel", "Features"], "Method"], [6, 8, ["Pedestrian", "Detection"], "Task"], [16, 20, ["top", "performing", "pedestrian", "detectors"], "Method"], [26, 33, ["intermediate", "layer", "filtering", "low", "-", "level", "features"], "Method"], [37, 40, ["boosted", "decision", "forest"], "Method"], [48, 50, ["unifying", "framework"], "Method"], [54, 56, ["filter", "families"], "Method"], [77, 78, ["Caltech"], "Material"], [102, 104, ["detection", "quality"], "Metric"], [112, 114, ["Caltech", "dataset"], "Material"], [118, 119, ["recall"], "Metric"], [121, 122, ["FPPI"], "Metric"], [129, 131, ["Pedestrian", "detection"], "Task"], [160, 162, ["object", "detection"], "Task"], [176, 179, ["generic", "object", "detection"], "Task"], [189, 191, ["detection", "quality"], "Metric"], [208, 210, ["pedestrian", "detection"], "Task"], [222, 224, ["features", "learning"], "Method"], [232, 235, ["pedestrian", "detection", "methods"], "Method"], [244, 247, ["max", "-", "pooling"], "Method"], [248, 249, ["filtering"], "Method"], [259, 261, ["classification", "layer"], "Method"], [274, 276, ["linear", "transformation"], "Method"], [278, 279, ["convolution"], "Method"], [281, 283, ["filter", "bank"], "Method"], [314, 316, ["filter", "banks"], "Method"], [327, 329, ["filter", "bank"], "Method"], [334, 337, ["top", "detection", "quality"], "Metric"], [365, 367, ["detection", "quality"], "Metric"], [380, 383, ["sliding", "window", "algorithm"], "Method"], [441, 442, ["Caltech"], "Material"], [443, 444, ["KITTI"], "Material"], [478, 480, ["pedestrian", "detection"], "Task"], [488, 490, ["classifier", "choice"], "Method"], [492, 498, ["linear", "/", "non", "-", "linear", "SVM"], "Method"], [499, 501, ["decision", "forest"], "Method"], [508, 509, ["quality"], "Metric"], [533, 535, ["HOG", "descriptor"], "Method"], [543, 547, ["pooling", "and", "normalization", "steps"], "Method"], [559, 562, ["deformable", "parts", "model"], "Method"], [571, 575, ["integral", "channel", "features", "detector"], "Method"], [578, 580, ["HOG", "variant"], "Method"], [581, 583, ["sum", "pooling"], "Method"], [585, 586, ["normalizations"], "Method"], [590, 591, ["HOG"], "Method"], [610, 611, ["LBP"], "Method"], [649, 651, ["pedestrian", "detection"], "Task"], [696, 698, ["segmentation", "methods"], "Method"], [710, 711, ["Covariance"], "Method"], [758, 761, ["sparse", "dictionary", "encoders"], "Method"], [767, 770, ["convolutional", "neural", "network"], "Method"], [775, 777, ["stereo", "depth"], "Task"], [778, 780, ["optical", "flow"], "Task"], [810, 811, ["SDt"], "Method"], [824, 829, ["integral", "channel", "features", "detector", "framework"], "Method"], [855, 857, ["decision", "trees"], "Method"], [859, 860, ["Adaboost"], "Method"], [904, 907, ["pedestrian", "detection", "methods"], "Method"], [911, 912, ["Caltech"], "Material"], [913, 914, ["KITTI"], "Material"], [922, 926, ["integral", "channel", "features", "detector"], "Method"], [928, 929, ["ChnFtrs"], "Method"], [940, 941, ["SquaresChnFtrs"], "Method"], [942, 943, ["InformedHaar"], "Method"], [945, 946, ["LDCF"], "Method"], [961, 963, ["optical", "flow"], "Method"], [967, 970, ["SpatialPooling", "(+", ")"], "Method"], [971, 974, ["max", "-", "pooling"], "Method"], [977, 980, ["sum", "-", "pooling"], "Method"], [997, 998, ["Regionlets"], "Method"], [1003, 1006, ["max", "-", "pooling"], "Method"], [1010, 1012, ["weak", "classifiers"], "Method"], [1015, 1018, ["cascade", "of", "classifiers"], "Method"], [1023, 1024, ["Regionlets"], "Method"], [1075, 1077, ["pedestrian", "detection"], "Task"], [1087, 1088, ["ACF"], "Method"], [1090, 1093, ["Squares", ")", "ChnFtrs"], "Method"], [1094, 1095, ["InformedHaar"], "Method"], [1097, 1098, ["LDCF"], "Method"], [1116, 1119, ["filtered", "integral", "channels"], "Method"], [1140, 1142, ["machine", "computation"], "Task"], [1162, 1163, ["detection"], "Task"], [1168, 1169, ["Caltech"], "Material"], [1170, 1171, ["KITTI"], "Material"], [1186, 1187, ["Caltech"], "Material"], [1215, 1216, ["ChnFtrs"], "Method"], [1217, 1218, ["SquaresChnFtrs"], "Method"], [1219, 1220, ["ACF"], "Method"], [1264, 1267, ["sum", "-", "pooling"], "Method"], [1284, 1286, ["decision", "forest"], "Method"], [1288, 1289, ["Adaboost"], "Method"], [1322, 1324, ["decision", "forest"], "Method"], [1325, 1326, ["Adaboost"], "Method"], [1329, 1331, ["feature", "selection"], "Task"], [1358, 1361, ["sum", "-", "pooling"], "Method"], [1367, 1368, ["convolution"], "Method"], [1370, 1372, ["filter", "bank"], "Method"], [1387, 1391, ["convolution", "’s", "response", "map"], "Method"], [1395, 1398, ["filter", "+", "pick"], "Method"], [1403, 1407, ["integral", "channel", "features", "detectors"], "Method"], [1412, 1414, ["filter", "bank"], "Method"], [1428, 1432, ["filtered", "channel", "features", "detectors"], "Method"], [1439, 1440, ["ACF"], "Method"], [1443, 1444, ["filter"], "Method"], [1455, 1456, ["ChnFtrs"], "Method"], [1465, 1467, ["filter", "bank"], "Method"], [1473, 1474, ["SquaresChnFtrs"], "Method"], [1522, 1524, ["InformedHaar", "method"], "Method"], [1530, 1534, ["filtered", "channel", "features", "detector"], "Method"], [1537, 1539, ["filter", "bank"], "Method"], [1562, 1563, ["LDCF"], "Method"], [1574, 1576, ["filter", "bank"], "Method"], [1578, 1580, ["PCA", "bases"], "Method"], [1639, 1641, ["linear", "filters"], "Method"], [1642, 1644, ["decision", "trees"], "Method"], [1666, 1669, ["SpatialPooling", "and", "Regionlets"], "Method"], [1696, 1697, ["Caltech"], "Material"], [1698, 1699, ["KITTI"], "Material"], [1720, 1721, ["Caltech"], "Material"], [1732, 1738, ["log", "-", "average", "miss", "-", "rate"], "Metric"], [1739, 1740, ["MR"], "Metric"], [1757, 1758, ["KITTI"], "Material"], [1763, 1766, ["online", "evaluation", "portal"], "Metric"], [1770, 1772, ["average", "precision"], "Metric"], [1773, 1774, ["AP"], "Metric"], [1790, 1791, ["Caltech10x"], "Material"], [1871, 1872, ["Caltech10x"], "Material"], [1875, 1876, ["LDCF"], "Method"], [1896, 1898, ["optical", "flow"], "Method"], [1959, 1961, ["pedestrian", "detector"], "Method"], [1964, 1967, ["Caltech", "validation", "set"], "Material"], [2073, 2074, ["ACF"], "Method"], [2079, 2084, ["vanilla", "ACF", "re", "-", "trained"], "Method"], [2087, 2089, ["Caltech", "set"], "Material"], [2096, 2099, ["Caltech", "test", "set"], "Material"], [2142, 2143, ["Caltech"], "Material"], [2196, 2197, ["InformedHaar"], "Method"], [2216, 2217, ["ACF"], "Method"], [2226, 2229, ["InformedHaar", "-", "Ours"], "Method"], [2232, 2235, ["Caltech", "test", "set"], "Material"], [2260, 2264, ["InformedHaar", "-", "Ours", "baseline"], "Method"], [2280, 2283, ["Caltech", "training", "set"], "Material"], [2297, 2298, ["LDCF"], "Method"], [2319, 2322, ["HOG", "+", "LUV"], "Method"], [2326, 2327, ["classifier"], "Method"], [2328, 2333, ["level", "-", "2", "decision", "trees"], "Method"], [2344, 2347, ["vanilla", "discrete", "Adaboost"], "Method"], [2354, 2357, ["exhaustive", "greedy", "search"], "Method"], [2377, 2380, ["hard", "negative", "mining"], "Method"], [2414, 2417, ["hard", "negative", "mining"], "Task"], [2482, 2487, ["level", "-", "4", "decision", "trees"], "Method"], [2492, 2493, ["Realboost"], "Method"], [2495, 2497, ["discrete", "Adaboost"], "Method"], [2507, 2510, ["Filter", "bank", "families"], "Method"], [2532, 2534, ["filter", "banks"], "Method"], [2555, 2559, ["data", "-", "driven", "techniques"], "Method"], [2575, 2577, ["detection", "quality"], "Metric"], [2584, 2586, ["InformedHaar", "baseline"], "Method"], [2618, 2620, ["boosting", "learning"], "Method"], [2644, 2646, ["InformedHaar", "filters"], "Method"], [2688, 2690, ["decision", "trees"], "Method"], [2784, 2786, ["decision", "trees"], "Method"], [2812, 2813, ["Adaboost"], "Method"], [2818, 2820, ["feature", "selection"], "Task"], [2831, 2832, ["InformedHaar"], "Method"], [2861, 2862, ["Checkerboards"], "Method"], [2867, 2868, ["filters"], "Method"], [2921, 2922, ["InformedFilters"], "Method"], [2982, 2986, ["hand", "-", "crafted", "design"], "Method"], [2989, 2991, ["random", "filters"], "Method"], [3075, 3076, ["filters"], "Method"], [3082, 3083, ["LDCF"], "Method"], [3086, 3087, ["PCANet"], "Method"], [3097, 3102, ["PCA", "projections", "of", "image", "patches"], "Method"], [3107, 3109, ["image", "classification"], "Task"], [3113, 3114, ["LDCF"], "Method"], [3137, 3141, ["ACF", "’s", "source", "code"], "Method"], [3245, 3248, ["LDCF", "-", "Ours"], "Method"], [3267, 3268, ["PcaForeground"], "Method"], [3272, 3273, ["LDCF8"], "Method"], [3288, 3289, ["PcaForeground"], "Method"], [3290, 3291, ["LDCF"], "Method"], [3310, 3311, ["filters"], "Method"], [3330, 3332, ["learning", "filters"], "Method"], [3350, 3353, ["Linear", "Discriminant", "Analysis"], "Method"], [3376, 3380, ["Fisher", "’s", "Discriminant", "Analysis"], "Method"], [3383, 3384, ["PcaForeground"], "Method"], [3386, 3387, ["filters"], "Method"], [3414, 3415, ["LDCF8"], "Method"], [3434, 3437, ["PcaForeground", "/", "LDCF8"], "Method"], [3452, 3453, ["LDCF"], "Method"], [3504, 3505, ["LDCF"], "Method"], [3507, 3511, ["LDCF", "/", "PcaForeground", "filters"], "Method"], [3520, 3521, ["LDCF8"], "Method"], [3543, 3546, ["feature", "vector", "size"], "Metric"], [3564, 3566, ["filter", "bank"], "Method"], [3587, 3589, ["filter", "banks"], "Method"], [3607, 3609, ["detection", "quality"], "Metric"], [3614, 3617, ["filter", "bank", "family"], "Method"], [3633, 3636, ["Caltech", "validation", "set"], "Material"], [3639, 3641, ["detection", "quality"], "Metric"], [3671, 3674, ["InformedHaar", "-", "Ours"], "Method"], [3675, 3676, ["InformedFilters"], "Method"], [3693, 3694, ["Adaboost"], "Method"], [3709, 3712, ["InformedHaar", "-", "Ours"], "Method"], [3738, 3739, ["filters"], "Method"], [3767, 3769, ["full", "model"], "Method"], [3770, 3771, ["filters"], "Method"], [3797, 3799, ["reduced", "model"], "Method"], [3823, 3826, ["per", "channel", "selection"], "Task"], [3847, 3848, ["selection"], "Task"], [3919, 3922, ["miss", "-", "rate"], "Metric"], [3956, 3960, ["minimum", "miss", "-", "rate"], "Metric"], [4034, 4036, ["reduced", "models"], "Method"], [4047, 4049, ["random", "filters"], "Method"], [4055, 4057, ["filter", "families"], "Method"], [4066, 4067, ["filtering"], "Method"], [4073, 4075, ["detection", "quality"], "Metric"], [4092, 4093, ["LDCF"], "Method"], [4094, 4095, ["PcaForeground"], "Method"], [4100, 4103, ["filter", "bank", "families"], "Method"], [4104, 4105, ["LDCF"], "Method"], [4122, 4125, ["Caltech", "training", "set"], "Material"], [4133, 4134, ["PcaForeground"], "Method"], [4137, 4138, ["LDCF8"], "Method"], [4162, 4163, ["filters"], "Method"], [4167, 4170, ["Caltech", "training", "data"], "Material"], [4226, 4227, ["ACF"], "Method"], [4231, 4232, ["filter"], "Method"], [4356, 4361, ["level", "-", "4", "decision", "trees"], "Method"], [4369, 4371, ["detection", "quality"], "Metric"], [4373, 4375, ["training", "times"], "Metric"], [4383, 4385, ["ChnFtrs", "detectors"], "Task"], [4388, 4390, ["boosting", "variants"], "Method"], [4402, 4403, ["Realboost"], "Method"], [4407, 4409, ["discrete", "Adaboost"], "Method"], [4423, 4425, ["Caltech10x", "setup"], "Material"], [4445, 4446, ["Realboost"], "Method"], [4450, 4453, ["detection", "quality", "boost"], "Metric"], [4472, 4473, ["RealBoost"], "Method"], [4476, 4482, ["level", "-", "2", "and", "discrete", "Adaboost"], "Method"], [4484, 4486, ["Caltech1x", "models"], "Method"], [4492, 4494, ["Caltech", "data"], "Material"], [4506, 4507, ["Checkerboards4x4"], "Method"], [4515, 4516, ["Caltech10x"], "Material"], [4517, 4519, ["training", "times"], "Metric"], [4532, 4534, ["training", "time"], "Metric"], [4545, 4548, ["hard", "negative", "mining"], "Method"], [4615, 4618, ["Caltech", "validation", "set"], "Material"], [4649, 4652, ["LDCF", "-", "Ours"], "Method"], [4659, 4660, ["Checkerboards"], "Method"], [4662, 4666, ["manual", "design", "of", "InformedFilters"], "Method"], [4717, 4718, ["complementarity"], "Task"], [4729, 4731, ["detection", "quality"], "Metric"], [4752, 4757, ["2Ped", "re", "-", "scoring", "method"], "Method"], [4769, 4771, ["detection", "scores"], "Metric"], [4776, 4779, ["two", "person", "DPM"], "Method"], [4803, 4805, ["Caltech", "set"], "Material"], [4819, 4820, ["detector"], "Method"], [4821, 4822, ["SquaresChnFtrs"], "Method"], [4823, 4824, ["DCT"], "Method"], [4825, 4826, ["SDt"], "Method"], [4845, 4847, ["2Ped", "detection"], "Task"], [4855, 4857, ["fusion", "method"], "Method"], [4893, 4895, ["2Ped", "approach"], "Method"], [4907, 4909, ["DPM", "template"], "Method"], [4922, 4924, ["Optical", "flow"], "Method"], [4947, 4948, ["SDt"], "Method"], [4958, 4959, ["Caltech"], "Material"], [4967, 4968, ["ACF"], "Method"], [4993, 4994, ["SDt"], "Method"], [5000, 5001, ["Katamari"], "Method"], [5002, 5004, ["SpatialPooling", "+"], "Method"], [5014, 5015, ["Checkerboards"], "Method"], [5016, 5017, ["SDt"], "Method"], [5048, 5050, ["Checkerboards", "+"], "Method"], [5050, 5051, ["SDt"], "Method"], [5094, 5096, ["detection", "quality"], "Metric"], [5114, 5115, ["2Ped"], "Method"], [5116, 5117, ["SDt"], "Method"], [5129, 5130, ["Caltech"], "Material"], [5135, 5139, ["context", "and", "optical", "flow"], "Task"], [5153, 5155, ["core", "detector"], "Method"], [5199, 5200, ["Caltech"], "Material"], [5207, 5210, ["Caltech", "test", "set"], "Material"], [5224, 5227, ["Caltech", "test", "set"], "Material"], [5276, 5278, ["Caltech", "images"], "Material"], [5279, 5282, ["MT", "-", "DPM"], "Method"], [5290, 5291, ["DPM"], "Method"], [5293, 5294, ["SDN"], "Method"], [5301, 5304, ["convolutional", "neural", "networks"], "Method"], [5313, 5315, ["SpatialPooling", "+"], "Method"], [5332, 5333, ["LDCF"], "Method"], [5433, 5434, ["Caltech"], "Material"], [5441, 5445, ["pre", "-", "processing", "step"], "Method"], [5447, 5451, ["HOG", "+", "LUV", "computation"], "Task"], [5471, 5472, ["Caltech"], "Material"], [5507, 5508, ["ACF"], "Method"], [5544, 5547, ["InformedHaar", "-", "Ours"], "Method"], [5612, 5613, ["Checkerboards4x3"], "Method"], [5622, 5623, ["RandomFilters"], "Method"], [5641, 5644, ["miss", "-", "rate"], "Metric"], [5648, 5649, ["convnet"], "Method"], [5650, 5652, ["pedestrian", "detection"], "Task"], [5653, 5654, ["SDN"], "Method"], [5670, 5671, ["filtering"], "Method"], [5674, 5676, ["optical", "flow"], "Method"], [5694, 5697, ["optical", "flow", "method"], "Method"], [5698, 5700, ["SpatialPooling", "+"], "Method"], [5741, 5742, ["Regionlets"], "Method"], [5745, 5746, ["SpatialPooling"], "Method"], [5749, 5750, ["HOG"], "Method"], [5782, 5785, ["Fusion", "-", "DPM"], "Method"], [5810, 5813, ["disconnected", "methods", "ACF"], "Method"], [5814, 5818, ["(", "Squares", ")", "ChnFtrs"], "Method"], [5819, 5820, ["InformedHaar"], "Method"], [5822, 5823, ["LDCF"], "Method"], [5829, 5834, ["filtered", "channel", "features", "detectors", "umbrella"], "Method"], [5840, 5842, ["filter", "banks"], "Method"], [5855, 5857, ["pedestrian", "detection"], "Task"], [5877, 5878, ["Caltech"], "Material"], [5879, 5880, ["KITTI"], "Material"], [5904, 5906, ["Caltech", "dataset"], "Material"], [5909, 5910, ["recall"], "Metric"], [5911, 5915, ["false", "positive", "per", "image"], "Metric"], [5937, 5939, ["detection", "architecture"], "Method"], [5941, 5944, ["convolutional", "neural", "networks"], "Method"], [6003, 6004, ["Checkerboards4x3"], "Method"], [6005, 6006, ["RandomFilters"], "Method"], [6071, 6074, ["weak", "Roerei", "model"], "Method"], [6168, 6169, ["detection"], "Task"], [6186, 6189, ["filter", "bank", "families"], "Method"], [6193, 6198, ["Roerei", "(", "scale", ")", "model"], "Method"], [6204, 6206, ["Checkerboards4x3", "model"], "Method"], [6209, 6211, ["RandomFilters", "model"], "Method"], [6218, 6220, ["Checkerboards4x3", "model"], "Method"], [6227, 6229, ["RandomFilters", "model"], "Method"], [79, 80, ["KITTI"], "Material"], [1507, 1508, ["SquaresChnFtrs"], "Method"], [1793, 1795, ["Caltech", "dataset"], "Material"], [2025, 2026, ["KITTI"], "Material"], [2091, 2092, ["Caltech10x"], "Material"], [2476, 2477, ["Caltech10x"], "Material"], [2656, 2657, ["SquaresChnFtrs"], "Method"], [3207, 3208, ["Caltech10x"], "Material"], [4238, 4239, ["Caltech10x"], "Material"], [4464, 4465, ["Caltech10x"], "Material"], [4594, 4595, ["Caltech10x"], "Material"], [4982, 4983, ["SquaresChnFtrs"], "Method"], [4984, 4985, ["DCT"], "Method"], [5040, 5041, ["SDt"], "Method"], [5119, 5120, ["KITTI"], "Material"], [5201, 5202, ["KITTI"], "Material"], [5324, 5325, ["Caltech10x"], "Material"], [5343, 5344, ["KITTI"], "Material"], [5355, 5356, ["KITTI"], "Material"], [5387, 5388, ["KITTI"], "Material"], [5424, 5425, ["KITTI"], "Material"], [5453, 5454, ["KITTI"], "Material"], [5488, 5489, ["KITTI"], "Material"], [5562, 5563, ["Caltech10x"], "Material"], [5717, 5718, ["KITTI"], "Material"]]} {"doc_id": "0c47cad9729c38d9db1f75491b1ee4bd883a5d4e", "text": ["document", ":", "Semi", "-", "Supervised", "Sequence", "Modeling", "with", "Cross", "-", "View", "Training", "kevclark@cs.stanford.edu", ",", "thangluong@google.com", ",", "manning@cs.stanford.edu", ",", "qvl@google.com", "Unsupervised", "representation", "learning", "algorithms", "such", "as", "word2vec", "and", "ELMo", "improve", "the", "accuracy", "of", "many", "supervised", "NLP", "models", ",", "mainly", "because", "they", "can", "take", "advantage", "of", "large", "amounts", "of", "unlabeled", "text", ".", "However", ",", "the", "supervised", "models", "only", "learn", "from", "task", "-", "specific", "labeled", "data", "during", "the", "main", "training", "phase", ".", "We", "therefore", "propose", "Cross", "-", "View", "Training", "(", "CVT", ")", ",", "a", "semi", "-", "supervised", "learning", "algorithm", "that", "improves", "the", "representations", "of", "a", "Bi", "-", "LSTM", "sentence", "encoder", "using", "a", "mix", "of", "labeled", "and", "unlabeled", "data", ".", "On", "labeled", "examples", ",", "standard", "supervised", "learning", "is", "used", ".", "On", "unlabeled", "examples", ",", "CVT", "teaches", "auxiliary", "prediction", "modules", "that", "see", "restricted", "views", "of", "the", "input", "(", "e.g.", ",", "only", "part", "of", "a", "sentence", ")", "to", "match", "the", "predictions", "of", "the", "full", "model", "seeing", "the", "whole", "input", ".", "Since", "the", "auxiliary", "modules", "and", "the", "full", "model", "share", "intermediate", "representations", ",", "this", "in", "turn", "improves", "the", "full", "model", ".", "Moreover", ",", "we", "show", "that", "CVT", "is", "particularly", "effective", "when", "combined", "with", "multi", "-", "task", "learning", ".", "We", "evaluate", "CVT", "on", "five", "sequence", "tagging", "tasks", ",", "machine", "translation", ",", "and", "dependency", "parsing", ",", "achieving", "state", "-", "of", "-", "the", "-", "art", "results", ".", "section", ":", "Introduction", "Deep", "learning", "models", "work", "best", "when", "trained", "on", "large", "amounts", "of", "labeled", "data", ".", "However", ",", "acquiring", "labels", "is", "costly", ",", "motivating", "the", "need", "for", "effective", "semi", "-", "supervised", "learning", "techniques", "that", "leverage", "unlabeled", "examples", ".", "A", "widely", "successful", "semi", "-", "supervised", "learning", "strategy", "for", "neural", "NLP", "is", "pre", "-", "training", "word", "vectors", "Mikolov2013DistributedRO", ".", "More", "recent", "work", "trains", "a", "Bi", "-", "LSTM", "sentence", "encoder", "to", "do", "language", "modeling", "and", "then", "incorporates", "its", "context", "-", "sensitive", "representations", "into", "supervised", "models", "dai2015semi", ",", "peters2018deep", ".", "Such", "pre", "-", "training", "methods", "perform", "unsupervised", "representation", "learning", "on", "a", "large", "corpus", "of", "unlabeled", "data", "followed", "by", "supervised", "training", ".", "A", "key", "disadvantage", "of", "pre", "-", "training", "is", "that", "the", "first", "representation", "learning", "phase", "does", "not", "take", "advantage", "of", "labeled", "data", "–", "the", "model", "attempts", "to", "learn", "generally", "effective", "representations", "rather", "than", "ones", "that", "are", "targeted", "towards", "a", "particular", "task", ".", "Older", "semi", "-", "supervised", "learning", "algorithms", "like", "self", "-", "training", "do", "not", "suffer", "from", "this", "problem", "because", "they", "continually", "learn", "about", "a", "task", "on", "a", "mix", "of", "labeled", "and", "unlabeled", "data", ".", "Self", "-", "training", "has", "historically", "been", "effective", "for", "NLP", "yarowsky1995unsupervised", ",", "mcclosky2006effective", ",", "but", "is", "less", "commonly", "used", "with", "neural", "models", ".", "This", "paper", "presents", "Cross", "-", "View", "Training", "(", "CVT", ")", ",", "a", "new", "self", "-", "training", "algorithm", "that", "works", "well", "for", "neural", "sequence", "models", ".", "In", "self", "-", "training", ",", "the", "model", "learns", "as", "normal", "on", "labeled", "examples", ".", "On", "unlabeled", "examples", ",", "the", "model", "acts", "as", "both", "a", "“teacher”", "that", "makes", "predictions", "about", "the", "examples", "and", "a", "“student”", "that", "is", "trained", "on", "those", "predictions", ".", "Although", "this", "process", "has", "shown", "value", "for", "some", "tasks", ",", "it", "is", "somewhat", "tautological", ":", "the", "model", "already", "produces", "the", "predictions", "it", "is", "being", "trained", "on", ".", "Recent", "research", "on", "computer", "vision", "addresses", "this", "by", "adding", "noise", "to", "the", "student", "’s", "input", ",", "training", "the", "model", "so", "it", "is", "robust", "to", "input", "perturbations", "sajjadi2016regularization", ",", "wei2018improving", ".", "However", ",", "applying", "noise", "is", "difficult", "for", "discrete", "inputs", "like", "text", ".", "As", "a", "solution", ",", "we", "take", "inspiration", "from", "multi", "-", "view", "learning", "blum1998combining", ",", "Xu2013ASO", "and", "train", "the", "model", "to", "produce", "consistent", "predictions", "across", "different", "views", "of", "the", "input", ".", "Instead", "of", "only", "training", "the", "full", "model", "as", "a", "student", ",", "CVT", "adds", "auxiliary", "prediction", "modules", "–", "neural", "networks", "that", "transform", "vector", "representations", "into", "predictions", "–", "to", "the", "model", "and", "also", "trains", "them", "as", "students", ".", "The", "input", "to", "each", "student", "prediction", "module", "is", "a", "subset", "of", "the", "model", "’s", "intermediate", "representations", "corresponding", "to", "a", "restricted", "view", "of", "the", "input", "example", ".", "For", "example", ",", "one", "auxiliary", "prediction", "module", "for", "sequence", "tagging", "is", "attached", "to", "only", "the", "“", "forward", "”", "LSTM", "in", "the", "model", "’s", "first", "Bi", "-", "LSTM", "layer", ",", "so", "it", "makes", "predictions", "without", "seeing", "any", "tokens", "to", "the", "right", "of", "the", "current", "one", ".", "CVT", "works", "by", "improving", "the", "model", "’s", "representation", "learning", ".", "The", "auxiliary", "prediction", "modules", "can", "learn", "from", "the", "full", "model", "’s", "predictions", "because", "the", "full", "model", "has", "a", "better", ",", "unrestricted", "view", "of", "the", "input", ".", "As", "the", "auxiliary", "modules", "learn", "to", "make", "accurate", "predictions", "despite", "their", "restricted", "views", "of", "the", "input", ",", "they", "improve", "the", "quality", "of", "the", "representations", "they", "are", "built", "on", "top", "of", ".", "This", "in", "turn", "improves", "the", "full", "model", ",", "which", "uses", "the", "same", "shared", "representations", ".", "In", "short", ",", "our", "method", "combines", "the", "idea", "of", "representation", "learning", "on", "unlabeled", "data", "with", "classic", "self", "-", "training", ".", "CVT", "can", "be", "applied", "to", "a", "variety", "of", "tasks", "and", "neural", "architectures", ",", "but", "we", "focus", "on", "sequence", "modeling", "tasks", "where", "the", "prediction", "modules", "are", "attached", "to", "a", "shared", "Bi", "-", "LSTM", "encoder", ".", "We", "propose", "auxiliary", "prediction", "modules", "that", "work", "well", "for", "sequence", "taggers", ",", "graph", "-", "based", "dependency", "parsers", ",", "and", "sequence", "-", "to", "-", "sequence", "models", ".", "We", "evaluate", "our", "approach", "on", "English", "dependency", "parsing", ",", "combinatory", "categorial", "grammar", "supertagging", ",", "named", "entity", "recognition", ",", "part", "-", "of", "-", "speech", "tagging", ",", "and", "text", "chunking", ",", "as", "well", "as", "English", "to", "Vietnamese", "machine", "translation", ".", "CVT", "improves", "over", "previously", "published", "results", "on", "all", "these", "tasks", ".", "Furthermore", ",", "CVT", "can", "easily", "and", "effectively", "be", "combined", "with", "multi", "-", "task", "learning", ":", "we", "just", "add", "additional", "prediction", "modules", "for", "the", "different", "tasks", "on", "top", "of", "the", "shared", "Bi", "-", "LSTM", "encoder", ".", "Training", "a", "unified", "model", "to", "jointly", "perform", "all", "of", "the", "tasks", "except", "machine", "translation", "improves", "results", "(", "outperforming", "a", "multi", "-", "task", "ELMo", "model", ")", "while", "decreasing", "the", "total", "training", "time", ".", "section", ":", "Cross", "-", "View", "Training", "We", "first", "present", "Cross", "-", "View", "Training", "and", "describe", "how", "it", "can", "be", "combined", "effectively", "with", "multi", "-", "task", "learning", ".", "See", "Figure", "[", "reference", "]", "for", "an", "overview", "of", "the", "training", "method", ".", "subsection", ":", "Method", "Let", "represent", "a", "labeled", "dataset", "and", "represent", "an", "unlabeled", "dataset", "We", "use", "to", "denote", "the", "output", "distribution", "over", "classes", "produced", "by", "the", "model", "with", "parameters", "on", "input", ".", "During", "CVT", ",", "the", "model", "alternates", "learning", "on", "a", "minibatch", "of", "labeled", "examples", "and", "learning", "on", "a", "minibatch", "of", "unlabeled", "examples", ".", "For", "labeled", "examples", ",", "CVT", "uses", "standard", "cross", "-", "entropy", "loss", ":", "CVT", "adds", "auxiliary", "prediction", "modules", "to", "the", "model", ",", "which", "are", "used", "when", "learning", "on", "unlabeled", "examples", ".", "A", "prediction", "module", "is", "usually", "a", "small", "neural", "network", "(", "e.g.", ",", "a", "hidden", "layer", "followed", "by", "a", "softmax", "layer", ")", ".", "Each", "one", "takes", "as", "input", "an", "intermediate", "representation", "produced", "by", "the", "model", "(", "e.g.", ",", "the", "outputs", "of", "one", "of", "the", "LSTMs", "in", "a", "Bi", "-", "LSTM", "model", ")", ".", "It", "outputs", "a", "distribution", "over", "labels", ".", "Each", "is", "chosen", "such", "that", "it", "only", "uses", "a", "part", "of", "the", "input", ";", "the", "particular", "choice", "can", "depend", "on", "the", "task", "and", "model", "architecture", ".", "We", "propose", "variants", "for", "several", "tasks", "in", "Section", "[", "reference", "]", ".", "The", "auxiliary", "prediction", "modules", "are", "only", "used", "during", "training", ";", "the", "test", "-", "time", "prediction", "come", "from", "the", "primary", "prediction", "module", "that", "produces", ".", "On", "an", "unlabeled", "example", ",", "the", "model", "first", "produces", "soft", "targets", "by", "performing", "inference", ".", "CVT", "trains", "the", "auxiliary", "prediction", "modules", "to", "match", "the", "primary", "prediction", "module", "on", "the", "unlabeled", "data", "by", "minimizing", "where", "is", "a", "distance", "function", "between", "probability", "distributions", "(", "we", "use", "KL", "divergence", ")", ".", "We", "hold", "the", "primary", "module", "’s", "prediction", "fixed", "during", "training", "(", "i.e.", ",", "we", "do", "not", "back", "-", "propagate", "through", "it", ")", "so", "the", "auxiliary", "modules", "learn", "to", "imitate", "the", "primary", "one", ",", "but", "not", "vice", "versa", ".", "CVT", "works", "by", "enhancing", "the", "model", "’s", "representation", "learning", ".", "As", "the", "auxiliary", "modules", "train", ",", "the", "representations", "they", "take", "as", "input", "improve", "so", "they", "are", "useful", "for", "making", "predictions", "even", "when", "some", "of", "the", "model", "’s", "inputs", "are", "not", "available", ".", "This", "in", "turn", "improves", "the", "primary", "prediction", "module", ",", "which", "is", "built", "on", "top", "of", "the", "same", "shared", "representations", ".", "We", "combine", "the", "supervised", "and", "CVT", "losses", "into", "the", "total", "loss", ",", ",", "and", "minimize", "it", "with", "stochastic", "gradient", "descent", ".", "In", "particular", ",", "we", "alternate", "minimizing", "over", "a", "minibatch", "of", "labeled", "examples", "and", "minimizing", "over", "a", "minibatch", "of", "unlabeled", "examples", ".", "For", "most", "neural", "networks", ",", "adding", "a", "few", "additional", "prediction", "modules", "is", "computationally", "cheap", "compared", "to", "the", "portion", "of", "the", "model", "building", "up", "representations", "(", "such", "as", "an", "RNN", "or", "CNN", ")", ".", "Therefore", "our", "method", "contributes", "little", "overhead", "to", "training", "time", "over", "other", "self", "-", "training", "approaches", "for", "most", "tasks", ".", "CVT", "does", "not", "change", "inference", "time", "or", "the", "number", "of", "parameters", "in", "the", "fully", "-", "trained", "model", "because", "the", "auxiliary", "prediction", "modules", "are", "only", "used", "during", "training", ".", "subsection", ":", "Combining", "CVT", "with", "Multi", "-", "Task", "Learning", "CVT", "can", "easily", "be", "combined", "with", "multi", "-", "task", "learning", "by", "adding", "additional", "prediction", "modules", "for", "the", "other", "tasks", "on", "top", "of", "the", "shared", "Bi", "-", "LSTM", "encoder", ".", "During", "supervised", "learning", ",", "we", "randomly", "select", "a", "task", "and", "then", "update", "using", "a", "minibatch", "of", "labeled", "data", "for", "that", "task", ".", "When", "learning", "on", "the", "unlabeled", "data", ",", "we", "optimize", "jointly", "across", "all", "tasks", "at", "once", ",", "first", "running", "inference", "with", "all", "the", "primary", "prediction", "modules", "and", "then", "learning", "from", "the", "predictions", "with", "all", "the", "auxiliary", "prediction", "modules", ".", "As", "before", ",", "the", "model", "alternates", "training", "on", "minibatches", "of", "labeled", "and", "unlabeled", "examples", ".", "Examples", "labeled", "across", "many", "tasks", "are", "useful", "for", "multi", "-", "task", "systems", "to", "learn", "from", ",", "but", "most", "datasets", "are", "only", "labeled", "with", "one", "task", ".", "A", "benefit", "of", "multi", "-", "task", "CVT", "is", "that", "the", "model", "creates", "(", "artificial", ")", "all", "-", "tasks", "-", "labeled", "examples", "from", "unlabeled", "data", ".", "This", "significantly", "improves", "the", "model", "’s", "data", "efficiency", "and", "training", "time", ".", "Since", "running", "prediction", "modules", "is", "computationally", "cheap", ",", "computing", "is", "not", "much", "slower", "for", "many", "tasks", "than", "it", "is", "for", "a", "single", "one", ".", "However", ",", "we", "find", "the", "all", "-", "tasks", "-", "labeled", "examples", "substantially", "speed", "up", "model", "convergence", ".", "For", "example", ",", "our", "model", "trained", "on", "six", "tasks", "takes", "about", "three", "times", "as", "long", "to", "converge", "as", "the", "average", "model", "trained", "on", "one", "task", ",", "a", "50", "%", "decrease", "in", "total", "training", "time", ".", "section", ":", "Cross", "-", "View", "Training", "Models", "CVT", "relies", "on", "auxiliary", "prediction", "modules", "that", "have", "restricted", "views", "of", "the", "input", ".", "In", "this", "section", ",", "we", "describe", "specific", "constructions", "of", "the", "auxiliary", "prediction", "modules", "that", "are", "effective", "for", "sequence", "tagging", ",", "dependency", "parsing", ",", "and", "sequence", "-", "to", "-", "sequence", "learning", ".", "subsection", ":", "Bi", "-", "LSTM", "Sentence", "Encoder", "All", "of", "our", "models", "use", "a", "two", "-", "layer", "CNN", "-", "BiLSTM", "chiu2015named", ",", "ma2016end", "sentence", "encoder", ".", "It", "takes", "as", "input", "a", "sequence", "of", "words", ".", "First", ",", "each", "word", "is", "represented", "as", "the", "sum", "of", "an", "embedding", "vector", "and", "the", "output", "of", "a", "character", "-", "level", "Convolutional", "Neural", "Network", ",", "resulting", "in", "a", "sequence", "of", "vectors", ".", "The", "encoder", "applies", "a", "two", "-", "layer", "bidirectional", "LSTM", "graves2005framewise", "to", "these", "representations", ".", "The", "first", "layer", "runs", "a", "Long", "Short", "-", "Term", "Memory", "unit", "hochreiter1997long", "in", "the", "forward", "direction", "(", "taking", "as", "input", "at", "each", "step", ")", "and", "the", "backward", "direction", "(", "taking", "at", "each", "step", ")", "to", "produce", "vector", "sequences", "and", ".", "The", "output", "of", "the", "Bi", "-", "LSTM", "is", "the", "concatenation", "of", "these", "vectors", ":", "The", "second", "Bi", "-", "LSTM", "layer", "works", "the", "same", ",", "producing", "outputs", ",", "except", "it", "takes", "as", "input", "instead", "of", ".", "subsection", ":", "CVT", "for", "Sequence", "Tagging", "In", "sequence", "tagging", ",", "each", "token", "has", "a", "corresponding", "label", ".", "The", "primary", "prediction", "module", "for", "sequence", "tagging", "produces", "a", "probability", "distribution", "over", "classes", "for", "the", "label", "using", "a", "one", "-", "hidden", "-", "layer", "neural", "network", "applied", "to", "the", "corresponding", "encoder", "outputs", ":", "The", "auxiliary", "prediction", "modules", "take", "and", ",", "the", "outputs", "of", "the", "forward", "and", "backward", "LSTMs", "in", "the", "first", "Bi", "-", "LSTM", "layer", ",", "as", "inputs", ".", "We", "add", "the", "following", "four", "auxiliary", "prediction", "modules", "to", "the", "model", "(", "see", "Figure", "[", "reference", "]", ")", ":", "The", "“", "forward", "”", "module", "makes", "each", "prediction", "without", "seeing", "the", "right", "context", "of", "the", "current", "token", ".", "The", "“", "future", "”", "module", "makes", "each", "prediction", "without", "the", "right", "context", "or", "the", "current", "token", "itself", ".", "Therefore", "it", "works", "like", "a", "neural", "language", "model", "that", ",", "instead", "of", "predicting", "which", "token", "comes", "next", "in", "the", "sequence", ",", "predicts", "which", "class", "of", "token", "comes", "next", ".", "The", "“", "backward", "”", "and", "“", "past", "”", "modules", "are", "analogous", ".", "subsection", ":", "CVT", "for", "Dependency", "Parsing", "In", "a", "dependency", "parse", ",", "words", "in", "a", "sentence", "are", "treated", "as", "nodes", "in", "a", "graph", ".", "Typed", "directed", "edges", "connect", "the", "words", ",", "forming", "a", "tree", "structure", "describing", "the", "syntactic", "structure", "of", "the", "sentence", ".", "In", "particular", ",", "each", "word", "in", "a", "sentence", "receives", "exactly", "one", "in", "-", "going", "edge", "going", "from", "word", "(", "called", "the", "“", "head", "”", ")", "to", "it", "(", "the", "“", "dependent", "”", ")", "of", "type", "(", "the", "“", "relation", "”", ")", ".", "We", "use", "a", "graph", "-", "based", "dependency", "parser", "similar", "to", "the", "one", "from", "Dozat2017Deep", ".", "This", "treats", "dependency", "parsing", "as", "a", "classification", "task", "where", "the", "goal", "is", "to", "predict", "which", "in", "-", "going", "edge", "connects", "to", "each", "word", ".", "First", ",", "the", "representations", "produced", "by", "the", "encoder", "for", "the", "candidate", "head", "and", "dependent", "are", "passed", "through", "separate", "hidden", "layers", ".", "A", "bilinear", "classifier", "applied", "to", "these", "representations", "produces", "a", "score", "for", "each", "candidate", "edge", ".", "Lastly", ",", "these", "scores", "are", "passed", "through", "a", "softmax", "layer", "to", "produce", "probabilities", ".", "Mathematically", ",", "the", "probability", "of", "an", "edge", "is", "given", "as", ":", "where", "is", "the", "scoring", "function", ":", "The", "bilinear", "classifier", "uses", "a", "weight", "matrix", "specific", "to", "the", "candidate", "relation", "as", "well", "as", "a", "weight", "matrix", "shared", "across", "all", "relations", ".", "Note", "that", "unlike", "in", "most", "prior", "work", ",", "our", "dependency", "parser", "only", "takes", "words", "as", "inputs", ",", "not", "words", "and", "part", "-", "of", "-", "speech", "tags", ".", "We", "add", "four", "auxiliary", "prediction", "modules", "to", "our", "model", "for", "cross", "-", "view", "training", ":", "Each", "one", "has", "some", "missing", "context", "(", "not", "seeing", "either", "the", "preceding", "or", "following", "words", ")", "for", "the", "candidate", "head", "and", "candidate", "dependent", ".", "subsection", ":", "CVT", "for", "Sequence", "-", "to", "-", "Sequence", "Learning", "We", "use", "an", "encoder", "-", "decoder", "sequence", "-", "to", "-", "sequence", "model", "with", "attention", "sutskever2014sequence", ",", "bahdanau2014neural", ".", "Each", "example", "consists", "of", "an", "input", "(", "source", ")", "sequence", "and", "output", "(", "target", ")", "sequence", ".", "The", "encoder", "’s", "representations", "are", "passed", "into", "an", "LSTM", "decoder", "using", "a", "bilinear", "attention", "mechanism", "Luong2015EffectiveAT", ".", "In", "particular", ",", "at", "each", "time", "step", "the", "decoder", "computes", "an", "attention", "distribution", "over", "source", "sequence", "hidden", "states", "as", "where", "is", "the", "decoder", "’s", "current", "hidden", "state", ".", "The", "source", "hidden", "states", "weighted", "by", "the", "attention", "distribution", "form", "a", "context", "vector", ":", ".", "Next", ",", "the", "context", "vector", "and", "current", "hidden", "state", "are", "combined", "into", "an", "attention", "vector", ".", "Lastly", ",", "a", "softmax", "layer", "predicts", "the", "next", "token", "in", "the", "output", "sequence", ":", ".", "We", "add", "two", "auxiliary", "decoders", "when", "applying", "CVT", ".", "The", "auxiliary", "decoders", "share", "embedding", "and", "LSTM", "parameters", "with", "the", "primary", "decoder", ",", "but", "have", "different", "parameters", "for", "the", "attention", "mechanisms", "and", "softmax", "layers", ".", "For", "the", "first", "one", ",", "we", "restrict", "its", "view", "of", "the", "input", "by", "applying", "attention", "dropout", ",", "randomly", "zeroing", "out", "a", "fraction", "of", "its", "attention", "weights", ".", "The", "second", "one", "is", "trained", "to", "predict", "the", "next", "word", "in", "the", "target", "sequence", "rather", "than", "the", "current", "one", ":", ".", "Since", "there", "is", "no", "target", "sequence", "for", "unlabeled", "examples", ",", "we", "can", "not", "apply", "teacher", "forcing", "to", "get", "an", "output", "distribution", "over", "the", "vocabulary", "from", "the", "primary", "decoder", "at", "each", "time", "step", ".", "Instead", ",", "we", "produce", "hard", "targets", "for", "the", "auxiliary", "modules", "by", "running", "the", "primary", "decoder", "with", "beam", "search", "on", "the", "input", "sequence", ".", "This", "idea", "has", "previously", "been", "applied", "to", "sequence", "-", "level", "knowledge", "distillation", "by", "Kim2016SequenceLevelKD", "and", "makes", "the", "training", "procedure", "similar", "to", "back", "-", "translation", "Sennrich2016ImprovingNM", ".", "section", ":", "Experiments", "We", "compare", "Cross", "-", "View", "Training", "against", "several", "strong", "baselines", "on", "seven", "tasks", ":", "Combinatory", "Categorial", "Grammar", "(", "CCG", ")", "Supertagging", ":", "We", "use", "data", "from", "CCGBank", "hockenmaier2007ccgbank", ".", "Text", "Chunking", ":", "We", "use", "the", "CoNLL", "-", "2000", "data", "tjong2000introduction", ".", "Named", "Entity", "Recognition", "(", "NER", ")", ":", "We", "use", "the", "CoNLL", "-", "2003", "data", "tjong2003introduction", ".", "Fine", "-", "Grained", "NER", "(", "FGN", ")", ":", "We", "use", "the", "OntoNotes", "hovy2006ontonotes", "dataset", ".", "Part", "-", "of", "-", "Speech", "(", "POS", ")", "Tagging", ":", "We", "use", "the", "Wall", "Street", "Journal", "portion", "of", "the", "Penn", "Treebank", "marcus1993building", ".", "Dependency", "Parsing", ":", "We", "use", "the", "Penn", "Treebank", "converted", "to", "Stanford", "Dependencies", "version", "3.3.0", ".", "Machine", "Translation", ":", "We", "use", "the", "English", "-", "Vietnamese", "translation", "dataset", "from", "IWSLT", "2015", "iwslt15", ".", "We", "report", "(", "tokenized", ")", "BLEU", "scores", "on", "the", "tst2013", "test", "set", ".", "We", "use", "the", "1", "Billion", "Word", "Language", "Model", "Benchmark", "chelba2013one", "as", "a", "pool", "of", "unlabeled", "sentences", "for", "semi", "-", "supervised", "learning", ".", "subsection", ":", "Model", "Details", "and", "Baselines", "We", "apply", "dropout", "during", "training", ",", "but", "not", "when", "running", "the", "primary", "prediction", "module", "to", "produce", "soft", "targets", "on", "unlabeled", "examples", ".", "In", "addition", "to", "the", "auxiliary", "prediction", "modules", "listed", "in", "Section", "[", "reference", "]", ",", "we", "find", "it", "slightly", "improves", "results", "to", "add", "another", "one", "that", "sees", "the", "whole", "input", "rather", "than", "a", "subset", "(", "but", "unlike", "the", "primary", "prediction", "module", ",", "does", "have", "dropout", "applied", "to", "its", "representations", ")", ".", "Unless", "indicated", "otherwise", ",", "our", "models", "have", "LSTMs", "with", "1024", "-", "sized", "hidden", "states", "and", "512", "-", "sized", "projection", "layers", ".", "See", "the", "appendix", "for", "full", "training", "details", "and", "hyperparameters", ".", "We", "compare", "CVT", "with", "the", "following", "other", "semi", "-", "supervised", "learning", "algorithms", ":", "Word", "Dropout", ".", "In", "this", "method", ",", "we", "only", "train", "the", "primary", "prediction", "module", ".", "When", "acting", "as", "a", "teacher", "it", "is", "run", "as", "normal", ",", "but", "when", "acting", "as", "a", "student", ",", "we", "randomly", "replace", "some", "of", "the", "input", "words", "with", "a", "REMOVED", "token", ".", "This", "is", "similar", "to", "CVT", "in", "that", "it", "exposes", "the", "model", "to", "a", "restricted", "view", "of", "the", "input", ".", "However", ",", "it", "is", "less", "data", "efficient", ".", "By", "carefully", "designing", "the", "auxiliary", "prediction", "modules", ",", "it", "is", "possible", "to", "train", "the", "auxiliary", "prediction", "modules", "to", "match", "the", "primary", "one", "across", "many", "different", "views", "of", "the", "input", "a", "once", ",", "rather", "than", "just", "one", "view", "at", "a", "time", ".", "Virtual", "Adversarial", "Training", "(", "VAT", ")", ".", "VAT", "miyato2015distributional", "works", "like", "word", "dropout", ",", "but", "adds", "noise", "to", "the", "word", "embeddings", "of", "the", "student", "instead", "of", "dropping", "out", "words", ".", "Notably", ",", "the", "noise", "is", "chosen", "adversarially", "so", "it", "most", "changes", "the", "model", "’s", "prediction", ".", "This", "method", "was", "applied", "successfully", "to", "semi", "-", "supervised", "text", "classification", "by", "miyato2016adversarial", ".", "ELMo", ".", "ELMo", "incorporates", "the", "representations", "from", "a", "large", "separately", "-", "trained", "language", "model", "into", "a", "task", "-", "specific", "model", ".", "Our", "implementaiton", "follows", "peters2018deep", ".", "When", "combining", "ELMo", "with", "multi", "-", "task", "learning", ",", "we", "allow", "each", "task", "to", "learn", "its", "own", "weights", "for", "the", "ELMo", "embeddings", "going", "into", "each", "prediction", "module", ".", "We", "found", "applying", "dropout", "to", "the", "ELMo", "embeddings", "was", "crucial", "for", "achieving", "good", "performance", ".", "subsection", ":", "Results", "Results", "are", "shown", "in", "Table", "[", "reference", "]", ".", "CVT", "on", "its", "own", "outperforms", "or", "is", "comparable", "to", "the", "best", "previously", "published", "results", "on", "all", "tasks", ".", "Figure", "[", "reference", "]", "shows", "an", "example", "win", "for", "CVT", "over", "supervised", "learning", ".", ".", "Of", "the", "prior", "results", "listed", "in", "Table", "[", "reference", "]", ",", "only", "TagLM", "and", "ELMo", "are", "semi", "-", "supervised", ".", "These", "methods", "first", "train", "an", "enormous", "language", "model", "on", "unlabeled", "data", "and", "incorporate", "the", "representations", "produced", "by", "the", "language", "model", "into", "a", "supervised", "classifier", ".", "Our", "base", "models", "use", "1024", "hidden", "units", "in", "their", "LSTMs", "(", "compared", "to", "4096", "in", "ELMo", ")", ",", "require", "fewer", "training", "steps", "(", "around", "one", "pass", "over", "the", "billion", "-", "word", "benchmark", "rather", "than", "many", "passes", ")", ",", "and", "do", "not", "require", "a", "pipelined", "training", "procedure", ".", "Therefore", ",", "although", "they", "perform", "on", "par", "with", "ELMo", ",", "they", "are", "faster", "and", "simpler", "to", "train", ".", "Increasing", "the", "size", "of", "our", "CVT", "+", "Multi", "-", "task", "model", "so", "it", "has", "4096", "units", "in", "its", "LSTMs", "like", "ELMo", "improves", "results", "further", "so", "they", "are", "significantly", "better", "than", "the", "ELMo", "+", "Multi", "-", "task", "ones", ".", "We", "suspect", "there", "could", "be", "further", "gains", "from", "combining", "our", "method", "with", "language", "model", "pre", "-", "training", ",", "which", "we", "leave", "for", "future", "work", ".", "CVT", "+", "Multi", "-", "Task", ".", "We", "train", "a", "single", "shared", "-", "encoder", "CVT", "model", "to", "perform", "all", "of", "the", "tasks", "except", "machine", "translation", "(", "as", "it", "is", "quite", "different", "and", "requires", "more", "training", "time", "than", "the", "other", "ones", ")", ".", "Multi", "-", "task", "learning", "improves", "results", "on", "all", "of", "the", "tasks", "except", "fine", "-", "grained", "NER", ",", "sometimes", "by", "large", "margins", ".", "Prior", "work", "on", "many", "-", "task", "NLP", "such", "as", "hashimoto2016joint", "uses", "complicated", "architectures", "and", "training", "algorithms", ".", "Our", "result", "shows", "that", "simple", "parameter", "sharing", "can", "be", "enough", "for", "effective", "many", "-", "task", "learning", "when", "the", "model", "is", "big", "and", "trained", "on", "a", "large", "amount", "of", "data", ".", "Interestingly", ",", "multi", "-", "task", "learning", "works", "better", "in", "conjunction", "with", "CVT", "than", "with", "ELMo", ".", "We", "hypothesize", "that", "the", "ELMo", "models", "quickly", "fit", "to", "the", "data", "primarily", "using", "the", "ELMo", "vectors", ",", "which", "perhaps", "hinders", "the", "model", "from", "learning", "effective", "representations", "that", "transfer", "across", "tasks", ".", "We", "also", "believe", "CVT", "alleviates", "the", "danger", "of", "the", "model", "“", "forgetting", "”", "one", "task", "while", "training", "on", "the", "other", "ones", ",", "a", "well", "-", "known", "problem", "in", "many", "-", "task", "learning", "Kirkpatrick2017OvercomingCF", ".", "During", "multi", "-", "task", "CVT", ",", "the", "model", "makes", "predictions", "about", "unlabeled", "examples", "across", "all", "tasks", ",", "creating", "(", "artificial", ")", "all", "-", "tasks", "-", "labeled", "examples", ",", "so", "the", "model", "does", "not", "only", "see", "one", "task", "at", "a", "time", ".", "In", "fact", ",", "multi", "-", "task", "learning", "plus", "self", "training", "is", "similar", "to", "the", "Learning", "without", "Forgetting", "algorithm", "Li2016LearningWF", ",", "which", "trains", "the", "model", "to", "keep", "its", "predictions", "on", "an", "old", "task", "unchanged", "when", "learning", "a", "new", "task", ".", "To", "test", "the", "value", "of", "all", "-", "tasks", "-", "labeled", "examples", ",", "we", "trained", "a", "multi", "-", "task", "CVT", "model", "that", "only", "computes", "on", "one", "task", "at", "a", "time", "(", "chosen", "randomly", "for", "each", "unlabeled", "minibatch", ")", "instead", "of", "for", "all", "tasks", "in", "parallel", ".", "The", "one", "-", "at", "-", "a", "-", "time", "model", "performs", "substantially", "worse", "(", "see", "Table", "[", "reference", "]", ")", ".", "Model", "Generalization", ".", "In", "order", "to", "evaluate", "how", "our", "models", "generalize", "to", "the", "dev", "set", "from", "the", "train", "set", ",", "we", "plot", "the", "dev", "vs.", "train", "accuracy", "for", "our", "different", "methods", "as", "they", "learn", "(", "see", "Figure", "[", "reference", "]", ")", ".", "Both", "CVT", "and", "multi", "-", "task", "learning", "improve", "model", "generalization", ":", "for", "the", "same", "train", "accuracy", ",", "the", "models", "get", "better", "dev", "accuracy", "than", "purely", "supervised", "learning", ".", "Interestingly", ",", "CVT", "continues", "to", "improve", "in", "dev", "set", "accuracy", "while", "close", "to", "100", "%", "train", "accuracy", "for", "CCG", ",", "Chunking", ",", "and", "NER", ",", "perhaps", "because", "the", "model", "is", "still", "learning", "from", "unlabeled", "data", "even", "when", "it", "has", "completely", "fit", "to", "the", "train", "set", ".", "We", "also", "show", "results", "for", "a", "smaller", "multi", "-", "task", "+", "CVT", "model", ".", "Although", "it", "generalizes", "at", "least", "as", "well", "as", "the", "larger", "one", ",", "it", "halts", "making", "progress", "on", "the", "train", "set", "earlier", ".", "This", "suggests", "it", "is", "important", "to", "use", "sufficiently", "large", "neural", "networks", "for", "multi", "-", "task", "learning", ":", "otherwise", "the", "model", "does", "not", "have", "the", "capacity", "to", "fit", "to", "all", "the", "training", "data", ".", "Auxiliary", "Prediction", "Module", "Ablation", ".", "We", "briefly", "explore", "which", "auxiliary", "prediction", "modules", "are", "more", "important", "for", "the", "sequence", "tagging", "tasks", "in", "Table", "[", "reference", "]", ".", "We", "find", "that", "both", "kinds", "of", "auxiliary", "prediction", "modules", "improve", "performance", ",", "but", "that", "the", "future", "and", "past", "modules", "improve", "results", "more", "than", "the", "forward", "and", "backward", "ones", ",", "perhaps", "because", "they", "see", "a", "more", "restricted", "and", "challenging", "view", "of", "the", "input", ".", "Training", "Models", "on", "Small", "Datasets", ".", "We", "explore", "how", "CVT", "scales", "with", "dataset", "size", "by", "varying", "the", "amount", "of", "training", "data", "the", "model", "has", "access", "to", ".", "Unsurprisingly", ",", "the", "improvement", "of", "CVT", "over", "purely", "supervised", "learning", "grows", "larger", "as", "the", "amount", "of", "labeled", "data", "decreases", "(", "see", "Figure", "[", "reference", "]", ",", "left", ")", ".", "Using", "only", "25", "%", "of", "the", "labeled", "data", ",", "our", "approach", "already", "performs", "as", "well", "or", "better", "than", "a", "fully", "supervised", "model", "using", "100", "%", "of", "the", "training", "data", ",", "demonstrating", "that", "CVT", "is", "particularly", "useful", "on", "small", "datasets", ".", "Training", "Larger", "Models", ".", "Most", "sequence", "taggers", "and", "dependency", "parsers", "in", "prior", "work", "use", "small", "LSTMs", "(", "hidden", "state", "sizes", "of", "around", "300", ")", "because", "larger", "models", "yield", "little", "to", "no", "gains", "in", "performance", "reimers2017reporting", ".", "We", "found", "our", "own", "supervised", "approaches", "also", "do", "not", "benefit", "greatly", "from", "increasing", "the", "model", "size", ".", "In", "contrast", ",", "when", "using", "CVT", "accuracy", "scales", "better", "with", "model", "size", "(", "see", "Figure", "[", "reference", "]", ",", "right", ")", ".", "This", "finding", "suggests", "the", "appropriate", "semi", "-", "supervised", "learning", "methods", "may", "enable", "the", "development", "of", "larger", ",", "more", "sophisticated", "models", "for", "NLP", "tasks", "with", "limited", "amounts", "of", "labeled", "data", ".", "Generalizable", "Representations", ".", "Lastly", ",", "we", "explore", "training", "the", "CVT", "+", "multi", "-", "task", "model", "on", "five", "tasks", ",", "freezing", "the", "encoder", ",", "and", "then", "only", "training", "a", "prediction", "module", "on", "the", "sixth", "task", ".", "This", "tests", "whether", "the", "encoder", "’s", "representations", "generalize", "to", "a", "new", "task", "not", "seen", "during", "its", "training", ".", "Only", "training", "the", "prediction", "module", "is", "very", "fast", "because", "(", "1", ")", "the", "encoder", "(", "which", "is", "by", "far", "the", "slowest", "part", "of", "the", "model", ")", "has", "to", "be", "run", "over", "each", "example", "only", "once", "and", "(", "2", ")", "we", "do", "not", "back", "-", "propagate", "into", "the", "encoder", ".", "Results", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Training", "only", "a", "prediction", "module", "on", "top", "of", "multi", "-", "task", "representations", "works", "remarkably", "well", ",", "outperforming", "ELMo", "embeddings", "and", "sometimes", "even", "a", "vanilla", "supervised", "model", ",", "showing", "the", "multi", "-", "task", "model", "is", "building", "up", "effective", "representations", "for", "language", ".", "In", "particular", ",", "the", "representations", "could", "be", "used", "like", "skip", "-", "thought", "vectors", "kiros2015skip", "to", "quickly", "train", "models", "on", "new", "tasks", "without", "slow", "representation", "learning", ".", "section", ":", "Related", "Work", "Unsupervised", "Representation", "Learning", ".", "Early", "approaches", "to", "deep", "semi", "-", "supervised", "learning", "pre", "-", "train", "neural", "models", "on", "unlabeled", "data", ",", "which", "has", "been", "successful", "for", "applications", "in", "computer", "vision", "jarrett2009best", ",", "lecun2010convolutional", "and", "NLP", ".", "Particularly", "noteworthy", "for", "NLP", "are", "algorithms", "for", "learning", "effective", "word", "embeddings", "and", "language", "model", "pretraining", "dai2015semi", ",", "ramachandran2016unsupervised", ",", "peters2018deep", ",", "howard2018universal", ",", "radford2018improving", ".", "Pre", "-", "training", "on", "other", "tasks", "such", "as", "machine", "translation", "has", "also", "been", "studied", "McCann2017LearnedIT", ".", "Other", "approaches", "train", "“", "thought", "vectors", "”", "representing", "sentences", "through", "unsupervised", "or", "supervised", "learning", ".", "Self", "-", "Training", ".", "One", "of", "the", "earliest", "approaches", "to", "semi", "-", "supervised", "learning", "is", "self", "-", "training", "scudder1965probability", ",", "which", "has", "been", "successfully", "applied", "to", "NLP", "tasks", "such", "as", "word", "-", "sense", "disambiguation", "yarowsky1995unsupervised", "and", "parsing", "mcclosky2006effective", ".", "In", "each", "round", "of", "training", ",", "the", "classifier", ",", "acting", "as", "a", "“", "teacher", ",", "”", "labels", "some", "of", "the", "unlabeled", "data", "and", "adds", "it", "to", "the", "training", "set", ".", "Then", ",", "acting", "as", "a", "“", "student", ",", "”", "it", "is", "retrained", "on", "the", "new", "training", "set", ".", "Many", "recent", "approaches", "(", "including", "the", "consistentency", "regularization", "methods", "discussed", "below", "and", "our", "own", "method", ")", "train", "the", "student", "with", "soft", "targets", "from", "the", "teacher", "’s", "output", "distribution", "rather", "than", "a", "hard", "label", ",", "making", "the", "procedure", "more", "akin", "to", "knowledge", "distillation", "hinton2015distilling", ".", "It", "is", "also", "possible", "to", "use", "multiple", "models", "or", "prediction", "modules", "for", "the", "teacher", ",", "such", "as", "in", "tri", "-", "training", "zhou2005tri", ",", "ruder2018strong", ".", "Consistency", "Regularization", ".", "Recent", "works", "add", "noise", "(", "e.g.", ",", "drawn", "from", "a", "Gaussian", "distribution", ")", "or", "apply", "stochastic", "transformations", "(", "e.g.", ",", "horizontally", "flipping", "an", "image", ")", "to", "the", "student", "’s", "inputs", ".", "This", "trains", "the", "model", "to", "give", "consistent", "predictions", "to", "nearby", "data", "points", ",", "encouraging", "distributional", "smoothness", "in", "the", "model", ".", "Consistency", "regularization", "has", "been", "very", "successful", "for", "computer", "vision", "applications", "bachman2014learning", ",", "laine2016temporal", ",", "tarvainen2017weight", ".", "However", ",", "stochastic", "input", "alterations", "are", "more", "difficult", "to", "apply", "to", "discrete", "data", "like", "text", ",", "making", "consistency", "regularization", "less", "used", "for", "natural", "language", "processing", ".", "One", "solution", "is", "to", "add", "noise", "to", "the", "model", "’s", "word", "embeddings", "miyato2016adversarial", ";", "we", "compare", "against", "this", "approach", "in", "our", "experiments", ".", "CVT", "is", "easily", "applicable", "to", "text", "because", "it", "does", "not", "require", "changing", "the", "student", "’s", "inputs", ".", "Multi", "-", "View", "Learning", ".", "Multi", "-", "view", "learning", "on", "data", "where", "features", "can", "be", "separated", "into", "distinct", "subsets", "has", "been", "well", "studied", "Xu2013ASO", ".", "Particularly", "relevant", "are", "co", "-", "training", "blum1998combining", "and", "co", "-", "regularization", "Sindhwani2005ACA", ",", "which", "trains", "two", "models", "with", "disjoint", "views", "of", "the", "input", ".", "On", "unlabeled", "data", ",", "each", "one", "acts", "as", "a", "“", "teacher", "”", "for", "the", "other", "model", ".", "In", "contrast", "to", "these", "methods", ",", "our", "approach", "trains", "a", "single", "unified", "model", "where", "auxiliary", "prediction", "modules", "see", "different", ",", "but", "not", "necessarily", "independent", "views", "of", "the", "input", ".", "Self", "Supervision", ".", "Self", "-", "supervised", "learning", "methods", "train", "auxiliary", "prediction", "modules", "on", "tasks", "where", "performance", "can", "be", "measured", "without", "human", "-", "provided", "labels", ".", "Recent", "work", "has", "jointly", "trained", "image", "classifiers", "with", "tasks", "like", "relative", "position", "and", "colorization", "doersch2017multi", ",", "sequence", "taggers", "with", "language", "modeling", "rei2017semi", ",", "and", "reinforcement", "learning", "agents", "with", "predicting", "changes", "in", "the", "environment", "jaderberg2016reinforcement", ".", "Unlike", "these", "approaches", ",", "our", "auxiliary", "losses", "are", "based", "on", "self", "-", "labeling", ",", "not", "labels", "deterministically", "constructed", "from", "the", "input", ".", "Multi", "-", "Task", "Learning", ".", "There", "has", "been", "extensive", "prior", "work", "on", "multi", "-", "task", "learning", "Caruana1997MultitaskL", ",", "Ruder2017AnOO", ".", "For", "NLP", ",", "most", "work", "has", "focused", "on", "a", "small", "number", "of", "closely", "related", "tasks", "Luong2015MultitaskST", ",", "zhang2016stack", ",", "Sgaard2016DeepML", ",", "Peng2017DeepML", ".", "Many", "-", "task", "systems", "are", "less", "commonly", "developed", ".", "Collobert2008AUA", "propose", "a", "many", "-", "task", "system", "sharing", "word", "embeddings", "between", "the", "tasks", ",", "hashimoto2016joint", "train", "a", "many", "-", "task", "model", "where", "the", "tasks", "are", "arranged", "hierarchically", "according", "to", "their", "linguistic", "level", ",", "and", "subramanian2018learning", "train", "a", "shared", "-", "encoder", "many", "-", "task", "model", "for", "the", "purpose", "of", "learning", "better", "sentence", "representations", "for", "use", "in", "downstream", "tasks", ",", "not", "for", "improving", "results", "on", "the", "original", "tasks", ".", "section", ":", "Conclusion", "We", "propose", "Cross", "-", "View", "Training", ",", "a", "new", "method", "for", "semi", "-", "supervised", "learning", ".", "Our", "approach", "allows", "models", "to", "effectively", "leverage", "their", "own", "predictions", "on", "unlabeled", "data", ",", "training", "them", "to", "produce", "effective", "representations", "that", "yield", "accurate", "predictions", "even", "when", "some", "of", "the", "input", "is", "not", "available", ".", "We", "achieve", "excellent", "results", "across", "seven", "NLP", "tasks", ",", "especially", "when", "CVT", "is", "combined", "with", "multi", "-", "task", "learning", ".", "section", ":", "Acknowledgements", "We", "thank", "Abi", "See", ",", "Christopher", "Clark", ",", "He", "He", ",", "Peng", "Qi", ",", "Reid", "Pryzant", ",", "Yuaho", "Zhang", ",", "and", "the", "anonymous", "reviewers", "for", "their", "thoughtful", "comments", "and", "suggestions", ".", "We", "thank", "Takeru", "Miyato", "for", "help", "with", "his", "virtual", "adversarial", "training", "code", "and", "Emma", "Strubell", "for", "answering", "our", "questions", "about", "OntoNotes", "NER", ".", "Kevin", "is", "supported", "by", "a", "Google", "PhD", "Fellowship", ".", "bibliography", ":", "References", "appendix", ":", "Detailed", "Results", "We", "provide", "a", "more", "detailed", "version", "of", "the", "test", "set", "results", "in", "the", "paper", ",", "adding", "two", "decimals", "of", "precision", ",", "standard", "deviations", "of", "the", "5", "runs", "for", "each", "model", ",", "and", "more", "prior", "work", ",", "in", "Table", "[", "reference", "]", ".", "appendix", ":", "Model", "Details", "Our", "models", "use", "two", "layer", "CNN", "-", "BiLSTM", "encoders", "chiu2015named", ",", "ma2016end", ",", "lample2016neural", "and", "task", "-", "specific", "prediction", "modules", ".", "See", "Section", "[", "reference", "]", "of", "the", "paper", "for", "details", ".", "We", "provide", "a", "few", "minor", "details", "not", "covered", "there", "below", ".", "Sequence", "Tagging", ".", "For", "Chunking", "and", "Named", "Entity", "Recognition", ",", "we", "use", "a", "BIOES", "tagging", "scheme", ".", "We", "apply", "label", "smoothing", "szegedy2016rethinking", ",", "pereyra2017regularizing", "with", "a", "rate", "of", "0.1", "to", "the", "target", "labels", "when", "training", "on", "the", "labeled", "data", ".", "Dependency", "Parsing", ".", "We", "omit", "punctuation", "from", "evaluation", ",", "which", "is", "standard", "practice", "for", "the", "PTB", "-", "SD", "3.3.0", "dataset", ".", "ROOT", "is", "represented", "with", "a", "fixed", "vector", "instead", "of", "using", "a", "vector", "from", "the", "encoder", ",", "but", "otherwise", "dependencies", "coming", "from", "ROOT", "are", "scored", "the", "same", "way", "as", "the", "other", "dependencies", ".", "Machine", "Translation", ".", "We", "apply", "dropout", "to", "the", "output", "of", "each", "LSTM", "layer", "in", "the", "decoder", ".", "Our", "implementation", "is", "heavily", "based", "off", "of", "the", "Google", "NMT", "Tutorial", "luong17", ".", "We", "attribute", "our", "significantly", "better", "results", "to", "using", "pre", "-", "trained", "word", "embeddings", ",", "a", "character", "-", "level", "CNN", ",", "a", "larger", "model", ",", "stronger", "regularization", ",", "and", "better", "hyperparameter", "tuning", ".", "Target", "words", "occurring", "5", "or", "fewer", "times", "in", "the", "train", "set", "are", "replaced", "with", "a", "UNK", "token", "(", "but", "not", "during", "evaluation", ")", ".", "We", "use", "a", "beam", "size", "of", "10", "when", "performing", "beam", "search", ".", "We", "found", "it", "slightly", "beneficial", "to", "apply", "label", "smoothing", "with", "a", "rate", "of", "0.1", "to", "the", "teacher", "’s", "predictions", "(", "unlike", "our", "other", "tasks", ",", "the", "teacher", "only", "provides", "hard", "targets", "to", "the", "students", "for", "translation", ")", ".", "Multi", "-", "Task", "Learning", ".", "Several", "of", "our", "datasets", "are", "constructed", "from", "the", "Penn", "Treebank", ".", "However", ",", "we", "treat", "them", "as", "separate", "rather", "than", "providing", "examples", "labeled", "across", "multiple", "tasks", "to", "our", "model", "during", "supervised", "training", ".", "Furthermore", ",", "the", "Penn", "Treebank", "tasks", "do", "not", "all", "use", "the", "same", "train", "/", "dev", "/", "test", "splits", ".", "We", "ensure", "the", "training", "split", "of", "one", "task", "never", "overlaps", "the", "evaluation", "split", "of", "another", "by", "discarding", "the", "overlapping", "examples", "from", "the", "train", "sets", ".", "Other", "Details", ".", "We", "apply", "dropout", "hinton2012improving", "to", "the", "word", "embeddings", "and", "outputs", "of", "each", "Bi", "-", "LSTM", ".", "We", "use", "an", "exponential", "-", "moving", "-", "average", "(", "EMA", ")", "of", "the", "model", "weights", "from", "training", "for", "the", "final", "model", ";", "we", "found", "this", "to", "slightly", "improve", "accuracy", "and", "significantly", "reduce", "the", "variance", "in", "accuracy", "between", "models", "trained", "with", "different", "random", "initializations", ".", "The", "model", "is", "trained", "using", "SGD", "with", "momentum", "polyak1964some", ",", "sutskever2013importance", ".", "Word", "embeddings", "are", "initialized", "with", "GloVe", "vectors", "pennington2014glove", "and", "fine", "-", "tuned", "during", "training", ".", "The", "full", "set", "of", "model", "hyperparameters", "are", "listed", "in", "Table", "[", "reference", "]", ".", "Baselines", ".", "Baselines", "were", "run", "with", "the", "same", "architecture", "and", "hyperparameters", "as", "the", "CVT", "model", ".", "For", "the", "“", "word", "dropout", "”", "model", ",", "we", "randomly", "replace", "words", "in", "the", "input", "sentence", "with", "a", "REMOVED", "token", "with", "probability", "0.1", "(", "this", "value", "worked", "well", "on", "the", "dev", "sets", ")", ".", "For", "Virtual", "Adversarial", "Training", ",", "we", "set", "the", "norm", "of", "the", "perturbation", "to", "be", "1.5", "for", "CCG", ",", "1.0", "for", "Dependency", "Parsing", ",", "and", "0.5", "for", "the", "other", "tasks", "(", "these", "values", "worked", "best", "on", "the", "dev", "sets", ")", ".", "Otherwise", ",", "the", "implementation", "is", "as", "described", "in", "miyato2016adversarial", ";", "we", "based", "our", "implementation", "off", "of", "their", "code", ".", "We", "were", "unable", "to", "successfully", "apply", "VAT", "to", "machine", "translation", ",", "perhaps", "because", "the", "student", "is", "provided", "hard", "targets", "for", "that", "task", ".", "For", "ELMo", ",", "we", "applied", "dropout", "to", "the", "ELMo", "embeddings", "before", "they", "are", "incorporated", "into", "the", "rest", "of", "the", "model", ".", "When", "training", "the", "multi", "-", "task", "ELMo", "model", ",", "each", "prediction", "module", "has", "its", "own", "set", "of", "softmax", "-", "normalized", "weights", "(", "in", "peters2018deep", ")", "for", "the", "ELMo", "emeddings", "going", "into", "the", "task", "-", "specific", "prediction", "modules", ".", "All", "tasks", "share", "the", "same", "weights", "for", "the", "ELMo", "embeddings", "going", "into", "the", "shared", "Bi", "-", "LSTM", "encoder", ".", "appendix", ":", "CVT", "for", "Image", "Recognition", "Although", "the", "focus", "of", "our", "work", "is", "on", "NLP", ",", "we", "also", "applied", "CVT", "to", "image", "recognition", "and", "found", "it", "performs", "competitively", "with", "existing", "methods", ".", "Most", "of", "the", "semi", "-", "supervised", "image", "recognition", "approaches", "we", "compare", "against", "rely", "on", "the", "inputs", "being", "continuous", ",", "so", "they", "would", "be", "difficult", "to", "apply", "to", "text", ".", "More", "specifically", ",", "consistency", "regularization", "methods", "sajjadi2016regularization", ",", "laine2016temporal", ",", "miyato2017virtual", "rely", "on", "adding", "continuous", "noise", "and", "applying", "image", "-", "specific", "transformations", "like", "cropping", "to", "inputs", ",", "GANs", "salimans2016improved", ",", "wei2018improving", "are", "very", "difficult", "to", "train", "on", "text", "due", "to", "its", "discrete", "nature", ",", "and", "mixup", "zhang2017mixup", ",", "verma2018manifold", "requires", "a", "way", "of", "smoothly", "interpolating", "between", "different", "inputs", ".", "Approach", ".", "Our", "image", "recognition", "models", "are", "based", "on", "Convolutional", "Neural", "Networks", ",", "which", "produce", "a", "set", "of", "features", "from", "an", "image", ".", "The", "first", "two", "dimensions", "of", "index", "into", "the", "spatial", "coordinates", "of", "feature", "vectors", "and", "is", "the", "size", "of", "the", "feature", "vectors", ".", "For", "shallower", "CNNs", ",", "a", "particular", "feature", "vector", "corresponds", "to", "a", "region", "of", "the", "input", "image", ".", "For", "example", ",", "would", "be", "a", "-", "dimensional", "vector", "of", "features", "extracted", "from", "the", "upper", "left", "corner", ".", "For", "deeper", "CNNs", ",", "a", "particular", "feature", "vector", "would", "be", "extracted", "from", "the", "whole", "image", ",", "but", "still", "only", "use", "a", "“", "region", "”", "of", "the", "representations", "from", "an", "earlier", "layer", ".", "The", "CNNs", "in", "our", "experiments", "are", "all", "in", "the", "first", "category", ".", "The", "primary", "prediction", "layers", "of", "our", "CNNs", "take", "as", "input", "the", "mean", "of", "over", "the", "first", "two", "dimensions", ",", "which", "results", "in", "a", "-", "dimensional", "vector", "that", "is", "fed", "into", "a", "softmax", "layer", ":", "We", "add", "auxiliary", "prediction", "layers", "to", "the", "top", "of", "the", "CNN", ".", "The", "th", "layer", "takes", "a", "single", "feature", "vector", "as", "input", ":", "Data", ".", "We", "evaluated", "our", "models", "on", "the", "CIFAR", "-", "10", "krizhevsky2009learning", "dataset", ".", "Following", "previous", "work", ",", "we", "make", "the", "datasets", "semi", "-", "supervised", "by", "only", "using", "the", "provided", "labels", "for", "a", "subset", "of", "the", "examples", "in", "the", "training", "set", ";", "the", "rest", "are", "treated", "as", "unlabeled", "examples", ".", "Model", ".", "We", "use", "the", "convolutional", "neural", "network", "from", "miyato2017virtual", ",", "adapting", "their", "TensorFlow", "implementation", ".", "Their", "model", "contains", "9", "convolutional", "layers", "and", "2", "max", "pooling", "layers", ".", "See", "Appendix", "D", "of", "Miyato", "et", "al", ".", "’s", "paper", "for", "more", "details", ".", "We", "add", "36", "auxiliary", "softmax", "layers", "to", "the", "collection", "of", "feature", "vectors", "produced", "by", "the", "CNN", ".", "Each", "auxiliary", "layer", "sees", "a", "patch", "of", "the", "image", "ranging", "in", "size", "from", "pixels", "(", "the", "corner", ")", "to", "pixels", "(", "the", "center", ")", "of", "the", "pixel", "images", ".", "For", "some", "experiments", ",", "we", "combine", "CVT", "with", "standard", "consistency", "regularization", "by", "adding", "a", "perturbation", "(", "e.g.", ",", "a", "small", "random", "vector", ")", "to", "the", "student", "’s", "inputs", "when", "computing", ".", "Results", ".", "The", "results", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Unsurprisingly", ",", "adding", "continuous", "noise", "to", "the", "inputs", "works", "much", "better", "with", "images", ",", "where", "the", "inputs", "are", "naturally", "continuous", ",", "than", "with", "language", ".", "Therefore", "we", "see", "much", "better", "results", "from", "VAT", "on", "semi", "-", "supervised", "CIFAR", "-", "10", "compared", "to", "on", "our", "NLP", "tasks", ".", "However", ",", "we", "still", "find", "incorporating", "CVT", "improves", "over", "models", "without", "CVT", ".", "Our", "CVT", "+", "VAT", "models", "are", "competitive", "with", "current", "start", "-", "of", "-", "the", "-", "art", "approaches", ".", "We", "found", "the", "gains", "from", "CVT", "are", "larger", "when", "no", "data", "augmentation", "is", "applied", ",", "perhaps", "because", "random", "translations", "of", "the", "input", "expose", "the", "model", "to", "different", "“", "views", "”", "in", "a", "similar", "manner", "as", "with", "CVT", ".", "appendix", ":", "Negative", "Results", "We", "briefly", "describe", "a", "few", "ideas", "we", "implemented", "that", "did", "not", "seem", "to", "be", "effective", "in", "initial", "experiments", ".", "Note", "these", "findings", "are", "from", "early", "one", "-", "off", "experiments", ".", "We", "did", "not", "pursue", "them", "further", "after", "our", "first", "attempts", "did", "not", "pan", "out", ",", "so", "it", "is", "possible", "that", "some", "of", "these", "approaches", "could", "be", "effective", "with", "the", "proper", "adjustments", "and", "tuning", ".", "Hard", "vs", "soft", "targets", ":", "Classic", "self", "-", "training", "algorithms", "train", "the", "student", "model", "with", "one", "-", "hot", "“", "hard", "”", "targets", "corresponding", "to", "the", "teacher", "’s", "highest", "probability", "prediction", ".", "In", "our", "experiments", ",", "this", "decreased", "performance", "compared", "to", "using", "soft", "targets", ".", "This", "finding", "is", "consistent", "with", "research", "on", "knowledge", "distillation", "hinton2015distilling", ",", "furlanello2018born", "where", "soft", "targets", "also", "work", "notably", "better", "than", "hard", "targets", ".", "Confidence", "thresholding", ":", "Classic", "self", "-", "training", "often", "only", "trains", "the", "student", "on", "a", "subset", "of", "the", "unlabeled", "examples", "on", "which", "the", "teacher", "has", "confident", "predictions", "(", "i.e.", ",", "the", "output", "distribution", "has", "low", "entropy", ")", ".", "We", "tried", "both", "“", "hard", "”", "(", "where", "the", "student", "ignores", "low", "-", "confidence", "examples", ")", "and", "“", "soft", "”", "(", "where", "examples", "are", "weighted", "according", "to", "the", "teacher", "’s", "confidence", ")", "versions", "of", "this", "for", "training", "our", "models", ",", "but", "they", "did", "not", "seem", "to", "improve", "performance", ".", "Mean", "Teacher", ":", "The", "Mean", "Teacher", "method", "tarvainen2017weight", "tracks", "an", "exponential", "moving", "average", "(", "EMA", ")", "of", "model", "weights", ",", "which", "are", "used", "to", "produce", "targets", "for", "the", "students", ".", "The", "idea", "is", "that", "these", "targets", "may", "be", "better", "quality", "due", "to", "a", "self", "-", "ensembling", "effect", ".", "However", ",", "we", "found", "this", "approach", "to", "have", "little", "to", "no", "benefit", "in", "our", "experiments", ",", "although", "using", "EMA", "model", "weights", "at", "test", "time", "did", "improve", "results", "slightly", ".", "Purely", "supervised", "CVT", ":", "Lastly", ",", "we", "explored", "adding", "cross", "-", "view", "losses", "to", "purely", "supervised", "classifiers", ".", "We", "hoped", "that", "adding", "auxiliary", "softmax", "layers", "with", "different", "views", "of", "the", "input", "would", "act", "as", "a", "regularizer", "on", "the", "model", ".", "However", ",", "we", "found", "little", "to", "no", "benefit", "from", "this", "approach", ".", "This", "negative", "result", "suggests", "that", "the", "gains", "from", "CVT", "are", "from", "the", "improved", "semi", "-", "supervised", "learning", "mechanism", ",", "not", "the", "additional", "prediction", "layers", "regularizing", "the", "model", "."], "entity": [[2, 7, ["Semi", "-", "Supervised", "Sequence", "Modeling"], "Method"], [8, 12, ["Cross", "-", "View", "Training"], "Method"], [19, 23, ["Unsupervised", "representation", "learning", "algorithms"], "Method"], [25, 26, ["word2vec"], "Method"], [27, 28, ["ELMo"], "Method"], [30, 31, ["accuracy"], "Metric"], [33, 36, ["supervised", "NLP", "models"], "Method"], [53, 55, ["supervised", "models"], "Method"], [72, 76, ["Cross", "-", "View", "Training"], "Method"], [77, 78, ["CVT"], "Method"], [81, 86, ["semi", "-", "supervised", "learning", "algorithm"], "Method"], [92, 97, ["Bi", "-", "LSTM", "sentence", "encoder"], "Method"], [111, 113, ["supervised", "learning"], "Method"], [120, 121, ["CVT"], "Method"], [122, 125, ["auxiliary", "prediction", "modules"], "Method"], [147, 149, ["full", "model"], "Method"], [156, 158, ["auxiliary", "modules"], "Method"], [160, 162, ["full", "model"], "Method"], [163, 165, ["intermediate", "representations"], "Method"], [171, 173, ["full", "model"], "Method"], [179, 180, ["CVT"], "Method"], [193, 194, ["CVT"], "Method"], [196, 199, ["sequence", "tagging", "tasks"], "Task"], [200, 202, ["machine", "translation"], "Task"], [204, 206, ["dependency", "parsing"], "Task"], [220, 223, ["Deep", "learning", "models"], "Method"], [236, 238, ["acquiring", "labels"], "Task"], [246, 251, ["semi", "-", "supervised", "learning", "techniques"], "Method"], [259, 264, ["semi", "-", "supervised", "learning", "strategy"], "Method"], [265, 267, ["neural", "NLP"], "Task"], [280, 285, ["Bi", "-", "LSTM", "sentence", "encoder"], "Method"], [287, 289, ["language", "modeling"], "Task"], [293, 297, ["context", "-", "sensitive", "representations"], "Method"], [298, 300, ["supervised", "models"], "Method"], [305, 309, ["pre", "-", "training", "methods"], "Method"], [310, 313, ["unsupervised", "representation", "learning"], "Method"], [322, 324, ["supervised", "training"], "Method"], [329, 332, ["pre", "-", "training"], "Task"], [336, 339, ["representation", "learning", "phase"], "Method"], [367, 372, ["semi", "-", "supervised", "learning", "algorithms"], "Method"], [373, 376, ["self", "-", "training"], "Method"], [398, 401, ["Self", "-", "training"], "Method"], [406, 407, ["NLP"], "Method"], [417, 419, ["neural", "models"], "Method"], [423, 427, ["Cross", "-", "View", "Training"], "Method"], [428, 429, ["CVT"], "Method"], [433, 437, ["self", "-", "training", "algorithm"], "Method"], [441, 444, ["neural", "sequence", "models"], "Method"], [446, 449, ["self", "-", "training"], "Task"], [469, 470, ["“teacher”"], "Method"], [516, 518, ["computer", "vision"], "Task"], [563, 567, ["multi", "-", "view", "learning"], "Method"], [569, 570, ["Xu2013ASO"], "Method"], [590, 592, ["full", "model"], "Method"], [596, 597, ["CVT"], "Method"], [598, 601, ["auxiliary", "prediction", "modules"], "Method"], [602, 604, ["neural", "networks"], "Method"], [606, 608, ["vector", "representations"], "Method"], [625, 628, ["student", "prediction", "module"], "Method"], [635, 637, ["intermediate", "representations"], "Method"], [651, 654, ["auxiliary", "prediction", "module"], "Method"], [655, 657, ["sequence", "tagging"], "Task"], [663, 666, ["forward", "”", "LSTM"], "Method"], [671, 675, ["Bi", "-", "LSTM", "layer"], "Method"], [692, 693, ["CVT"], "Method"], [699, 701, ["representation", "learning"], "Method"], [703, 706, ["auxiliary", "prediction", "modules"], "Method"], [710, 712, ["full", "model"], "Method"], [716, 718, ["full", "model"], "Method"], [730, 732, ["auxiliary", "modules"], "Method"], [764, 766, ["full", "model"], "Method"], [783, 785, ["representation", "learning"], "Method"], [790, 793, ["self", "-", "training"], "Method"], [794, 795, ["CVT"], "Method"], [804, 806, ["neural", "architectures"], "Method"], [811, 814, ["sequence", "modeling", "tasks"], "Task"], [816, 818, ["prediction", "modules"], "Method"], [822, 827, ["shared", "Bi", "-", "LSTM", "encoder"], "Method"], [830, 833, ["auxiliary", "prediction", "modules"], "Method"], [837, 839, ["sequence", "taggers"], "Task"], [840, 845, ["graph", "-", "based", "dependency", "parsers"], "Method"], [847, 853, ["sequence", "-", "to", "-", "sequence", "models"], "Method"], [863, 867, ["combinatory", "categorial", "grammar", "supertagging"], "Task"], [868, 871, ["named", "entity", "recognition"], "Task"], [872, 878, ["part", "-", "of", "-", "speech", "tagging"], "Task"], [880, 882, ["text", "chunking"], "Task"], [886, 891, ["English", "to", "Vietnamese", "machine", "translation"], "Task"], [892, 893, ["CVT"], "Method"], [905, 906, ["CVT"], "Method"], [922, 924, ["prediction", "modules"], "Method"], [932, 937, ["shared", "Bi", "-", "LSTM", "encoder"], "Method"], [940, 942, ["unified", "model"], "Method"], [950, 952, ["machine", "translation"], "Task"], [967, 969, ["training", "time"], "Metric"], [972, 976, ["Cross", "-", "View", "Training"], "Method"], [979, 983, ["Cross", "-", "View", "Training"], "Method"], [1007, 1009, ["training", "method"], "Method"], [1042, 1043, ["CVT"], "Method"], [1047, 1048, ["learning"], "Method"], [1067, 1068, ["CVT"], "Method"], [1070, 1074, ["cross", "-", "entropy", "loss"], "Method"], [1075, 1076, ["CVT"], "Method"], [1077, 1080, ["auxiliary", "prediction", "modules"], "Method"], [1094, 1096, ["prediction", "module"], "Method"], [1099, 1102, ["small", "neural", "network"], "Method"], [1111, 1113, ["softmax", "layer"], "Method"], [1136, 1137, ["LSTMs"], "Method"], [1139, 1143, ["Bi", "-", "LSTM", "model"], "Method"], [1175, 1177, ["model", "architecture"], "Method"], [1191, 1194, ["auxiliary", "prediction", "modules"], "Method"], [1201, 1205, ["test", "-", "time", "prediction"], "Task"], [1208, 1211, ["primary", "prediction", "module"], "Method"], [1227, 1228, ["inference"], "Task"], [1229, 1230, ["CVT"], "Method"], [1232, 1235, ["auxiliary", "prediction", "modules"], "Method"], [1238, 1241, ["primary", "prediction", "module"], "Method"], [1286, 1288, ["auxiliary", "modules"], "Method"], [1300, 1301, ["CVT"], "Method"], [1307, 1309, ["representation", "learning"], "Method"], [1312, 1314, ["auxiliary", "modules"], "Method"], [1328, 1330, ["making", "predictions"], "Task"], [1347, 1350, ["primary", "prediction", "module"], "Method"], [1359, 1361, ["shared", "representations"], "Method"], [1371, 1373, ["total", "loss"], "Metric"], [1379, 1382, ["stochastic", "gradient", "descent"], "Method"], [1388, 1389, ["minimizing"], "Task"], [1406, 1408, ["neural", "networks"], "Method"], [1413, 1415, ["prediction", "modules"], "Method"], [1432, 1433, ["RNN"], "Method"], [1434, 1435, ["CNN"], "Method"], [1444, 1446, ["training", "time"], "Metric"], [1448, 1452, ["self", "-", "training", "approaches"], "Method"], [1456, 1457, ["CVT"], "Method"], [1460, 1462, ["inference", "time"], "Metric"], [1469, 1473, ["fully", "-", "trained", "model"], "Method"], [1475, 1478, ["auxiliary", "prediction", "modules"], "Method"], [1487, 1488, ["CVT"], "Method"], [1493, 1494, ["CVT"], "Method"], [1506, 1508, ["prediction", "modules"], "Method"], [1516, 1521, ["shared", "Bi", "-", "LSTM", "encoder"], "Method"], [1523, 1525, ["supervised", "learning"], "Task"], [1562, 1563, ["inference"], "Task"], [1566, 1569, ["primary", "prediction", "modules"], "Method"], [1578, 1581, ["auxiliary", "prediction", "modules"], "Method"], [1654, 1656, ["data", "efficiency"], "Metric"], [1657, 1659, ["training", "time"], "Metric"], [1662, 1664, ["prediction", "modules"], "Method"], [1668, 1669, ["computing"], "Task"], [1698, 1700, ["model", "convergence"], "Task"], [1733, 1735, ["training", "time"], "Metric"], [1743, 1744, ["CVT"], "Method"], [1746, 1749, ["auxiliary", "prediction", "modules"], "Method"], [1767, 1770, ["auxiliary", "prediction", "modules"], "Method"], [1774, 1776, ["sequence", "tagging"], "Task"], [1777, 1779, ["dependency", "parsing"], "Task"], [1781, 1787, ["sequence", "-", "to", "-", "sequence", "learning"], "Task"], [1790, 1795, ["Bi", "-", "LSTM", "Sentence", "Encoder"], "Method"], [1801, 1807, ["two", "-", "layer", "CNN", "-", "BiLSTM"], "Method"], [1810, 1812, ["sentence", "encoder"], "Method"], [1840, 1846, ["character", "-", "level", "Convolutional", "Neural", "Network"], "Method"], [1858, 1862, ["two", "-", "layer", "bidirectional"], "Method"], [1873, 1880, ["Long", "Short", "-", "Term", "Memory", "unit", "hochreiter1997long"], "Method"], [1912, 1915, ["Bi", "-", "LSTM"], "Method"], [1924, 1928, ["Bi", "-", "LSTM", "layer"], "Method"], [1945, 1946, ["CVT"], "Method"], [1947, 1949, ["Sequence", "Tagging"], "Task"], [1950, 1952, ["sequence", "tagging"], "Task"], [1961, 1964, ["primary", "prediction", "module"], "Method"], [1965, 1967, ["sequence", "tagging"], "Task"], [1978, 1985, ["one", "-", "hidden", "-", "layer", "neural", "network"], "Method"], [1993, 1996, ["auxiliary", "prediction", "modules"], "Method"], [2003, 2007, ["forward", "and", "backward", "LSTMs"], "Method"], [2010, 2014, ["Bi", "-", "LSTM", "layer"], "Method"], [2024, 2026, ["prediction", "modules"], "Method"], [2039, 2042, ["forward", "”", "module"], "Method"], [2057, 2060, ["future", "”", "module"], "Method"], [2078, 2081, ["neural", "language", "model"], "Method"], [2108, 2111, ["past", "”", "modules"], "Method"], [2116, 2117, ["CVT"], "Method"], [2118, 2120, ["Dependency", "Parsing"], "Task"], [2122, 2124, ["dependency", "parse"], "Task"], [2201, 2206, ["graph", "-", "based", "dependency", "parser"], "Method"], [2215, 2217, ["dependency", "parsing"], "Task"], [2219, 2221, ["classification", "task"], "Task"], [2255, 2257, ["hidden", "layers"], "Method"], [2259, 2261, ["bilinear", "classifier"], "Method"], [2281, 2283, ["softmax", "layer"], "Method"], [2305, 2307, ["bilinear", "classifier"], "Method"], [2336, 2338, ["dependency", "parser"], "Method"], [2357, 2360, ["auxiliary", "prediction", "modules"], "Method"], [2364, 2368, ["cross", "-", "view", "training"], "Task"], [2395, 2396, ["CVT"], "Method"], [2397, 2403, ["Sequence", "-", "to", "-", "Sequence", "Learning"], "Task"], [2406, 2415, ["encoder", "-", "decoder", "sequence", "-", "to", "-", "sequence", "model"], "Method"], [2416, 2417, ["attention"], "Method"], [2439, 2442, ["encoder", "’s", "representations"], "Method"], [2446, 2448, ["LSTM", "decoder"], "Method"], [2450, 2453, ["bilinear", "attention", "mechanism"], "Method"], [2463, 2464, ["decoder"], "Method"], [2466, 2468, ["attention", "distribution"], "Method"], [2517, 2519, ["softmax", "layer"], "Method"], [2532, 2534, ["auxiliary", "decoders"], "Method"], [2536, 2537, ["CVT"], "Method"], [2539, 2541, ["auxiliary", "decoders"], "Method"], [2548, 2550, ["primary", "decoder"], "Method"], [2557, 2559, ["attention", "mechanisms"], "Method"], [2577, 2579, ["attention", "dropout"], "Method"], [2625, 2627, ["teacher", "forcing"], "Method"], [2637, 2639, ["primary", "decoder"], "Method"], [2652, 2654, ["auxiliary", "modules"], "Method"], [2657, 2659, ["primary", "decoder"], "Method"], [2660, 2662, ["beam", "search"], "Method"], [2674, 2679, ["sequence", "-", "level", "knowledge", "distillation"], "Task"], [2684, 2686, ["training", "procedure"], "Method"], [2688, 2691, ["back", "-", "translation"], "Method"], [2698, 2702, ["Cross", "-", "View", "Training"], "Method"], [2710, 2713, ["Combinatory", "Categorial", "Grammar"], "Task"], [2714, 2715, ["CCG"], "Task"], [2716, 2717, ["Supertagging"], "Task"], [2722, 2723, ["CCGBank"], "Material"], [2725, 2727, ["Text", "Chunking"], "Task"], [2737, 2740, ["Named", "Entity", "Recognition"], "Task"], [2741, 2742, ["NER"], "Task"], [2747, 2750, ["CoNLL", "-", "2003"], "Material"], [2753, 2757, ["Fine", "-", "Grained", "NER"], "Task"], [2758, 2759, ["FGN"], "Task"], [2764, 2765, ["OntoNotes"], "Material"], [2768, 2773, ["Part", "-", "of", "-", "Speech"], "Metric"], [2774, 2775, ["POS"], "Metric"], [2787, 2789, ["Penn", "Treebank"], "Material"], [2791, 2793, ["Dependency", "Parsing"], "Task"], [2797, 2799, ["Penn", "Treebank"], "Material"], [2806, 2808, ["Machine", "Translation"], "Task"], [2825, 2829, ["tokenized", ")", "BLEU", "scores"], "Metric"], [2852, 2856, ["semi", "-", "supervised", "learning"], "Task"], [2865, 2866, ["dropout"], "Method"], [2874, 2877, ["primary", "prediction", "module"], "Method"], [2890, 2892, ["prediction", "modules"], "Method"], [2922, 2925, ["primary", "prediction", "module"], "Method"], [2942, 2943, ["LSTMs"], "Method"], [2950, 2955, ["512", "-", "sized", "projection", "layers"], "Method"], [2968, 2969, ["CVT"], "Method"], [2973, 2978, ["semi", "-", "supervised", "learning", "algorithms"], "Method"], [2979, 2981, ["Word", "Dropout"], "Method"], [2990, 2993, ["primary", "prediction", "module"], "Method"], [3029, 3030, ["CVT"], "Method"], [3056, 3059, ["auxiliary", "prediction", "modules"], "Method"], [3066, 3069, ["auxiliary", "prediction", "modules"], "Method"], [3093, 3096, ["Virtual", "Adversarial", "Training"], "Method"], [3097, 3098, ["VAT"], "Method"], [3100, 3101, ["VAT"], "Method"], [3104, 3106, ["word", "dropout"], "Method"], [3137, 3138, ["prediction"], "Task"], [3145, 3150, ["semi", "-", "supervised", "text", "classification"], "Task"], [3153, 3154, ["ELMo"], "Method"], [3155, 3156, ["ELMo"], "Method"], [3162, 3167, ["separately", "-", "trained", "language", "model"], "Method"], [3169, 3173, ["task", "-", "specific", "model"], "Method"], [3181, 3182, ["ELMo"], "Method"], [3204, 3206, ["prediction", "module"], "Method"], [3210, 3211, ["dropout"], "Method"], [3234, 3235, ["CVT"], "Method"], [3261, 3262, ["CVT"], "Method"], [3263, 3265, ["supervised", "learning"], "Method"], [3279, 3280, ["TagLM"], "Method"], [3281, 3282, ["ELMo"], "Method"], [3293, 3295, ["language", "model"], "Method"], [3305, 3307, ["language", "model"], "Method"], [3309, 3311, ["supervised", "classifier"], "Method"], [3321, 3322, ["LSTMs"], "Method"], [3327, 3328, ["ELMo"], "Method"], [3355, 3358, ["pipelined", "training", "procedure"], "Method"], [3367, 3368, ["ELMo"], "Method"], [3382, 3383, ["CVT"], "Method"], [3384, 3387, ["Multi", "-", "task"], "Method"], [3395, 3396, ["LSTMs"], "Method"], [3397, 3398, ["ELMo"], "Method"], [3408, 3409, ["ELMo"], "Method"], [3410, 3413, ["Multi", "-", "task"], "Method"], [3427, 3432, ["language", "model", "pre", "-", "training"], "Method"], [3440, 3441, ["CVT"], "Method"], [3442, 3445, ["Multi", "-", "Task"], "Method"], [3462, 3464, ["machine", "translation"], "Task"], [3493, 3497, ["fine", "-", "grained", "NER"], "Task"], [3506, 3510, ["many", "-", "task", "NLP"], "Task"], [3515, 3516, ["architectures"], "Method"], [3517, 3519, ["training", "algorithms"], "Method"], [3525, 3527, ["parameter", "sharing"], "Method"], [3532, 3536, ["many", "-", "task", "learning"], "Task"], [3561, 3562, ["CVT"], "Method"], [3564, 3565, ["ELMo"], "Method"], [3600, 3601, ["CVT"], "Method"], [3625, 3629, ["many", "-", "task", "learning"], "Method"], [3680, 3682, ["self", "training"], "Task"], [3686, 3690, ["Learning", "without", "Forgetting", "algorithm"], "Method"], [3757, 3765, ["one", "-", "at", "-", "a", "-", "time", "model"], "Method"], [3776, 3778, ["Model", "Generalization"], "Task"], [3799, 3800, ["dev"], "Metric"], [3801, 3803, ["train", "accuracy"], "Metric"], [3819, 3820, ["CVT"], "Method"], [3826, 3828, ["model", "generalization"], "Method"], [3832, 3834, ["train", "accuracy"], "Metric"], [3843, 3845, ["supervised", "learning"], "Method"], [3848, 3849, ["CVT"], "Method"], [3861, 3863, ["train", "accuracy"], "Metric"], [3864, 3865, ["CCG"], "Task"], [3866, 3867, ["Chunking"], "Task"], [3869, 3870, ["NER"], "Task"], [3899, 3902, ["multi", "-", "task"], "Method"], [3903, 3904, ["CVT"], "Method"], [3937, 3939, ["neural", "networks"], "Method"], [3961, 3965, ["Auxiliary", "Prediction", "Module", "Ablation"], "Method"], [3970, 3973, ["auxiliary", "prediction", "modules"], "Method"], [3978, 3981, ["sequence", "tagging", "tasks"], "Task"], [3993, 3996, ["auxiliary", "prediction", "modules"], "Method"], [4004, 4006, ["past", "modules"], "Method"], [4030, 4032, ["Training", "Models"], "Method"], [4039, 4040, ["CVT"], "Method"], [4062, 4063, ["CVT"], "Method"], [4065, 4067, ["supervised", "learning"], "Method"], [4105, 4108, ["fully", "supervised", "model"], "Method"], [4118, 4119, ["CVT"], "Method"], [4131, 4133, ["sequence", "taggers"], "Method"], [4134, 4136, ["dependency", "parsers"], "Method"], [4141, 4142, ["LSTMs"], "Method"], [4166, 4168, ["supervised", "approaches"], "Method"], [4206, 4211, ["semi", "-", "supervised", "learning", "methods"], "Method"], [4222, 4224, ["NLP", "tasks"], "Task"], [4231, 4233, ["Generalizable", "Representations"], "Method"], [4240, 4241, ["CVT"], "Method"], [4242, 4246, ["multi", "-", "task", "model"], "Method"], [4252, 4253, ["encoder"], "Method"], [4259, 4261, ["prediction", "module"], "Method"], [4270, 4273, ["encoder", "’s", "representations"], "Method"], [4287, 4289, ["prediction", "module"], "Method"], [4297, 4298, ["encoder"], "Method"], [4331, 4332, ["encoder"], "Method"], [4345, 4347, ["prediction", "module"], "Method"], [4365, 4368, ["vanilla", "supervised", "model"], "Method"], [4371, 4375, ["multi", "-", "task", "model"], "Method"], [4381, 4382, ["language"], "Task"], [4392, 4396, ["skip", "-", "thought", "vectors"], "Method"], [4405, 4408, ["slow", "representation", "learning"], "Method"], [4413, 4416, ["Unsupervised", "Representation", "Learning"], "Method"], [4420, 4425, ["deep", "semi", "-", "supervised", "learning"], "Task"], [4428, 4430, ["neural", "models"], "Method"], [4441, 4443, ["computer", "vision"], "Task"], [4447, 4448, ["NLP"], "Task"], [4452, 4453, ["NLP"], "Method"], [4457, 4460, ["effective", "word", "embeddings"], "Task"], [4461, 4464, ["language", "model", "pretraining"], "Method"], [4474, 4477, ["Pre", "-", "training"], "Task"], [4482, 4484, ["machine", "translation"], "Task"], [4500, 4504, ["unsupervised", "or", "supervised", "learning"], "Method"], [4505, 4508, ["Self", "-", "Training"], "Method"], [4515, 4519, ["semi", "-", "supervised", "learning"], "Task"], [4520, 4523, ["self", "-", "training"], "Method"], [4531, 4533, ["NLP", "tasks"], "Task"], [4535, 4539, ["word", "-", "sense", "disambiguation"], "Task"], [4541, 4542, ["parsing"], "Task"], [4551, 4552, ["classifier"], "Method"], [4598, 4601, ["consistentency", "regularization", "methods"], "Method"], [4632, 4634, ["knowledge", "distillation"], "Task"], [4645, 4647, ["prediction", "modules"], "Method"], [4654, 4657, ["tri", "-", "training"], "Task"], [4661, 4663, ["Consistency", "Regularization"], "Method"], [4674, 4676, ["Gaussian", "distribution"], "Method"], [4679, 4681, ["stochastic", "transformations"], "Method"], [4715, 4717, ["Consistency", "regularization"], "Method"], [4722, 4725, ["computer", "vision", "applications"], "Task"], [4748, 4750, ["consistency", "regularization"], "Method"], [4753, 4756, ["natural", "language", "processing"], "Task"], [4767, 4769, ["word", "embeddings"], "Method"], [4780, 4781, ["CVT"], "Method"], [4797, 4801, ["Multi", "-", "View", "Learning"], "Method"], [4802, 4806, ["Multi", "-", "view", "learning"], "Method"], [4825, 4828, ["co", "-", "training"], "Method"], [4830, 4833, ["co", "-", "regularization"], "Method"], [4874, 4876, ["unified", "model"], "Method"], [4877, 4880, ["auxiliary", "prediction", "modules"], "Method"], [4892, 4894, ["Self", "Supervision"], "Task"], [4895, 4900, ["Self", "-", "supervised", "learning", "methods"], "Method"], [4901, 4904, ["auxiliary", "prediction", "modules"], "Method"], [4922, 4924, ["image", "classifiers"], "Method"], [4933, 4935, ["sequence", "taggers"], "Method"], [4936, 4938, ["language", "modeling"], "Method"], [4941, 4944, ["reinforcement", "learning", "agents"], "Method"], [4957, 4959, ["auxiliary", "losses"], "Method"], [4962, 4965, ["self", "-", "labeling"], "Method"], [4995, 4996, ["NLP"], "Task"], [5017, 5021, ["Many", "-", "task", "systems"], "Method"], [5029, 5033, ["many", "-", "task", "system"], "Method"], [5034, 5036, ["word", "embeddings"], "Method"], [5043, 5047, ["many", "-", "task", "model"], "Method"], [5063, 5070, ["shared", "-", "encoder", "many", "-", "task", "model"], "Method"], [5076, 5078, ["sentence", "representations"], "Task"], [5081, 5083, ["downstream", "tasks"], "Task"], [5098, 5102, ["Cross", "-", "View", "Training"], "Method"], [5107, 5111, ["semi", "-", "supervised", "learning"], "Task"], [5152, 5154, ["NLP", "tasks"], "Task"], [5157, 5158, ["CVT"], "Method"], [5208, 5212, ["virtual", "adversarial", "training", "code"], "Method"], [5220, 5222, ["OntoNotes", "NER"], "Task"], [5288, 5294, ["two", "layer", "CNN", "-", "BiLSTM", "encoders"], "Method"], [5300, 5305, ["task", "-", "specific", "prediction", "modules"], "Method"], [5328, 5330, ["Sequence", "Tagging"], "Task"], [5332, 5333, ["Chunking"], "Task"], [5334, 5337, ["Named", "Entity", "Recognition"], "Task"], [5341, 5344, ["BIOES", "tagging", "scheme"], "Method"], [5348, 5349, ["smoothing"], "Method"], [5351, 5352, ["pereyra2017regularizing"], "Method"], [5368, 5370, ["Dependency", "Parsing"], "Task"], [5403, 5404, ["encoder"], "Method"], [5421, 5423, ["Machine", "Translation"], "Task"], [5426, 5427, ["dropout"], "Method"], [5432, 5434, ["LSTM", "layer"], "Method"], [5436, 5437, ["decoder"], "Method"], [5446, 5450, ["Google", "NMT", "Tutorial", "luong17"], "Method"], [5462, 5464, ["word", "embeddings"], "Method"], [5466, 5470, ["character", "-", "level", "CNN"], "Method"], [5476, 5477, ["regularization"], "Method"], [5480, 5482, ["hyperparameter", "tuning"], "Method"], [5516, 5518, ["beam", "search"], "Method"], [5526, 5528, ["label", "smoothing"], "Method"], [5554, 5555, ["translation"], "Task"], [5570, 5572, ["Penn", "Treebank"], "Material"], [5592, 5594, ["supervised", "training"], "Task"], [5644, 5645, ["dropout"], "Method"], [5654, 5657, ["Bi", "-", "LSTM"], "Method"], [5671, 5673, ["model", "weights"], "Method"], [5686, 5687, ["accuracy"], "Metric"], [5693, 5694, ["accuracy"], "Metric"], [5699, 5701, ["random", "initializations"], "Method"], [5707, 5708, ["SGD"], "Method"], [5709, 5710, ["momentum"], "Method"], [5753, 5754, ["hyperparameters"], "Method"], [5762, 5766, ["word", "dropout", "”", "model"], "Method"], [5794, 5797, ["Virtual", "Adversarial", "Training"], "Task"], [5809, 5810, ["CCG"], "Task"], [5813, 5815, ["Dependency", "Parsing"], "Task"], [5858, 5859, ["VAT"], "Method"], [5860, 5862, ["machine", "translation"], "Task"], [5876, 5877, ["ELMo"], "Method"], [5880, 5881, ["dropout"], "Method"], [5906, 5908, ["prediction", "module"], "Method"], [5931, 5933, ["prediction", "modules"], "Method"], [5947, 5952, ["shared", "Bi", "-", "LSTM", "encoder"], "Method"], [5955, 5956, ["CVT"], "Method"], [5957, 5959, ["Image", "Recognition"], "Task"], [5967, 5968, ["NLP"], "Task"], [5972, 5973, ["CVT"], "Method"], [5974, 5976, ["image", "recognition"], "Task"], [5988, 5994, ["semi", "-", "supervised", "image", "recognition", "approaches"], "Method"], [6017, 6020, ["consistency", "regularization", "methods"], "Method"], [6041, 6042, ["GANs"], "Method"], [6076, 6079, ["image", "recognition", "models"], "Method"], [6082, 6085, ["Convolutional", "Neural", "Networks"], "Method"], [6119, 6121, ["shallower", "CNNs"], "Method"], [6154, 6156, ["deeper", "CNNs"], "Method"], [6186, 6187, ["CNNs"], "Method"], [6199, 6201, ["prediction", "layers"], "Method"], [6203, 6204, ["CNNs"], "Method"], [6228, 6230, ["softmax", "layer"], "Method"], [6233, 6236, ["auxiliary", "prediction", "layers"], "Method"], [6241, 6242, ["CNN"], "Method"], [6276, 6279, ["semi", "-", "supervised"], "Task"], [6309, 6312, ["convolutional", "neural", "network"], "Method"], [6317, 6319, ["TensorFlow", "implementation"], "Method"], [6324, 6326, ["convolutional", "layers"], "Method"], [6328, 6331, ["max", "pooling", "layers"], "Method"], [6361, 6362, ["CNN"], "Method"], [6364, 6366, ["auxiliary", "layer"], "Method"], [6398, 6399, ["CVT"], "Method"], [6401, 6403, ["consistency", "regularization"], "Method"], [6467, 6468, ["VAT"], "Method"], [6469, 6475, ["semi", "-", "supervised", "CIFAR", "-", "10"], "Task"], [6479, 6481, ["NLP", "tasks"], "Task"], [6488, 6489, ["CVT"], "Method"], [6493, 6494, ["CVT"], "Method"], [6496, 6497, ["CVT"], "Method"], [6498, 6500, ["VAT", "models"], "Method"], [6518, 6519, ["CVT"], "Method"], [6523, 6525, ["data", "augmentation"], "Task"], [6549, 6550, ["CVT"], "Method"], [6625, 6629, ["self", "-", "training", "algorithms"], "Method"], [6631, 6633, ["student", "model"], "Method"], [6670, 6672, ["knowledge", "distillation"], "Task"], [6686, 6688, ["Confidence", "thresholding"], "Method"], [6690, 6693, ["self", "-", "training"], "Method"], [6776, 6779, ["Mean", "Teacher", "method"], "Method"], [6782, 6785, ["exponential", "moving", "average"], "Method"], [6786, 6787, ["EMA"], "Method"], [6864, 6866, ["supervised", "classifiers"], "Method"], [6909, 6910, ["CVT"], "Method"], [6914, 6919, ["semi", "-", "supervised", "learning", "mechanism"], "Method"], [6923, 6926, ["prediction", "layers", "regularizing"], "Method"], [186, 189, ["multi", "-", "task"], "Method"], [860, 862, ["dependency", "parsing"], "Task"], [913, 916, ["multi", "-", "task"], "Method"], [957, 960, ["multi", "-", "task"], "Method"], [960, 961, ["ELMo"], "Method"], [992, 995, ["multi", "-", "task"], "Method"], [1367, 1368, ["CVT"], "Method"], [1489, 1492, ["Multi", "-", "Task"], "Method"], [1499, 1502, ["multi", "-", "task"], "Method"], [1605, 1608, ["multi", "-", "task"], "Method"], [1626, 1629, ["multi", "-", "task"], "Method"], [1629, 1630, ["CVT"], "Method"], [1738, 1742, ["Cross", "-", "View", "Training"], "Method"], [3183, 3186, ["multi", "-", "task"], "Method"], [3199, 3200, ["ELMo"], "Method"], [3213, 3214, ["ELMo"], "Method"], [3453, 3454, ["CVT"], "Method"], [3481, 3484, ["Multi", "-", "task"], "Method"], [3552, 3555, ["multi", "-", "task"], "Method"], [3570, 3571, ["ELMo"], "Method"], [3580, 3581, ["ELMo"], "Method"], [3632, 3635, ["multi", "-", "task"], "Method"], [3635, 3636, ["CVT"], "Method"], [3675, 3678, ["multi", "-", "task"], "Method"], [3726, 3729, ["multi", "-", "task"], "Method"], [3729, 3730, ["CVT"], "Method"], [3789, 3790, ["dev"], "Metric"], [3821, 3824, ["multi", "-", "task"], "Method"], [3839, 3840, ["dev"], "Metric"], [3853, 3854, ["dev"], "Metric"], [3940, 3943, ["multi", "-", "task"], "Method"], [4184, 4185, ["CVT"], "Method"], [4350, 4353, ["multi", "-", "task"], "Method"], [4359, 4360, ["ELMo"], "Method"], [4974, 4977, ["Multi", "-", "Task"], "Method"], [4986, 4989, ["multi", "-", "task"], "Method"], [5161, 5164, ["multi", "-", "task"], "Method"], [5557, 5560, ["Multi", "-", "Task"], "Method"], [5598, 5600, ["Penn", "Treebank"], "Material"], [5609, 5610, ["dev"], "Metric"], [5667, 5668, ["EMA"], "Method"], [5756, 5757, ["CVT"], "Method"], [5789, 5790, ["dev"], "Metric"], [5829, 5830, ["dev"], "Metric"], [5883, 5884, ["ELMo"], "Method"], [5899, 5902, ["multi", "-", "task"], "Method"], [5902, 5903, ["ELMo"], "Method"], [5923, 5924, ["ELMo"], "Method"], [5942, 5943, ["ELMo"], "Method"], [6838, 6839, ["EMA"], "Method"], [6851, 6852, ["CVT"], "Method"]]} {"doc_id": "0dc9eb7d17f2def56ad930945f2521653f04c3fa", "text": ["Skip", "-", "gram", "Language", "Modeling", "Using", "Sparse", "Non", "-", "negative", "Matrix", "Probability", "Estimation", "section", ":", "Abstract", "We", "present", "a", "novel", "family", "of", "language", "model", "(", "LM", ")", "estimation", "techniques", "named", "Sparse", "Non", "-", "negative", "Matrix", "(", "SNM", ")", "estimation", ".", "A", "first", "set", "of", "experiments", "empirically", "evaluating", "it", "on", "the", "One", "Billion", "Word", "Benchmark", "[", "reference", "]", "shows", "that", "SNM", "n", "-", "gram", "LMs", "perform", "almost", "as", "well", "as", "the", "well", "-", "established", "Kneser", "-", "Ney", "(", "KN", ")", "models", ".", "When", "using", "skip", "-", "gram", "features", "the", "models", "are", "able", "to", "match", "the", "state", "-", "of", "-", "the", "-", "art", "recurrent", "neural", "network", "(", "RNN", ")", "LMs", ";", "combining", "the", "two", "modeling", "techniques", "yields", "the", "best", "known", "result", "on", "the", "benchmark", ".", "The", "computational", "advantages", "of", "SNM", "over", "both", "maximum", "entropy", "and", "RNN", "LM", "estimation", "are", "probably", "its", "main", "strength", ",", "promising", "an", "approach", "that", "has", "the", "same", "flexibility", "in", "combining", "arbitrary", "features", "effectively", "and", "yet", "should", "scale", "to", "very", "large", "amounts", "of", "data", "as", "gracefully", "as", "n", "-", "gram", "LMs", "do", ".", "section", ":", "Introduction", "A", "statistical", "language", "model", "estimates", "probability", "values", "P", "(", "W", ")", "for", "strings", "of", "words", "W", "in", "a", "vocabulary", "V", "whose", "size", "is", "in", "the", "tens", ",", "hundreds", "of", "thousands", "and", "sometimes", "even", "millions", ".", "Typically", "the", "string", "W", "is", "broken", "into", "sentences", ",", "or", "other", "segments", "such", "as", "utterances", "in", "automatic", "speech", "recognition", ",", "which", "are", "often", "assumed", "to", "be", "conditionally", "independent", ";", "we", "will", "assume", "that", "W", "is", "such", "a", "segment", ",", "or", "sentence", ".", "Estimating", "full", "sentence", "language", "models", "is", "computationally", "hard", "if", "one", "seeks", "a", "properly", "normalized", "probability", "model", "1", "over", "strings", "of", "words", "of", "finite", "length", "in", "[", "reference", "]", "We", "note", "that", "in", "some", "practical", "systems", "the", "constraint", "on", "using", "a", "properly", "normalized", "language", "V", "*", ".", "A", "simple", "and", "sufficient", "way", "to", "ensure", "proper", "normalization", "of", "the", "model", "is", "to", "decompose", "the", "sentence", "probability", "according", "to", "the", "chain", "rule", "and", "make", "sure", "that", "the", "end", "-", "of", "-", "sentence", "symbol", "<", "/", "s", ">", "is", "predicted", "with", "non", "-", "zero", "probability", "in", "any", "context", ".", "With", "W", "=", "w", "1", ",", "w", "2", ",", ".", ".", ".", ",", "w", "n", "we", "get", ":", "P", "(", "w", "i", "|w", "1", ",", "w", "2", ",", ".", ".", ".", ",", "w", "i−1", ")", "Since", "the", "parameter", "space", "of", "P", "(", "w", "k", "|w", "1", ",", "w", "2", ",", ".", ".", ".", ",", "w", "k−1", ")", "is", "too", "large", ",", "the", "language", "model", "is", "forced", "to", "put", "the", "context", "W", "k−1", "=", "w", "1", ",", "w", "2", ",", ".", ".", ".", ",", "w", "k−1", "into", "an", "equivalence", "class", "determined", "by", "a", "function", "Φ", "(", "W", "k−1", ")", ".", "As", "a", "result", ",", "Research", "in", "language", "modeling", "consists", "of", "finding", "appropriate", "equivalence", "classifiers", "Φ", "and", "methods", "to", "estimate", "P", "(", "w", "k", "|Φ", "(", "W", "k−1", ")", ")", ".", "The", "most", "successful", "paradigm", "in", "language", "modeling", "uses", "the", "(", "n", "−", "1", ")-", "gram", "equivalence", "classification", ",", "that", "is", ",", "defines", "Φ", "(", "W", "k−1", ")", ".", "=", "w", "k−n", "+", "1", ",", "w", "k−n", "+", "2", ",", ".", ".", ".", ",", "w", "k−1", "Once", "the", "form", "Φ", "(", "W", "k−1", ")", "is", "specified", ",", "only", "the", "problem", "of", "estimating", "P", "(", "w", "k", "|Φ", "(", "W", "k−1", ")", ")", "from", "training", "data", "remains", ".", "section", ":", "Perplexity", "as", "a", "Measure", "of", "Language", "Model", "Quality", "A", "statistical", "language", "model", "can", "be", "evaluated", "by", "how", "well", "it", "predicts", "a", "string", "of", "symbols", "W", "t", "-", "commonly", "referred", "to", "as", "test", "data", "-", "generated", "by", "the", "source", "to", "be", "modeled", ".", "A", "commonly", "used", "quality", "measure", "for", "a", "given", "model", "M", "is", "related", "to", "the", "entropy", "of", "the", "underlying", "source", "and", "was", "introduced", "under", "the", "name", "of", "perplexity", "(", "PPL", ")", ":", "For", "an", "excellent", "discussion", "on", "the", "use", "of", "perplexity", "in", "statistical", "language", "modeling", ",", "as", "well", "as", "various", "estimates", "for", "the", "entropy", "of", "English", "the", "reader", "is", "referred", "to", "[", "reference", "]", ",", "Section", "8.4", ",", "pages", "141", "-", "142", "and", "the", "additional", "reading", "suggested", "in", "Section", "8.5", "of", "the", "same", "book", ".", "model", "is", "side", "-", "stepped", "at", "a", "gain", "in", "modeling", "power", "and", "simplicity", ".", "Very", "likely", ",", "not", "all", "words", "in", "the", "test", "string", "W", "t", "are", "part", "of", "the", "language", "model", "vocabulary", ".", "It", "is", "common", "practice", "to", "map", "all", "words", "that", "are", "out", "-", "of", "-", "vocabulary", "to", "a", "distinguished", "unknown", "word", "symbol", ",", "and", "report", "the", "out", "-", "of", "-", "vocabulary", "(", "OOV", ")", "rate", "on", "test", "data", "-", "the", "rate", "at", "which", "one", "encounters", "OOV", "words", "in", "the", "test", "string", "W", "tas", "yet", "another", "language", "model", "performance", "metric", "besides", "perplexity", ".", "Usually", "the", "unknown", "word", "is", "assumed", "to", "be", "part", "of", "the", "language", "model", "vocabulary", "-", "open", "vocabulary", "language", "models", "-", "and", "its", "occurrences", "are", "counted", "in", "the", "language", "model", "perplexity", "calculation", ",", "Eq", ".", "(", "3", ")", ".", "A", "situation", "less", "common", "in", "practice", "is", "that", "of", "closed", "vocabulary", "language", "models", "where", "all", "words", "in", "the", "test", "data", "will", "always", "be", "part", "of", "the", "vocabulary", "V.", "section", ":", "Skip", "-", "gram", "Language", "Modeling", "Recently", ",", "neural", "network", "(", "NN", ")", "smoothing", "[", "reference", "]", ",", "[", "reference", "]", ",", "[", "reference", "]", ",", "and", "in", "particular", "recurrent", "neural", "networks", "[", "reference", "]", "(", "RNN", ")", "have", "shown", "excellent", "performance", "in", "language", "modeling", "[", "reference", "]", ".", "Their", "excellent", "performance", "is", "attributed", "to", "a", "combination", "of", "leveraging", "long", "-", "distance", "context", ",", "and", "training", "a", "vector", "representation", "for", "words", ".", "Another", "simple", "way", "of", "leveraging", "long", "distance", "context", "is", "to", "use", "skip", "-", "grams", ".", "In", "our", "approach", ",", "a", "skip", "-", "gram", "feature", "extracted", "from", "the", "context", "W", "k−1", "is", "characterized", "by", "the", "tuple", "(", "r", ",", "s", ",", "a", ")", "where", ":", "•", "r", "denotes", "number", "of", "remote", "context", "words", "•", "s", "denotes", "the", "number", "of", "skipped", "words", "•", "a", "denotes", "the", "number", "of", "adjacent", "context", "words", "relative", "to", "the", "target", "word", "w", "k", "being", "predicted", ".", "For", "example", ",", "in", "the", "sentence", ",", "<", "S", ">", "The", "quick", "brown", "fox", "jumps", "over", "the", "lazy", "dog", "<", "/", "S", ">", "a", "(", "1", ",", "2", ",", "3", ")", "skip", "-", "gram", "feature", "for", "the", "target", "word", "dog", "is", ":", "[", "brown", "skip", "-", "2", "over", "the", "lazy", "]", "For", "performance", "reasons", ",", "it", "is", "recommended", "to", "limit", "s", "and", "to", "limit", "either", "(", "r", "+", "a", ")", "or", "limit", "both", "r", "and", "s", ";", "not", "setting", "any", "limits", "will", "result", "in", "events", "containing", "a", "set", "of", "skip", "-", "gram", "features", "whose", "total", "representation", "size", "is", "quintic", "in", "the", "length", "of", "the", "sentence", ".", "We", "configure", "the", "skip", "-", "gram", "feature", "extractor", "to", "produce", "all", "features", "f", ",", "defined", "by", "the", "equivalence", "class", "Φ", "(", "W", "k−1", ")", ",", "that", "meet", "constraints", "on", "the", "minimum", "and", "maximum", "values", "for", ":", "•", "the", "number", "of", "context", "words", "used", "r", "+", "a", ";", "•", "the", "number", "of", "remote", "words", "r", ";", "•", "the", "number", "of", "adjacent", "words", "a", ";", "•", "the", "skip", "length", "s.", "We", "also", "allow", "the", "option", "of", "not", "including", "the", "exact", "value", "of", "s", "in", "the", "feature", "representation", ";", "this", "may", "help", "with", "smoothing", "by", "sharing", "counts", "for", "various", "skip", "features", ".", "Tied", "skip", "-", "gram", "features", "will", "look", "like", ":", "[", "curiousity", "skip", "-", "*", "the", "cat", "]", "In", "order", "to", "build", "a", "good", "probability", "estimate", "for", "the", "target", "word", "w", "k", "in", "a", "context", "W", "k−1", "we", "need", "a", "way", "of", "combining", "an", "arbitrary", "number", "of", "skip", "-", "gram", "features", "f", "k−1", ",", "which", "do", "not", "fall", "into", "a", "simple", "hierarchy", "like", "regular", "n", "-", "gram", "features", ".", "The", "following", "section", "describes", "a", "simple", ",", "yet", "novel", "approach", "for", "combining", "such", "predictors", "in", "a", "way", "that", "is", "computationally", "easy", ",", "scales", "up", "gracefully", "to", "large", "amounts", "of", "data", "and", "as", "it", "turns", "out", "is", "also", "very", "effective", "from", "a", "modeling", "point", "of", "view", ".", "section", ":", "Sparse", "Non", "-", "negative", "Matrix", "Modeling", "section", ":", "Model", "definition", "In", "the", "Sparse", "Non", "-", "negative", "Matrix", "(", "SNM", ")", "paradigm", ",", "we", "represent", "the", "training", "data", "as", "a", "sequence", "of", "events", "E", "=", "e", "1", ",", "e", "2", ",", "...", "where", "each", "event", "e", "∈", "E", "consists", "of", "a", "sparse", "non", "-", "negative", "feature", "vector", "f", "and", "a", "sparse", "non", "-", "negative", "target", "word", "vector", "t.", "Both", "vectors", "are", "binary", "-", "valued", ",", "indicating", "the", "presence", "or", "absence", "of", "a", "feature", "or", "target", "words", ",", "respectively", ".", "Hence", ",", "the", "training", "data", "consists", "of", "|E||P", "os", "(", "f", ")", "|", "positive", "and", "|E||P", "os", "(", "f", ")", "|", "(", "|V|", "−", "1", ")", "negative", "training", "examples", ",", "where", "P", "os", "(", "f", ")", "denotes", "the", "number", "of", "positive", "elements", "in", "the", "vector", "f", ".", "A", "language", "model", "is", "represented", "by", "a", "non", "-", "negative", "matrix", "M", "that", ",", "when", "applied", "to", "a", "given", "feature", "vector", "f", ",", "produces", "a", "dense", "prediction", "vector", "y", ":", "Upon", "evaluation", ",", "we", "normalize", "y", "such", "that", "we", "end", "up", "with", "a", "conditional", "probability", "distribution", "P", "M", "(", "t|f", ")", "for", "a", "model", "M.", "For", "each", "word", "w", "∈", "V", "that", "corresponds", "to", "index", "j", "in", "t", ",", "and", "its", "feature", "vector", "f", "that", "is", "defined", "by", "the", "equivalence", "class", "Φ", "applied", "to", "the", "history", "h", "(", "w", ")", "of", "that", "word", "in", "a", "text", ",", "the", "conditional", "probability", "P", "M", "(", "w|Φ", "(", "h", "(", "w", ")", ")", ")", "then", "becomes", ":", "For", "convenience", ",", "we", "will", "write", "P", "(", "t", "j", "|f", ")", "instead", "of", "P", "M", "(", "t", "j", "|f", ")", "in", "the", "rest", "of", "the", "paper", ".", "As", "required", "by", "the", "denominator", "in", "Eq", ".", "(", "5", ")", ",", "this", "computation", "involves", "summing", "over", "all", "of", "the", "present", "features", "for", "the", "entire", "vocabulary", ".", "However", ",", "if", "we", "precompute", "the", "row", "sums", "|V|", "u=1", "M", "iu", "and", "store", "them", "together", "with", "the", "model", ",", "the", "evaluation", "can", "be", "done", "very", "efficiently", "in", "only", "|P", "os", "(", "f", ")", "|", "time", ".", "Moreover", ",", "only", "the", "positive", "entries", "in", "M", "i", "need", "to", "be", "considered", ",", "making", "the", "range", "of", "the", "sum", "sparse", ".", "section", ":", "Adjustment", "function", "and", "metafeatures", "We", "let", "the", "entries", "of", "M", "be", "a", "slightly", "modified", "version", "of", "the", "relative", "frequencies", ":", "where", "C", "is", "a", "feature", "-", "target", "count", "matrix", ",", "computed", "over", "the", "entire", "training", "corpus", "and", "A", "(", "i", ",", "j", ")", "is", "a", "real", "-", "valued", "function", ",", "dubbed", "adjustment", "function", ".", "For", "each", "featuretarget", "pair", "(", "f", "i", ",", "t", "j", ")", ",", "the", "adjustment", "function", "extracts", "k", "new", "features", "α", "k", ",", "called", "metafeatures", ",", "which", "are", "hashed", "as", "keys", "to", "store", "corresponding", "weights", "θ", "(", "hash", "(", "α", "k", ")", ")", "in", "a", "huge", "hash", "table", ".", "To", "limit", "memory", "usage", ",", "we", "use", "a", "flat", "hash", "table", "and", "allow", "collisions", ",", "although", "this", "has", "the", "potentially", "undesirable", "effect", "of", "tying", "together", "the", "weights", "of", "different", "metafeatures", ".", "Computing", "the", "adjustment", "function", "for", "any", "(", "f", "i", ",", "t", "j", ")", "then", "amounts", "to", "summing", "the", "weights", "that", "correspond", "to", "its", "metafeatures", ":", "From", "the", "given", "input", "features", ",", "such", "as", "regular", "n", "-", "grams", "and", "skip", "n", "-", "grams", ",", "we", "construct", "our", "metafeatures", "as", "conjunctions", "of", "any", "or", "all", "of", "the", "following", "elementary", "metafeatures", ":", "•", "feature", "identity", ",", "e.g.", "[", "brown", "skip", "-", "2", "over", "the", "lazy", "]", "•", "feature", "type", ",", "e.g.", "(", "1", ",", "2", ",", "3", ")", "skip", "-", "grams", "•", "feature", "count", "C", "i", "*", "•", "target", "identity", ",", "e.g.", "dog", "where", "we", "reused", "the", "example", "from", "Section", "2", ".", "Note", "that", "the", "seemingly", "absent", "feature", "-", "target", "identity", "is", "represented", "by", "the", "conjunction", "of", "the", "feature", "identity", "and", "the", "target", "identity", ".", "Since", "the", "metafeatures", "may", "involve", "the", "feature", "count", "and", "feature", "-", "target", "count", ",", "in", "the", "rest", "of", "the", "paper", "we", "will", "write", "α", "k", "(", "i", ",", "j", ",", "C", "i", "*", ",", "C", "ij", ")", ".", "This", "will", "become", "important", "later", "when", "we", "discuss", "leave", "-", "one", "-", "out", "training", ".", "Each", "elementary", "metafeature", "is", "joined", "with", "the", "others", "to", "form", "more", "complex", "metafeatures", "which", "in", "turn", "are", "joined", "with", "all", "the", "other", "elementary", "and", "complex", "metafeatures", ",", "ultimately", "ending", "up", "with", "all", "2", "5", "−", "1", "possible", "combinations", "of", "metafeatures", ".", "Before", "they", "are", "joined", ",", "count", "metafeatures", "are", "bucketed", "together", "according", "to", "their", "(", "floored", ")", "log", "2", "value", ".", "As", "this", "effectively", "puts", "the", "lowest", "count", "values", ",", "of", "which", "there", "are", "many", ",", "into", "a", "different", "bucket", ",", "we", "optionally", "introduce", "a", "second", "(", "ceiled", ")", "bucket", "to", "assure", "smoother", "transitions", ".", "Both", "buckets", "are", "then", "weighted", "according", "to", "the", "log", "2", "fraction", "lost", "by", "the", "corresponding", "rounding", "operation", ".", "Note", "that", "if", "we", "apply", "double", "bucketing", "to", "both", "the", "feature", "and", "feature", "-", "target", "count", ",", "the", "amount", "of", "metafeatures", "per", "input", "feature", "becomes", "2", "7", "−", "1", ".", "We", "will", "come", "back", "to", "these", "metafeatures", "in", "Section", "4.4", "where", "we", "examine", "their", "individual", "effect", "on", "the", "model", ".", "section", ":", "Loss", "function", "Estimating", "a", "model", "M", "corresponds", "to", "finding", "optimal", "weights", "θ", "k", "for", "all", "the", "metafeatures", "for", "all", "events", "in", "such", "a", "way", "that", "the", "average", "loss", "over", "all", "events", "between", "the", "target", "vector", "t", "and", "the", "prediction", "vector", "y", "is", "minimized", ",", "according", "to", "some", "loss", "function", "L.", "The", "most", "natural", "choice", "of", "loss", "function", "is", "one", "that", "is", "based", "on", "the", "multinomial", "distribution", ".", "That", "is", ",", "we", "consider", "t", "to", "be", "multinomially", "distributed", "with", "|V|", "possible", "outcomes", ".", "The", "loss", "function", "L", "multi", "then", "is", ":", "Another", "possibility", "is", "the", "loss", "function", "based", "on", "the", "Poisson", "distribution", "2", ":", "we", "consider", "each", "t", "j", "in", "t", "to", "be", "Poisson", "distributed", "with", "parameter", "y", "j", ".", "The", "conditional", "probability", "of", "P", "P", "oisson", "(", "t|f", ")", "then", "is", ":", "and", "the", "corresponding", "Poisson", "loss", "function", "is", ":", "where", "we", "dropped", "the", "last", "term", ",", "since", "t", "j", "is", "binary", "-", "valued", "3", ".", "Although", "this", "choice", "is", "not", "obvious", "in", "the", "context", "of", "language", "modeling", ",", "it", "is", "well", "suited", "to", "gradient", "-", "based", "optimization", "and", ",", "as", "we", "will", "see", ",", "the", "experimental", "results", "are", "in", "fact", "excellent", ".", "section", ":", "Model", "Estimation", "The", "adjustment", "function", "is", "learned", "by", "applying", "stochastic", "gradient", "descent", "on", "the", "loss", "function", ".", "That", "is", ",", "for", "each", "feature", "-", "target", "pair", "(", "f", "i", ",", "t", "j", ")", "in", "each", "event", "we", "need", "to", "update", "the", "parameters", "of", "the", "metafeatures", "by", "calculating", "the", "gradient", "with", "respect", "to", "the", "adjustment", "function", ".", "For", "the", "multinomial", "loss", ",", "this", "gradient", "is", ":", "The", "problem", "with", "this", "update", "rule", "is", "that", "we", "need", "to", "sum", "over", "the", "entire", "vocabulary", "V", "in", "the", "denominator", ".", "For", "most", "features", "f", "i", ",", "this", "is", "not", "a", "big", "deal", "as", "C", "iu", "=", "0", ",", "but", "some", "features", "occur", "with", "many", "if", "not", "all", "targets", "e.g.", "the", "empty", "feature", "for", "unigrams", ".", "Although", "we", "might", "be", "able", "to", "get", "away", "with", "this", "by", "re", "-", "using", "these", "sums", "and", "applying", "them", "to", "many", "/", "all", "events", "in", "a", "mini", "batch", ",", "we", "chose", "to", "work", "with", "the", "Poisson", "loss", "in", "our", "first", "implementation", ".", "If", "we", "calculate", "the", "gradient", "of", "the", "Poisson", "loss", ",", "we", "get", "the", "following", ":", "If", "we", "were", "to", "apply", "this", "gradient", "to", "each", "(", "positive", "and", "negative", ")", "training", "example", ",", "it", "would", "be", "computationally", "too", "expensive", ",", "because", "even", "though", "the", "second", "term", "is", "zero", "for", "all", "the", "negative", "training", "examples", ",", "the", "first", "term", "needs", "to", "be", "computed", "for", "all", "|E||P", "os", "(", "f", ")", "||V|", "training", "examples", ".", "However", ",", "since", "the", "first", "term", "does", "not", "depend", "on", "y", "j", ",", "we", "are", "able", "to", "distribute", "the", "updates", "for", "the", "negative", "examples", "over", "the", "positive", "ones", "by", "adding", "in", "gradients", "for", "a", "fraction", "of", "the", "events", "where", "f", "i", "=", "1", ",", "but", "t", "j", "=", "0", ".", "In", "particular", ",", "instead", "of", "adding", "the", "term", "f", "i", "M", "ij", ",", "we", "add", "f", "i", "t", "j", "which", "lets", "us", "update", "the", "gradient", "only", "on", "positive", "examples", ".", "We", "note", "that", "this", "update", "is", "only", "strictly", "correct", "for", "batch", "training", ",", "and", "not", "for", "online", "training", "since", "M", "ij", "changes", "after", "each", "update", ".", "Nonetheless", ",", "we", "found", "this", "to", "yield", "good", "results", "as", "well", "as", "seriously", "reducing", "the", "computational", "cost", ".", "The", "online", "gradient", "applied", "to", "each", "training", "example", "then", "becomes", ":", "which", "is", "non", "-", "zero", "only", "for", "positive", "training", "examples", ",", "hence", "speeding", "up", "computation", "by", "a", "factor", "of", "|V|.", "These", "aggregated", "gradients", "however", "do", "not", "allow", "us", "to", "use", "additional", "data", "to", "train", "the", "adjustment", "function", ",", "since", "they", "tie", "the", "update", "computation", "to", "the", "relative", "frequencies", ".", "Instead", ",", "we", "have", "to", "resort", "to", "leave", "-", "one", "-", "out", "training", "to", "prevent", "the", "model", "from", "overfitting", "the", "training", "data", ".", "We", "do", "this", "by", "excluding", "the", "event", ",", "generating", "the", "gradients", ",", "from", "the", "counts", "used", "to", "compute", "those", "gradients", ".", "So", ",", "for", "each", "positive", "example", "(", "f", "i", ",", "t", "j", ")", "of", "each", "event", "e", "=", "(", "f", ",", "t", ")", ",", "we", "compute", "the", "gradient", ",", "excluding", "f", "i", "from", "C", "i", "*", "and", "f", "i", "t", "j", "from", "C", "ij", ".", "For", "the", "gradients", "of", "the", "negative", "examples", "on", "the", "other", "hand", "we", "only", "exclude", "f", "i", "from", "C", "i", "*", "and", "we", "leave", "C", "ij", "untouched", ",", "since", "here", "we", "did", "not", "observe", "t", "j", ".", "In", "order", "to", "keep", "the", "aggregate", "computation", "of", "the", "gradients", "for", "the", "negative", "examples", ",", "we", "distribute", "them", "uniformly", "over", "all", "the", "positive", "examples", "with", "the", "same", "feature", ";", "each", "of", "the", "C", "ij", "positive", "examples", "will", "then", "compute", "the", "gradient", "of", "negative", "examples", ".", "To", "summarize", ",", "when", "we", "do", "leave", "-", "one", "-", "out", "training", "we", "apply", "the", "following", "gradient", "update", "rule", "on", "all", "positive", "training", "examples", ":", "where", "y", "′", "j", "is", "the", "product", "of", "leaving", "one", "out", "for", "all", "the", "relevant", "features", "i.e.", "section", ":", "Experiments", "section", ":", "Corpus", ":", "One", "Billion", "Benchmark", "Our", "experimental", "setup", "used", "the", "One", "Billion", "Word", "Benchmark", "corpus", "4", "made", "available", "by", "[", "reference", "]", ".", "For", "completeness", ",", "here", "is", "a", "short", "description", "of", "the", "corpus", ",", "containing", "only", "monolingual", "English", "data", ":", "•", "Total", "number", "of", "training", "tokens", "is", "about", "0.8", "billion", "•", "The", "vocabulary", "provided", "consists", "of", "793471", "words", "including", "sentence", "boundary", "markers", "<", "S", ">", ",", "<", "\\S", ">", ",", "and", "was", "constructed", "by", "discarding", "all", "words", "with", "count", "below", "3", "•", "Words", "outside", "of", "the", "vocabulary", "were", "mapped", "to", "<", "UNK", ">", "token", ",", "also", "part", "of", "the", "vocabulary", "•", "Sentence", "order", "was", "randomized", "•", "The", "test", "data", "consisted", "of", "159658", "words", "(", "without", "counting", "the", "sentence", "beginning", "marker", "<", "S", ">", "which", "is", "never", "predicted", "by", "the", "language", "model", ")", "•", "The", "out", "-", "of", "-", "vocabulary", "(", "OoV", ")", "rate", "on", "the", "test", "set", "was", "0.28", "%", ".", "section", ":", "SNM", "for", "n", "-", "gram", "LMs", "When", "trained", "using", "solely", "n", "-", "gram", "features", ",", "SNM", "comes", "very", "close", "to", "the", "stateof", "-", "the", "-", "art", "Kneser", "-", "Ney", "[", "reference", "]", "(", "KN", ")", "models", ".", "Table", "1", "shows", "that", "Katz", "[", "reference", "]", "performs", "considerably", "worse", "than", "both", "SNM", "and", "KN", "which", "only", "differ", "by", "about", "5", "%", ".", "When", "we", "interpolate", "these", "two", "models", "linearly", ",", "the", "added", "gain", "is", "only", "about", "1", "%", ",", "suggesting", "that", "they", "are", "approximately", "modeling", "the", "same", "things", ".", "The", "difference", "between", "KN", "and", "SNM", "becomes", "smaller", "when", "we", "increase", "the", "size", "of", "the", "context", ",", "going", "from", "5", "%", "for", "5", "-", "grams", "to", "3", "%", "for", "8", "-", "grams", ",", "which", "indicates", "that", "SNM", "is", "better", "suited", "to", "a", "large", "number", "of", "features", ".", "section", ":", "Sparse", "Non", "-", "negative", "Modeling", "for", "Skip", "n", "-", "grams", "When", "we", "incorporate", "skip", "-", "gram", "features", ",", "we", "can", "either", "build", "a", "'", "pure", "'", "skip", "-", "gram", "SNM", "that", "contains", "no", "regular", "n", "-", "gram", "features", ",", "except", "for", "unigrams", ",", "and", "interpolate", "this", "model", "with", "KN", ",", "or", "we", "can", "build", "a", "single", "SNM", "that", "has", "both", "the", "regular", "ngram", "features", "and", "the", "skip", "-", "gram", "features", ".", "We", "compared", "the", "two", "approaches", "by", "choosing", "skip", "-", "gram", "features", "that", "can", "be", "considered", "the", "skip", "-", "equivalent", "of", "5", "-", "grams", "i.e.", "they", "contain", "at", "most", "4", "words", ".", "In", "particular", ",", "we", "used", "skip", "-", "gram", "features", "where", "the", "remote", "span", "is", "limited", "to", "at", "most", "3", "words", "for", "skips", "of", "length", "between", "1", "and", "3", "(", "r", "=", "[", "1", ".", "section", ":", "]", ")", ".", "We", "then", "built", "a", "model", "that", "uses", "both", "these", "features", "and", "regular", "5", "-", "grams", "(", "SNM5", "-", "skip", ")", ",", "as", "well", "as", "one", "that", "only", "uses", "the", "skip", "-", "gram", "features", "(", "SNM5", "-", "skip", "(", "no", "n", "-", "grams", ")", ")", ".", "section", ":", "Model", "Num", ".", "Params", "PPL", "SNM5", "-", "skip", "(", "no", "n", "-", "grams", ")", "61", "B", "69.8", "SNM5", "-", "skip", "62", "B", "54.2", "KN5", "+", "SNM5", "-", "skip", "(", "no", "n", "-", "grams", ")", "56.5", "KN5", "+", "SNM5", "-", "skip", "53.6", "Table", "2", ":", "Number", "of", "parameters", "(", "in", "billions", ")", "and", "perplexity", "results", "for", "SNM5", "-", "skip", "models", "with", "and", "without", "n", "-", "grams", ",", "as", "well", "as", "perplexity", "results", "for", "the", "interpolation", "with", "KN5", ".", "As", "it", "turns", "out", "and", "as", "can", "be", "seen", "from", "Table", "2", ",", "it", "is", "better", "to", "incorporate", "all", "the", "features", "into", "one", "single", "SNM", "model", "than", "to", "interpolate", "with", "a", "KN", "5", "-", "gram", "model", "(", "KN5", ")", ".", "Interpolating", "the", "all", "-", "in", "-", "one", "SNM5", "-", "skip", "with", "KN5", "yields", "almost", "no", "additional", "gain", ".", "The", "best", "SNM", "results", "so", "far", "(", "SNM10", "-", "skip", ")", "were", "achieved", "using", "10", "-", "grams", ",", "together", "with", "untied", "skip", "features", "of", "at", "most", "5", "words", "with", "a", "skip", "of", "exactly", "1", "word", "(", "s", "=", "1", ",", "r", "+", "a", "=", "[", "1", ".", "section", ":", "]", ")", ".", "This", "mixture", "of", "rich", "short", "-", "distance", "and", "shallow", "long", "-", "distance", "features", "enables", "the", "model", "to", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", ",", "as", "can", "be", "seen", "in", "Table", "3", ".", "When", "we", "compare", "the", "perplexity", "of", "this", "model", "with", "the", "state", "-", "of", "-", "the", "art", "RNN", "results", "in", "[", "reference", "]", ",", "the", "difference", "is", "only", "3", "%", ".", "Moreover", ",", "although", "our", "model", "has", "more", "parameters", "than", "the", "RNN", "(", "33", "vs", "20", "billion", ")", ",", "training", "takes", "about", "a", "tenth", "of", "the", "time", "(", "24", "hours", "vs", "240", "hours", ")", ".", "Interestingly", ",", "when", "we", "interpolate", "the", "two", "models", ",", "we", "have", "an", "additional", "gain", "of", "20", "%", ",", "and", "as", "far", "as", "we", "know", ",", "the", "perplexity", "of", "41.3", "is", "already", "the", "best", "ever", "reported", "on", "this", "database", ",", "beating", "the", "previous", "best", "by", "6", "%", "[", "reference", "]", ".", "Finally", ",", "when", "we", "optimize", "interpolation", "weights", "over", "all", "models", "in", "[", "reference", "]", ",", "including", "SNM5", "-", "skip", "and", "SNM10", "-", "skip", ",", "the", "contribution", "of", "the", "other", "models", "as", "well", "as", "the", "perplexity", "reduction", "is", "negligible", ",", "as", "can", "be", "seen", "in", "Table", "3", ",", "which", "also", "summarizes", "the", "perplexity", "results", "for", "each", "of", "the", "individual", "models", ".", "section", ":", "Ablation", "Experiments", "To", "find", "out", "how", "much", ",", "if", "anything", "at", "all", ",", "each", "metafeature", "contributes", "to", "the", "adjustment", "function", ",", "we", "ran", "a", "series", "of", "ablation", "experiments", "in", "which", "we", "ablated", "one", "metafeature", "at", "a", "time", ".", "When", "we", "experimented", "on", "SNM5", ",", "we", "found", ",", "unsurprisingly", ",", "that", "the", "most", "important", "metafeature", "is", "the", "feature", "-", "target", "count", ".", "At", "first", "glance", ",", "it", "does", "not", "seem", "to", "matter", "much", "whether", "the", "counts", "are", "stored", "in", "1", "or", "2", "buckets", ",", "but", "the", "second", "bucket", "really", "starts", "to", "pay", "off", "for", "models", "with", "a", "large", "number", "of", "singleton", "features", "e.g.", "SNM10", "-", "skip", "5", ".", "This", "is", "not", "the", "case", "for", "the", "feature", "counts", ",", "where", "having", "a", "single", "bucket", "is", "always", "better", ",", "although", "in", "general", "the", "feature", "counts", "do", "not", "contribute", "much", ".", "In", "any", "case", ",", "feature", "counts", "are", "definitely", "the", "least", "important", "for", "the", "model", ".", "The", "remaining", "metafeatures", "all", "contribute", "more", "or", "less", "equally", ",", "all", "of", "which", "can", "be", "seen", "in", "Table", "4", ".", "section", ":", "Related", "Work", "SNM", "estimation", "is", "closely", "related", "to", "all", "n", "-", "gram", "LM", "smoothing", "techniques", "that", "rely", "on", "mixing", "relative", "frequencies", "at", "various", "orders", ".", "Unlike", "most", "of", "those", ",", "it", "combines", "the", "predictors", "at", "various", "orders", "without", "relying", "on", "a", "hierarchical", "nesting", "of", "the", "contexts", ",", "setting", "it", "closer", "to", "the", "family", "of", "maximum", "entropy", "(", "ME", ")", "[", "reference", "]", ",", "or", "exponential", "models", ".", "We", "are", "not", "the", "first", "ones", "to", "highlight", "the", "effectiveness", "of", "skip", "n", "-", "grams", "at", "capturing", "dependencies", "across", "longer", "contexts", ",", "similar", "to", "RNN", "LMs", ";", "previous", "such", "results", "were", "reported", "in", "[", "reference", "]", ".", "[", "reference", "]", "attempts", "to", "capture", "long", "range", "dependencies", "in", "language", "where", "the", "skip", "n", "-", "grams", "are", "identified", "using", "a", "left", "-", "to", "-", "right", "syntactic", "parser", ".", "Approaches", "such", "as", "[", "reference", "]", "leverage", "latent", "semantic", "information", ",", "whereas", "[", "reference", "]", "integrates", "both", "syntactic", "and", "topic", "-", "based", "modeling", "in", "a", "unified", "approach", ".", "The", "speed", "-", "ups", "to", "ME", ",", "and", "RNN", "LM", "training", "provided", "by", "hierarchically", "predicting", "words", "at", "the", "output", "layer", "[", "reference", "]", ",", "and", "subsampling", "[", "reference", "]", "still", "require", "updates", "that", "are", "linear", "in", "the", "vocabulary", "size", "times", "the", "number", "of", "words", "in", "the", "training", "data", ",", "whereas", "the", "SNM", "updates", "in", "Eq", ".", "(", "15", ")", "for", "the", "much", "smaller", "adjustment", "function", "eliminate", "the", "dependency", "on", "the", "vocabulary", "size", ".", "Scaling", "up", "RNN", "LM", "training", "is", "described", "in", "[", "reference", "]", "and", "[", "reference", "]", ".", "The", "computational", "advantages", "of", "SNM", "over", "both", "Maximum", "Entropy", "and", "RNN", "LM", "estimation", "are", "probably", "its", "main", "strength", ",", "promising", "an", "approach", "that", "has", "the", "same", "flexibility", "in", "combining", "arbitrary", "features", "effectively", "and", "yet", "should", "scale", "to", "very", "large", "amounts", "of", "data", "as", "gracefully", "as", "n", "-", "gram", "LMs", "do", ".", "section", ":", "section", ":", "Conclusions", "and", "Future", "Work", "We", "have", "presented", "SNM", ",", "a", "new", "family", "of", "LM", "estimation", "techniques", ".", "A", "first", "empirical", "evaluation", "on", "the", "One", "Billion", "Word", "Benchmark", "[", "reference", "]", "shows", "that", "SNM", "n", "-", "gram", "LMs", "perform", "almost", "as", "well", "as", "the", "well", "-", "established", "KN", "models", ".", "When", "using", "skip", "-", "gram", "features", "the", "models", "are", "able", "to", "match", "the", "stat", "-", "of", "-", "the", "-", "art", "RNN", "LMs", ";", "combining", "the", "two", "modeling", "techniques", "yields", "the", "best", "known", "result", "on", "the", "benchmark", ".", "Future", "work", "items", "include", "model", "pruning", ",", "exploring", "richer", "features", "similar", "to", "[", "reference", "]", ",", "as", "well", "as", "richer", "metafeatures", "in", "the", "adjustment", "model", ",", "mixing", "SNM", "models", "trained", "on", "various", "data", "sources", "such", "that", "they", "perform", "best", "on", "a", "given", "development", "set", ",", "and", "estimation", "techniques", "that", "are", "more", "flexible", "in", "this", "respect", ".", "section", ":"], "entity": [[0, 5, ["Skip", "-", "gram", "Language", "Modeling"], "Task"], [6, 13, ["Sparse", "Non", "-", "negative", "Matrix", "Probability", "Estimation"], "Method"], [22, 24, ["language", "model"], "Method"], [25, 29, ["LM", ")", "estimation", "techniques"], "Method"], [30, 35, ["Sparse", "Non", "-", "negative", "Matrix"], "Method"], [36, 37, ["SNM"], "Method"], [50, 54, ["One", "Billion", "Word", "Benchmark"], "Material"], [59, 64, ["SNM", "n", "-", "gram", "LMs"], "Method"], [73, 76, ["Kneser", "-", "Ney"], "Method"], [77, 78, ["KN"], "Method"], [101, 104, ["recurrent", "neural", "network"], "Method"], [105, 106, ["RNN"], "Method"], [112, 114, ["modeling", "techniques"], "Method"], [124, 125, ["computational"], "Metric"], [127, 128, ["SNM"], "Method"], [130, 132, ["maximum", "entropy"], "Method"], [168, 172, ["n", "-", "gram", "LMs"], "Method"], [178, 181, ["statistical", "language", "model"], "Method"], [228, 231, ["automatic", "speech", "recognition"], "Task"], [254, 259, ["Estimating", "full", "sentence", "language", "models"], "Task"], [267, 270, ["normalized", "probability", "model"], "Method"], [321, 323, ["chain", "rule"], "Method"], [411, 413, ["language", "model"], "Method"], [454, 456, ["language", "modeling"], "Task"], [460, 463, ["equivalence", "classifiers", "Φ"], "Method"], [483, 485, ["language", "modeling"], "Task"], [488, 495, ["n", "−", "1", ")-", "gram", "equivalence", "classification"], "Task"], [556, 557, ["Perplexity"], "Metric"], [561, 564, ["Language", "Model", "Quality"], "Metric"], [565, 568, ["statistical", "language", "model"], "Method"], [601, 603, ["quality", "measure"], "Metric"], [624, 625, ["perplexity"], "Metric"], [626, 627, ["PPL"], "Metric"], [637, 638, ["perplexity"], "Metric"], [639, 642, ["statistical", "language", "modeling"], "Task"], [691, 693, ["modeling", "power"], "Metric"], [694, 695, ["simplicity"], "Metric"], [747, 750, ["OOV", ")", "rate"], "Metric"], [770, 774, ["language", "model", "performance", "metric"], "Metric"], [775, 776, ["perplexity"], "Metric"], [788, 791, ["language", "model", "vocabulary"], "Method"], [792, 796, ["open", "vocabulary", "language", "models"], "Method"], [824, 828, ["closed", "vocabulary", "language", "models"], "Method"], [845, 850, ["Skip", "-", "gram", "Language", "Modeling"], "Task"], [852, 854, ["neural", "network"], "Method"], [855, 856, ["NN"], "Method"], [873, 876, ["recurrent", "neural", "networks"], "Method"], [880, 881, ["RNN"], "Method"], [887, 889, ["language", "modeling"], "Task"], [911, 913, ["vector", "representation"], "Method"], [1184, 1186, ["feature", "representation"], "Method"], [1191, 1192, ["smoothing"], "Task"], [1223, 1225, ["probability", "estimate"], "Method"], [1328, 1333, ["Sparse", "Non", "-", "negative", "Matrix"], "Method"], [1334, 1335, ["SNM"], "Method"], [1452, 1454, ["language", "model"], "Method"], [1458, 1463, ["non", "-", "negative", "matrix", "M"], "Method"], [1494, 1499, ["conditional", "probability", "distribution", "P", "M"], "Method"], [1748, 1750, ["adjustment", "function"], "Method"], [1860, 1861, ["metafeatures"], "Method"], [1992, 1998, ["leave", "-", "one", "-", "out", "training"], "Task"], [2109, 2111, ["rounding", "operation"], "Method"], [2164, 2166, ["Loss", "function"], "Method"], [2211, 2213, ["loss", "function"], "Method"], [2258, 2260, ["loss", "function"], "Method"], [2263, 2265, ["Poisson", "distribution"], "Method"], [2299, 2302, ["Poisson", "loss", "function"], "Method"], [2330, 2332, ["language", "modeling"], "Task"], [2338, 2342, ["gradient", "-", "based", "optimization"], "Method"], [2359, 2361, ["Model", "Estimation"], "Task"], [2368, 2371, ["stochastic", "gradient", "descent"], "Method"], [2417, 2419, ["multinomial", "loss"], "Task"], [2428, 2430, ["update", "rule"], "Method"], [2684, 2686, ["batch", "training"], "Task"], [2690, 2692, ["online", "training"], "Task"], [2715, 2717, ["computational", "cost"], "Metric"], [2743, 2744, ["computation"], "Task"], [2764, 2766, ["adjustment", "function"], "Method"], [2771, 2773, ["update", "computation"], "Task"], [2785, 2791, ["leave", "-", "one", "-", "out", "training"], "Method"], [2954, 2960, ["leave", "-", "one", "-", "out", "training"], "Task"], [2964, 2967, ["gradient", "update", "rule"], "Method"], [2997, 3000, ["One", "Billion", "Benchmark"], "Material"], [3125, 3127, ["language", "model"], "Method"], [3130, 3135, ["out", "-", "of", "-", "vocabulary"], "Metric"], [3136, 3139, ["OoV", ")", "rate"], "Metric"], [3149, 3150, ["SNM"], "Method"], [3151, 3155, ["n", "-", "gram", "LMs"], "Method"], [3164, 3165, ["SNM"], "Method"], [3190, 3191, ["Katz"], "Method"], [3199, 3200, ["SNM"], "Method"], [3201, 3202, ["KN"], "Method"], [3240, 3241, ["KN"], "Method"], [3242, 3243, ["SNM"], "Method"], [3273, 3274, ["SNM"], "Method"], [3286, 3291, ["Sparse", "Non", "-", "negative", "Modeling"], "Method"], [3292, 3296, ["Skip", "n", "-", "grams"], "Task"], [3310, 3316, ["pure", "'", "skip", "-", "gram", "SNM"], "Method"], [3334, 3335, ["KN"], "Method"], [3342, 3343, ["SNM"], "Method"], [3443, 3446, ["SNM5", "-", "skip"], "Method"], [3461, 3464, ["SNM5", "-", "skip"], "Method"], [3465, 3469, ["no", "n", "-", "grams"], "Method"], [3477, 3478, ["Params"], "Metric"], [3478, 3479, ["PPL"], "Metric"], [3479, 3482, ["SNM5", "-", "skip"], "Method"], [3491, 3494, ["SNM5", "-", "skip"], "Method"], [3509, 3510, ["KN5"], "Method"], [3511, 3512, ["SNM5"], "Method"], [3513, 3514, ["skip"], "Method"], [3547, 3548, ["interpolation"], "Method"], [3549, 3550, ["KN5"], "Method"], [3575, 3577, ["SNM", "model"], "Method"], [3588, 3589, ["KN5"], "Method"], [3602, 3603, ["KN5"], "Method"], [3611, 3612, ["SNM"], "Method"], [3616, 3619, ["SNM10", "-", "skip"], "Method"], [3623, 3626, ["10", "-", "grams"], "Method"], [3712, 3713, ["RNN"], "Method"], [3736, 3737, ["RNN"], "Method"], [3744, 3745, ["training"], "Task"], [3826, 3829, ["SNM5", "-", "skip"], "Method"], [3830, 3833, ["SNM10", "-", "skip"], "Method"], [3872, 3873, ["Ablation"], "Task"], [3914, 3915, ["SNM5"], "Method"], [4048, 4050, ["SNM", "estimation"], "Method"], [4055, 4061, ["n", "-", "gram", "LM", "smoothing", "techniques"], "Method"], [4100, 4102, ["maximum", "entropy"], "Method"], [4110, 4112, ["exponential", "models"], "Method"], [4171, 4178, ["left", "-", "to", "-", "right", "syntactic", "parser"], "Method"], [4196, 4202, ["syntactic", "and", "topic", "-", "based", "modeling"], "Method"], [4212, 4213, ["ME"], "Task"], [4258, 4260, ["SNM", "updates"], "Method"], [4297, 4298, ["computational"], "Metric"], [4300, 4301, ["SNM"], "Method"], [4341, 4345, ["n", "-", "gram", "LMs"], "Method"], [4358, 4359, ["SNM"], "Method"], [4364, 4367, ["LM", "estimation", "techniques"], "Method"], [4374, 4378, ["One", "Billion", "Word", "Benchmark"], "Material"], [4383, 4388, ["SNM", "n", "-", "gram", "LMs"], "Method"], [4426, 4428, ["modeling", "techniques"], "Method"], [4441, 4443, ["model", "pruning"], "Task"], [4460, 4462, ["adjustment", "model"], "Method"], [4464, 4466, ["SNM", "models"], "Method"], [4483, 4485, ["estimation", "techniques"], "Method"], [83, 84, ["skip"], "Method"], [133, 134, ["RNN"], "Method"], [806, 807, ["perplexity"], "Metric"], [927, 928, ["skip"], "Method"], [936, 937, ["skip"], "Method"], [1026, 1027, ["skip"], "Method"], [1039, 1040, ["skip"], "Method"], [1084, 1085, ["skip"], "Method"], [1104, 1105, ["skip"], "Method"], [1166, 1167, ["skip"], "Method"], [1197, 1198, ["skip"], "Method"], [1201, 1202, ["skip"], "Method"], [1211, 1212, ["skip"], "Method"], [1246, 1247, ["skip"], "Method"], [1316, 1321, ["Sparse", "Non", "-", "negative", "Matrix"], "Method"], [1852, 1853, ["skip"], "Method"], [1880, 1881, ["skip"], "Method"], [1899, 1900, ["skip"], "Method"], [3005, 3009, ["One", "Billion", "Word", "Benchmark"], "Material"], [3175, 3178, ["Kneser", "-", "Ney"], "Method"], [3182, 3183, ["KN"], "Method"], [3299, 3300, ["skip"], "Method"], [3352, 3353, ["skip"], "Method"], [3364, 3365, ["skip"], "Method"], [3373, 3374, ["skip"], "Method"], [3393, 3394, ["skip"], "Method"], [3456, 3457, ["skip"], "Method"], [3497, 3498, ["KN5"], "Method"], [3499, 3500, ["SNM5"], "Method"], [3501, 3502, ["skip"], "Method"], [3526, 3527, ["perplexity"], "Metric"], [3529, 3530, ["SNM5"], "Method"], [3531, 3532, ["skip"], "Method"], [3543, 3544, ["perplexity"], "Metric"], [3582, 3583, ["KN"], "Method"], [3598, 3599, ["SNM5"], "Method"], [3600, 3601, ["skip"], "Method"], [3630, 3631, ["skip"], "Method"], [3639, 3640, ["skip"], "Method"], [3700, 3701, ["perplexity"], "Metric"], [3786, 3787, ["perplexity"], "Metric"], [3844, 3845, ["perplexity"], "Metric"], [3861, 3862, ["perplexity"], "Metric"], [3976, 3977, ["skip"], "Method"], [4103, 4104, ["ME"], "Task"], [4124, 4125, ["skip"], "Method"], [4137, 4138, ["RNN"], "Method"], [4163, 4164, ["skip"], "Method"], [4215, 4216, ["RNN"], "Method"], [4282, 4283, ["RNN"], "Method"], [4306, 4307, ["RNN"], "Method"], [4397, 4398, ["KN"], "Method"], [4402, 4403, ["skip"], "Method"], [4420, 4421, ["RNN"], "Method"]]} {"doc_id": "107010b7f2abe3c0c9df62bcef35eb77f6fc76df", "text": ["Domain", "-", "Adversarial", "Training", "of", "Neural", "Networks", "section", ":", "Abstract", "We", "introduce", "a", "new", "representation", "learning", "approach", "for", "domain", "adaptation", ",", "in", "which", "data", "at", "training", "and", "test", "time", "come", "from", "similar", "but", "different", "distributions", ".", "Our", "approach", "is", "directly", "inspired", "by", "the", "theory", "on", "domain", "adaptation", "suggesting", "that", ",", "for", "effective", "domain", "transfer", "to", "be", "achieved", ",", "predictions", "must", "be", "made", "based", "on", "features", "that", "can", "not", "discriminate", "between", "the", "training", "(", "source", ")", "and", "test", "(", "target", ")", "domains", ".", "The", "approach", "implements", "this", "idea", "in", "the", "context", "of", "neural", "network", "architectures", "that", "are", "trained", "on", "labeled", "data", "from", "the", "source", "domain", "and", "unlabeled", "data", "from", "the", "target", "domain", "(", "no", "labeled", "target", "-", "domain", "data", "is", "necessary", ")", ".", "As", "the", "training", "progresses", ",", "the", "approach", "promotes", "the", "emergence", "of", "features", "that", "are", "(", "i", ")", "discriminative", "for", "the", "main", "learning", "task", "on", "the", "source", "domain", "and", "(", "ii", ")", "indiscriminate", "with", "respect", "to", "the", "shift", "between", "the", "domains", ".", "We", "show", "that", "this", "adaptation", "behaviour", "can", "be", "achieved", "in", "almost", "any", "feed", "-", "forward", "model", "by", "augmenting", "it", "with", "few", "standard", "layers", "and", "a", "new", "gradient", "reversal", "layer", ".", "The", "resulting", "augmented", "architecture", "can", "be", "trained", "using", "standard", "backpropagation", "and", "stochastic", "gradient", "descent", ",", "and", "can", "thus", "be", "implemented", "with", "little", "effort", "using", "any", "of", "the", "deep", "learning", "packages", ".", "We", "demonstrate", "the", "success", "of", "our", "approach", "for", "two", "distinct", "classification", "problems", "(", "document", "sentiment", "analysis", "and", "image", "classification", ")", ",", "where", "state", "-", "of", "-", "the", "-", "art", "domain", "adaptation", "performance", "on", "standard", "benchmarks", "is", "achieved", ".", "We", "also", "validate", "the", "approach", "for", "descriptor", "learning", "task", "in", "the", "context", "of", "person", "re", "-", "identification", "application", ".", "section", ":", "Introduction", "The", "cost", "of", "generating", "labeled", "data", "for", "a", "new", "machine", "learning", "task", "is", "often", "an", "obstacle", "for", "applying", "machine", "learning", "methods", ".", "In", "particular", ",", "this", "is", "a", "limiting", "factor", "for", "the", "further", "progress", "of", "deep", "neural", "network", "architectures", ",", "that", "have", "already", "brought", "impressive", "advances", "to", "the", "state", "-", "of", "-", "the", "-", "art", "across", "a", "wide", "variety", "of", "machine", "-", "learning", "tasks", "and", "applications", ".", "For", "problems", "lacking", "labeled", "data", ",", "it", "may", "be", "still", "possible", "to", "obtain", "training", "sets", "that", "are", "big", "enough", "for", "training", "large", "-", "scale", "deep", "models", ",", "but", "that", "suffer", "from", "the", "shift", "in", "data", "distribution", "from", "the", "actual", "data", "encountered", "at", "\"", "test", "time", "\"", ".", "One", "important", "example", "is", "training", "an", "image", "classifier", "on", "synthetic", "or", "semi", "-", "synthetic", "images", ",", "which", "may", "come", "in", "abundance", "and", "be", "fully", "labeled", ",", "but", "which", "inevitably", "have", "a", "distribution", "that", "is", "different", "from", "real", "images", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "Another", "example", "is", "in", "the", "context", "of", "sentiment", "analysis", "in", "written", "reviews", ",", "where", "one", "might", "have", "labeled", "data", "for", "reviews", "of", "one", "type", "of", "product", "(", "e.g.", ",", "movies", ")", ",", "while", "having", "the", "need", "to", "classify", "reviews", "of", "other", "products", "(", "e.g.", ",", "books", ")", ".", "Learning", "a", "discriminative", "classifier", "or", "other", "predictor", "in", "the", "presence", "of", "a", "shift", "between", "training", "and", "test", "distributions", "is", "known", "as", "domain", "adaptation", "(", "DA", ")", ".", "The", "proposed", "approaches", "build", "mappings", "between", "the", "source", "(", "training", "-", "time", ")", "and", "the", "target", "(", "test", "-", "time", ")", "domains", ",", "so", "that", "the", "classifier", "learned", "for", "the", "source", "domain", "can", "also", "be", "applied", "to", "the", "target", "domain", ",", "when", "composed", "with", "the", "learned", "mapping", "between", "domains", ".", "The", "appeal", "of", "the", "domain", "adaptation", "approaches", "is", "the", "ability", "to", "learn", "a", "mapping", "between", "domains", "in", "the", "situation", "when", "the", "target", "domain", "data", "are", "either", "fully", "unlabeled", "(", "unsupervised", "domain", "annotation", ")", "or", "have", "few", "labeled", "samples", "(", "semi", "-", "supervised", "domain", "adaptation", ")", ".", "Below", ",", "we", "focus", "on", "the", "harder", "unsupervised", "case", ",", "although", "the", "proposed", "approach", "(", "domain", "-", "adversarial", "learning", ")", "can", "be", "generalized", "to", "the", "semi", "-", "supervised", "case", "rather", "straightforwardly", ".", "Unlike", "many", "previous", "papers", "on", "domain", "adaptation", "that", "worked", "with", "fixed", "feature", "representations", ",", "we", "focus", "on", "combining", "domain", "adaptation", "and", "deep", "feature", "learning", "within", "one", "training", "process", ".", "Our", "goal", "is", "to", "embed", "domain", "adaptation", "into", "the", "process", "of", "learning", "representation", ",", "so", "that", "the", "final", "classification", "decisions", "are", "made", "based", "on", "features", "that", "are", "both", "discriminative", "and", "invariant", "to", "the", "change", "of", "domains", ",", "i.e.", ",", "have", "the", "same", "or", "very", "similar", "distributions", "in", "the", "source", "and", "the", "target", "domains", ".", "In", "this", "way", ",", "the", "obtained", "feed", "-", "forward", "network", "can", "be", "applicable", "to", "the", "target", "domain", "without", "being", "hindered", "by", "the", "shift", "between", "the", "two", "domains", ".", "Our", "approach", "is", "motivated", "by", "the", "theory", "on", "domain", "adaptation", "[", "reference", "]", ",", "that", "suggests", "that", "a", "good", "representation", "for", "cross", "-", "domain", "transfer", "is", "one", "for", "which", "an", "algorithm", "can", "not", "learn", "to", "identify", "the", "domain", "of", "origin", "of", "the", "input", "observation", ".", "We", "thus", "focus", "on", "learning", "features", "that", "combine", "(", "i", ")", "discriminativeness", "and", "(", "ii", ")", "domaininvariance", ".", "This", "is", "achieved", "by", "jointly", "optimizing", "the", "underlying", "features", "as", "well", "as", "two", "discriminative", "classifiers", "operating", "on", "these", "features", ":", "(", "i", ")", "the", "label", "predictor", "that", "predicts", "class", "labels", "and", "is", "used", "both", "during", "training", "and", "at", "test", "time", "and", "(", "ii", ")", "the", "domain", "classifier", "that", "discriminates", "between", "the", "source", "and", "the", "target", "domains", "during", "training", ".", "While", "the", "parameters", "of", "the", "classifiers", "are", "optimized", "in", "order", "to", "minimize", "their", "error", "on", "the", "training", "set", ",", "the", "parameters", "of", "the", "underlying", "deep", "feature", "mapping", "are", "optimized", "in", "order", "to", "minimize", "the", "loss", "of", "the", "label", "classifier", "and", "to", "maximize", "the", "loss", "of", "the", "domain", "classifier", ".", "The", "latter", "update", "thus", "works", "adversarially", "to", "the", "domain", "classifier", ",", "and", "it", "encourages", "domain", "-", "invariant", "features", "to", "emerge", "in", "the", "course", "of", "the", "optimization", ".", "Crucially", ",", "we", "show", "that", "all", "three", "training", "processes", "can", "be", "embedded", "into", "an", "appropriately", "composed", "deep", "feed", "-", "forward", "network", ",", "called", "domain", "-", "adversarial", "neural", "network", "(", "DANN", ")", "(", "illustrated", "by", "Figure", "1", ",", "page", "12", ")", "that", "uses", "standard", "layers", "and", "loss", "functions", ",", "and", "can", "be", "trained", "using", "standard", "backpropagation", "algorithms", "based", "on", "stochastic", "gradient", "descent", "or", "its", "modifications", "(", "e.g.", ",", "SGD", "with", "momentum", ")", ".", "The", "approach", "is", "generic", "as", "a", "DANN", "version", "can", "be", "created", "for", "almost", "any", "existing", "feed", "-", "forward", "architecture", "that", "is", "trainable", "by", "backpropagation", ".", "In", "practice", ",", "the", "only", "non", "-", "standard", "component", "of", "the", "proposed", "architecture", "is", "a", "rather", "trivial", "gradient", "reversal", "layer", "that", "leaves", "the", "input", "unchanged", "during", "forward", "propagation", "and", "reverses", "the", "gradient", "by", "multiplying", "it", "by", "a", "negative", "scalar", "during", "the", "backpropagation", ".", "We", "provide", "an", "experimental", "evaluation", "of", "the", "proposed", "domain", "-", "adversarial", "learning", "idea", "over", "a", "range", "of", "deep", "architectures", "and", "applications", ".", "We", "first", "consider", "the", "simplest", "DANN", "architecture", "where", "the", "three", "parts", "(", "label", "predictor", ",", "domain", "classifier", "and", "feature", "extractor", ")", "are", "linear", ",", "and", "demonstrate", "the", "success", "of", "domain", "-", "adversarial", "learning", "for", "such", "architecture", ".", "The", "evaluation", "is", "performed", "for", "synthetic", "data", "as", "well", "as", "for", "the", "sentiment", "analysis", "problem", "in", "natural", "language", "processing", ",", "where", "DANN", "improves", "the", "state", "-", "of", "-", "the", "-", "art", "marginalized", "Stacked", "Autoencoders", "(", "mSDA", ")", "of", "[", "reference", "]", "on", "the", "common", "Amazon", "reviews", "benchmark", ".", "We", "further", "evaluate", "the", "approach", "extensively", "for", "an", "image", "classification", "task", ",", "and", "present", "results", "on", "traditional", "deep", "learning", "image", "data", "sets", "-", "such", "as", "MNIST", "[", "reference", "]", "and", "SVHN", "[", "reference", "])-", "as", "well", "as", "on", "Office", "benchmarks", "[", "reference", "]", ",", "where", "domain", "-", "adversarial", "learning", "allows", "obtaining", "a", "deep", "architecture", "that", "considerably", "improves", "over", "previous", "state", "-", "of", "-", "the", "-", "art", "accuracy", ".", "Finally", ",", "we", "evaluate", "domain", "-", "adversarial", "descriptor", "learning", "in", "the", "context", "of", "person", "re", "-", "identification", "application", "[", "reference", "]", ",", "where", "the", "task", "is", "to", "obtain", "good", "pedestrian", "image", "descriptors", "that", "are", "suitable", "for", "retrieval", "and", "verification", ".", "We", "apply", "domainadversarial", "learning", ",", "as", "we", "consider", "a", "descriptor", "predictor", "trained", "with", "a", "Siamese", "-", "like", "loss", "instead", "of", "the", "label", "predictor", "trained", "with", "a", "classification", "loss", ".", "In", "a", "series", "of", "experiments", ",", "we", "demonstrate", "that", "domain", "-", "adversarial", "learning", "can", "improve", "cross", "-", "data", "-", "set", "re", "-", "identification", "considerably", ".", "section", ":", "Related", "work", "The", "general", "approach", "of", "achieving", "domain", "adaptation", "explored", "under", "many", "facets", ".", "Over", "the", "years", ",", "a", "large", "part", "of", "the", "literature", "has", "focused", "mainly", "on", "linear", "hypothesis", "(", "see", "for", "instance", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "More", "recently", ",", "non", "-", "linear", "representations", "have", "become", "increasingly", "studied", ",", "including", "neural", "network", "representations", "[", "reference", "]", "and", "most", "notably", "the", "state", "-", "of", "-", "the", "-", "art", "mSDA", "[", "reference", "]", ".", "That", "literature", "has", "mostly", "focused", "on", "exploiting", "the", "principle", "of", "robust", "representations", ",", "based", "on", "the", "denoising", "autoencoder", "paradigm", "[", "reference", "]", ".", "Concurrently", ",", "multiple", "methods", "of", "matching", "the", "feature", "distributions", "in", "the", "source", "and", "the", "target", "domains", "have", "been", "proposed", "for", "unsupervised", "domain", "adaptation", ".", "Some", "ap", "-", "proaches", "perform", "this", "by", "reweighing", "or", "selecting", "samples", "from", "the", "source", "domain", "[", "reference", "]", ",", "while", "others", "seek", "an", "explicit", "feature", "space", "transformation", "that", "would", "map", "source", "distribution", "into", "the", "target", "one", "[", "reference", "][", "reference", "][", "reference", "]", ")", ".", "An", "important", "aspect", "of", "the", "distribution", "matching", "approach", "is", "the", "way", "the", "(", "dis", ")", "similarity", "between", "distributions", "is", "measured", ".", "Here", ",", "one", "popular", "choice", "is", "matching", "the", "distribution", "means", "in", "the", "kernelreproducing", "Hilbert", "space", "[", "reference", "]", ",", "whereas", "[", "reference", "]", "and", "[", "reference", "]", "map", "the", "principal", "axes", "associated", "with", "each", "of", "the", "distributions", ".", "Our", "approach", "also", "attempts", "to", "match", "feature", "space", "distributions", ",", "however", "this", "is", "accomplished", "by", "modifying", "the", "feature", "representation", "itself", "rather", "than", "by", "reweighing", "or", "geometric", "transformation", ".", "Also", ",", "our", "method", "uses", "a", "rather", "different", "way", "to", "measure", "the", "disparity", "between", "distributions", "based", "on", "their", "separability", "by", "a", "deep", "discriminatively", "-", "trained", "classifier", ".", "Note", "also", "that", "several", "approaches", "perform", "transition", "from", "the", "source", "to", "the", "target", "domain", "[", "reference", "][", "reference", "]", "by", "changing", "gradually", "the", "training", "distribution", ".", "Among", "these", "methods", ",", "[", "reference", "]", "does", "this", "in", "a", "\"", "deep", "\"", "way", "by", "the", "layerwise", "training", "of", "a", "sequence", "of", "deep", "autoencoders", ",", "while", "gradually", "replacing", "source", "-", "domain", "samples", "with", "target", "-", "domain", "samples", ".", "This", "improves", "over", "a", "similar", "approach", "of", "[", "reference", "]", "that", "simply", "trains", "a", "single", "deep", "autoencoder", "for", "both", "domains", ".", "In", "both", "approaches", ",", "the", "actual", "classifier", "/", "predictor", "is", "learned", "in", "a", "separate", "step", "using", "the", "feature", "representation", "learned", "by", "autoencoder", "(", "s", ")", ".", "In", "contrast", "to", "[", "reference", "][", "reference", "]", ",", "our", "approach", "performs", "feature", "learning", ",", "domain", "adaptation", "and", "classifier", "learning", "jointly", ",", "in", "a", "unified", "architecture", ",", "and", "using", "a", "single", "learning", "algorithm", "(", "backpropagation", ")", ".", "We", "therefore", "argue", "that", "our", "approach", "is", "simpler", "(", "both", "conceptually", "and", "in", "terms", "of", "its", "implementation", ")", ".", "Our", "method", "also", "achieves", "considerably", "better", "results", "on", "the", "popular", "Office", "benchmark", ".", "While", "the", "above", "approaches", "perform", "unsupervised", "domain", "adaptation", ",", "there", "are", "approaches", "that", "perform", "supervised", "domain", "adaptation", "by", "exploiting", "labeled", "data", "from", "the", "target", "domain", ".", "In", "the", "context", "of", "deep", "feed", "-", "forward", "architectures", ",", "such", "data", "can", "be", "used", "to", "\"", "fine", "-", "tune", "\"", "the", "network", "trained", "on", "the", "source", "domain", "[", "reference", "][", "reference", "][", "reference", "]", ".", "Our", "approach", "does", "not", "require", "labeled", "target", "-", "domain", "data", ".", "At", "the", "same", "time", ",", "it", "can", "easily", "incorporate", "such", "data", "when", "they", "are", "available", ".", "An", "idea", "related", "to", "ours", "is", "described", "in", "[", "reference", "]", ".", "While", "their", "goal", "is", "quite", "different", "(", "building", "generative", "deep", "networks", "that", "can", "synthesize", "samples", ")", ",", "the", "way", "they", "measure", "and", "minimize", "the", "discrepancy", "between", "the", "distribution", "of", "the", "training", "data", "and", "the", "distribution", "of", "the", "synthesized", "data", "is", "very", "similar", "to", "the", "way", "our", "architecture", "measures", "and", "minimizes", "the", "discrepancy", "between", "feature", "distributions", "for", "the", "two", "domains", ".", "Moreover", ",", "the", "authors", "mention", "the", "problem", "of", "saturating", "sigmoids", "which", "may", "arise", "at", "the", "early", "stages", "of", "training", "due", "to", "the", "significant", "dissimilarity", "of", "the", "domains", ".", "The", "technique", "they", "use", "to", "circumvent", "this", "issue", "(", "the", "\"", "adversarial", "\"", "part", "of", "the", "gradient", "is", "replaced", "by", "a", "gradient", "computed", "with", "respect", "to", "a", "suitable", "cost", ")", "is", "directly", "applicable", "to", "our", "method", ".", "Also", ",", "recent", "and", "concurrent", "reports", "by", "[", "reference", "]", "focus", "on", "domain", "adaptation", "in", "feed", "-", "forward", "networks", ".", "Their", "set", "of", "techniques", "measures", "and", "minimizes", "the", "distance", "between", "the", "data", "distribution", "means", "across", "domains", "(", "potentially", ",", "after", "embedding", "distributions", "into", "RKHS", ")", ".", "Their", "approach", "is", "thus", "different", "from", "our", "idea", "of", "matching", "distributions", "by", "making", "them", "indistinguishable", "for", "a", "discriminative", "classifier", ".", "Below", ",", "we", "compare", "our", "approach", "to", ";", "[", "reference", "]", "on", "the", "Office", "benchmark", ".", "Another", "approach", "to", "deep", "domain", "adaptation", ",", "which", "is", "arguably", "more", "different", "from", "ours", ",", "has", "been", "developed", "in", "parallel", "by", "[", "reference", "]", ".", "From", "a", "theoretical", "standpoint", ",", "our", "approach", "is", "directly", "derived", "from", "the", "seminal", "theoretical", "works", "of", "[", "reference", "]", ".", "Indeed", ",", "DANN", "directly", "optimizes", "the", "notion", "of", "H", "-", "divergence", ".", "We", "do", "note", "the", "work", "of", "[", "reference", "]", ",", "in", "which", "HMM", "representations", "are", "learned", "for", "word", "tagging", "using", "a", "posterior", "regularizer", "that", "is", "also", "inspired", "by", "Ben", "-", "David", "et", "al", ".", "'s", "work", ".", "In", "addition", "to", "the", "tasks", "being", "different", "-", "Huang", "and", "Yates", "(", "2012", ")", "focus", "on", "word", "tagging", "problems", "-", ",", "we", "would", "argue", "that", "DANN", "learning", "objective", "more", "closely", "optimizes", "the", "H", "-", "divergence", ",", "with", "Huang", "and", "Yates", "(", "2012", ")", "relying", "on", "cruder", "approximations", "for", "efficiency", "reasons", ".", "A", "part", "of", "this", "paper", "has", "been", "published", "as", "a", "conference", "paper", "[", "reference", "]", ".", "This", "version", "extends", "[", "reference", "]", "very", "considerably", "by", "incorporating", "the", "report", "[", "reference", "]", "(", "presented", "as", "part", "of", "the", "Second", "Workshop", "on", "Transfer", "and", "Multi", "-", "Task", "Learning", ")", ",", "which", "brings", "in", "new", "terminology", ",", "in", "-", "depth", "theoretical", "analysis", "and", "justification", "of", "the", "approach", ",", "extensive", "experiments", "with", "the", "shallow", "DANN", "case", "on", "synthetic", "data", "as", "well", "as", "on", "a", "natural", "language", "processing", "task", "(", "sentiment", "analysis", ")", ".", "Furthermore", ",", "in", "this", "version", "we", "go", "beyond", "classification", "and", "evaluate", "domain", "-", "adversarial", "learning", "for", "descriptor", "learning", "setting", "within", "the", "person", "re", "-", "identification", "application", ".", "section", ":", "Domain", "Adaptation", "We", "consider", "classification", "tasks", "where", "X", "is", "the", "input", "space", "and", "Y", "=", "{", "0", ",", "1", ",", ".", ".", ".", ",", "L−1", "}", "is", "the", "set", "of", "L", "possible", "labels", ".", "Moreover", ",", "we", "have", "two", "different", "distributions", "over", "X×Y", ",", "called", "the", "source", "domain", "D", "S", "and", "the", "target", "domain", "D", "T", ".", "An", "unsupervised", "domain", "adaptation", "learning", "algorithm", "is", "then", "provided", "with", "a", "labeled", "source", "sample", "S", "drawn", "i.i.d", ".", "from", "D", "S", ",", "and", "an", "unlabeled", "target", "sample", "with", "N", "=", "n", "+", "n", "being", "the", "total", "number", "of", "samples", ".", "The", "goal", "of", "the", "learning", "algorithm", "is", "to", "build", "a", "classifier", "η", ":", "X", "→", "Y", "with", "a", "low", "target", "risk", "while", "having", "no", "information", "about", "the", "labels", "of", "D", "T", ".", "section", ":", "Domain", "Divergence", "To", "tackle", "the", "challenging", "domain", "adaptation", "task", ",", "many", "approaches", "bound", "the", "target", "error", "by", "the", "sum", "of", "the", "source", "error", "and", "a", "notion", "of", "distance", "between", "the", "source", "and", "the", "target", "distributions", ".", "These", "methods", "are", "intuitively", "justified", "by", "a", "simple", "assumption", ":", "the", "source", "risk", "is", "expected", "to", "be", "a", "good", "indicator", "of", "the", "target", "risk", "when", "both", "distributions", "are", "similar", ".", "Several", "notions", "of", "distance", "have", "been", "proposed", "for", "domain", "adaptation", "[", "reference", "][", "reference", "][", "reference", "]", ".", "In", "this", "paper", ",", "we", "focus", "on", "the", "H", "-", "divergence", "used", "by", "[", "reference", "]", ",", "and", "based", "on", "the", "earlier", "work", "of", "[", "reference", "]", ".", "Note", "that", "we", "assume", "in", "definition", "1", "below", "that", "the", "hypothesis", "class", "H", "is", "a", "(", "discrete", "or", "continuous", ")", "set", "of", "binary", "classifiers", "η", ":", "X", "→", "{", "0", ",", "1}.", "1", "Definition", "1", "[", "reference", "][", "reference", "]", "That", "is", ",", "the", "H", "-", "divergence", "relies", "on", "the", "capacity", "of", "the", "hypothesis", "class", "H", "to", "distinguish", "between", "examples", "generated", "by", "D", "X", "S", "from", "examples", "generated", "by", "D", "X", "T", ".", "[", "reference", "]", "proved", "that", ",", "for", "a", "symmetric", "hypothesis", "class", "H", ",", "one", "can", "compute", "the", "empirical", "where", "I", "[", "a", "]", "is", "the", "indicator", "function", "which", "is", "1", "if", "predicate", "a", "is", "true", ",", "and", "0", "otherwise", ".", "section", ":", "Proxy", "Distance", "Ben", "-", "David", "et", "al", ".", "[", "reference", "]", "suggested", "that", ",", "even", "if", "it", "is", "generally", "hard", "to", "computed", "H", "(", "S", ",", "T", ")", "exactly", "(", "e.g.", ",", "when", "H", "is", "the", "space", "of", "linear", "classifiers", "on", "X", ")", ",", "we", "can", "easily", "approximate", "it", "by", "running", "a", "learning", "algorithm", "on", "the", "problem", "of", "discriminating", "between", "source", "and", "target", "examples", ".", "To", "do", "so", ",", "we", "construct", "a", "new", "data", "set", "where", "the", "examples", "of", "the", "source", "sample", "are", "labeled", "0", "and", "the", "examples", "of", "the", "target", "sample", "are", "labeled", "1", ".", "Then", ",", "the", "risk", "of", "the", "classifier", "trained", "on", "the", "new", "data", "set", "U", "approximates", "the", "\"", "min", "\"", "part", "of", "Equation", "(", "1", ")", ".", "Given", "a", "generalization", "error", "on", "the", "problem", "of", "discriminating", "between", "source", "and", "target", "examples", ",", "the", "H", "-", "divergence", "is", "then", "approximated", "bŷ", "In", "[", "reference", "]", ",", "the", "valued", "A", "is", "called", "the", ",", "where", "A", "is", "a", "subset", "of", "X.", "Note", "that", ",", "by", "choosing", "A", "=", "{", "A", "η", "|η", "∈", "H", "}", ",", "with", "A", "η", "the", "set", "represented", "by", "the", "characteristic", "function", "η", ",", "the", "A", "-", "distance", "and", "the", "H", "-", "divergence", "of", "Definition", "1", "are", "identical", ".", "In", "the", "experiments", "section", "of", "this", "paper", ",", "we", "compute", "the", "PAD", "value", "following", "the", "approach", "of", "[", "reference", "][", "reference", "]", ",", "i.e.", ",", "we", "train", "either", "a", "linear", "SVM", "or", "a", "deeper", "MLP", "classifier", "on", "a", "subset", "of", "U", "(", "Equation", "2", ")", ",", "and", "we", "use", "the", "obtained", "classifier", "error", "on", "the", "other", "subset", "as", "the", "value", "of", "in", "Equation", "(", "3", ")", ".", "More", "details", "and", "illustrations", "of", "the", "linear", "SVM", "case", "are", "provided", "in", "Section", "5.1.5", ".", "section", ":", "Generalization", "Bound", "on", "the", "Target", "Risk", "The", "work", "of", "[", "reference", "]", "is", "upper", "bounded", "by", "its", "empirical", "estimated", "H", "(", "S", ",", "T", ")", "plus", "a", "constant", "complexity", "term", "that", "depends", "on", "the", "VC", "dimension", "of", "H", "and", "the", "size", "of", "samples", "S", "and", "T", ".", "By", "combining", "this", "result", "with", "a", "similar", "bound", "on", "the", "source", "risk", ",", "the", "following", "theorem", "is", "obtained", ".", ",", "and", "is", "the", "empirical", "source", "risk", ".", "The", "previous", "result", "tells", "us", "that", "R", "D", "T", "(", "η", ")", "can", "be", "low", "only", "when", "the", "β", "term", "is", "low", ",", "i.e.", ",", "only", "when", "there", "exists", "a", "classifier", "that", "can", "achieve", "a", "low", "risk", "on", "both", "distributions", ".", "It", "also", "tells", "us", "that", ",", "to", "find", "a", "classifier", "with", "a", "small", "R", "D", "T", "(", "η", ")", "in", "a", "given", "class", "of", "fixed", "VC", "dimension", ",", "the", "learning", "algorithm", "should", "minimize", "(", "in", "that", "class", ")", "a", "trade", "-", "off", "between", "the", "source", "risk", "R", "S", "(", "η", ")", "and", "the", "empirical", "H", "-", "divergenced", "H", "(", "S", ",", "T", ")", ".", "As", "pointed", "-", "out", "by", "[", "reference", "]", ",", "a", "strategy", "to", "control", "the", "H", "-", "divergence", "is", "to", "find", "a", "representation", "of", "the", "examples", "where", "both", "the", "source", "and", "the", "target", "domain", "are", "as", "indistinguishable", "as", "possible", ".", "Under", "such", "a", "representation", ",", "a", "hypothesis", "with", "a", "low", "source", "risk", "will", ",", "according", "to", "Theorem", "2", ",", "perform", "well", "on", "the", "target", "data", ".", "In", "this", "paper", ",", "we", "present", "an", "algorithm", "that", "directly", "exploits", "this", "idea", ".", "section", ":", "Domain", "-", "Adversarial", "Neural", "Networks", "(", "DANN", ")", "An", "original", "aspect", "of", "our", "approach", "is", "to", "explicitly", "implement", "the", "idea", "exhibited", "by", "Theorem", "2", "into", "a", "neural", "network", "classifier", ".", "That", "is", ",", "to", "learn", "a", "model", "that", "can", "generalize", "well", "from", "one", "domain", "to", "another", ",", "we", "ensure", "that", "the", "internal", "representation", "of", "the", "neural", "network", "contains", "no", "discriminative", "information", "about", "the", "origin", "of", "the", "input", "(", "source", "or", "target", ")", ",", "while", "preserving", "a", "low", "risk", "on", "the", "source", "(", "labeled", ")", "examples", ".", "In", "this", "section", ",", "we", "detail", "the", "proposed", "approach", "for", "incorporating", "a", "\"", "domain", "adaptation", "component", "\"", "to", "neural", "networks", ".", "In", "Subsection", "4.1", ",", "we", "start", "by", "developing", "the", "idea", "for", "the", "simplest", "possible", "case", ",", "i.e.", ",", "a", "single", "hidden", "layer", ",", "fully", "connected", "neural", "network", ".", "We", "then", "describe", "how", "to", "generalize", "the", "approach", "to", "arbitrary", "(", "deep", ")", "network", "architectures", ".", "section", ":", "Example", "Case", "with", "a", "Shallow", "Neural", "Network", "Let", "us", "first", "consider", "a", "standard", "neural", "network", "(", "NN", ")", "architecture", "with", "a", "single", "hidden", "layer", ".", "For", "simplicity", ",", "we", "suppose", "that", "the", "input", "space", "is", "formed", "by", "m", "-", "dimensional", "real", "vectors", ".", "Thus", ",", "X", "=", "R", "m", ".", "The", "hidden", "layer", "G", "f", "learns", "a", "function", "G", "f", ":", "X", "→", "R", "D", "that", "maps", "an", "example", "into", "a", "new", "D", "-", "dimensional", "representation", "2", ",", "and", "is", "parameterized", "by", "a", "matrix", "-", "vector", "pair", "(", "W", ",", "b", ")", "∈", "R", "D×m", "×", "R", "D", ":", "with", "sigm", "(", "a", ")", "=", "Similarly", ",", "the", "prediction", "layer", "G", "y", "learns", "a", "function", "Here", "we", "have", "L", "=", "|Y", "|.", "By", "using", "the", "softmax", "function", ",", "each", "component", "of", "vector", "G", "y", "(", "G", "f", "(", "x", ")", ")", "denotes", "the", "conditional", "probability", "that", "the", "neural", "network", "assigns", "x", "to", "the", "class", "in", "Y", "represented", "by", "that", "component", ".", "Given", "a", "source", "example", "(", "x", "i", ",", "y", "i", ")", ",", "the", "natural", "classification", "loss", "to", "use", "is", "the", "negative", "log", "-", "probability", "of", "the", "correct", "label", ":", "Training", "the", "neural", "network", "then", "leads", "to", "the", "following", "optimization", "problem", "on", "the", "source", "domain", ":", "where", ",", "y", "i", "is", "a", "shorthand", "notation", "for", "the", "prediction", "loss", "on", "the", "i", "-", "th", "example", ",", "and", "R", "(", "W", ",", "b", ")", "is", "an", "optional", "regularizer", "that", "is", "weighted", "by", "hyper", "-", "parameter", "λ", ".", "The", "heart", "of", "our", "approach", "is", "to", "design", "a", "domain", "regularizer", "directly", "derived", "from", "the", "H", "-", "divergence", "of", "Definition", "1", ".", "To", "this", "end", ",", "we", "view", "the", "output", "of", "the", "hidden", "layer", "G", "f", "(", "·", ")", "(", "Equation", "4", ")", "as", "the", "internal", "representation", "of", "the", "neural", "network", ".", "Thus", ",", "we", "denote", "the", "source", "sample", "representations", "as", "Similarly", ",", "given", "an", "unlabeled", "sample", "from", "the", "target", "domain", "we", "denote", "the", "corresponding", "representations", "Based", "on", "Equation", "(", "1", ")", ",", "the", "empirical", "H", "-", "divergence", "of", "a", "symmetric", "hypothesis", "class", "H", "between", "samples", "S", "(", "G", "f", ")", "and", "Let", "us", "consider", "H", "as", "the", "class", "of", "hyperplanes", "in", "the", "representation", "space", ".", "Inspired", "by", "the", "Proxy", "A", "-", "distance", "(", "see", "Section", "3.2", ")", ",", "we", "suggest", "estimating", "the", "\"", "min", "\"", "part", "of", "Equation", "(", "6", ")", "by", "a", "domain", "classification", "layer", "G", "d", "that", "learns", "a", "logistic", "regressor", "parameterized", "by", "a", "vector", "-", "scalar", "pair", "(", "u", ",", "z", ")", "∈", "R", "D", "×", "R", ",", "that", "models", "the", "probability", "that", "a", "given", "input", "is", "from", "the", "source", "domain", "D", "Hence", ",", "the", "function", "G", "d", "(", "·", ")", "is", "a", "domain", "regressor", ".", "We", "define", "its", "loss", "by", "where", "d", "i", "denotes", "the", "binary", "variable", "(", "domain", "label", ")", "for", "the", "i", "-", "th", "example", ",", "which", "indicates", "whether", "x", "i", "come", "from", "the", "source", "distribution", "(", "Recall", "that", "for", "the", "examples", "from", "the", "source", "distribution", "(", "d", "i", "=", "0", ")", ",", "the", "corresponding", "labels", "y", "i", "∈", "Y", "are", "known", "at", "training", "time", ".", "For", "the", "examples", "from", "the", "target", "domains", ",", "we", "do", "not", "know", "the", "labels", "at", "training", "time", ",", "and", "we", "want", "to", "predict", "such", "labels", "at", "test", "time", ".", "This", "enables", "us", "to", "add", "a", "domain", "adaptation", "term", "to", "the", "objective", "of", "Equation", "(", "5", ")", ",", "giving", "the", "following", "regularizer", ":", "where", "This", "regularizer", "seeks", "to", "approximate", "the", "H", "-", "divergence", "of", "Equation", "(", "6", ")", ",", "as", "2", "(", "1−R", "(", "W", ",", "b", ")", ")", "is", "a", "surrogate", "ford", "H", "S", "(", "G", "f", ")", ",", "T", "(", "G", "f", ")", ".", "In", "line", "with", "Theorem", "2", ",", "the", "optimization", "problem", "given", "by", "Equations", "(", "5", ")", "and", "(", "8", ")", "implements", "a", "trade", "-", "off", "between", "the", "minimization", "of", "the", "source", "risk", "R", "S", "(", "·", ")", "and", "the", "divergenced", "H", "(", "·", ",", "·", ")", ".", "The", "hyper", "-", "parameter", "λ", "is", "then", "used", "to", "tune", "the", "trade", "-", "off", "between", "these", "two", "quantities", "during", "the", "learning", "process", ".", "For", "learning", ",", "we", "first", "note", "that", "we", "can", "rewrite", "the", "complete", "optimization", "objective", "of", "Equation", "(", "5", ")", "as", "follows", ":", "where", "we", "are", "seeking", "the", "parametersŴ", ",", "V", ",", "b", ",", "ĉ", ",", "û", ",", "ẑ", "that", "deliver", "a", "saddle", "point", "given", "by", "Thus", ",", "the", "optimization", "problem", "involves", "a", "minimization", "with", "respect", "to", "some", "parameters", ",", "as", "well", "as", "a", "maximization", "with", "respect", "to", "the", "others", ".", "Algorithm", "1", "Shallow", "DANN", "-", "Stochastic", "training", "update", "while", "stopping", "criterion", "is", "not", "met", "do", "6", ":", "for", "i", "from", "1", "to", "n", "do", "7", ":", "#", "Forward", "propagation", "8", ":", "32", ":", "#", "Update", "neural", "network", "parameters", "33", ":", "end", "for", "41", ":", "end", "while", "Note", ":", "In", "this", "pseudo", "-", "code", ",", "e", "(", "y", ")", "refers", "to", "a", "\"", "one", "-", "hot", "\"", "vector", ",", "consisting", "of", "all", "0s", "except", "for", "a", "1", "at", "position", "y", ",", "and", "is", "the", "element", "-", "wise", "product", ".", "We", "propose", "to", "tackle", "this", "problem", "with", "a", "simple", "stochastic", "gradient", "procedure", ",", "in", "which", "updates", "are", "made", "in", "the", "opposite", "direction", "of", "the", "gradient", "of", "Equation", "(", "9", ")", "for", "the", "minimizing", "parameters", ",", "and", "in", "the", "direction", "of", "the", "gradient", "for", "the", "maximizing", "parameters", ".", "Stochastic", "estimates", "of", "the", "gradient", "are", "made", ",", "using", "a", "subset", "of", "the", "training", "samples", "to", "compute", "the", "averages", ".", "Algorithm", "1", "provides", "the", "complete", "pseudo", "-", "code", "of", "this", "learning", "procedure", ".", "3", "In", "words", ",", "during", "training", ",", "the", "neural", "network", "(", "parameterized", "by", "W", ",", "b", ",", "V", ",", "c", ")", "and", "the", "domain", "regressor", "(", "parameterized", "by", "u", ",", "z", ")", "are", "competing", "against", "each", "other", ",", "in", "an", "adversarial", "way", ",", "over", "the", "objective", "of", "Equation", "(", "9", ")", ".", "For", "this", "reason", ",", "we", "refer", "to", "networks", "trained", "according", "to", "this", "objective", "as", "Domain", "-", "Adversarial", "Neural", "Networks", "(", "DANN", ")", ".", "DANN", "will", "effectively", "attempt", "to", "learn", "a", "hidden", "layer", "G", "f", "(", "·", ")", "that", "maps", "an", "example", "(", "either", "source", "or", "target", ")", "into", "a", "representation", "allowing", "the", "output", "layer", "G", "y", "(", "·", ")", "to", "accurately", "classify", "source", "samples", ",", "but", "crippling", "the", "ability", "of", "the", "domain", "regressor", "G", "d", "(", "·", ")", "to", "detect", "whether", "each", "example", "belongs", "to", "the", "source", "or", "target", "domains", ".", "section", ":", "Generalization", "to", "Arbitrary", "Architectures", "For", "illustration", "purposes", ",", "we", "'", "ve", "so", "far", "focused", "on", "the", "case", "of", "a", "single", "hidden", "layer", "DANN", ".", "However", ",", "it", "is", "straightforward", "to", "generalize", "to", "other", "sophisticated", "architectures", ",", "which", "might", "be", "more", "appropriate", "for", "the", "data", "at", "hand", ".", "For", "example", ",", "deep", "convolutional", "neural", "networks", "are", "well", "known", "for", "being", "state", "-", "of", "-", "the", "-", "art", "models", "for", "learning", "discriminative", "features", "of", "images", "[", "reference", "]", ".", "Let", "us", "now", "use", "a", "more", "general", "notation", "for", "the", "different", "components", "of", "DANN", ".", "Namely", ",", "let", "G", "f", "(", "·", ";", "θ", "f", ")", "be", "the", "D", "-", "dimensional", "neural", "network", "feature", "extractor", ",", "with", "parameters", "θ", "f", ".", "Also", ",", "let", "G", "y", "(", "·", ";", "θ", "y", ")", "be", "the", "part", "of", "DANN", "that", "computes", "the", "network", "'s", "label", "prediction", "output", "layer", ",", "with", "parameters", "θ", "y", ",", "while", "G", "d", "(", "·", ";", "θ", "d", ")", "now", "corresponds", "to", "the", "computation", "of", "the", "domain", "prediction", "output", "of", "the", "network", ",", "with", "parameters", "θ", "d", ".", "Note", "that", "for", "preserving", "the", "theoretical", "guarantees", "of", "Theorem", "2", ",", "the", "hypothesis", "class", "H", "d", "generated", "by", "the", "domain", "prediction", "component", "G", "d", "should", "include", "the", "hypothesis", "class", "H", "y", "generated", "by", "the", "label", "prediction", "component", "G", "y", ".", "Thus", ",", "We", "will", "note", "the", "prediction", "loss", "and", "the", "domain", "loss", "respectively", "by", "Training", "DANN", "then", "parallels", "the", "single", "layer", "case", "and", "consists", "in", "optimizing", "by", "finding", "the", "saddle", "pointθ", "f", ",", "θ", "y", ",", "θ", "d", "such", "that", "As", "suggested", "previously", ",", "a", "saddle", "point", "defined", "by", "Equations", "(", "11", "-", "12", ")", "can", "be", "found", "as", "a", "stationary", "point", "of", "the", "following", "gradient", "updates", ":", "where", "µ", "is", "the", "learning", "rate", ".", "We", "use", "stochastic", "estimates", "of", "these", "gradients", ",", "by", "sampling", "examples", "from", "the", "data", "set", ".", "The", "updates", "of", "Equations", "(", "13", "-", "15", ")", "are", "very", "similar", "to", "stochastic", "gradient", "descent", "(", "SGD", ")", "updates", "for", "a", "feed", "-", "forward", "deep", "model", "that", "comprises", "feature", "extractor", "fed", "into", "the", "label", "Figure", "1", ":", "The", "proposed", "architecture", "includes", "a", "deep", "feature", "extractor", "(", "green", ")", "and", "a", "deep", "label", "predictor", "(", "blue", ")", ",", "which", "together", "form", "a", "standard", "feed", "-", "forward", "architecture", ".", "Unsupervised", "domain", "adaptation", "is", "achieved", "by", "adding", "a", "domain", "classifier", "(", "red", ")", "connected", "to", "the", "feature", "extractor", "via", "a", "gradient", "reversal", "layer", "that", "multiplies", "the", "gradient", "by", "a", "certain", "negative", "constant", "during", "the", "backpropagation", "-", "based", "training", ".", "Otherwise", ",", "the", "training", "proceeds", "standardly", "and", "minimizes", "the", "label", "prediction", "loss", "(", "for", "source", "examples", ")", "and", "the", "domain", "classification", "loss", "(", "for", "all", "samples", ")", ".", "Gradient", "reversal", "ensures", "that", "the", "feature", "distributions", "over", "the", "two", "domains", "are", "made", "similar", "(", "as", "indistinguishable", "as", "possible", "for", "the", "domain", "classifier", ")", ",", "thus", "resulting", "in", "the", "domain", "-", "invariant", "features", ".", "predictor", "and", "into", "the", "domain", "classifier", "(", "with", "loss", "weighted", "by", "λ", ")", ".", "The", "only", "difference", "is", "that", "in", "(", "13", ")", ",", "the", "gradients", "from", "the", "class", "and", "domain", "predictors", "are", "subtracted", ",", "instead", "of", "being", "summed", "(", "the", "difference", "is", "important", ",", "as", "otherwise", "SGD", "would", "try", "to", "make", "features", "dissimilar", "across", "domains", "in", "order", "to", "minimize", "the", "domain", "classification", "loss", ")", ".", "Since", "SGDand", "its", "many", "variants", ",", "such", "as", "ADAGRAD", "[", "reference", "]", "or", "ADADELTA", "[", "reference", "]-", "is", "the", "main", "learning", "algorithm", "implemented", "in", "most", "libraries", "for", "deep", "learning", ",", "it", "would", "be", "convenient", "to", "frame", "an", "implementation", "of", "our", "stochastic", "saddle", "point", "procedure", "as", "SGD", ".", "Fortunately", ",", "such", "a", "reduction", "can", "be", "accomplished", "by", "introducing", "a", "special", "gradient", "reversal", "layer", "(", "GRL", ")", ",", "defined", "as", "follows", ".", "The", "gradient", "reversal", "layer", "has", "no", "parameters", "associated", "with", "it", ".", "During", "the", "forward", "propagation", ",", "the", "GRL", "acts", "as", "an", "identity", "transformation", ".", "During", "the", "backpropagation", "however", ",", "the", "GRL", "takes", "the", "gradient", "from", "the", "subsequent", "level", "and", "changes", "its", "sign", ",", "i.e.", ",", "multiplies", "it", "by", "−1", ",", "before", "passing", "it", "to", "the", "preceding", "layer", ".", "Implementing", "such", "a", "layer", "using", "existing", "object", "-", "oriented", "packages", "for", "deep", "learning", "is", "simple", ",", "requiring", "only", "to", "define", "procedures", "for", "the", "forward", "propagation", "(", "identity", "transformation", ")", ",", "and", "backpropagation", "(", "multiplying", "by", "−1", ")", ".", "The", "layer", "requires", "no", "parameter", "update", ".", "The", "GRL", "as", "defined", "above", "is", "inserted", "between", "the", "feature", "extractor", "G", "f", "and", "the", "domain", "classifier", "G", "d", ",", "resulting", "in", "the", "architecture", "depicted", "in", "Figure", "1", ".", "As", "the", "backpropagation", "process", "passes", "through", "the", "GRL", ",", "the", "partial", "derivatives", "of", "the", "loss", "that", "is", "downstream", "the", "GRL", "(", "i.e.", ",", "L", "d", ")", "w.r.t", ".", "the", "layer", "parameters", "that", "are", "upstream", "the", "GRL", "(", "i.e.", ",", "θ", "f", ")", "get", "multiplied", "by", "−1", ",", "i.e.", ",", ".", "Therefore", ",", "running", "SGD", "in", "the", "resulting", "model", "implements", "the", "updates", "of", "Equations", "(", "13", "-", "15", ")", "and", "converges", "to", "a", "saddle", "point", "of", "Equation", "(", "10", ")", ".", "Mathematically", ",", "we", "can", "formally", "treat", "the", "gradient", "reversal", "layer", "as", "a", "\"", "pseudo", "-", "function", "\"", "R", "(", "x", ")", "defined", "by", "two", "(", "incompatible", ")", "equations", "describing", "its", "forward", "and", "backpropagation", "behaviour", ":", "where", "I", "is", "an", "identity", "matrix", ".", "We", "can", "then", "define", "the", "objective", "\"", "pseudo", "-", "function", "\"", "of", "(", "θ", "f", ",", "θ", "y", ",", "θ", "d", ")", "that", "is", "being", "optimized", "by", "the", "stochastic", "gradient", "descent", "within", "our", "method", ":", "Running", "updates", "(", "13", "-", "15", ")", "can", "then", "be", "implemented", "as", "doing", "SGD", "for", "(", "18", ")", "and", "leads", "to", "the", "emergence", "of", "features", "that", "are", "domain", "-", "invariant", "and", "discriminative", "at", "the", "same", "time", ".", "After", "the", "learning", ",", "the", "label", "predictor", "G", "y", "(", "G", "f", "(", "x", ";", "θ", "f", ")", ";", "θ", "y", ")", "can", "be", "used", "to", "predict", "labels", "for", "samples", "from", "the", "target", "domain", "(", "as", "well", "as", "from", "the", "source", "domain", ")", ".", "Note", "that", "we", "release", "the", "source", "code", "for", "the", "Gradient", "Reversal", "layer", "along", "with", "the", "usage", "examples", "as", "an", "extension", "to", "Caffe", ".", "4", "section", ":", "Experiments", "In", "this", "section", ",", "we", "present", "a", "variety", "of", "empirical", "results", "for", "both", "shallow", "domain", "adversarial", "neural", "networks", "(", "Subsection", "5.1", ")", "and", "deep", "ones", "(", "Subsections", "5.2", "and", "5.3", ")", ".", "section", ":", "Experiments", "with", "Shallow", "Neural", "Networks", "In", "this", "first", "experiment", "section", ",", "we", "evaluate", "the", "behavior", "of", "the", "simple", "version", "of", "DANN", "described", "by", "Subsection", "4.1", ".", "Note", "that", "the", "results", "reported", "in", "the", "present", "subsection", "are", "obtained", "using", "Algorithm", "1", ".", "Thus", ",", "the", "stochastic", "gradient", "descent", "approach", "here", "consists", "of", "sampling", "a", "pair", "of", "source", "and", "target", "examples", "and", "performing", "a", "gradient", "step", "update", "of", "all", "parameters", "of", "DANN", ".", "Crucially", ",", "while", "the", "update", "of", "the", "regular", "parameters", "follows", "as", "usual", "the", "opposite", "direction", "of", "the", "gradient", ",", "for", "the", "adversarial", "parameters", "the", "step", "must", "follow", "the", "gradient", "'s", "direction", "(", "since", "we", "maximize", "with", "respect", "to", "them", ",", "instead", "of", "minimizing", ")", ".", "section", ":", "Experiments", "on", "a", "Toy", "Problem", "As", "a", "first", "experiment", ",", "we", "study", "the", "behavior", "of", "the", "proposed", "algorithm", "on", "a", "variant", "of", "the", "inter", "-", "twinning", "moons", "2D", "problem", ",", "where", "the", "target", "distribution", "is", "a", "rotation", "of", "the", "source", "one", ".", "As", "the", "source", "sample", "S", ",", "we", "generate", "a", "lower", "moon", "and", "an", "upper", "moon", "labeled", "0", "and", "1", "respectively", ",", "each", "of", "which", "containing", "150", "examples", ".", "The", "target", "sample", "T", "is", "obtained", "by", "the", "following", "procedure", ":", "(", "1", ")", "we", "generate", "a", "sample", "S", "the", "same", "way", "S", "has", "been", "generated", ";", "(", "2", ")", "we", "rotate", "each", "example", "by", "35", "•", ";", "and", "(", "3", ")", "we", "remove", "all", "the", "labels", ".", "Thus", ",", "T", "contains", "300", "unlabeled", "examples", ".", "We", "have", "represented", "those", "examples", "in", "Figure", "2", ".", "We", "study", "the", "adaptation", "capability", "of", "DANN", "by", "comparing", "it", "to", "the", "standard", "neural", "network", "(", "NN", ")", ".", "In", "these", "toy", "experiments", ",", "both", "algorithms", "share", "the", "same", "network", "architecture", ",", "with", "a", "hidden", "layer", "size", "of", "15", "neurons", ".", "We", "train", "the", "NN", "using", "the", "same", "procedure", "as", "the", "DANN", ".", "That", "is", ",", "we", "keep", "updating", "the", "domain", "regressor", "component", "using", "target", "sample", "T", "(", "with", "a", "hyper", "-", "parameter", "λ", "=", "6", ";", "the", "same", "value", "is", "used", "for", "DANN", ")", ",", "but", "we", "disable", "the", "adversarial", "back", "-", "propagation", "into", "the", "hidden", "layer", ".", "To", "do", "so", ",", "we", "execute", "Algorithm", "1", "by", "omitting", "the", "lines", "numbered", "22", "and", "31", ".", "This", "allows", "recovering", "the", "NN", "learning", "algorithm", "-", "based", "on", "the", "source", "risk", "minimization", "of", "Equation", "(", "5", ")", "without", "any", "regularizer", "-", "and", "simultaneously", "train", "the", "domain", "regressor", "of", "Equation", "(", "7", ")", "to", "discriminate", "between", "source", "and", "target", "domains", ".", "With", "this", "toy", "experience", ",", "we", "will", "first", "illustrate", "how", "DANN", "adapts", "its", "decision", "boundary", "when", "compared", "to", "NN", ".", "Moreover", ",", "we", "will", "also", "illustrate", "how", "the", "representation", "given", "by", "the", "hidden", "layer", "is", "less", "adapted", "to", "the", "source", "domain", "task", "with", "DANN", "than", "with", "NN", "(", "this", "is", "why", "we", "need", "a", "domain", "regressor", "in", "the", "NN", "experiment", ")", ".", "We", "recall", "that", "this", "is", "the", "founding", "idea", "behind", "our", "proposed", "algorithm", ".", "The", "analysis", "of", "the", "experiment", "appears", "in", "Figure", "2", ",", "where", "upper", "graphs", "relate", "to", "standard", "NN", ",", "and", "lower", "graphs", "relate", "to", "DANN", ".", "By", "looking", "at", "the", "lower", "and", "upper", "graphs", "pairwise", ",", "we", "compare", "NN", "and", "DANN", "from", "four", "different", "perspectives", ",", "described", "in", "details", "below", ".", "The", "column", "\"", "Label", "Classification", "\"", "of", "Figure", "2", "shows", "the", "decision", "boundaries", "of", "DANN", "and", "NN", "on", "the", "problem", "of", "predicting", "the", "labels", "of", "both", "source", "and", "the", "target", "examples", ".", "As", "expected", ",", "NN", "accurately", "classifies", "the", "two", "classes", "of", "the", "source", "sample", "S", ",", "but", "is", "not", "fully", "adapted", "to", "the", "target", "sample", "T", ".", "On", "the", "contrary", ",", "the", "decision", "boundary", "of", "DANN", "perfectly", "classifies", "examples", "from", "both", "source", "and", "target", "samples", ".", "In", "the", "studied", "task", ",", "DANN", "clearly", "adapts", "to", "the", "target", "distribution", ".", "The", "column", "\"", "Representation", "PCA", "\"", "studies", "how", "the", "domain", "adaptation", "regularizer", "affects", "the", "representation", "G", "f", "(", "·", ")", "provided", "by", "the", "network", "hidden", "layer", ".", "The", "graphs", "are", "obtained", "by", "applying", "a", "Principal", "component", "analysis", "(", "PCA", ")", "on", "the", "set", "of", "all", "representation", "of", "source", "and", "target", "data", "points", ",", "i.e.", ",", "S", "(", "G", "f", ")", "∪", "T", "(", "G", "f", ")", ".", "Thus", ",", "given", "the", "trained", "network", "(", "NN", "or", "DANN", ")", ",", "every", "point", "from", "S", "and", "T", "is", "mapped", "into", "a", "15", "-", "dimensional", "feature", "space", "through", "the", "hidden", "layer", ",", "and", "projected", "back", "into", "a", "two", "-", "dimensional", "plane", "by", "the", "PCA", "transformation", ".", "In", "the", "DANN", "-", "PCA", "representation", ",", "we", "observe", "that", "target", "points", "are", "homogeneously", "spread", "out", "among", "source", "points", ";", "In", "the", "NN", "-", "PCA", "representation", ",", "a", "number", "of", "target", "points", "belong", "to", "clusters", "containing", "no", "source", "points", ".", "Hence", ",", "labeling", "the", "target", "points", "seems", "an", "easier", "task", "given", "the", "DANN", "-", "PCA", "representation", ".", "To", "push", "the", "analysis", "further", ",", "the", "PCA", "graphs", "tag", "four", "crucial", "data", "points", "by", "the", "letters", "A", ",", "B", ",", "C", "and", "D", ",", "that", "correspond", "to", "the", "moon", "extremities", "in", "the", "original", "space", "(", "note", "that", "the", "original", "point", "locations", "are", "tagged", "in", "the", "first", "column", "graphs", ")", ".", "We", "observe", "that", "points", "A", "and", "B", "are", "very", "close", "to", "each", "other", "in", "the", "NN", "-", "PCA", "representation", ",", "while", "they", "clearly", "belong", "to", "different", "classes", ".", "The", "same", "happens", "to", "points", "C", "and", "D.", "Conversely", ",", "these", "four", "points", "are", "at", "the", "opposite", "four", "corners", "in", "the", "DANN", "-", "PCA", "representation", ".", "Note", "also", "that", "the", "target", "point", "A", "(", "resp", ".", "D", ")-", "that", "is", "difficult", "to", "classify", "in", "the", "original", "space", "-", "is", "located", "in", "the", "\"", "+", "\"", "cluster", "(", "resp", ".", "\"", "−", "−", "−\"cluster", ")", "in", "the", "DANN", "-", "PCA", "representation", ".", "Therefore", ",", "the", "representation", "promoted", "by", "DANN", "is", "better", "suited", "to", "the", "adaptation", "problem", ".", "The", "column", "\"", "Domain", "Classification", "\"", "shows", "the", "decision", "boundary", "on", "the", "domain", "classification", "problem", ",", "which", "is", "given", "by", "the", "domain", "regressor", "G", "d", "of", "Equation", "(", "7", ")", ".", "More", "precisely", ",", "an", "example", "x", "is", "classified", "as", "a", "source", "example", "when", "G", "d", "(", "G", "f", "(", "x", ")", ")", "≥", "0.5", ",", "and", "is", "classified", "as", "a", "domain", "example", "otherwise", ".", "Remember", "that", ",", "during", "the", "learning", "process", "of", "DANN", ",", "the", "G", "d", "regressor", "struggles", "to", "discriminate", "between", "source", "and", "target", "domains", ",", "while", "the", "hidden", "representation", "G", "f", "(", "·", ")", "is", "adversarially", "updated", "to", "prevent", "it", "to", "succeed", ".", "As", "explained", "above", ",", "we", "trained", "a", "domain", "regressor", "during", "the", "learning", "process", "of", "NN", ",", "but", "without", "allowing", "it", "to", "influence", "the", "learned", "representation", "G", "f", "(", "·", ")", ".", "On", "one", "hand", ",", "the", "DANN", "domain", "regressor", "clearly", "fails", "to", "generalize", "source", "and", "target", "distribution", "topologies", ".", "On", "the", "other", "hand", ",", "the", "NN", "domain", "regressor", "shows", "a", "better", "(", "although", "imperfect", ")", "generalization", "capability", ".", "Inter", "alia", ",", "it", "seems", "to", "roughly", "capture", "the", "rotation", "angle", "of", "the", "target", "distribution", ".", "This", "again", "corroborates", "that", "the", "DANN", "representation", "does", "not", "allow", "discriminating", "between", "domains", ".", "The", "column", "\"", "Hidden", "Neurons", "\"", "shows", "the", "configuration", "of", "hidden", "layer", "neurons", "(", "by", "Equation", "4", ",", "we", "have", "that", "each", "neuron", "is", "indeed", "a", "linear", "regressor", ")", ".", "In", "other", "words", ",", "each", "of", "the", "fifteen", "plot", "line", "corresponds", "to", "the", "coordinates", "x", "∈", "R", "2", "for", "which", "the", "i", "-", "th", "component", "of", "G", "f", "(", "x", ")", "equals", "1", "2", ",", "for", "i", "∈", "{", "1", ",", ".", ".", ".", ",", "15}.", "We", "observe", "that", "the", "standard", "NN", "neurons", "are", "grouped", "in", "three", "clusters", ",", "each", "one", "allowing", "to", "generate", "a", "straight", "line", "of", "the", "zigzag", "decision", "boundary", "for", "the", "label", "classification", "problem", ".", "However", ",", "most", "of", "these", "neurons", "are", "also", "able", "to", "(", "roughly", ")", "capture", "the", "rotation", "angle", "of", "the", "domain", "classification", "problem", ".", "Hence", ",", "we", "observe", "that", "the", "adaptation", "regularizer", "of", "DANN", "prevents", "these", "kinds", "of", "neurons", "to", "be", "produced", ".", "It", "is", "indeed", "striking", "to", "see", "that", "the", "two", "predominant", "patterns", "in", "the", "NN", "neurons", "(", "i.e.", ",", "the", "two", "parallel", "lines", "crossing", "the", "plane", "from", "lower", "left", "to", "upper", "right", ")", "are", "vanishing", "in", "the", "DANN", "neurons", ".", "section", ":", "Unsupervised", "Hyper", "-", "Parameter", "Selection", "To", "perform", "unsupervised", "domain", "adaption", ",", "one", "should", "provide", "ways", "to", "set", "hyper", "-", "parameters", "(", "such", "as", "the", "domain", "regularization", "parameter", "λ", ",", "the", "learning", "rate", ",", "the", "network", "architecture", "for", "our", "method", ")", "in", "an", "unsupervised", "way", ",", "i.e.", ",", "without", "referring", "to", "labeled", "data", "in", "the", "target", "domain", ".", "In", "the", "following", "experiments", "of", "Sections", "5.1.3", "and", "5.1.4", ",", "we", "select", "the", "hyper", "-", "parameters", "of", "each", "algorithm", "by", "using", "a", "variant", "of", "reverse", "cross", "-", "validation", "approach", "proposed", "by", "[", "reference", "]", ",", "that", "we", "call", "reverse", "validation", ".", "To", "evaluate", "the", "reverse", "validation", "risk", "associated", "to", "a", "tuple", "of", "hyper", "-", "parameters", ",", "we", "proceed", "as", "follows", ".", "Given", "the", "labeled", "source", "sample", "S", "and", "the", "unlabeled", "target", "sample", "T", ",", "we", "split", "each", "set", "into", "training", "sets", "(", "S", "and", "T", "respectively", ",", "containing", "90", "%", "of", "the", "original", "examples", ")", "and", "the", "validation", "sets", "(", "S", "V", "and", "T", "V", "respectively", ")", ".", "We", "use", "the", "labeled", "set", "S", "and", "the", "unlabeled", "target", "set", "T", "to", "learn", "a", "classifier", "η", ".", "Then", ",", "using", "the", "same", "algorithm", ",", "we", "learn", "a", "reverse", "classifier", "η", "r", "using", "the", "self", "-", "labeled", "set", "{", "(", "x", ",", "η", "(", "x", ")", ")", "}", "x∈T", "and", "the", "unlabeled", "part", "of", "S", "as", "target", "sample", ".", "Finally", ",", "the", "reverse", "classifier", "η", "r", "is", "evaluated", "on", "the", "validation", "set", "S", "V", "of", "source", "sample", ".", "We", "then", "say", "that", "the", "classifier", "η", "has", "a", "reverse", "validation", "risk", "of", "R", "S", "V", "(", "η", "r", ")", ".", "The", "process", "is", "repeated", "with", "multiple", "values", "of", "hyper", "-", "parameters", "and", "the", "selected", "parameters", "are", "those", "corresponding", "to", "the", "classifier", "with", "the", "lowest", "reverse", "validation", "risk", ".", "Note", "that", "when", "we", "train", "neural", "network", "architectures", ",", "the", "validation", "set", "S", "V", "is", "also", "used", "as", "an", "early", "stopping", "criterion", "during", "the", "learning", "of", "η", ",", "and", "self", "-", "labeled", "validation", "set", "{", "(", "x", ",", "η", "(", "x", ")", ")", "}", "x∈T", "V", "is", "used", "as", "an", "early", "stopping", "criterion", "during", "the", "learning", "of", "η", "r", ".", "We", "also", "observed", "better", "accuracies", "when", "we", "initialized", "the", "learning", "of", "the", "reverse", "classifier", "η", "r", "with", "the", "configuration", "learned", "by", "the", "network", "η", ".", "section", ":", "Experiments", "on", "Sentiment", "Analysis", "Data", "Sets", "We", "now", "compare", "the", "performance", "of", "our", "proposed", "DANN", "algorithm", "to", "a", "standard", "neural", "network", "with", "one", "hidden", "layer", "(", "NN", ")", "described", "by", "Equation", "(", "5", ")", ",", "and", "a", "Support", "Vector", "Machine", "(", "SVM", ")", "with", "a", "linear", "kernel", ".", "We", "compare", "the", "algorithms", "on", "the", "Amazon", "reviews", "data", "set", ",", "as", "pre", "-", "processed", "by", "[", "reference", "]", ".", "This", "data", "set", "includes", "four", "domains", ",", "each", "one", "composed", "of", "reviews", "of", "a", "specific", "kind", "of", "product", "(", "books", ",", "dvd", "disks", ",", "electronics", ",", "and", "kitchen", "appliances", ")", ".", "Reviews", "are", "encoded", "in", "5", "000", "dimensional", "feature", "vectors", "of", "unigrams", "and", "bigrams", ",", "and", "labels", "are", "binary", ":", "\"", "0", "\"", "if", "the", "product", "is", "ranked", "up", "to", "3", "stars", ",", "and", "\"", "1", "\"", "if", "the", "product", "is", "ranked", "4", "or", "5", "stars", ".", "We", "perform", "twelve", "domain", "adaptation", "tasks", ".", "All", "learning", "algorithms", "are", "given", "2", "000", "labeled", "source", "examples", "and", "2", "000", "unlabeled", "target", "examples", ".", "Then", ",", "we", "evaluate", "them", "on", "separate", "target", "test", "sets", "(", "between", "3", "000", "and", "6", "000", "examples", ")", ".", "Note", "that", "NN", "and", "SVM", "do", "not", "use", "the", "unlabeled", "target", "sample", "for", "learning", ".", "Here", "are", "more", "details", "about", "the", "procedure", "used", "for", "each", "learning", "algorithms", "leading", "to", "the", "empirical", "results", "of", "Table", "1", ":", "Classification", "accuracy", "on", "the", "Amazon", "reviews", "data", "set", ",", "and", "Pairwise", "Poisson", "binomial", "test", ".", "•", "For", "the", "DANN", "algorithm", ",", "the", "adaptation", "parameter", "λ", "is", "chosen", "among", "9", "values", "between", "10", "−2", "and", "1", "on", "a", "logarithmic", "scale", ".", "The", "hidden", "layer", "size", "l", "is", "either", "50", "or", "100", ".", "Finally", ",", "the", "learning", "rate", "µ", "is", "fixed", "at", "10", "−3", ".", "•", "For", "the", "NN", "algorithm", ",", "we", "use", "exactly", "the", "same", "hyper", "-", "parameters", "grid", "and", "training", "procedure", "as", "DANN", "above", ",", "except", "that", "we", "do", "not", "need", "an", "adaptation", "parameter", ".", "Note", "that", "one", "can", "train", "NN", "by", "using", "the", "DANN", "implementation", "(", "Algorithm", "1", ")", "with", "λ", "=", "0", ".", "•", "For", "the", "SVM", "algorithm", ",", "the", "hyper", "-", "parameter", "C", "is", "chosen", "among", "10", "values", "between", "10", "−5", "and", "1", "on", "a", "logarithmic", "scale", ".", "This", "range", "of", "values", "is", "the", "same", "as", "used", "by", "[", "reference", "]", "in", "their", "experiments", ".", "As", "presented", "at", "Section", "5.1.2", ",", "we", "used", "reverse", "cross", "validation", "selecting", "the", "hyper", "-", "parameters", "for", "all", "three", "learning", "algorithms", ",", "with", "early", "stopping", "as", "the", "stopping", "criterion", "for", "DANN", "and", "NN", ".", "The", "\"", "Original", "data", "\"", "part", "of", "Table", "1a", "shows", "the", "target", "test", "accuracy", "of", "all", "algorithms", ",", "and", "Table", "1b", "reports", "the", "probability", "that", "one", "algorithm", "is", "significantly", "better", "than", "the", "others", "according", "to", "the", "Poisson", "binomial", "test", "[", "reference", "]", ".", "We", "note", "that", "DANN", "has", "a", "significantly", "better", "performance", "than", "NN", "and", "SVM", ",", "with", "respective", "probabilities", "0.87", "and", "0.83", ".", "As", "the", "only", "difference", "between", "DANN", "and", "NN", "is", "the", "domain", "adaptation", "regularizer", ",", "we", "conclude", "that", "our", "approach", "successfully", "helps", "to", "find", "a", "representation", "suitable", "for", "the", "target", "domain", ".", "section", ":", "Combining", "DANN", "with", "Denoising", "Autoencoders", "We", "now", "investigate", "on", "whether", "the", "DANN", "algorithm", "can", "improve", "on", "the", "representation", "learned", "by", "the", "state", "-", "of", "-", "the", "-", "art", "Marginalized", "Stacked", "Denoising", "Autoencoders", "(", "mSDA", ")", "proposed", "by", "[", "reference", "]", ".", "In", "brief", ",", "mSDA", "is", "an", "unsupervised", "algorithm", "that", "learns", "a", "new", "robust", "feature", "representation", "of", "the", "training", "samples", ".", "It", "takes", "the", "unlabeled", "parts", "of", "both", "source", "and", "target", "samples", "to", "learn", "a", "feature", "map", "from", "input", "space", "X", "to", "a", "new", "representation", "space", ".", "As", "a", "denoising", "autoencoders", "algorithm", ",", "it", "finds", "a", "feature", "representation", "from", "which", "one", "can", "(", "approximately", ")", "reconstruct", "the", "original", "features", "of", "an", "example", "from", "its", "noisy", "counterpart", ".", "[", "reference", "]", "showed", "that", "using", "mSDA", "with", "a", "linear", "SVM", "classifier", "reaches", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "the", "Amazon", "reviews", "data", "sets", ".", "As", "an", "alternative", "to", "the", "SVM", ",", "we", "propose", "to", "apply", "our", "Shallow", "DANN", "algorithm", "on", "the", "same", "representations", "generated", "by", "mSDA", "(", "using", "representations", "of", "both", "source", "and", "target", "samples", ")", ".", "Note", "that", ",", "even", "if", "mSDA", "and", "DANN", "are", "two", "representation", "learning", "approaches", ",", "they", "optimize", "different", "objectives", ",", "which", "can", "be", "complementary", ".", "We", "perform", "this", "experiment", "on", "the", "same", "Amazon", "reviews", "data", "set", "described", "in", "the", "previous", "subsection", ".", "For", "each", "source", "-", "target", "domain", "pair", ",", "we", "generate", "the", "mSDA", "representations", "using", "a", "corruption", "probability", "of", "50", "%", "and", "a", "number", "of", "layers", "of", "5", ".", "We", "then", "execute", "the", "three", "learning", "algorithms", "(", "DANN", ",", "NN", ",", "and", "SVM", ")", "on", "these", "representations", ".", "More", "precisely", ",", "following", "the", "experimental", "procedure", "of", "[", "reference", "]", ",", "we", "use", "the", "concatenation", "of", "the", "output", "of", "the", "5", "layers", "and", "the", "original", "input", "as", "the", "new", "representation", ".", "Thus", ",", "each", "example", "is", "now", "encoded", "in", "a", "vector", "of", "30", "000", "dimensions", ".", "Note", "that", "we", "use", "the", "same", "grid", "search", "as", "in", "the", "previous", "Subsection", "5.1.3", ",", "but", "use", "a", "learning", "rate", "µ", "of", "10", "−4", "for", "both", "DANN", "and", "the", "NN", ".", "The", "results", "of", "\"", "mSDA", "representation", "\"", "columns", "in", "Table", "1a", "confirm", "that", "combining", "mSDA", "and", "DANN", "is", "a", "sound", "approach", ".", "Indeed", ",", "the", "Poisson", "binomial", "test", "shows", "that", "DANN", "has", "a", "better", "performance", "than", "the", "NN", "and", "the", "SVM", ",", "with", "probabilities", "0.92", "and", "0.88", "respectively", ",", "as", "reported", "in", "Table", "1b", ".", "We", "note", "however", "that", "the", "standard", "NN", "and", "the", "SVM", "find", "the", "best", "solution", "on", "respectively", "the", "second", "and", "the", "fourth", "tasks", ".", "This", "suggests", "that", "DANN", "and", "mSDA", "adaptation", "strategies", "are", "not", "fully", "complementary", ".", "section", ":", "Proxy", "Distance", "The", "theoretical", "foundation", "of", "the", "DANN", "algorithm", "is", "the", "domain", "adaptation", "theory", "of", "[", "reference", "][", "reference", "]", ".", "We", "claimed", "that", "DANN", "finds", "a", "representation", "in", "which", "the", "source", "and", "the", "target", "example", "are", "hardly", "distinguishable", ".", "Our", "toy", "experiment", "of", "Section", "5.1.1", "already", "points", "out", "some", "evidence", "for", "that", "and", "here", "we", "provide", "analysis", "on", "real", "data", ".", "To", "do", "so", ",", "we", "compare", "the", "Proxy", "A", "-", "distance", "(", "PAD", ")", "on", "various", "representations", "of", "the", "Amazon", "Reviews", "data", "set", ";", "these", "representations", "are", "obtained", "by", "running", "either", "NN", ",", "DANN", ",", "mSDA", ",", "or", "mSDA", "and", "DANN", "combined", ".", "Recall", "that", "PAD", ",", "as", "described", "in", "Section", "3.2", ",", "is", "a", "metric", "estimating", "the", "similarity", "of", "the", "source", "and", "the", "target", "representations", ".", "More", "precisely", ",", "to", "obtain", "a", "PAD", "value", ",", "we", "use", "the", "following", "procedure", ":", "(", "1", ")", "we", "construct", "the", "data", "set", "U", "of", "Equation", "(", "2", ")", "using", "both", "source", "and", "target", "representations", "of", "the", "training", "samples", ";", "(", "2", ")", "we", "randomly", "split", "U", "in", "two", "subsets", "of", "equal", "size", ";", "(", "3", ")", "we", "train", "linear", "SVMs", "on", "the", "first", "subset", "of", "U", "using", "a", "large", "range", "of", "C", "values", ";", "(", "4", ")", "we", "compute", "the", "error", "of", "all", "obtained", "classifiers", "on", "the", "second", "subset", "of", "U", ";", "and", "(", "5", ")", "we", "use", "the", "lowest", "error", "to", "compute", "the", "PAD", "value", "of", "Equation", "(", "3", ")", ".", "Firstly", ",", "Figure", "3a", "compares", "the", "PAD", "of", "DANN", "representations", "obtained", "in", "the", "experiments", "of", "Section", "5.1.3", "(", "using", "the", "hyper", "-", "parameters", "values", "leading", "to", "the", "results", "of", "Table", "1", ")", "to", "the", "PAD", "computed", "on", "raw", "data", ".", "As", "expected", ",", "the", "PAD", "values", "are", "driven", "down", "by", "the", "DANN", "representations", ".", "Secondly", ",", "Figure", "3b", "compares", "the", "PAD", "of", "DANN", "representations", "to", "the", "PAD", "of", "standard", "NN", "representations", ".", "As", "the", "PAD", "is", "influenced", "by", "the", "hidden", "layer", "size", "(", "the", "discriminating", "power", "tends", "to", "increase", "with", "the", "representation", "length", ")", ",", "we", "fix", "here", "the", "size", "to", "100", "neurons", "for", "both", "algorithms", ".", "We", "also", "fix", "the", "adaptation", "parameter", "of", "DANN", "to", "λ", "0.31", ";", "it", "was", "the", "value", "that", "has", "been", "selected", "most", "of", "the", "time", "during", "our", "preceding", "experiments", "on", "the", "Amazon", "Reviews", "data", "set", ".", "Again", ",", "DANN", "is", "clearly", "leading", "to", "the", "lowest", "PAD", "values", ".", "Lastly", ",", "Figure", "3c", "presents", "two", "sets", "of", "results", "related", "to", "Section", "5.1.4", "experiments", ".", "On", "one", "hand", ",", "we", "reproduce", "the", "results", "of", "[", "reference", "]", ",", "which", "noticed", "that", "the", "mSDA", "representations", "have", "greater", "PAD", "values", "than", "original", "(", "raw", ")", "data", ".", "Although", "the", "mSDA", "approach", "clearly", "helps", "to", "adapt", "to", "the", "target", "task", ",", "it", "seems", "to", "contradict", "the", "theory", "of", "BenDavid", "et", "al", "..", "On", "the", "other", "hand", ",", "we", "observe", "that", ",", "when", "running", "DANN", "on", "top", "of", "mSDA", "(", "using", "the", "hyper", "-", "parameters", "values", "leading", "to", "the", "results", "of", "Table", "1", ")", ",", "the", "obtained", "representations", "have", "much", "lower", "PAD", "values", ".", "These", "observations", "might", "explain", "the", "improvements", "provided", "by", "DANN", "when", "combined", "with", "the", "mSDA", "procedure", ".", "section", ":", "Experiments", "with", "Deep", "Networks", "on", "Image", "Classification", "We", "now", "perform", "extensive", "evaluation", "of", "a", "deep", "version", "of", "DANN", "(", "see", "Subsection", "4.2", ")", "on", "a", "number", "of", "popular", "image", "data", "sets", "and", "their", "modifications", ".", "These", "include", "large", "-", "scale", "data", "sets", "of", "small", "images", "popular", "with", "deep", "learning", "methods", ",", "and", "the", "Office", "data", "sets", "[", "reference", "]", ",", "which", "are", "a", "de", "facto", "standard", "for", "domain", "adaptation", "in", "computer", "vision", ",", "but", "have", "much", "fewer", "images", ".", "section", ":", "Baselines", "The", "following", "baselines", "are", "evaluated", "in", "the", "experiments", "of", "this", "subsection", ".", "The", "source", "-", "only", "model", "is", "trained", "without", "consideration", "for", "target", "-", "domain", "data", "(", "no", "domain", "classifier", "branch", "included", "into", "the", "network", ")", ".", "The", "train", "-", "on", "-", "target", "model", "is", "trained", "on", "the", "target", "domain", "with", "class", "labels", "revealed", ".", "This", "model", "serves", "as", "an", "upper", "bound", "on", "DA", "methods", ",", "assuming", "that", "target", "data", "are", "abundant", "and", "the", "shift", "between", "the", "domains", "is", "considerable", ".", "In", "addition", ",", "we", "compare", "our", "approach", "against", "the", "recently", "proposed", "unsupervised", "DA", "method", "based", "on", "subspace", "alignment", "(", "SA", ")", "[", "reference", "]", ",", "which", "is", "simple", "to", "setup", "and", "test", "on", "new", "data", "sets", ",", "but", "has", "also", "been", "shown", "to", "perform", "very", "well", "in", "experimental", "comparisons", "with", "other", "\"", "shallow", "\"", "DA", "methods", ".", "To", "boost", "the", "performance", "of", "this", "baseline", ",", "we", "pick", "its", "most", "important", "free", "parameter", "(", "the", "number", "of", "principal", "components", ")", "from", "the", "range", "{", "2", ",", ".", ".", ".", ",", "60", "}", ",", "so", "that", "the", "test", "performance", "on", "the", "target", "domain", "is", "maximized", ".", "To", "apply", "SA", "in", "our", "setting", ",", "we", "train", "a", "source", "-", "only", "model", "and", "then", "consider", "the", "activations", "of", "the", "last", "hidden", "layer", "in", "the", "label", "predictor", "(", "before", "the", "final", "linear", "classifier", ")", "as", "descriptors", "/", "features", ",", "and", "learn", "the", "mapping", "between", "the", "source", "and", "the", "target", "domains", "[", "reference", "]", ".", "Since", "the", "SA", "baseline", "requires", "training", "a", "new", "classifier", "after", "adapting", "the", "features", ",", "and", "in", "order", "to", "put", "all", "the", "compared", "settings", "on", "an", "equal", "footing", ",", "we", "retrain", "the", "last", "layer", "of", "the", "label", "predictor", "using", "a", "standard", "linear", "SVM", "[", "reference", "]", "for", "all", "four", "considered", "methods", "(", "including", "ours", ";", "the", "performance", "on", "the", "target", "domain", "remains", "approximately", "the", "same", "after", "the", "retraining", ")", ".", "For", "the", "Office", "data", "set", "[", "reference", "]", ",", "we", "directly", "compare", "the", "performance", "of", "our", "full", "network", "(", "feature", "extractor", "and", "label", "predictor", ")", "against", "recent", "DA", "approaches", "using", "previously", "published", "results", ".", "section", ":", "CNN", "architectures", "and", "Training", "Procedure", "In", "general", ",", "we", "compose", "feature", "extractor", "from", "two", "or", "three", "convolutional", "layers", ",", "picking", "their", "exact", "configurations", "from", "previous", "works", ".", "More", "precisely", ",", "four", "different", "architectures", "were", "used", "in", "our", "experiments", ".", "The", "first", "three", "are", "shown", "in", "Figure", "4", ".", "For", "the", "Office", "domains", ",", "we", "use", "pre", "-", "trained", "AlexNet", "from", "the", "Caffe", "-", "package", ".", "The", "adaptation", "architecture", "is", "identical", "to", ".", "[", "reference", "]", "For", "the", "domain", "adaption", "component", ",", "we", "use", "three", "(", "x→1024→1024→2", ")", "fully", "connected", "layers", ",", "except", "for", "MNIST", "where", "we", "used", "a", "simpler", "(", "x→100→2", ")", "architecture", "to", "speed", "up", "the", "experiments", ".", "Admittedly", "these", "choices", "for", "domain", "classifier", "are", "arbitrary", ",", "and", "better", "adaptation", "performance", "might", "be", "attained", "if", "this", "part", "of", "the", "architecture", "is", "tuned", ".", "(", "x→1024→1024→2", ")", "For", "the", "loss", "functions", ",", "we", "set", "L", "y", "and", "L", "d", "to", "be", "the", "logistic", "regression", "loss", "and", "the", "binomial", "cross", "-", "entropy", "respectively", ".", "Following", "[", "reference", "]", "we", "also", "use", "dropout", "and", "2", "-", "norm", "restriction", "when", "we", "train", "the", "SVHN", "architecture", ".", "section", ":", "A", "2", "-", "layer", "domain", "classifier", "The", "other", "hyper", "-", "parameters", "are", "not", "selected", "through", "a", "grid", "search", "as", "in", "the", "small", "scale", "experiments", "of", "Section", "5.1", ",", "which", "would", "be", "computationally", "costly", ".", "Instead", ",", "the", "learning", "rate", "is", "adjusted", "during", "the", "stochastic", "gradient", "descent", "using", "the", "following", "formula", ":", "where", "p", "is", "the", "training", "progress", "linearly", "changing", "from", "0", "to", "1", ",", "µ", "0", "=", "0.01", ",", "α", "=", "10", "and", "β", "=", "0.75", "(", "the", "schedule", "was", "optimized", "to", "promote", "convergence", "and", "low", "error", "on", "the", "source", "domain", ")", ".", "A", "momentum", "term", "of", "0.9", "is", "also", "used", ".", "The", "domain", "adaptation", "parameter", "λ", "is", "initiated", "at", "0", "and", "is", "gradually", "changed", "to", "1", "using", "the", "following", "schedule", ":", "where", "γ", "was", "set", "to", "10", "in", "all", "experiments", "(", "the", "schedule", "was", "not", "optimized", "/", "tweaked", ")", ".", "This", "strategy", "allows", "the", "domain", "classifier", "to", "be", "less", "sensitive", "to", "noisy", "signal", "at", "the", "early", "stages", "of", "the", "training", "procedure", ".", "Note", "however", "that", "these", "λ", "p", "were", "used", "only", "for", "updating", "the", "feature", "extractor", "component", "G", "f", ".", "For", "updating", "the", "domain", "classification", "component", ",", "we", "used", "a", "fixed", "λ", "=", "1", ",", "to", "ensure", "that", "the", "latter", "trains", "as", "fast", "as", "the", "label", "predictor", "G", "y", ".", "[", "reference", "]", "Finally", ",", "note", "that", "the", "model", "is", "trained", "on", "128", "-", "sized", "batches", "(", "images", "are", "preprocessed", "by", "the", "mean", "subtraction", ")", ".", "A", "half", "of", "each", "batch", "is", "populated", "by", "the", "samples", "from", "the", "source", "domain", "(", "with", "known", "labels", ")", ",", "the", "rest", "constitutes", "the", "target", "domain", "(", "with", "labels", "not", "revealed", "to", "the", "algorithms", "except", "for", "the", "train", "-", "on", "-", "target", "baseline", ")", ".", "section", ":", "Visualizations", "We", "use", "t", "-", "SNE", "(", "van", "der", "Maaten", ",", "2013", ")", "projection", "to", "visualize", "feature", "distributions", "at", "different", "points", "of", "the", "network", ",", "while", "color", "-", "coding", "the", "domains", "(", "Figure", "5", ")", ".", "As", "we", "already", "observed", "with", "the", "shallow", "version", "of", "DANN", "(", "see", "Figure", "2", ")", ",", "there", "is", "a", "strong", "correspondence", "6", ".", "Equivalently", ",", "one", "can", "use", "the", "same", "λp", "for", "both", "feature", "extractor", "and", "domain", "classification", "components", ",", "but", "use", "a", "learning", "rate", "of", "µ", "/", "λp", "for", "the", "latter", ".", "between", "the", "success", "of", "the", "adaptation", "in", "terms", "of", "the", "classification", "accuracy", "for", "the", "target", "domain", ",", "and", "the", "overlap", "between", "the", "domain", "distributions", "in", "such", "visualizations", ".", "section", ":", "Results", "On", "Image", "Data", "Sets", "We", "now", "discuss", "the", "experimental", "settings", "and", "the", "results", ".", "In", "each", "case", ",", "we", "train", "on", "the", "source", "data", "set", "and", "test", "on", "a", "different", "target", "domain", "data", "set", ",", "with", "considerable", "shifts", "between", "domains", "(", "see", "Figure", "6", ")", ".", "The", "results", "are", "summarized", "in", "Table", "2", "and", "Table", "3", ".", "MNIST", "→", "MNIST", "-", "M.", "Our", "first", "experiment", "deals", "with", "the", "MNIST", "data", "set", "[", "reference", "]", ".", "In", "order", "to", "obtain", "the", "target", "domain", "(", "MNIST", "-", "M", ")", "we", "blend", "digits", "from", "the", "original", "set", "over", "patches", "randomly", "extracted", "from", "color", "photos", "from", "BSDS500", "[", "reference", "]", ".", "This", "operation", "is", "formally", "defined", "for", "two", "images", "I", "1", ",", "I", "2", "as", "I", "out", "ijk", "=", "|I", "1", "ijk", "−", "I", "2", "ijk", "|", ",", "where", "i", ",", "j", "are", "the", "coordinates", "of", "a", "pixel", "and", "k", "is", "a", "channel", "index", ".", "In", "other", "words", ",", "an", "output", "sample", "is", "produced", "by", "taking", "a", "patch", "from", "a", "photo", "and", "inverting", "its", "pixels", "at", "positions", "corresponding", "to", "the", "pixels", "of", "a", "digit", ".", "For", "a", "human", "the", "classification", "task", "becomes", "only", "slightly", "harder", "compared", "to", "the", "original", "data", "set", "(", "the", "digits", "are", "still", "clearly", "distinguishable", ")", "whereas", "for", "a", "CNN", "trained", "on", "MNIST", "this", "domain", "is", "quite", "distinct", ",", "as", "the", "background", "and", "the", "strokes", "are", "no", "longer", "constant", ".", "Consequently", ",", "the", "source", "-", "only", "model", "performs", "poorly", ".", "Our", "approach", "succeeded", "at", "aligning", "feature", "distributions", "(", "Figure", "5", ")", ",", "which", "led", "to", "successful", "adaptation", "results", "(", "considering", "that", "the", "adaptation", "is", "unsupervised", ")", ".", "At", "the", "same", "time", ",", "the", "improvement", "over", "source", "-", "only", "model", "achieved", "by", "subspace", "alignment", "(", "SA", ")", "[", "reference", "]", ")", "is", "quite", "modest", ",", "thus", "highlighting", "the", "difficulty", "of", "the", "adaptation", "task", ".", "Synthetic", "numbers", "→", "SVHN", ".", "To", "address", "a", "common", "scenario", "of", "training", "on", "synthetic", "data", "and", "testing", "on", "real", "data", ",", "we", "use", "Street", "-", "View", "House", "Number", "data", "set", "SVHN", "[", "reference", "]", "as", "the", "target", "domain", "and", "synthetic", "digits", "as", "the", "source", ".", "The", "latter", "(", "Syn", "Numbers", ")", "consists", "of", "≈", "500", ",", "000", "images", "generated", "by", "ourselves", "from", "Windows", "TM", "fonts", "by", "varying", "the", "text", "(", "that", "includes", "different", "one", "-", ",", "two", "-", ",", "and", "three", "-", "digit", "numbers", ")", ",", "positioning", ",", "orientation", ",", "background", "and", "stroke", "colors", ",", "and", "the", "amount", "of", "blur", ".", "The", "degrees", "of", "variation", "were", "chosen", "manually", "to", "simulate", "SVHN", ",", "however", "the", "two", "data", "sets", "are", "still", "rather", "distinct", ",", "the", "biggest", "difference", "being", "the", "structured", "clutter", "in", "the", "background", "of", "SVHN", "images", ".", "The", "proposed", "backpropagation", "-", "based", "technique", "works", "well", "covering", "almost", "80", "%", "of", "the", "gap", "between", "training", "with", "source", "data", "only", "and", "training", "on", "target", "domain", "data", "with", "known", "target", "labels", ".", "In", "contrast", ",", "SA", "[", "reference", "]", "results", "in", "a", "slight", "classification", "accuracy", "drop", "(", "probably", "due", "to", "the", "information", "loss", "during", "the", "dimensionality", "reduction", ")", ",", "indicating", "that", "the", "adaptation", "task", "is", "even", "more", "challenging", "than", "in", "the", "case", "of", "the", "MNIST", "experiment", ".", "MNIST", "↔", "SVHN", ".", "In", "this", "experiment", ",", "we", "further", "increase", "the", "gap", "between", "distributions", ",", "and", "test", "on", "MNIST", "and", "SVHN", ",", "which", "are", "significantly", "different", "in", "appearance", ".", "Training", "on", "SVHN", "even", "without", "adaptation", "is", "challenging", "-", "classification", "error", "stays", "high", "during", "the", "first", "150", "epochs", ".", "In", "order", "to", "avoid", "ending", "up", "in", "a", "poor", "local", "minimum", "we", ",", "therefore", ",", "do", "not", "use", "learning", "rate", "annealing", "here", ".", "Obviously", ",", "the", "two", "directions", "(", "MNIST", "→", "SVHN", "and", "SVHN", "→", "MNIST", ")", "are", "not", "equally", "difficult", ".", "As", "SVHN", "is", "more", "diverse", ",", "a", "model", "trained", "on", "SVHN", "is", "expected", "to", "be", "more", "generic", "and", "to", "perform", "reasonably", "on", "the", "MNIST", "data", "set", ".", "This", ",", "indeed", ",", "turns", "out", "to", "be", "the", "case", "and", "is", "supported", "by", "the", "appearance", "of", "the", "Table", "2", ":", "Classification", "accuracies", "for", "digit", "image", "classifications", "for", "different", "source", "and", "target", "domains", ".", "MNIST", "-", "M", "corresponds", "to", "difference", "-", "blended", "digits", "over", "nonuniform", "background", ".", "The", "first", "row", "corresponds", "to", "the", "lower", "performance", "bound", "(", "i.e.", ",", "if", "no", "adaptation", "is", "performed", ")", ".", "The", "last", "row", "corresponds", "to", "training", "on", "the", "target", "domain", "data", "with", "known", "class", "labels", "(", "upper", "bound", "on", "the", "DA", "performance", ")", ".", "For", "each", "of", "the", "two", "DA", "methods", "[", "reference", "]", "we", "show", "how", "much", "of", "the", "gap", "between", "the", "lower", "and", "the", "upper", "bounds", "was", "covered", "(", "in", "brackets", ")", ".", "For", "all", "five", "cases", ",", "our", "approach", "outperforms", "[", "reference", "]", "considerably", ",", "and", "covers", "a", "big", "portion", "of", "the", "gap", ".", "Table", "3", ":", "Accuracy", "evaluation", "of", "different", "DA", "approaches", "on", "the", "standard", "Office", "[", "reference", "]", ")", "data", "set", ".", "All", "methods", "(", "except", "SA", ")", "are", "evaluated", "in", "the", "\"", "fullytransductive", "\"", "protocol", "(", "some", "results", "are", "reproduced", "from", "[", "reference", "]", ".", "Our", "method", "(", "last", "row", ")", "outperforms", "competitors", "setting", "the", "new", "state", "-", "of", "-", "the", "-", "art", ".", "Syn", "and", "Real", "denote", "available", "labeled", "data", "(", "100", ",", "000", "synthetic", "and", "430", "real", "images", "respectively", ")", ";", "Adapted", "means", "that", "≈", "31", ",", "000", "unlabeled", "target", "domain", "images", "were", "used", "for", "adaptation", ".", "The", "best", "performance", "is", "achieved", "by", "employing", "both", "the", "labeled", "samples", "and", "the", "large", "unlabeled", "corpus", "in", "the", "target", "domain", ".", "feature", "distributions", ".", "We", "observe", "a", "quite", "strong", "separation", "between", "the", "domains", "when", "we", "feed", "them", "into", "the", "CNN", "trained", "solely", "on", "MNIST", ",", "whereas", "for", "the", "SVHN", "-", "trained", "network", "the", "features", "are", "much", "more", "intermixed", ".", "This", "difference", "probably", "explains", "why", "our", "method", "succeeded", "in", "improving", "the", "performance", "by", "adaptation", "in", "the", "SVHN", "→", "MNIST", "scenario", "(", "see", "Table", "2", ")", "but", "not", "in", "the", "opposite", "direction", "(", "SA", "is", "not", "able", "to", "perform", "adaptation", "in", "this", "case", "either", ")", ".", "Unsupervised", "adaptation", "from", "MNIST", "to", "SVHN", "gives", "a", "failure", "example", "for", "our", "approach", ":", "it", "does", "n't", "manage", "to", "improve", "upon", "the", "performance", "of", "the", "non", "-", "adapted", "model", "which", "achieves", "≈", "0.25", "accuracy", "(", "we", "are", "unaware", "of", "any", "unsupervised", "DA", "methods", "capable", "of", "performing", "such", "adaptation", ")", ".", "Synthetic", "Signs", "→", "GTSRB", ".", "Overall", ",", "this", "setting", "is", "similar", "to", "the", "Syn", "Numbers", "→", "SVHN", "experiment", ",", "except", "the", "distribution", "of", "the", "features", "is", "more", "complex", "due", "to", "the", "significantly", "larger", "number", "of", "classes", "(", "43", "instead", "of", "10", ")", ".", "For", "the", "source", "domain", "we", "obtained", "100", ",", "000", "synthetic", "images", "(", "which", "we", "call", "Syn", "Signs", ")", "simulating", "various", "imaging", "conditions", ".", "In", "the", "target", "domain", ",", "we", "use", "31", ",", "367", "random", "training", "samples", "for", "unsupervised", "adaptation", "and", "the", "rest", "for", "evaluation", ".", "Once", "again", ",", "our", "method", "achieves", "a", "sensible", "increase", "in", "performance", "proving", "its", "suitability", "for", "the", "synthetic", "-", "to", "-", "real", "data", "adaptation", ".", "As", "an", "additional", "experiment", ",", "we", "also", "evaluate", "the", "proposed", "algorithm", "for", "semi", "-", "supervised", "domain", "adaptation", ",", "i.e.", ",", "when", "one", "is", "additionally", "provided", "with", "a", "small", "amount", "of", "labeled", "target", "data", ".", "Here", ",", "we", "reveal", "430", "labeled", "examples", "(", "10", "samples", "per", "class", ")", "and", "add", "them", "to", "the", "training", "set", "for", "the", "label", "predictor", ".", "Figure", "7", "shows", "the", "change", "of", "the", "validation", "error", "throughout", "the", "training", ".", "While", "the", "graph", "clearly", "suggests", "that", "our", "method", "can", "be", "beneficial", "in", "the", "semi", "-", "supervised", "setting", ",", "thorough", "verification", "of", "semi", "-", "supervised", "setting", "is", "left", "for", "future", "work", ".", "Office", "data", "set", ".", "We", "finally", "evaluate", "our", "method", "on", "Office", "data", "set", ",", "which", "is", "a", "collection", "of", "three", "distinct", "domains", ":", "Amazon", ",", "DSLR", ",", "and", "Webcam", ".", "Unlike", "previously", "discussed", "data", "sets", ",", "Office", "is", "rather", "small", "-", "scale", "with", "only", "2817", "labeled", "images", "spread", "across", "31", "different", "categories", "in", "the", "largest", "domain", ".", "The", "amount", "of", "available", "data", "is", "crucial", "for", "a", "successful", "training", "of", "a", "deep", "model", ",", "hence", "we", "opted", "for", "the", "fine", "-", "tuning", "of", "the", "CNN", "pre", "-", "trained", "on", "the", "ImageNet", "(", "AlexNet", "from", "the", "Caffe", "package", ",", "see", "as", "it", "is", "done", "in", "some", "recent", "DA", "works", "[", "reference", "][", "reference", "]", ".", "We", "make", "our", "approach", "more", "comparable", "with", "by", "using", "exactly", "the", "same", "network", "architecture", "replacing", "domain", "mean", "-", "based", "regularization", "with", "the", "domain", "classifier", ".", "Following", "previous", "works", ",", "we", "assess", "the", "performance", "of", "our", "method", "across", "three", "transfer", "tasks", "most", "commonly", "used", "for", "evaluation", ".", "Our", "training", "protocol", "is", "adopted", "from", ";", "[", "reference", "]", ";", "[", "reference", "]", "as", "during", "adaptation", "we", "use", "all", "available", "labeled", "source", "examples", "and", "unlabeled", "target", "examples", "(", "the", "premise", "of", "our", "method", "is", "the", "abundance", "of", "unlabeled", "data", "in", "the", "target", "domain", ")", ".", "Also", ",", "all", "source", "domain", "data", "are", "used", "for", "training", ".", "Under", "this", "\"", "fully", "-", "transductive", "\"", "setting", ",", "our", "method", "is", "able", "to", "improve", "previously", "-", "reported", "state", "-", "of", "-", "the", "-", "art", "accuracy", "for", "unsupervised", "adaptation", "very", "considerably", "(", "Table", "3", ")", ",", "especially", "in", "the", "most", "challenging", "Amazon", "→", "Webcam", "scenario", "(", "the", "two", "domains", "with", "the", "largest", "domain", "shift", ")", ".", "Interestingly", ",", "in", "all", "three", "experiments", "we", "observe", "a", "slight", "over", "-", "fitting", "(", "performance", "on", "the", "target", "domain", "degrades", "while", "accuracy", "on", "the", "source", "continues", "to", "improve", ")", "as", "training", "progresses", ",", "however", ",", "it", "does", "n't", "ruin", "the", "validation", "accuracy", ".", "Moreover", ",", "switching", "off", "the", "domain", "classifier", "branch", "makes", "this", "effect", "far", "more", "apparent", ",", "from", "which", "we", "conclude", "that", "our", "technique", "serves", "as", "a", "regularizer", ".", "section", ":", "Experiments", "with", "Deep", "Image", "Descriptors", "for", "Re", "-", "Identification", "In", "this", "section", "we", "discuss", "the", "application", "of", "the", "described", "adaptation", "method", "to", "person", "re", "-", "identification", "(", "re", "-", "i", "d", ")", "problem", ".", "The", "task", "of", "person", "re", "-", "identification", "is", "to", "associate", "people", "seen", "from", "different", "camera", "views", ".", "More", "formally", ",", "it", "can", "be", "defined", "as", "follows", ":", "given", "two", "sets", "of", "images", "from", "different", "cameras", "(", "probe", "and", "gallery", ")", "such", "that", "each", "person", "depicted", "in", "the", "probe", "set", "has", "an", "image", "in", "the", "gallery", "set", ",", "for", "each", "image", "of", "a", "person", "from", "the", "probe", "set", "find", "an", "image", "of", "the", "same", "person", "in", "the", "gallery", "set", ".", "Disjoint", "camera", "views", ",", "different", "illumination", "conditions", ",", "various", "poses", "and", "low", "quality", "of", "data", "make", "this", "problem", "difficult", "even", "for", "humans", "(", "e.g.", ",", "[", "reference", "]", ",", "reports", "human", "performance", "at", "Rank1=71.08", "%", ")", ".", "Unlike", "classification", "problems", "that", "are", "discussed", "above", ",", "re", "-", "identification", "problem", "implies", "that", "each", "image", "is", "mapped", "to", "a", "vector", "descriptor", ".", "The", "distance", "between", "descriptors", "is", "then", "used", "to", "match", "images", "from", "the", "probe", "set", "and", "the", "gallery", "set", ".", "To", "evaluate", "results", "of", "re", "-", "i", "d", "methods", "the", "Cumulative", "Match", "Characteristic", "(", "CMC", ")", "curve", "is", "commonly", "used", ".", "It", "is", "a", "plot", "of", "the", "identification", "rate", "(", "recall", ")", "at", "rank", "-", "k", ",", "that", "is", "the", "probability", "of", "the", "matching", "gallery", "image", "to", "be", "within", "the", "closest", "k", "images", "(", "in", "terms", "of", "descriptor", "distance", ")", "to", "the", "probe", "image", ".", "Most", "existing", "works", "train", "descriptor", "mappings", "and", "evaluate", "them", "within", "the", "same", "data", "set", "containing", "images", "from", "a", "certain", "camera", "network", "with", "similar", "imaging", "conditions", ".", "Several", "papers", ",", "however", ",", "observed", "that", "the", "performance", "of", "the", "resulting", "re", "-", "identification", "systems", "drops", "very", "considerably", "when", "descriptors", "trained", "on", "one", "data", "set", "and", "tested", "on", "another", ".", "It", "is", "therefore", "natural", "to", "handle", "such", "cross", "-", "domain", "evaluation", "as", "a", "domain", "-", "adaptation", "problem", ",", "where", "each", "camera", "network", "(", "data", "set", ")", "constitutes", "a", "domain", ".", "Recently", ",", "several", "papers", "with", "significantly", "improved", "re", "-", "identification", "performance", "[", "reference", "][", "reference", "][", "reference", "]", "have", "been", "presented", ",", "with", "[", "reference", "]", "reporting", "good", "results", "in", "cross", "-", "data", "-", "set", "evaluation", "scenario", ".", "At", "the", "moment", ",", "deep", "learning", "methods", "[", "reference", "]", "do", "not", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", "probably", "because", "of", "the", "limited", "size", "of", "the", "training", "sets", ".", "Domain", "adaptation", "thus", "represents", "a", "viable", "direction", "for", "improving", "deep", "re", "-", "identification", "descriptors", ".", "section", ":", "Data", "Sets", "and", "Protocols", "Following", "[", "reference", "]", ",", "we", "use", "PRID", "[", "reference", "]", ",", "VIPeR", "[", "reference", "]", ",", "CUHK", "[", "reference", "]", "as", "target", "data", "sets", "for", "our", "experiments", ".", "The", "PRID", "data", "set", "exists", "in", "two", "versions", ",", "and", "as", "in", "[", "reference", "]", "we", "use", "a", "single", "-", "shot", "variant", ".", "It", "contains", "images", "of", "385", "persons", "viewed", "from", "camera", "A", "and", "images", "of", "749", "persons", "viewed", "from", "camera", "B", ",", "200", "persons", "appear", "in", "both", "cameras", ".", "The", "VIPeR", "data", "set", "also", "contains", "images", "taken", "with", "two", "cameras", ",", "and", "in", "total", "632", "persons", "are", "captured", ",", "for", "every", "person", "there", "is", "one", "image", "for", "each", "of", "the", "two", "camera", "views", ".", "The", "CUHK", "data", "set", "consists", "of", "images", "from", "five", "pairs", "of", "cameras", ",", "two", "images", "for", "each", "person", "from", "each", "of", "the", "two", "cameras", ".", "We", "refer", "to", "the", "subset", "of", "this", "data", "set", "that", "includes", "the", "first", "pair", "of", "cameras", "only", "as", "CUHK", "/", "p1", "(", "as", "most", "papers", "use", "this", "subset", ")", ".", "See", "Figure", "8", "for", "samples", "of", "these", "data", "sets", ".", "We", "perform", "extensive", "experiments", "for", "various", "pairs", "of", "data", "sets", ",", "where", "one", "data", "set", "serves", "as", "a", "source", "domain", ",", "i.e.", ",", "it", "is", "used", "to", "train", "a", "descriptor", "mapping", "in", "a", "supervised", "way", "with", "known", "correspondences", "between", "probe", "and", "gallery", "images", ".", "The", "second", "data", "set", "is", "used", "as", "a", "target", "domain", ",", "so", "that", "images", "from", "that", "data", "set", "are", "used", "without", "probe", "-", "gallery", "correspondence", ".", "In", "more", "detail", ",", "CUHK", "/", "p1", "is", "used", "for", "experiments", "when", "CUHK", "serves", "as", "a", "target", "domain", "and", "two", "settings", "(", "\"", "whole", "CUHK", "\"", "and", "CUHK", "/", "p1", ")", "are", "used", "for", "experiments", "when", "CUHK", "serves", "as", "a", "source", "domain", ".", "Given", "PRID", "as", "a", "target", "data", "set", ",", "we", "randomly", "choose", "100", "persons", "appearing", "in", "both", "camera", "views", "as", "training", "set", ".", "The", "images", "of", "the", "other", "100", "persons", "from", "camera", "A", "are", "used", "as", "probe", ",", "all", "images", "from", "camera", "B", "excluding", "those", "used", "in", "training", "(", "649", "in", "total", ")", "are", "used", "as", "gallery", "at", "test", "time", ".", "For", "VIPeR", ",", "we", "use", "random", "316", "persons", "for", "training", "and", "all", "others", "for", "testing", ".", "For", "CUHK", ",", "971", "persons", "are", "split", "into", "485", "for", "training", "and", "486", "for", "testing", ".", "Unlike", "[", "reference", "]", ",", "we", "use", "all", "images", "in", "the", "first", "pair", "of", "cameras", "of", "CUHK", "instead", "of", "choosing", "one", "image", "of", "a", "person", "from", "each", "camera", "view", ".", "We", "also", "performed", "two", "experiments", "with", "all", "images", "of", "the", "whole", "CUHK", "data", "set", "as", "source", "domain", "and", "VIPeR", "and", "PRID", "data", "sets", "as", "target", "domains", "as", "in", "the", "original", "paper", "[", "reference", "]", ".", "Following", "[", "reference", "]", ",", "we", "augmented", "our", "data", "with", "mirror", "images", ",", "and", "during", "test", "time", "we", "calculate", "similarity", "score", "between", "two", "images", "as", "the", "mean", "of", "the", "four", "scores", "corresponding", "to", "different", "flips", "of", "the", "two", "compared", "images", ".", "In", "case", "of", "CUHK", ",", "where", "there", "are", "4", "images", "(", "including", "mirror", "images", ")", "for", "each", "of", "the", "two", "camera", "views", "for", "each", "person", ",", "all", "16", "combinations", "'", "scores", "are", "averaged", ".", "section", ":", "CNN", "architectures", "and", "Training", "Procedure", "In", "our", "experiments", ",", "we", "use", "siamese", "architecture", "described", "in", "[", "reference", "]", "(", "Deep", "Metric", "Learning", "or", "DML", ")", "for", "learning", "deep", "image", "descriptors", "on", "the", "source", "data", "set", ".", "This", "architecture", "incorporates", "two", "convolution", "layers", "(", "with", "7", "×", "7", "and", "5", "×", "5", "filter", "banks", ")", ",", "followed", "by", "ReLU", "and", "max", "pooling", ",", "and", "one", "fully", "-", "connected", "layer", ",", "which", "gives", "500", "-", "dimensional", "descriptors", "as", "an", "output", ".", "There", "are", "three", "parallel", "flows", "within", "the", "CNN", "for", "processing", "three", "part", "of", "an", "image", ":", "the", "upper", ",", "the", "middle", ",", "and", "the", "lower", "one", ".", "The", "first", "convolution", "layer", "shares", "parameters", "between", "three", "parts", ",", "and", "the", "outputs", "of", "the", "second", "convolution", "layers", "are", "concatenated", ".", "During", "training", ",", "we", "follow", "[", "reference", "]", "and", "calculate", "pairwise", "cosine", "similarities", "between", "500", "-", "dimensional", "features", "within", "each", "batch", "and", "backpropagate", "the", "loss", "for", "all", "pairs", "within", "batch", ".", "To", "perform", "domain", "-", "adversarial", "training", ",", "we", "construct", "a", "DANN", "architecture", ".", "The", "feature", "extractor", "includes", "the", "two", "convolutional", "layers", "(", "followed", "by", "max", "-", "pooling", "and", "ReLU", ")", "discussed", "above", ".", "The", "label", "predictor", "in", "this", "case", "is", "replaced", "with", "descriptor", "predictor", "that", "includes", "one", "fully", "-", "connected", "layer", ".", "The", "domain", "classifier", "includes", "two", "fully", "-", "connected", "layers", "with", "500", "units", "in", "the", "intermediate", "representation", "(", "x→500→1", ")", ".", "For", "the", "verification", "loss", "function", "in", "the", "descriptor", "predictor", "we", "used", "Binomial", "Deviance", "loss", ",", "defined", "in", "[", "reference", "]", "with", "similar", "parameters", ":", "α", "=", "2", ",", "β", "=", "0.5", ",", "c", "=", "2", "(", "the", "asymmetric", "cost", "parameter", "for", "negative", "pairs", ")", ".", "The", "domain", "classifier", "is", "trained", "with", "logistic", "loss", "as", "in", "subsection", "5.2.2", ".", "We", "used", "learning", "rate", "fixed", "to", "0.001", "and", "momentum", "of", "0.9", ".", "The", "schedule", "of", "adaptation", "similar", "to", "the", "one", "described", "in", "subsection", "5.2.2", "was", "used", ".", "We", "also", "inserted", "dropout", "layer", "with", "rate", "0.5", "after", "the", "concatenation", "of", "outputs", "of", "the", "second", "max", "-", "pooling", "layer", ".", "128", "-", "sized", "batches", "were", "used", "for", "source", "data", "and", "128", "-", "sized", "batches", "for", "target", "data", ".", "Figure", "9", "shows", "results", "in", "the", "form", "of", "CMC", "-", "curves", "for", "eight", "pairs", "of", "data", "sets", ".", "Depending", "on", "the", "hardness", "of", "the", "annotation", "problem", "we", "trained", "either", "for", "50", ",", "000", "iterations", "(", "CUHK", "/", "p1", "→", "VIPeR", ",", "VIPeR", "→", "CUHK", "/", "p1", ",", "PRID", "→", "VIPeR", ")", "or", "for", "20", ",", "000", "iterations", "(", "the", "other", "five", "pairs", ")", ".", "section", ":", "Results", "on", "Re", "-", "identification", "data", "sets", "After", "the", "sufficient", "number", "of", "iterations", ",", "domain", "-", "adversarial", "training", "consistently", "improves", "the", "performance", "of", "re", "-", "identification", ".", "For", "the", "pairs", "that", "involve", "PRID", "data", "set", ",", "which", "is", "more", "dissimilar", "to", "the", "other", "two", "data", "sets", ",", "the", "improvement", "is", "considerable", ".", "Overall", ",", "this", "demonstrates", "the", "applicability", "of", "the", "domain", "-", "adversarial", "learning", "beyond", "classification", "problems", ".", "section", ":", "Conclusion", "The", "paper", "proposes", "a", "new", "approach", "to", "domain", "adaptation", "of", "feed", "-", "forward", "neural", "networks", ",", "which", "allows", "large", "-", "scale", "training", "based", "on", "large", "amount", "of", "annotated", "data", "in", "the", "source", "domain", "and", "large", "amount", "of", "unannotated", "data", "in", "the", "target", "domain", ".", "Similarly", "to", "many", "previous", "shallow", "and", "deep", "DA", "techniques", ",", "the", "adaptation", "is", "achieved", "through", "aligning", "the", "distributions", "of", "features", "across", "the", "two", "domains", ".", "However", ",", "unlike", "previous", "approaches", ",", "the", "alignment", "is", "accomplished", "through", "standard", "backpropagation", "training", ".", "The", "approach", "is", "motivated", "and", "supported", "by", "the", "domain", "adaptation", "theory", "of", "[", "reference", "]", ".", "The", "main", "idea", "behind", "DANN", "is", "to", "enjoin", "the", "network", "hidden", "layer", "to", "learn", "a", "representation", "which", "is", "predictive", "of", "the", "source", "example", "labels", ",", "but", "uninformative", "about", "the", "domain", "of", "the", "input", "(", "source", "or", "target", ")", ".", "We", "implement", "this", "new", "approach", "within", "both", "shallow", "and", "deep", "feed", "-", "forward", "architectures", ".", "The", "latter", "allows", "simple", "implementation", "within", "virtually", "any", "deep", "learning", "package", "through", "the", "introduction", "of", "a", "simple", "gradient", "reversal", "layer", ".", "We", "have", "shown", "that", "our", "approach", "is", "flexible", "and", "achieves", "state", "-", "of", "-", "the", "-", "art", "results", "on", "a", "variety", "of", "benchmark", "in", "domain", "adaptation", ",", "namely", "for", "sentiment", "analysis", "and", "image", "classification", "tasks", ".", "A", "convenient", "aspect", "of", "our", "approach", "is", "that", "the", "domain", "adaptation", "component", "can", "be", "added", "to", "almost", "any", "neural", "network", "architecture", "that", "is", "trainable", "with", "backpropagation", ".", "Towards", "this", "end", ",", "We", "have", "demonstrated", "experimentally", "that", "the", "approach", "is", "not", "confined", "to", "classification", "tasks", "but", "can", "be", "used", "in", "other", "feed", "-", "forward", "architectures", ",", "e.g.", ",", "for", "descriptor", "learning", "for", "person", "re", "-", "identification", ".", "section", ":", "section", ":", "Acknowledgments", "This", "work", "has", "been", "supported", "by", "National", "Science", "and", "Engineering", "Research", "Council", "(", "NSERC", ")", "Discovery", "grants", "262067", "and", "0122405", "as", "well", "as", "the", "Russian", "Ministry", "of", "Science", "and", "Education", "grant", "RFMEFI57914X0071", ".", "Computations", "were", "performed", "on", "the", "Colosse", "supercomputer", "grid", "at", "Université", "Laval", ",", "under", "the", "auspices", "of", "Calcul", "Québec", "and", "Compute", "Canada", ".", "The", "operations", "of", "Colosse", "are", "funded", "by", "the", "NSERC", ",", "the", "Canada", "Foundation", "for", "Innovation", "(", "CFI", ")", ",", "NanoQuébec", ",", "and", "the", "Fonds", "de", "recherche", "du", "Québec", "-", "Nature", "et", "technologies", "(", "FRQNT", ")", ".", "We", "also", "thank", "the", "Graphics", "&", "Media", "Lab", ",", "Faculty", "of", "Computational", "Mathematics", "and", "Cybernetics", ",", "Lomonosov", "Moscow", "State", "University", "for", "providing", "the", "synthetic", "road", "signs", "data", "set", ".", "section", ":"], "entity": [[0, 7, ["Domain", "-", "Adversarial", "Training", "of", "Neural", "Networks"], "Method"], [14, 17, ["representation", "learning", "approach"], "Method"], [18, 20, ["domain", "adaptation"], "Method"], [45, 47, ["domain", "adaptation"], "Method"], [52, 54, ["domain", "transfer"], "Task"], [58, 59, ["predictions"], "Task"], [91, 94, ["neural", "network", "architectures"], "Method"], [143, 145, ["learning", "task"], "Task"], [175, 179, ["feed", "-", "forward", "model"], "Method"], [184, 186, ["standard", "layers"], "Method"], [189, 192, ["gradient", "reversal", "layer"], "Method"], [195, 197, ["augmented", "architecture"], "Method"], [201, 203, ["standard", "backpropagation"], "Method"], [204, 207, ["stochastic", "gradient", "descent"], "Method"], [220, 223, ["deep", "learning", "packages"], "Method"], [234, 236, ["classification", "problems"], "Task"], [237, 240, ["document", "sentiment", "analysis"], "Task"], [241, 243, ["image", "classification"], "Task"], [253, 255, ["domain", "adaptation"], "Method"], [268, 271, ["descriptor", "learning", "task"], "Task"], [275, 280, ["person", "re", "-", "identification", "application"], "Task"], [293, 296, ["machine", "learning", "task"], "Task"], [302, 305, ["machine", "learning", "methods"], "Method"], [319, 323, ["deep", "neural", "network", "architectures"], "Method"], [344, 348, ["machine", "-", "learning", "tasks"], "Task"], [372, 377, ["large", "-", "scale", "deep", "models"], "Method"], [404, 406, ["image", "classifier"], "Method"], [453, 455, ["sentiment", "analysis"], "Task"], [456, 458, ["written", "reviews"], "Material"], [496, 498, ["discriminative", "classifier"], "Method"], [515, 517, ["domain", "adaptation"], "Method"], [518, 519, ["DA"], "Method"], [547, 548, ["classifier"], "Method"], [575, 578, ["domain", "adaptation", "approaches"], "Method"], [600, 603, ["unsupervised", "domain", "annotation"], "Task"], [624, 626, ["unsupervised", "case"], "Task"], [632, 636, ["domain", "-", "adversarial", "learning"], "Method"], [642, 646, ["semi", "-", "supervised", "case"], "Task"], [654, 656, ["domain", "adaptation"], "Method"], [659, 662, ["fixed", "feature", "representations"], "Method"], [667, 669, ["domain", "adaptation"], "Method"], [670, 673, ["deep", "feature", "learning"], "Method"], [683, 685, ["domain", "adaptation"], "Method"], [689, 691, ["learning", "representation"], "Task"], [696, 698, ["classification", "decisions"], "Task"], [738, 742, ["feed", "-", "forward", "network"], "Method"], [768, 770, ["domain", "adaptation"], "Method"], [781, 785, ["cross", "-", "domain", "transfer"], "Task"], [836, 838, ["discriminative", "classifiers"], "Method"], [847, 849, ["label", "predictor"], "Method"], [868, 870, ["domain", "classifier"], "Method"], [887, 888, ["classifiers"], "Method"], [906, 909, ["deep", "feature", "mapping"], "Method"], [919, 921, ["label", "classifier"], "Method"], [928, 930, ["domain", "classifier"], "Method"], [939, 941, ["domain", "classifier"], "Method"], [956, 957, ["optimization"], "Task"], [965, 967, ["training", "processes"], "Method"], [974, 979, ["deep", "feed", "-", "forward", "network"], "Method"], [981, 986, ["domain", "-", "adversarial", "neural", "network"], "Method"], [987, 988, ["DANN"], "Method"], [1016, 1019, ["stochastic", "gradient", "descent"], "Method"], [1025, 1028, ["SGD", "with", "momentum"], "Method"], [1045, 1049, ["feed", "-", "forward", "architecture"], "Method"], [1053, 1054, ["backpropagation"], "Method"], [1072, 1075, ["gradient", "reversal", "layer"], "Method"], [1081, 1083, ["forward", "propagation"], "Method"], [1106, 1111, ["domain", "-", "adversarial", "learning", "idea"], "Method"], [1115, 1117, ["deep", "architectures"], "Method"], [1132, 1134, ["label", "predictor"], "Method"], [1135, 1137, ["domain", "classifier"], "Method"], [1138, 1140, ["feature", "extractor"], "Method"], [1149, 1153, ["domain", "-", "adversarial", "learning"], "Method"], [1169, 1172, ["sentiment", "analysis", "problem"], "Task"], [1173, 1176, ["natural", "language", "processing"], "Task"], [1178, 1179, ["DANN"], "Method"], [1188, 1191, ["marginalized", "Stacked", "Autoencoders"], "Method"], [1192, 1193, ["mSDA"], "Method"], [1200, 1204, ["common", "Amazon", "reviews", "benchmark"], "Material"], [1213, 1216, ["image", "classification", "task"], "Task"], [1250, 1254, ["domain", "-", "adversarial", "learning"], "Method"], [1257, 1259, ["deep", "architecture"], "Method"], [1271, 1272, ["accuracy"], "Metric"], [1277, 1282, ["domain", "-", "adversarial", "descriptor", "learning"], "Method"], [1286, 1291, ["person", "re", "-", "identification", "application"], "Task"], [1302, 1305, ["pedestrian", "image", "descriptors"], "Task"], [1309, 1312, ["retrieval", "and", "verification"], "Task"], [1315, 1317, ["domainadversarial", "learning"], "Method"], [1322, 1324, ["descriptor", "predictor"], "Method"], [1327, 1331, ["Siamese", "-", "like", "loss"], "Method"], [1334, 1336, ["label", "predictor"], "Method"], [1339, 1341, ["classification", "loss"], "Method"], [1351, 1355, ["domain", "-", "adversarial", "learning"], "Method"], [1357, 1365, ["cross", "-", "data", "-", "set", "re", "-", "identification"], "Task"], [1376, 1378, ["domain", "adaptation"], "Method"], [1416, 1420, ["non", "-", "linear", "representations"], "Method"], [1426, 1429, ["neural", "network", "representations"], "Method"], [1443, 1444, ["mSDA"], "Method"], [1458, 1460, ["robust", "representations"], "Method"], [1464, 1467, ["denoising", "autoencoder", "paradigm"], "Method"], [1519, 1522, ["feature", "space", "transformation"], "Method"], [1545, 1548, ["distribution", "matching", "approach"], "Method"], [1616, 1618, ["feature", "representation"], "Method"], [1624, 1626, ["geometric", "transformation"], "Method"], [1648, 1653, ["deep", "discriminatively", "-", "trained", "classifier"], "Method"], [1697, 1699, ["layerwise", "training"], "Method"], [1703, 1705, ["deep", "autoencoders"], "Method"], [1734, 1736, ["deep", "autoencoder"], "Method"], [1746, 1749, ["classifier", "/", "predictor"], "Method"], [1757, 1759, ["feature", "representation"], "Method"], [1761, 1764, ["autoencoder", "(", "s"], "Method"], [1778, 1780, ["feature", "learning"], "Task"], [1781, 1783, ["domain", "adaptation"], "Method"], [1784, 1786, ["classifier", "learning"], "Task"], [1790, 1792, ["unified", "architecture"], "Method"], [1797, 1799, ["learning", "algorithm"], "Method"], [1800, 1801, ["backpropagation"], "Method"], [1865, 1870, ["deep", "feed", "-", "forward", "architectures"], "Method"], [1944, 1947, ["generative", "deep", "networks"], "Method"], [2014, 2015, ["training"], "Task"], [2073, 2075, ["domain", "adaptation"], "Method"], [2076, 2080, ["feed", "-", "forward", "networks"], "Method"], [2089, 2090, ["distance"], "Metric"], [2104, 2105, ["RKHS"], "Method"], [2124, 2126, ["discriminative", "classifier"], "Method"], [2190, 2191, ["DANN"], "Method"], [2212, 2214, ["HMM", "representations"], "Method"], [2217, 2219, ["word", "tagging"], "Task"], [2221, 2223, ["posterior", "regularizer"], "Method"], [2253, 2257, ["word", "tagging", "problems", "-"], "Task"], [2263, 2265, ["learning", "objective"], "Metric"], [2328, 2334, ["Transfer", "and", "Multi", "-", "Task", "Learning"], "Task"], [2368, 2372, ["natural", "language", "processing", "task"], "Task"], [2373, 2375, ["sentiment", "analysis"], "Task"], [2385, 2386, ["classification"], "Task"], [2388, 2392, ["domain", "-", "adversarial", "learning"], "Method"], [2393, 2396, ["descriptor", "learning", "setting"], "Task"], [2398, 2403, ["person", "re", "-", "identification", "application"], "Task"], [2406, 2408, ["Domain", "Adaptation"], "Task"], [2410, 2412, ["classification", "tasks"], "Task"], [2464, 2469, ["unsupervised", "domain", "adaptation", "learning", "algorithm"], "Method"], [2507, 2509, ["learning", "algorithm"], "Method"], [2513, 2515, ["classifier", "η"], "Method"], [2551, 2553, ["target", "error"], "Metric"], [2558, 2560, ["source", "error"], "Metric"], [2584, 2586, ["source", "risk"], "Metric"], [2611, 2613, ["domain", "adaptation"], "Method"], [2802, 2804, ["linear", "classifiers"], "Method"], [2816, 2818, ["learning", "algorithm"], "Method"], [2866, 2867, ["classifier"], "Method"], [2888, 2890, ["generalization", "error"], "Metric"], [2961, 2964, ["H", "-", "divergence"], "Metric"], [2981, 2983, ["PAD", "value"], "Metric"], [3003, 3006, ["deeper", "MLP", "classifier"], "Method"], [3021, 3023, ["classifier", "error"], "Metric"], [3058, 3060, ["Target", "Risk"], "Metric"], [3124, 3127, ["empirical", "source", "risk"], "Metric"], [3158, 3159, ["classifier"], "Method"], [3178, 3179, ["classifier"], "Method"], [3198, 3200, ["learning", "algorithm"], "Method"], [3213, 3216, ["source", "risk", "R"], "Metric"], [3222, 3226, ["empirical", "H", "-", "divergenced"], "Metric"], [3314, 3319, ["Domain", "-", "Adversarial", "Neural", "Networks"], "Method"], [3320, 3321, ["DANN"], "Method"], [3340, 3343, ["neural", "network", "classifier"], "Method"], [3365, 3367, ["internal", "representation"], "Method"], [3369, 3371, ["neural", "network"], "Method"], [3413, 3416, ["domain", "adaptation", "component"], "Method"], [3418, 3420, ["neural", "networks"], "Method"], [3441, 3448, ["hidden", "layer", ",", "fully", "connected", "neural", "network"], "Method"], [3460, 3464, ["deep", ")", "network", "architectures"], "Method"], [3471, 3474, ["Shallow", "Neural", "Network"], "Method"], [3489, 3491, ["hidden", "layer"], "Method"], [3567, 3570, ["sigm", "(", "a"], "Method"], [3575, 3579, ["prediction", "layer", "G", "y"], "Method"], [3592, 3594, ["softmax", "function"], "Method"], [3614, 3616, ["neural", "network"], "Method"], [3642, 3644, ["classification", "loss"], "Metric"], [3659, 3661, ["neural", "network"], "Method"], [3666, 3668, ["optimization", "problem"], "Task"], [3701, 3703, ["optional", "regularizer"], "Method"], [3721, 3723, ["domain", "regularizer"], "Method"], [3757, 3759, ["internal", "representation"], "Method"], [3761, 3763, ["neural", "network"], "Method"], [3769, 3772, ["source", "sample", "representations"], "Method"], [3796, 3800, ["empirical", "H", "-", "divergence"], "Metric"], [3856, 3859, ["domain", "classification", "layer"], "Method"], [3864, 3866, ["logistic", "regressor"], "Method"], [3909, 3911, ["domain", "regressor"], "Method"], [4010, 4013, ["domain", "adaptation", "term"], "Method"], [4025, 4026, ["regularizer"], "Method"], [4029, 4030, ["regularizer"], "Method"], [4077, 4079, ["optimization", "problem"], "Task"], [4099, 4103, ["source", "risk", "R", "S"], "Metric"], [4136, 4138, ["learning", "process"], "Method"], [4140, 4141, ["learning"], "Task"], [4151, 4153, ["optimization", "objective"], "Metric"], [4187, 4189, ["optimization", "problem"], "Task"], [4191, 4192, ["minimization"], "Task"], [4202, 4203, ["maximization"], "Task"], [4211, 4217, ["Shallow", "DANN", "-", "Stochastic", "training", "update"], "Method"], [4218, 4220, ["stopping", "criterion"], "Metric"], [4236, 4238, ["Forward", "propagation"], "Method"], [4244, 4247, ["neural", "network", "parameters"], "Method"], [4306, 4309, ["stochastic", "gradient", "procedure"], "Method"], [4344, 4346, ["Stochastic", "estimates"], "Method"], [4374, 4376, ["learning", "procedure"], "Method"], [4382, 4383, ["training"], "Task"], [4385, 4387, ["neural", "network"], "Method"], [4400, 4402, ["domain", "regressor"], "Method"], [4443, 4448, ["Domain", "-", "Adversarial", "Neural", "Networks"], "Method"], [4449, 4450, ["DANN"], "Method"], [4459, 4463, ["hidden", "layer", "G", "f"], "Method"], [4500, 4503, ["domain", "regressor", "G"], "Method"], [4572, 4576, ["deep", "convolutional", "neural", "networks"], "Method"], [4590, 4595, ["learning", "discriminative", "features", "of", "images"], "Task"], [4627, 4634, ["D", "-", "dimensional", "neural", "network", "feature", "extractor"], "Method"], [4655, 4656, ["DANN"], "Method"], [4718, 4721, ["domain", "prediction", "component"], "Method"], [4733, 4736, ["label", "prediction", "component"], "Method"], [4804, 4806, ["gradient", "updates"], "Method"], [4811, 4813, ["learning", "rate"], "Metric"], [4816, 4818, ["stochastic", "estimates"], "Method"], [4852, 4857, ["feed", "-", "forward", "deep", "model"], "Method"], [4859, 4861, ["feature", "extractor"], "Method"], [4873, 4876, ["deep", "feature", "extractor"], "Method"], [4881, 4884, ["deep", "label", "predictor"], "Method"], [4893, 4897, ["feed", "-", "forward", "architecture"], "Method"], [4906, 4908, ["domain", "classifier"], "Method"], [4909, 4910, ["red"], "Method"], [4914, 4916, ["feature", "extractor"], "Method"], [4918, 4921, ["gradient", "reversal", "layer"], "Method"], [4932, 4936, ["backpropagation", "-", "based", "training"], "Method"], [4946, 4949, ["label", "prediction", "loss"], "Metric"], [4956, 4959, ["domain", "classification", "loss"], "Metric"], [4965, 4967, ["Gradient", "reversal"], "Task"], [4986, 4988, ["domain", "classifier"], "Method"], [4999, 5000, ["predictor"], "Method"], [5003, 5005, ["domain", "classifier"], "Method"], [5046, 5047, ["SGD"], "Method"], [5060, 5063, ["domain", "classification", "loss"], "Task"], [5066, 5067, ["SGDand"], "Method"], [5073, 5074, ["ADAGRAD"], "Method"], [5078, 5079, ["ADADELTA"], "Method"], [5085, 5087, ["learning", "algorithm"], "Method"], [5092, 5094, ["deep", "learning"], "Method"], [5105, 5109, ["stochastic", "saddle", "point", "procedure"], "Method"], [5110, 5111, ["SGD"], "Method"], [5124, 5127, ["gradient", "reversal", "layer"], "Method"], [5128, 5129, ["GRL"], "Method"], [5136, 5139, ["gradient", "reversal", "layer"], "Method"], [5148, 5150, ["forward", "propagation"], "Method"], [5152, 5153, ["GRL"], "Method"], [5156, 5158, ["identity", "transformation"], "Method"], [5161, 5162, ["backpropagation"], "Task"], [5165, 5166, ["GRL"], "Method"], [5199, 5203, ["object", "-", "oriented", "packages"], "Method"], [5204, 5206, ["deep", "learning"], "Method"], [5216, 5218, ["forward", "propagation"], "Method"], [5219, 5221, ["identity", "transformation"], "Method"], [5224, 5225, ["backpropagation"], "Method"], [5239, 5240, ["GRL"], "Method"], [5247, 5251, ["feature", "extractor", "G", "f"], "Method"], [5253, 5257, ["domain", "classifier", "G", "d"], "Method"], [5269, 5271, ["backpropagation", "process"], "Method"], [5274, 5275, ["GRL"], "Method"], [5320, 5321, ["SGD"], "Method"], [5354, 5357, ["gradient", "reversal", "layer"], "Method"], [5377, 5381, ["forward", "and", "backpropagation", "behaviour"], "Method"], [5396, 5399, ["pseudo", "-", "function"], "Metric"], [5417, 5420, ["stochastic", "gradient", "descent"], "Method"], [5437, 5438, ["SGD"], "Method"], [5463, 5464, ["learning"], "Task"], [5466, 5468, ["label", "predictor"], "Method"], [5514, 5517, ["Gradient", "Reversal", "layer"], "Method"], [5526, 5527, ["Caffe"], "Task"], [5545, 5550, ["shallow", "domain", "adversarial", "neural", "networks"], "Method"], [5568, 5571, ["Shallow", "Neural", "Networks"], "Method"], [5586, 5587, ["DANN"], "Method"], [5628, 5631, ["gradient", "step", "update"], "Method"], [5687, 5689, ["Toy", "Problem"], "Task"], [5707, 5713, ["inter", "-", "twinning", "moons", "2D", "problem"], "Task"], [5822, 5824, ["adaptation", "capability"], "Metric"], [5825, 5826, ["DANN"], "Method"], [5832, 5834, ["neural", "network"], "Method"], [5835, 5836, ["NN"], "Method"], [5848, 5850, ["network", "architecture"], "Method"], [5863, 5864, ["NN"], "Method"], [5870, 5871, ["DANN"], "Method"], [5879, 5882, ["domain", "regressor", "component"], "Method"], [5902, 5903, ["DANN"], "Method"], [5909, 5913, ["adversarial", "back", "-", "propagation"], "Method"], [5915, 5917, ["hidden", "layer"], "Method"], [5946, 5949, ["source", "risk", "minimization"], "Method"], [5956, 5957, ["regularizer"], "Method"], [5962, 5964, ["domain", "regressor"], "Method"], [5995, 5996, ["NN"], "Method"], [6009, 6011, ["hidden", "layer"], "Method"], [6016, 6019, ["source", "domain", "task"], "Task"], [6020, 6021, ["DANN"], "Method"], [6023, 6024, ["NN"], "Method"], [6035, 6036, ["NN"], "Method"], [6068, 6069, ["NN"], "Method"], [6089, 6090, ["NN"], "Method"], [6091, 6092, ["DANN"], "Method"], [6105, 6107, ["Label", "Classification"], "Task"], [6116, 6117, ["DANN"], "Method"], [6118, 6119, ["NN"], "Method"], [6137, 6138, ["NN"], "Method"], [6168, 6169, ["DANN"], "Method"], [6184, 6185, ["DANN"], "Method"], [6195, 6197, ["Representation", "PCA"], "Method"], [6201, 6204, ["domain", "adaptation", "regularizer"], "Method"], [6215, 6218, ["network", "hidden", "layer"], "Method"], [6226, 6229, ["Principal", "component", "analysis"], "Method"], [6230, 6231, ["PCA"], "Method"], [6266, 6267, ["NN"], "Method"], [6268, 6269, ["DANN"], "Method"], [6302, 6304, ["PCA", "transformation"], "Method"], [6307, 6311, ["DANN", "-", "PCA", "representation"], "Method"], [6357, 6361, ["DANN", "-", "PCA", "representation"], "Method"], [6369, 6371, ["PCA", "graphs"], "Method"], [6462, 6466, ["DANN", "-", "PCA", "representation"], "Method"], [6507, 6511, ["DANN", "-", "PCA", "representation"], "Method"], [6518, 6519, ["DANN"], "Method"], [6524, 6526, ["adaptation", "problem"], "Task"], [6530, 6532, ["Domain", "Classification"], "Task"], [6539, 6542, ["domain", "classification", "problem"], "Task"], [6548, 6552, ["domain", "regressor", "G", "d"], "Method"], [6603, 6606, ["G", "d", "regressor"], "Method"], [6617, 6619, ["hidden", "representation"], "Method"], [6640, 6642, ["domain", "regressor"], "Method"], [6644, 6646, ["learning", "process"], "Method"], [6647, 6648, ["NN"], "Method"], [6698, 6700, ["generalization", "capability"], "Metric"], [6757, 6759, ["linear", "regressor"], "Method"], [6835, 6838, ["label", "classification", "problem"], "Task"], [6858, 6861, ["domain", "classification", "problem"], "Task"], [6922, 6927, ["Unsupervised", "Hyper", "-", "Parameter", "Selection"], "Method"], [6929, 6932, ["unsupervised", "domain", "adaption"], "Task"], [6952, 6954, ["learning", "rate"], "Metric"], [6956, 6958, ["network", "architecture"], "Method"], [7003, 7008, ["reverse", "cross", "-", "validation", "approach"], "Method"], [7017, 7019, ["reverse", "validation"], "Method"], [7023, 7026, ["reverse", "validation", "risk"], "Metric"], [7102, 7104, ["classifier", "η"], "Method"], [7115, 7119, ["reverse", "classifier", "η", "r"], "Method"], [7149, 7153, ["reverse", "classifier", "η", "r"], "Method"], [7170, 7172, ["classifier", "η"], "Method"], [7174, 7177, ["reverse", "validation", "risk"], "Metric"], [7206, 7207, ["classifier"], "Method"], [7210, 7213, ["reverse", "validation", "risk"], "Metric"], [7219, 7222, ["neural", "network", "architectures"], "Method"], [7233, 7236, ["early", "stopping", "criterion"], "Metric"], [7240, 7241, ["η"], "Task"], [7269, 7273, ["learning", "of", "η", "r"], "Task"], [7278, 7279, ["accuracies"], "Metric"], [7286, 7290, ["reverse", "classifier", "η", "r"], "Method"], [7296, 7298, ["network", "η"], "Method"], [7303, 7307, ["Sentiment", "Analysis", "Data", "Sets"], "Material"], [7320, 7322, ["neural", "network"], "Method"], [7324, 7326, ["hidden", "layer"], "Method"], [7327, 7328, ["NN"], "Method"], [7338, 7341, ["Support", "Vector", "Machine"], "Method"], [7342, 7343, ["SVM"], "Method"], [7346, 7348, ["linear", "kernel"], "Method"], [7355, 7359, ["Amazon", "reviews", "data", "set"], "Material"], [7388, 7389, ["books"], "Material"], [7390, 7392, ["dvd", "disks"], "Material"], [7393, 7394, ["electronics"], "Material"], [7396, 7398, ["kitchen", "appliances"], "Material"], [7400, 7401, ["Reviews"], "Material"], [7412, 7413, ["bigrams"], "Method"], [7454, 7456, ["learning", "algorithms"], "Method"], [7492, 7493, ["NN"], "Method"], [7494, 7495, ["SVM"], "Method"], [7503, 7504, ["learning"], "Task"], [7515, 7517, ["learning", "algorithms"], "Method"], [7526, 7528, ["Classification", "accuracy"], "Metric"], [7530, 7534, ["Amazon", "reviews", "data", "set"], "Material"], [7536, 7540, ["Pairwise", "Poisson", "binomial", "test"], "Task"], [7580, 7583, ["learning", "rate", "µ"], "Metric"], [7605, 7607, ["training", "procedure"], "Method"], [7626, 7627, ["NN"], "Method"], [7692, 7695, ["reverse", "cross", "validation"], "Method"], [7703, 7705, ["learning", "algorithms"], "Method"], [7711, 7713, ["stopping", "criterion"], "Metric"], [7714, 7715, ["DANN"], "Method"], [7716, 7717, ["NN"], "Method"], [7731, 7732, ["accuracy"], "Metric"], [7754, 7757, ["Poisson", "binomial", "test"], "Method"], [7764, 7765, ["DANN"], "Method"], [7771, 7772, ["NN"], "Method"], [7773, 7774, ["SVM"], "Method"], [7787, 7788, ["DANN"], "Method"], [7789, 7790, ["NN"], "Method"], [7792, 7795, ["domain", "adaptation", "regularizer"], "Method"], [7816, 7817, ["DANN"], "Method"], [7818, 7820, ["Denoising", "Autoencoders"], "Method"], [7843, 7847, ["Marginalized", "Stacked", "Denoising", "Autoencoders"], "Method"], [7859, 7860, ["mSDA"], "Method"], [7862, 7864, ["unsupervised", "algorithm"], "Method"], [7868, 7871, ["robust", "feature", "representation"], "Method"], [7904, 7907, ["denoising", "autoencoders", "algorithm"], "Method"], [7911, 7913, ["feature", "representation"], "Method"], [7938, 7939, ["mSDA"], "Method"], [7955, 7959, ["Amazon", "reviews", "data", "sets"], "Material"], [7965, 7966, ["SVM"], "Method"], [7981, 7982, ["mSDA"], "Method"], [7998, 7999, ["mSDA"], "Method"], [8000, 8001, ["DANN"], "Method"], [8003, 8006, ["representation", "learning", "approaches"], "Method"], [8024, 8028, ["Amazon", "reviews", "data", "set"], "Material"], [8045, 8047, ["mSDA", "representations"], "Method"], [8049, 8051, ["corruption", "probability"], "Metric"], [8067, 8069, ["learning", "algorithms"], "Method"], [8070, 8071, ["DANN"], "Method"], [8072, 8073, ["NN"], "Method"], [8075, 8076, ["SVM"], "Method"], [8134, 8136, ["grid", "search"], "Method"], [8146, 8149, ["learning", "rate", "µ"], "Metric"], [8154, 8155, ["DANN"], "Method"], [8157, 8158, ["NN"], "Method"], [8163, 8165, ["mSDA", "representation"], "Method"], [8173, 8174, ["mSDA"], "Method"], [8175, 8176, ["DANN"], "Method"], [8184, 8187, ["Poisson", "binomial", "test"], "Method"], [8189, 8190, ["DANN"], "Method"], [8196, 8197, ["NN"], "Method"], [8199, 8200, ["SVM"], "Method"], [8220, 8221, ["NN"], "Method"], [8223, 8224, ["SVM"], "Method"], [8240, 8241, ["DANN"], "Method"], [8242, 8243, ["mSDA"], "Method"], [8263, 8266, ["domain", "adaptation", "theory"], "Method"], [8326, 8327, ["PAD"], "Method"], [8333, 8337, ["Amazon", "Reviews", "data", "set"], "Material"], [8345, 8346, ["NN"], "Method"], [8347, 8348, ["DANN"], "Method"], [8349, 8350, ["mSDA"], "Method"], [8352, 8353, ["mSDA"], "Method"], [8354, 8355, ["DANN"], "Method"], [8359, 8360, ["PAD"], "Method"], [8372, 8373, ["similarity"], "Metric"], [8440, 8442, ["linear", "SVMs"], "Method"], [8466, 8467, ["classifiers"], "Method"], [8486, 8488, ["PAD", "value"], "Metric"], [8528, 8529, ["PAD"], "Method"], [8631, 8635, ["Amazon", "Reviews", "data", "set"], "Material"], [8638, 8639, ["DANN"], "Method"], [8680, 8682, ["mSDA", "representations"], "Method"], [8728, 8729, ["DANN"], "Method"], [8732, 8733, ["mSDA"], "Method"], [8755, 8757, ["PAD", "values"], "Metric"], [8766, 8767, ["DANN"], "Method"], [8778, 8780, ["Deep", "Networks"], "Method"], [8781, 8783, ["Image", "Classification"], "Task"], [8790, 8792, ["deep", "version"], "Method"], [8793, 8794, ["DANN"], "Method"], [8823, 8826, ["deep", "learning", "methods"], "Method"], [8843, 8845, ["domain", "adaptation"], "Method"], [8846, 8848, ["computer", "vision"], "Task"], [8871, 8875, ["source", "-", "only", "model"], "Method"], [8896, 8902, ["train", "-", "on", "-", "target", "model"], "Method"], [8955, 8960, ["subspace", "alignment", "(", "SA", ")"], "Method"], [9045, 9046, ["SA"], "Method"], [9053, 9057, ["source", "-", "only", "model"], "Method"], [9069, 9071, ["label", "predictor"], "Method"], [9075, 9077, ["linear", "classifier"], "Method"], [9100, 9102, ["SA", "baseline"], "Method"], [9106, 9107, ["classifier"], "Method"], [9133, 9135, ["label", "predictor"], "Method"], [9183, 9185, ["full", "network"], "Method"], [9186, 9188, ["feature", "extractor"], "Method"], [9189, 9191, ["label", "predictor"], "Method"], [9203, 9205, ["CNN", "architectures"], "Method"], [9213, 9215, ["feature", "extractor"], "Method"], [9219, 9221, ["convolutional", "layers"], "Method"], [9261, 9262, ["AlexNet"], "Method"], [9264, 9267, ["Caffe", "-", "package"], "Method"], [9269, 9271, ["adaptation", "architecture"], "Method"], [9280, 9283, ["domain", "adaption", "component"], "Method"], [9290, 9293, ["fully", "connected", "layers"], "Method"], [9296, 9297, ["MNIST"], "Method"], [9316, 9318, ["domain", "classifier"], "Method"], [9323, 9324, ["adaptation"], "Task"], [9355, 9358, ["logistic", "regression", "loss"], "Method"], [9360, 9364, ["binomial", "cross", "-", "entropy"], "Method"], [9383, 9385, ["SVHN", "architecture"], "Method"], [9389, 9394, ["2", "-", "layer", "domain", "classifier"], "Method"], [9404, 9406, ["grid", "search"], "Method"], [9425, 9427, ["learning", "rate"], "Metric"], [9431, 9434, ["stochastic", "gradient", "descent"], "Method"], [9473, 9475, ["low", "error"], "Metric"], [9533, 9535, ["domain", "classifier"], "Method"], [9563, 9566, ["feature", "extractor", "component"], "Method"], [9572, 9575, ["domain", "classification", "component"], "Method"], [9594, 9598, ["label", "predictor", "G", "y"], "Method"], [9621, 9623, ["mean", "subtraction"], "Method"], [9662, 9668, ["train", "-", "on", "-", "target", "baseline"], "Metric"], [9672, 9673, ["Visualizations"], "Task"], [9675, 9678, ["t", "-", "SNE"], "Method"], [9714, 9716, ["shallow", "version"], "Method"], [9717, 9718, ["DANN"], "Method"], [9738, 9739, ["λp"], "Method"], [9741, 9743, ["feature", "extractor"], "Method"], [9744, 9747, ["domain", "classification", "components"], "Method"], [9751, 9753, ["learning", "rate"], "Metric"], [9771, 9773, ["classification", "accuracy"], "Metric"], [9787, 9788, ["visualizations"], "Method"], [9977, 9979, ["classification", "task"], "Task"], [10024, 10028, ["source", "-", "only", "model"], "Method"], [10047, 10048, ["adaptation"], "Task"], [10053, 10054, ["adaptation"], "Task"], [10066, 10070, ["source", "-", "only", "model"], "Method"], [10072, 10074, ["subspace", "alignment"], "Method"], [10075, 10076, ["SA"], "Method"], [10091, 10093, ["adaptation", "task"], "Task"], [10105, 10106, ["training"], "Task"], [10232, 10236, ["backpropagation", "-", "based", "technique"], "Method"], [10265, 10266, ["SA"], "Method"], [10273, 10276, ["classification", "accuracy", "drop"], "Metric"], [10285, 10287, ["dimensionality", "reduction"], "Task"], [10292, 10294, ["adaptation", "task"], "Task"], [10342, 10343, ["adaptation"], "Method"], [10347, 10348, ["error"], "Metric"], [10374, 10377, ["learning", "rate", "annealing"], "Method"], [10446, 10448, ["Classification", "accuracies"], "Metric"], [10449, 10452, ["digit", "image", "classifications"], "Task"], [10478, 10481, ["lower", "performance", "bound"], "Metric"], [10571, 10572, ["Accuracy"], "Metric"], [10592, 10593, ["SA"], "Method"], [10664, 10665, ["adaptation"], "Task"], [10738, 10739, ["adaptation"], "Task"], [10763, 10764, ["adaptation"], "Task"], [10770, 10772, ["Unsupervised", "adaptation"], "Method"], [10795, 10799, ["non", "-", "adapted", "model"], "Method"], [10803, 10804, ["accuracy"], "Metric"], [10823, 10824, ["GTSRB"], "Method"], [10900, 10902, ["unsupervised", "adaptation"], "Task"], [10924, 10931, ["synthetic", "-", "to", "-", "real", "data", "adaptation"], "Task"], [10988, 10990, ["label", "predictor"], "Task"], [10998, 11000, ["validation", "error"], "Metric"], [11017, 11021, ["semi", "-", "supervised", "setting"], "Task"], [11025, 11029, ["semi", "-", "supervised", "setting"], "Task"], [11058, 11059, ["Amazon"], "Material"], [11071, 11072, ["Office"], "Method"], [11105, 11107, ["deep", "model"], "Method"], [11118, 11119, ["CNN"], "Method"], [11129, 11131, ["Caffe", "package"], "Method"], [11160, 11162, ["network", "architecture"], "Method"], [11163, 11168, ["domain", "mean", "-", "based", "regularization"], "Method"], [11170, 11172, ["domain", "classifier"], "Method"], [11186, 11188, ["transfer", "tasks"], "Task"], [11210, 11211, ["adaptation"], "Task"], [11243, 11246, ["source", "domain", "data"], "Material"], [11276, 11277, ["accuracy"], "Metric"], [11278, 11280, ["unsupervised", "adaptation"], "Task"], [11328, 11329, ["accuracy"], "Metric"], [11347, 11349, ["validation", "accuracy"], "Metric"], [11355, 11358, ["domain", "classifier", "branch"], "Method"], [11375, 11376, ["regularizer"], "Method"], [11381, 11384, ["Deep", "Image", "Descriptors"], "Method"], [11385, 11388, ["Re", "-", "Identification"], "Task"], [11398, 11400, ["adaptation", "method"], "Method"], [11401, 11412, ["person", "re", "-", "identification", "(", "re", "-", "i", "d", ")", "problem"], "Task"], [11416, 11420, ["person", "re", "-", "identification"], "Task"], [11525, 11526, ["Rank1=71.08"], "Metric"], [11530, 11532, ["classification", "problems"], "Task"], [11537, 11541, ["re", "-", "identification", "problem"], "Task"], [11549, 11551, ["vector", "descriptor"], "Method"], [11581, 11588, ["Cumulative", "Match", "Characteristic", "(", "CMC", ")", "curve"], "Method"], [11598, 11600, ["identification", "rate"], "Metric"], [11601, 11602, ["recall"], "Metric"], [11604, 11607, ["rank", "-", "k"], "Metric"], [11640, 11642, ["descriptor", "mappings"], "Method"], [11674, 11678, ["re", "-", "identification", "systems"], "Method"], [11700, 11704, ["cross", "-", "domain", "evaluation"], "Task"], [11706, 11710, ["domain", "-", "adaptation", "problem"], "Task"], [11713, 11715, ["camera", "network"], "Method"], [11730, 11733, ["re", "-", "identification"], "Metric"], [11753, 11760, ["cross", "-", "data", "-", "set", "evaluation", "scenario"], "Task"], [11765, 11768, ["deep", "learning", "methods"], "Method"], [11793, 11795, ["Domain", "adaptation"], "Method"], [11802, 11807, ["deep", "re", "-", "identification", "descriptors"], "Task"], [11861, 11865, ["single", "-", "shot", "variant"], "Method"], [12022, 12024, ["descriptor", "mapping"], "Method"], [12026, 12028, ["supervised", "way"], "Method"], [12282, 12284, ["similarity", "score"], "Metric"], [12340, 12342, ["CNN", "architectures"], "Method"], [12351, 12353, ["siamese", "architecture"], "Method"], [12359, 12362, ["Deep", "Metric", "Learning"], "Method"], [12363, 12364, ["DML"], "Method"], [12367, 12370, ["deep", "image", "descriptors"], "Task"], [12380, 12382, ["convolution", "layers"], "Method"], [12391, 12393, ["filter", "banks"], "Method"], [12397, 12401, ["ReLU", "and", "max", "pooling"], "Method"], [12404, 12408, ["fully", "-", "connected", "layer"], "Method"], [12422, 12424, ["parallel", "flows"], "Method"], [12426, 12427, ["CNN"], "Method"], [12448, 12450, ["convolution", "layer"], "Method"], [12462, 12464, ["convolution", "layers"], "Method"], [12500, 12504, ["domain", "-", "adversarial", "training"], "Method"], [12512, 12514, ["feature", "extractor"], "Method"], [12517, 12519, ["convolutional", "layers"], "Method"], [12522, 12525, ["max", "-", "pooling"], "Method"], [12526, 12527, ["ReLU"], "Method"], [12532, 12534, ["label", "predictor"], "Method"], [12540, 12542, ["descriptor", "predictor"], "Method"], [12545, 12549, ["fully", "-", "connected", "layer"], "Method"], [12551, 12553, ["domain", "classifier"], "Method"], [12555, 12559, ["fully", "-", "connected", "layers"], "Method"], [12564, 12566, ["intermediate", "representation"], "Method"], [12572, 12575, ["verification", "loss", "function"], "Metric"], [12577, 12579, ["descriptor", "predictor"], "Method"], [12616, 12618, ["domain", "classifier"], "Method"], [12621, 12623, ["logistic", "loss"], "Method"], [12630, 12632, ["learning", "rate"], "Metric"], [12643, 12644, ["adaptation"], "Task"], [12671, 12675, ["max", "-", "pooling", "layer"], "Method"], [12718, 12720, ["annotation", "problem"], "Task"], [12762, 12766, ["Re", "-", "identification", "data"], "Task"], [12774, 12778, ["domain", "-", "adversarial", "training"], "Method"], [12783, 12786, ["re", "-", "identification"], "Task"], [12820, 12824, ["domain", "-", "adversarial", "learning"], "Method"], [12825, 12827, ["classification", "problems"], "Task"], [12838, 12840, ["domain", "adaptation"], "Method"], [12841, 12846, ["feed", "-", "forward", "neural", "networks"], "Method"], [12849, 12853, ["large", "-", "scale", "training"], "Task"], [12907, 12908, ["alignment"], "Task"], [12923, 12926, ["domain", "adaptation", "theory"], "Method"], [12935, 12936, ["DANN"], "Method"], [12940, 12943, ["network", "hidden", "layer"], "Method"], [12977, 12984, ["shallow", "and", "deep", "feed", "-", "forward", "architectures"], "Method"], [12993, 12996, ["deep", "learning", "package"], "Method"], [13002, 13005, ["gradient", "reversal", "layer"], "Method"], [13030, 13032, ["domain", "adaptation"], "Method"], [13035, 13037, ["sentiment", "analysis"], "Task"], [13038, 13041, ["image", "classification", "tasks"], "Task"], [13051, 13054, ["domain", "adaptation", "component"], "Method"], [13060, 13063, ["neural", "network", "architecture"], "Method"], [13067, 13068, ["backpropagation"], "Method"], [13084, 13086, ["classification", "tasks"], "Task"], [13092, 13096, ["feed", "-", "forward", "architectures"], "Method"], [13100, 13102, ["descriptor", "learning"], "Task"], [13103, 13107, ["person", "re", "-", "identification"], "Task"], [491, 492, ["books"], "Material"], [613, 615, ["domain", "adaptation"], "Method"], [1011, 1013, ["standard", "backpropagation"], "Method"], [1036, 1037, ["DANN"], "Method"], [1125, 1126, ["DANN"], "Method"], [1222, 1224, ["deep", "learning"], "Method"], [1492, 1494, ["domain", "adaptation"], "Method"], [1841, 1843, ["domain", "adaptation"], "Method"], [1850, 1852, ["domain", "adaptation"], "Method"], [2147, 2149, ["domain", "adaptation"], "Method"], [2262, 2263, ["DANN"], "Method"], [2358, 2359, ["DANN"], "Method"], [2543, 2545, ["domain", "adaptation"], "Method"], [2564, 2565, ["distance"], "Metric"], [2606, 2607, ["distance"], "Metric"], [2958, 2959, ["distance"], "Metric"], [3000, 3001, ["SVM"], "Method"], [3044, 3045, ["SVM"], "Method"], [3483, 3484, ["NN"], "Method"], [3834, 3835, ["distance"], "Metric"], [4452, 4453, ["DANN"], "Method"], [4544, 4545, ["DANN"], "Method"], [4612, 4613, ["DANN"], "Method"], [4754, 4755, ["DANN"], "Method"], [4843, 4846, ["stochastic", "gradient", "descent"], "Method"], [4899, 4901, ["domain", "adaptation"], "Method"], [5610, 5613, ["stochastic", "gradient", "descent"], "Method"], [5635, 5636, ["DANN"], "Method"], [5939, 5940, ["NN"], "Method"], [5987, 5988, ["DANN"], "Method"], [6075, 6076, ["DANN"], "Method"], [6327, 6328, ["NN"], "Method"], [6428, 6429, ["NN"], "Method"], [6600, 6601, ["DANN"], "Method"], [6669, 6670, ["DANN"], "Method"], [6688, 6689, ["NN"], "Method"], [6722, 6723, ["DANN"], "Method"], [6812, 6813, ["NN"], "Method"], [6871, 6872, ["DANN"], "Method"], [6894, 6895, ["NN"], "Method"], [6917, 6918, ["DANN"], "Method"], [7315, 7316, ["DANN"], "Method"], [7449, 7451, ["domain", "adaptation"], "Method"], [7544, 7545, ["DANN"], "Method"], [7592, 7593, ["NN"], "Method"], [7608, 7609, ["DANN"], "Method"], [7630, 7631, ["DANN"], "Method"], [7644, 7645, ["SVM"], "Method"], [7826, 7827, ["DANN"], "Method"], [7848, 7849, ["mSDA"], "Method"], [7942, 7943, ["SVM"], "Method"], [7973, 7974, ["DANN"], "Method"], [8259, 8260, ["DANN"], "Method"], [8276, 8277, ["DANN"], "Method"], [8324, 8325, ["distance"], "Metric"], [8502, 8503, ["DANN"], "Method"], [8545, 8546, ["DANN"], "Method"], [8556, 8557, ["DANN"], "Method"], [8563, 8564, ["NN"], "Method"], [8608, 8609, ["DANN"], "Method"], [8695, 8696, ["mSDA"], "Method"], [8771, 8772, ["mSDA"], "Method"], [8886, 8889, ["domain", "classifier", "branch"], "Method"], [8921, 8922, ["DA"], "Method"], [8951, 8952, ["DA"], "Method"], [8993, 8994, ["DA"], "Method"], [9139, 9140, ["SVM"], "Method"], [9194, 9195, ["DA"], "Method"], [9491, 9493, ["domain", "adaptation"], "Method"], [10511, 10512, ["DA"], "Method"], [10520, 10521, ["DA"], "Method"], [10575, 10576, ["DA"], "Method"], [10811, 10812, ["DA"], "Method"], [10947, 10949, ["domain", "adaptation"], "Method"], [11140, 11141, ["DA"], "Method"], [11553, 11554, ["distance"], "Metric"], [11629, 11630, ["distance"], "Metric"], [12508, 12509, ["DANN"], "Method"], [12882, 12883, ["DA"], "Method"], [12911, 12913, ["standard", "backpropagation"], "Method"]]} {"doc_id": "1130d8fdd931225c2d7563c3808367726cfa1c3a", "text": ["document", ":", "PixelGAN", "Autoencoders", "In", "this", "paper", ",", "we", "describe", "the", "‘", "‘", "PixelGAN", "autoencoder", "’", "’", ",", "a", "generative", "autoencoder", "in", "which", "the", "generative", "path", "is", "a", "convolutional", "autoregressive", "neural", "network", "on", "pixels", "(", "PixelCNN", ")", "that", "is", "conditioned", "on", "a", "latent", "code", ",", "and", "the", "recognition", "path", "uses", "a", "generative", "adversarial", "network", "(", "GAN", ")", "to", "impose", "a", "prior", "distribution", "on", "the", "latent", "code", ".", "We", "show", "that", "different", "priors", "result", "in", "different", "decompositions", "of", "information", "between", "the", "latent", "code", "and", "the", "autoregressive", "decoder", ".", "For", "example", ",", "by", "imposing", "a", "Gaussian", "distribution", "as", "the", "prior", ",", "we", "can", "achieve", "a", "global", "vs.", "local", "decomposition", ",", "or", "by", "imposing", "a", "categorical", "distribution", "as", "the", "prior", ",", "we", "can", "disentangle", "the", "style", "and", "content", "information", "of", "images", "in", "an", "unsupervised", "fashion", ".", "We", "further", "show", "how", "the", "PixelGAN", "autoencoder", "with", "a", "categorical", "prior", "can", "be", "directly", "used", "in", "semi", "-", "supervised", "settings", "and", "achieve", "competitive", "semi", "-", "supervised", "classification", "results", "on", "the", "MNIST", ",", "SVHN", "and", "NORB", "datasets", ".", "numbers", ",", "compressnatbib", "section", ":", "Introduction", "In", "recent", "years", ",", "generative", "models", "that", "can", "be", "trained", "via", "direct", "back", "-", "propagation", "have", "enabled", "remarkable", "progress", "in", "modeling", "natural", "images", ".", "One", "of", "the", "most", "successful", "models", "is", "the", "generative", "adversarial", "network", "(", "GAN", ")", "gan", ",", "which", "employs", "a", "two", "player", "min", "-", "max", "game", ".", "The", "generative", "model", ",", ",", "samples", "the", "prior", "and", "generates", "the", "sample", ".", "The", "discriminator", ",", ",", "is", "trained", "to", "identify", "whether", "a", "point", "is", "a", "sample", "from", "the", "data", "distribution", "or", "a", "sample", "from", "the", "generative", "model", ".", "The", "generator", "is", "trained", "to", "maximally", "confuse", "the", "discriminator", "into", "believing", "that", "generated", "samples", "come", "from", "the", "data", "distribution", ".", "The", "cost", "function", "of", "GAN", "is", "GANs", "can", "be", "considered", "within", "the", "wider", "framework", "of", "implicit", "generative", "models", "mohamed2016learning", ",", "ference", ",", "dustin", ".", "Implicit", "distributions", "can", "be", "sampled", "through", "their", "generative", "path", ",", "but", "their", "likelihood", "function", "is", "not", "tractable", ".", "Recently", ",", "several", "papers", "have", "proposed", "another", "application", "of", "GAN", "-", "style", "algorithms", "for", "approximate", "inference", ",", "mohamed2016learning", ",", "ference", ",", "dustin", ",", "ranganath2016operator", ",", "aae", ",", "avb", ",", "ali", ",", "bigan", ".", "These", "algorithms", "use", "implicit", "distributions", "to", "learn", "posterior", "approximations", "that", "are", "more", "expressive", "than", "the", "distributions", "with", "tractable", "densities", "that", "are", "often", "used", "in", "variational", "inference", ".", "For", "example", ",", "adversarial", "autoencoders", "(", "AAE", ")", "aae", "use", "a", "universal", "approximator", "posterior", "as", "the", "implicit", "posterior", "distribution", "and", "use", "adversarial", "training", "to", "match", "the", "aggregated", "posterior", "of", "the", "latent", "code", "to", "the", "prior", "distribution", ".", "Adversarial", "variational", "Bayes", "ference", ",", "avb", "uses", "a", "more", "general", "amortized", "GAN", "inference", "framework", "within", "a", "maximum", "-", "likelihood", "learning", "setting", ".", "Another", "type", "of", "GAN", "inference", "technique", "is", "used", "in", "the", "ALI", "ali", "and", "BiGAN", "bigan", "models", ",", "which", "have", "been", "shown", "to", "approximate", "maximum", "likelihood", "learning", "ference", ".", "In", "these", "models", ",", "both", "the", "recognition", "and", "generative", "models", "are", "implicit", "and", "are", "jointly", "learnt", "by", "an", "adversarial", "training", "process", ".", "Variational", "autoencoders", "(", "VAE", ")", "vae", ",", "rezende", "are", "another", "state", "-", "of", "-", "the", "-", "art", "image", "modeling", "technique", "that", "use", "neural", "networks", "to", "parametrize", "the", "posterior", "distribution", "and", "pair", "it", "with", "a", "top", "-", "down", "generative", "network", ".", "Both", "networks", "are", "jointly", "trained", "to", "maximize", "a", "variational", "lower", "bound", "on", "the", "data", "log", "-", "likelihood", ".", "A", "different", "framework", "for", "learning", "density", "models", "is", "autoregressive", "neural", "networks", "such", "as", "NADE", "nade", ",", "MADE", "made", ",", "PixelRNN", "pixelrnn", "and", "PixelCNN", "pixelcnn", ".", "Unlike", "variational", "autoencoders", ",", "which", "capture", "the", "statistics", "of", "the", "data", "in", "hierarchical", "latent", "codes", ",", "the", "autoregressive", "models", "learn", "the", "image", "densities", "directly", "at", "the", "pixel", "level", "without", "learning", "a", "hierarchical", "latent", "representation", ".", "In", "this", "paper", ",", "we", "present", "the", "PixelGAN", "autoencoder", "as", "a", "generative", "autoencoder", "that", "combines", "the", "benefits", "of", "latent", "variable", "models", "with", "autoregressive", "architectures", ".", "The", "PixelGAN", "autoencoder", "is", "a", "generative", "autoencoder", "in", "which", "the", "generative", "path", "is", "a", "PixelCNN", "that", "is", "conditioned", "on", "a", "latent", "variable", ".", "The", "latent", "variable", "is", "inferred", "by", "matching", "the", "aggregated", "posterior", "distribution", "to", "the", "prior", "distribution", "by", "an", "adversarial", "training", "technique", "similar", "to", "that", "of", "the", "adversarial", "autoencoder", "aae", ".", "However", ",", "whereas", "in", "adversarial", "autoencoders", "the", "statistics", "of", "the", "data", "distribution", "are", "captured", "by", "the", "latent", "code", ",", "in", "the", "PixelGAN", "autoencoder", "they", "are", "captured", "jointly", "by", "the", "latent", "code", "and", "the", "autoregressive", "decoder", ".", "We", "show", "that", "imposing", "different", "distributions", "as", "the", "prior", "results", "in", "different", "factorizations", "of", "information", "between", "the", "latent", "code", "and", "the", "autoregressive", "decoder", ".", "For", "example", ",", "in", ":", "pixelgan_gaussian", "]", "Section", "[", "reference", "]", ",", "we", "show", "that", "by", "imposing", "a", "Gaussian", "distribution", "on", "the", "latent", "code", ",", "we", "can", "achieve", "a", "global", "vs.", "local", "decomposition", "of", "information", ".", "In", "this", "case", ",", "the", "global", "latent", "code", "no", "longer", "has", "to", "model", "all", "the", "irrelevant", "and", "fine", "details", "of", "the", "image", ",", "and", "can", "use", "its", "capacity", "to", "capture", "more", "relevant", "and", "global", "statistics", "of", "the", "image", ".", "Another", "type", "of", "decomposition", "of", "information", "that", "can", "be", "learnt", "by", "PixelGAN", "autoencoders", "is", "a", "discrete", "vs.", "continuous", "decomposition", ".", "In", ":", "pixelgan_cat", "]", "Section", "[", "reference", "]", ",", "we", "show", "that", "we", "can", "achieve", "this", "decomposition", "by", "imposing", "a", "categorical", "prior", "on", "the", "latent", "code", "using", "adversarial", "training", ".", "In", "this", "case", ",", "the", "categorical", "latent", "code", "captures", "the", "discrete", "underlying", "factors", "of", "variation", "in", "the", "data", ",", "such", "as", "class", "label", "information", ",", "and", "the", "autoregressive", "decoder", "captures", "the", "remaining", "continuous", "structure", ",", "such", "as", "style", "information", ",", "in", "an", "unsupervised", "fashion", ".", "We", "then", "show", "how", "PixelGAN", "autoencoders", "with", "categorical", "priors", "can", "be", "directly", "used", "in", "clustering", "and", "semi", "-", "supervised", "scenarios", "and", "achieve", "very", "competitive", "classification", "results", "on", "several", "datasets", "in", ":", "experiments", "]", "Section", "[", "reference", "]", ".", "Finally", ",", "we", "present", "one", "of", "the", "main", "potential", "applications", "of", "PixelGAN", "autoencoders", "in", "learning", "cross", "-", "domain", "relations", "between", "two", "different", "domains", "in", ":", "cross", "-", "domain", "]", "Section", "[", "reference", "]", ".", "section", ":", "PixelGAN", "Autoencoders", "Let", "be", "a", "datapoint", "that", "comes", "from", "the", "distribution", "and", "be", "the", "hidden", "code", ".", "The", "recognition", "path", "of", "the", "PixelGAN", "autoencoder", "(", ":", "pixelgan_gaussian", "]", "Figure", "[", "reference", "]", ")", "defines", "an", "implicit", "posterior", "distribution", "by", "using", "a", "deterministic", "neural", "function", "that", "takes", "the", "input", "along", "with", "random", "noise", "with", "a", "fixed", "distribution", "and", "outputs", ".", "The", "aggregated", "posterior", "of", "this", "model", "is", "defined", "as", "follows", ":", "This", "parametrization", "of", "the", "implicit", "posterior", "distribution", "was", "originally", "proposed", "in", "the", "adversarial", "autoencoder", "work", "aae", "as", "the", "universal", "approximator", "posterior", ".", "We", "can", "sample", "from", "this", "implicit", "distribution", ",", "by", "evaluating", "at", "different", "samples", "of", ",", "but", "the", "density", "function", "of", "this", "posterior", "distribution", "is", "intractable", ".", "endix", ":", "input_noise", "]", "Appendix", "[", "reference", "]", "discusses", "the", "importance", "of", "the", "input", "noise", "in", "training", "PixelGAN", "autoencoders", ".", "The", "generative", "path", "is", "a", "conditional", "PixelCNN", "pixelcnn", "that", "conditions", "on", "the", "latent", "vector", "using", "an", "adaptive", "bias", "in", "PixelCNN", "layers", ".", "The", "inference", "is", "done", "by", "an", "amortized", "GAN", "inference", "technique", "that", "was", "originally", "proposed", "in", "the", "adversarial", "autoencoder", "work", "aae", ".", "In", "this", "method", ",", "an", "adversarial", "network", "is", "attached", "on", "top", "of", "the", "hidden", "code", "vector", "of", "the", "autoencoder", "and", "matches", "the", "aggregated", "posterior", "distribution", ",", ",", "to", "an", "arbitrary", "prior", ",", ".", "Samples", "from", "and", "are", "provided", "to", "the", "adversarial", "network", "as", "the", "negative", "and", "positive", "examples", "respectively", ",", "and", "the", "generator", "of", "the", "adversarial", "network", ",", "which", "is", "also", "the", "encoder", "of", "the", "autoencoder", ",", "tries", "to", "match", "to", "by", "the", "gradient", "that", "comes", "through", "the", "discriminative", "adversarial", "network", ".", "The", "adversarial", "network", ",", "the", "PixelCNN", "decoder", "and", "the", "encoder", "are", "trained", "jointly", "in", "two", "phases", "–", "the", "reconstruction", "phase", "and", "the", "adversarial", "phase", "–", "executed", "on", "each", "mini", "-", "batch", ".", "In", "the", "reconstruction", "phase", ",", "the", "ground", "truth", "input", "along", "with", "the", "hidden", "code", "inferred", "by", "the", "encoder", "are", "provided", "to", "the", "PixelCNN", "decoder", ".", "The", "PixelCNN", "decoder", "weights", "are", "updated", "to", "maximize", "the", "log", "-", "likelihood", "of", "the", "input", ".", "The", "encoder", "weights", "are", "also", "updated", "at", "this", "stage", "by", "the", "gradient", "that", "comes", "through", "the", "conditioning", "vector", "of", "the", "PixelCNN", ".", "In", "the", "adversarial", "phase", ",", "the", "adversarial", "network", "updates", "both", "its", "discriminative", "network", "and", "its", "generative", "network", "(", "the", "encoder", ")", "to", "match", "to", ".", "Once", "the", "training", "is", "done", ",", "we", "can", "sample", "from", "the", "model", "by", "first", "sampling", "from", "the", "prior", "distribution", ",", "and", "then", "sampling", "from", "the", "conditional", "likelihood", "parametrized", "by", "the", "PixelCNN", "decoder", ".", "We", "now", "establish", "a", "connection", "between", "the", "PixelGAN", "autoencoder", "cost", "and", "maximum", "likelihood", "learning", "using", "a", "decomposition", "of", "the", "aggregated", "evidence", "lower", "bound", "(", "ELBO", ")", "proposed", "in", "surgery", ":", "The", "first", "term", "in", "elbo", "]", "Equation", "[", "reference", "]", "is", "the", "reconstruction", "term", "and", "the", "second", "term", "is", "the", "marginal", "KL", "divergence", "between", "the", "aggregated", "posterior", "and", "the", "prior", "distribution", ".", "The", "third", "term", "is", "the", "mutual", "information", "between", "the", "latent", "code", "and", "the", "input", ".", "This", "is", "a", "regularization", "term", "that", "encourages", "and", "to", "be", "decoupled", "by", "removing", "the", "information", "of", "the", "data", "distribution", "from", "the", "hidden", "code", ".", "If", "the", "training", "set", "has", "examples", ",", "is", "bounded", "as", "follows", "(", "see", "surgery", ")", ".", "In", "order", "to", "maximize", "the", "ELBO", ",", "we", "need", "to", "minimize", "all", "the", "three", "terms", "of", "elbo", "]", "Equation", "[", "reference", "]", ".", "We", "consider", "two", "cases", "for", "the", "decoder", ":", "Deterministic", "Decoder", ".", "If", "the", "decoder", "is", "deterministic", "or", "has", "very", "limited", "stochasticity", "such", "as", "the", "simple", "factorized", "decoder", "of", "the", "VAE", ",", "the", "mutual", "information", "term", "acts", "in", "the", "complete", "opposite", "direction", "of", "the", "reconstruction", "term", ".", "This", "is", "because", "the", "only", "way", "to", "minimize", "the", "reconstruction", "error", "of", "is", "to", "learn", "a", "hidden", "code", "that", "is", "relevant", "to", ",", "which", "results", "in", "maximizing", ".", "Indeed", ",", "it", "can", "be", "shown", "that", "minimizing", "the", "reconstruction", "term", "maximizes", "a", "variational", "lower", "bound", "on", "i", "m", ",", "infogan", ".", "For", "example", ",", "in", "the", "case", "of", "the", "VAE", "trained", "on", "MNIST", ",", "since", "the", "reconstruction", "is", "precise", ",", "the", "mutual", "information", "term", "is", "dominated", "and", "is", "close", "to", "its", "maximum", "value", "surgery", ".", "Stochastic", "Decoder", ".", "If", "we", "use", "a", "powerful", "decoder", "such", "as", "the", "PixelCNN", ",", "the", "reconstruction", "term", "and", "the", "mutual", "information", "term", "will", "not", "compete", "with", "each", "other", "anymore", "and", "the", "network", "can", "minimize", "both", "independently", ".", "In", "this", "case", ",", "the", "optimal", "solution", "for", "maximizing", "the", "ELBO", "would", "be", "to", "model", "solely", "by", "and", "thereby", "minimizing", "the", "reconstruction", "term", ",", "and", "at", "the", "same", "time", ",", "minimizing", "the", "mutual", "information", "term", "by", "ignoring", "the", "latent", "code", ".", "As", "a", "result", ",", "even", "though", "the", "model", "achieves", "a", "high", "likelihood", ",", "the", "latent", "code", "does", "not", "learn", "any", "useful", "representation", ",", "which", "is", "undesirable", ".", "This", "problem", "has", "been", "observed", "in", "several", "previous", "works", "bowman", ",", "vlae", "and", "different", "techniques", "such", "as", "annealing", "the", "weight", "of", "the", "KL", "term", "bowman", "or", "weakening", "the", "decoder", "vlae", "have", "been", "proposed", "to", "make", "and", "more", "dependent", ".", "As", "suggested", "in", "ference_ml", ",", "vlae", ",", "we", "think", "that", "the", "maximum", "likelihood", "objective", "by", "itself", "is", "not", "a", "useful", "objective", "for", "representation", "learning", "especially", "when", "a", "powerful", "decoder", "is", "used", ".", "In", "PixelGAN", "autoencoders", ",", "in", "order", "to", "encourage", "learning", "more", "useful", "representations", ",", "we", "modify", "the", "ELBO", "(", "elbo", "]", "Equation", "[", "reference", "]", ")", "by", "removing", "the", "mutual", "information", "term", "from", "it", ",", "since", "this", "term", "is", "explicitly", "encouraging", "to", "become", "independent", "of", ".", "So", "our", "cost", "function", "only", "includes", "the", "reconstruction", "term", "and", "the", "marginal", "KL", "term", ".", "The", "reconstruction", "term", "is", "optimized", "by", "the", "reconstruction", "phase", "of", "training", "and", "the", "marginal", "KL", "term", "is", "approximately", "optimized", "by", "the", "adversarial", "phase", ".", "Note", "that", "since", "the", "mutual", "information", "term", "is", "upper", "bounded", "by", "a", "constant", "(", ")", ",", "we", "are", "still", "maximizing", "a", "lower", "bound", "on", "the", "log", "-", "likelihood", "of", "data", ".", "However", ",", "this", "bound", "is", "weaker", "than", "the", "ELBO", ",", "which", "is", "the", "price", "that", "is", "paid", "for", "learning", "more", "useful", "latent", "representations", "by", "balancing", "the", "decomposition", "of", "information", "between", "the", "latent", "code", "and", "the", "autoregressive", "decoder", ".", "For", "implementing", "the", "conditioning", "adaptive", "bias", "in", "the", "PixelCNN", "decoder", ",", "we", "explore", "two", "different", "architectures", "pixelcnn", ".", "In", "the", "location", "-", "invariant", "bias", ",", "for", "each", "PixelCNN", "layer", ",", "we", "use", "the", "latent", "code", "to", "construct", "a", "vector", "that", "is", "broadcasted", "within", "each", "feature", "map", "of", "the", "layer", "and", "then", "added", "as", "an", "adaptive", "bias", "to", "that", "layer", ".", "In", "the", "location", "-", "dependent", "bias", ",", "we", "use", "the", "latent", "code", "to", "construct", "a", "spatial", "feature", "map", "that", "is", "broadcasted", "across", "different", "feature", "maps", "and", "then", "added", "only", "to", "the", "first", "layer", "of", "the", "decoder", "as", "an", "adaptive", "bias", ".", "We", "will", "discuss", "the", "effect", "of", "these", "architectures", "on", "the", "learnt", "representation", "in", ":", "mnist_code", "]", "Figure", "[", "reference", "]", "of", ":", "pixelgan_gaussian", "]", "Section", "[", "reference", "]", "and", "their", "implementation", "details", "in", "endix", ":", "conditioning_of_pixelcnn", "]", "Appendix", "[", "reference", "]", ".", "subsection", ":", "PixelGAN", "Autoencoders", "with", "Gaussian", "Priors", "Here", ",", "we", "show", "that", "PixelGAN", "autoencoders", "with", "Gaussian", "priors", "can", "decompose", "the", "global", "and", "local", "statistics", "of", "the", "images", "between", "the", "latent", "code", "and", "the", "autoregressive", "decoder", ".", ":", "mnist", "]", "Figure", "[", "reference", "]", "a", "shows", "the", "samples", "of", "a", "PixelGAN", "autoencoder", "model", "with", "the", "location", "-", "dependent", "bias", "trained", "on", "the", "MNIST", "dataset", ".", "For", "the", "purpose", "of", "better", "illustrating", "the", "decomposition", "of", "information", ",", "we", "have", "chosen", "a", "2", "-", "D", "Gaussian", "latent", "code", "and", "a", "limited", "the", "receptive", "field", "of", "size", "9", "for", "the", "PixelGAN", "autoencoder", ".", ":", "mnist", "]", "Figure", "[", "reference", "]", "b", "shows", "the", "samples", "of", "a", "PixelCNN", "model", "with", "the", "same", "limited", "receptive", "field", "size", "of", "9", "and", ":", "mnist", "]", "Figure", "[", "reference", "]", "c", "shows", "the", "samples", "of", "an", "adversarial", "autoencoder", "with", "the", "2", "-", "D", "Gaussian", "latent", "code", ".", "The", "PixelCNN", "can", "successfully", "capture", "the", "local", "statistics", ",", "but", "fails", "to", "capture", "the", "global", "statistics", "due", "to", "the", "limited", "receptive", "field", "size", ".", "In", "contrast", ",", "the", "adversarial", "autoencoder", ",", "whose", "sample", "quality", "is", "very", "similar", "to", "that", "of", "the", "VAE", ",", "can", "successfully", "capture", "the", "global", "statistics", ",", "but", "fails", "to", "generate", "the", "details", "of", "the", "images", ".", "However", ",", "the", "PixelGAN", "autoencoder", ",", "with", "the", "same", "receptive", "field", "and", "code", "size", ",", "can", "combine", "the", "best", "of", "both", "and", "generates", "sharp", "images", "with", "global", "statistics", ".", "In", "PixelGAN", "autoencoders", ",", "both", "the", "PixelCNN", "depth", "and", "the", "conditioning", "architecture", "affect", "the", "decomposition", "of", "information", "between", "the", "latent", "code", "and", "the", "autoregressive", "decoder", ".", "We", "investigate", "these", "effects", "in", ":", "mnist_code", "]", "Figure", "[", "reference", "]", "by", "training", "a", "PixelGAN", "autoencoder", "on", "MNIST", "where", "the", "code", "size", "is", "chosen", "to", "be", "for", "the", "visualization", "purpose", ".", "As", "shown", "in", ":", "mnist_code", "]", "Figure", "[", "reference", "]", "a", ",", "b", ",", "when", "a", "shallow", "decoder", "is", "used", ",", "most", "of", "the", "information", "will", "be", "encoded", "in", "the", "hidden", "code", "and", "there", "is", "a", "clean", "separation", "between", "the", "digit", "clusters", ".", "As", "we", "make", "the", "PixelCNN", "more", "powerful", "(", ":", "mnist_code", "]", "Figure", "[", "reference", "]", "c", ",", "d", ")", ",", "we", "can", "see", "that", "the", "hidden", "code", "is", "still", "used", "to", "capture", "some", "relevant", "information", "of", "the", "input", ",", "but", "the", "separation", "of", "digit", "clusters", "is", "not", "as", "sharp", "when", "the", "limited", "code", "size", "of", "2", "is", "used", ".", "In", "the", "next", "section", ",", "we", "will", "show", "that", "by", "using", "a", "larger", "code", "size", "(", "e.g.", ",", "30", ")", ",", "we", "can", "get", "a", "much", "better", "separation", "of", "digit", "clusters", "even", "when", "a", "powerful", "PixelCNN", "is", "used", ".", "The", "conditioning", "architecture", "also", "affects", "the", "decomposition", "of", "information", ".", "In", "the", "case", "of", "the", "location", "-", "invariant", "bias", ",", "the", "hidden", "code", "is", "encouraged", "to", "learn", "the", "global", "information", "that", "is", "location", "-", "invariant", "(", "the", "what", "information", "and", "not", "the", "where", "information", ")", "such", "as", "the", "class", "label", "information", ".", "For", "example", ",", "we", "can", "see", "in", ":", "mnist_code", "]", "Figure", "[", "reference", "]", "a", ",", "c", "that", "the", "network", "has", "learnt", "to", "use", "one", "of", "the", "axes", "of", "the", "2D", "Gaussian", "code", "to", "explicitly", "encode", "the", "digit", "label", "even", "though", "a", "continuous", "prior", "is", "imposed", ".", "In", "this", "case", ",", "we", "can", "potentially", "get", "a", "much", "better", "separation", "if", "we", "impose", "a", "discrete", "prior", ".", "This", "makes", "this", "architecture", "suitable", "for", "the", "discrete", "vs.", "continuous", "decomposition", "and", "we", "use", "it", "for", "our", "clustering", "and", "semi", "-", "supervised", "learning", "experiments", ".", "In", "the", "case", "of", "the", "location", "-", "dependent", "bias", "(", ":", "mnist_code", "]", "Figure", "[", "reference", "]", "b", ",", "d", ")", ",", "the", "hidden", "code", "is", "encouraged", "to", "learn", "the", "global", "information", "that", "has", "location", "dependent", "information", "such", "as", "low", "-", "frequency", "content", "of", "the", "image", ",", "similar", "to", "what", "the", "hidden", "code", "of", "an", "adversarial", "or", "variational", "autoencoder", "would", "learn", "(", ":", "mnist", "]", "Figure", "[", "reference", "]", "c", ")", ".", "This", "makes", "this", "architecture", "suitable", "for", "the", "global", "vs.", "local", "decomposition", "experiments", "such", "as", ":", "mnist", "]", "Figure", "[", "reference", "]", "a.", "From", ":", "mnist_code", "]", "Figure", "[", "reference", "]", ",", "we", "can", "see", "that", "the", "class", "label", "information", "is", "mostly", "captured", "by", "while", "the", "style", "information", "of", "the", "images", "is", "captured", "by", "both", "and", ".", "This", "decomposition", "of", "information", "has", "also", "been", "studied", "in", "other", "works", "that", "combine", "the", "latent", "variable", "models", "with", "autoregressive", "decoders", "such", "as", "PixelVAE", "pixelvae", "and", "variational", "lossy", "autoencoders", "(", "VLAE", ")", "vlae", ".", "For", "example", ",", "the", "VLAE", "model", "vlae", "proposes", "to", "use", "the", "depth", "of", "the", "PixelCNN", "decoder", "to", "control", "the", "decomposition", "of", "information", ".", "In", "their", "model", ",", "the", "PixelCNN", "decoder", "is", "designed", "to", "have", "a", "shallow", "depth", "(", "small", "local", "receptive", "field", ")", "so", "that", "the", "latent", "code", "is", "forced", "to", "capture", "more", "global", "information", ".", "This", "approach", "is", "very", "similar", "to", "our", "example", "of", "the", "PixelGAN", "autoencoder", "in", ":", "mnist", "]", "Figure", "[", "reference", "]", ".", "However", ",", "the", "question", "that", "has", "remained", "unanswered", "is", "whether", "it", "is", "possible", "to", "achieve", "a", "complete", "decomposition", "of", "content", "and", "style", "in", "an", "unsupervised", "fashion", ",", "where", "the", "class", "label", "or", "discrete", "structure", "information", "is", "encoded", "in", "the", "latent", "code", ",", "and", "the", "remaining", "continuous", "structure", "such", "as", "style", "is", "captured", "by", "a", "powerful", "and", "deep", "PixelCNN", "decoder", ".", "This", "kind", "of", "decomposition", "is", "particularly", "interesting", "as", "it", "can", "be", "directly", "used", "for", "clustering", "and", "semi", "-", "supervised", "classification", ".", "In", "the", "next", "section", ",", "we", "show", "that", "we", "can", "learn", "this", "decomposition", "of", "content", "and", "style", "by", "imposing", "a", "categorical", "distribution", "on", "the", "latent", "representation", "using", "adversarial", "training", ".", "Note", "that", "this", "discrete", "vs.", "continuous", "decomposition", "is", "very", "different", "from", "the", "global", "vs.", "local", "decomposition", ",", "because", "a", "continuous", "factor", "of", "variation", "such", "as", "style", "can", "have", "both", "global", "and", "local", "effect", "on", "the", "image", ".", "Indeed", ",", "in", "order", "to", "achieve", "the", "discrete", "vs.", "continuous", "decomposition", ",", "we", "have", "to", "use", "very", "deep", "and", "powerful", "PixelCNN", "decoders", "(", "up", "to", "20", "residual", "blocks", ")", "to", "capture", "both", "the", "global", "and", "local", "statistics", "of", "the", "style", "by", "the", "PixelCNN", "while", "the", "discrete", "content", "of", "the", "image", "is", "captured", "by", "the", "categorical", "latent", "variable", ".", "subsection", ":", "PixelGAN", "Autoencoders", "with", "Categorical", "Priors", "In", "this", "section", ",", "we", "present", "an", "architecture", "of", "the", "PixelGAN", "autoencoder", "that", "can", "separate", "the", "discrete", "information", "(", "e.g.", ",", "class", "label", ")", "from", "the", "continuous", "information", "(", "e.g.", ",", "style", "information", ")", "in", "the", "images", ".", "We", "then", "show", "how", "our", "architecture", "can", "be", "naturally", "adopted", "for", "the", "semi", "-", "supervised", "settings", ".", "The", "architecture", "that", "we", "use", "is", "similar", "to", ":", "pixelgan_gaussian", "]", "Figure", "[", "reference", "]", ",", "with", "the", "difference", "that", "we", "impose", "a", "categorical", "distribution", "as", "the", "prior", "rather", "the", "Gaussian", "distribution", "(", ":", "pixelgan_cat", "]", "Figure", "[", "reference", "]", ")", "and", "also", "use", "the", "location", "-", "independent", "bias", "architecture", ".", "Another", "difference", "is", "that", "we", "use", "a", "convolutional", "network", "as", "the", "inference", "network", "to", "encourage", "the", "encoder", "to", "preserve", "the", "content", "and", "lose", "the", "style", "information", "of", "the", "image", ".", "The", "inference", "network", "has", "a", "softmax", "output", "and", "predicts", "a", "one", "-", "hot", "vector", "whose", "dimension", "is", "the", "number", "of", "discrete", "labels", "or", "categories", "that", "we", "wish", "the", "data", "to", "be", "clustered", "into", ".", "The", "adversarial", "network", "is", "trained", "directly", "on", "the", "continuous", "probability", "outputs", "of", "the", "softmax", "layer", "of", "the", "encoder", ".", "Imposing", "a", "categorical", "distribution", "at", "the", "output", "of", "the", "encoder", "imposes", "two", "constraints", ".", "The", "first", "constraint", "is", "that", "the", "encoder", "has", "to", "make", "confident", "decisions", "about", "the", "class", "labels", "of", "the", "inputs", ".", "The", "adversarial", "training", "pushes", "the", "output", "of", "the", "encoder", "to", "the", "corners", "of", "the", "softmax", "simplex", ",", "by", "which", "it", "ensures", "that", "the", "autoencoder", "can", "not", "use", "the", "latent", "vector", "to", "carry", "any", "continuous", "style", "information", ".", "The", "second", "constraint", "imposed", "by", "adversarial", "training", "is", "that", "the", "aggregated", "posterior", "distribution", "of", "should", "match", "the", "categorical", "prior", "distribution", "with", "uniform", "outcome", "probabilities", ".", "This", "constraint", "enforces", "the", "encoder", "to", "evenly", "distribute", "the", "class", "labels", "across", "the", "corners", "of", "the", "softmax", "simplex", ".", "Because", "of", "these", "constraints", ",", "the", "latent", "variable", "will", "only", "capture", "the", "discrete", "content", "of", "the", "image", "and", "all", "the", "continuous", "style", "information", "will", "be", "captured", "by", "the", "autoregressive", "decoder", ".", "In", "order", "to", "better", "understand", "and", "visualize", "the", "effect", "of", "the", "adversarial", "training", "on", "shaping", "the", "hidden", "code", "distribution", ",", "we", "train", "a", "PixelGAN", "autoencoder", "on", "the", "first", "three", "digits", "of", "MNIST", "(", "18000", "training", "and", "3000", "test", "points", ")", "and", "choose", "the", "number", "of", "clusters", "to", "be", "3", ".", "Suppose", "is", "the", "hidden", "code", "which", "in", "this", "case", "is", "the", "output", "probabilities", "of", "the", "softmax", "layer", "of", "the", "inference", "network", ".", "In", ":", "pixelgan_cluster_toy", "]", "Figure", "[", "reference", "]", "a", ",", "we", "project", "the", "3D", "softmax", "simplex", "of", "onto", "a", "2D", "triangle", "and", "plot", "the", "hidden", "codes", "of", "the", "training", "examples", "when", "no", "distribution", "is", "imposed", "on", "the", "hidden", "code", ".", "We", "can", "see", "from", "this", "figure", "that", "the", "network", "has", "learnt", "to", "use", "the", "surface", "of", "the", "softmax", "simplex", "to", "encode", "style", "information", "of", "the", "digits", "and", "thus", "the", "three", "corners", "of", "the", "simplex", "do", "not", "have", "any", "meaningful", "interpretation", ".", ":", "pixelgan_cluster_toy", "]", "Figure", "[", "reference", "]", "b", "corresponds", "to", "the", "code", "space", "of", "the", "same", "network", "when", "a", "categorical", "distribution", "is", "imposed", "using", "the", "adversarial", "training", ".", "In", "this", "case", ",", "we", "can", "see", "the", "network", "has", "successfully", "learnt", "to", "encode", "the", "label", "information", "of", "the", "three", "digits", "in", "the", "three", "corners", "of", "the", "simplex", ",", "and", "all", "the", "style", "information", "has", "been", "separately", "captured", "by", "the", "autoregressive", "decoder", ".", "This", "network", "achieves", "an", "almost", "perfect", "test", "error", "-", "rate", "of", "on", "the", "first", "three", "digits", "of", "MNIST", ",", "even", "though", "it", "is", "trained", "in", "a", "purely", "unsupervised", "fashion", ".", "Once", "the", "PixelGAN", "autoencoder", "is", "trained", ",", "its", "encoder", "can", "be", "used", "for", "clustering", "new", "points", "and", "its", "decoder", "can", "be", "used", "to", "generate", "samples", "from", "each", "cluster", ".", "Figure", "[", "reference", "]", "illustrates", "the", "samples", "of", "the", "PixelGAN", "autoencoder", "trained", "on", "the", "full", "MNIST", "dataset", ".", "The", "number", "of", "clusters", "is", "set", "to", "be", "30", "and", "each", "row", "corresponds", "to", "the", "conditional", "samples", "of", "one", "of", "the", "clusters", "(", "only", "16", "are", "shown", ")", ".", "We", "can", "see", "that", "the", "discrete", "latent", "code", "of", "the", "network", "has", "learnt", "discrete", "factors", "of", "variation", "such", "as", "class", "label", "information", "and", "some", "discrete", "style", "information", ".", "For", "example", "digit", "s", "are", "put", "in", "different", "clusters", "based", "on", "how", "much", "tilted", "they", "are", ".", "The", "network", "is", "also", "assigning", "different", "clusters", "to", "digit", "s", "(", "based", "on", "whether", "they", "have", "a", "loop", ")", "and", "digit", "s", "(", "based", "on", "whether", "they", "have", "a", "dash", "in", "the", "middle", ")", ".", "In", ":", "experiments", ":", "unsup", "]", "Section", "[", "reference", "]", ",", "we", "will", "show", "that", "by", "using", "the", "encoder", "of", "this", "network", ",", "we", "can", "obtain", "about", "5", "%", "error", "rate", "in", "classifying", "digits", "in", "an", "unsupervised", "fashion", ",", "just", "by", "matching", "each", "cluster", "to", "a", "digit", "type", ".", "Semi", "-", "Supervised", "PixelGAN", "Autoencoders", ".", "The", "PixelGAN", "autoencoder", "can", "be", "used", "in", "a", "semi", "-", "supervised", "setting", ".", "In", "order", "to", "incorporate", "the", "label", "information", ",", "we", "add", "a", "semi", "-", "supervised", "training", "phase", ".", "Specifically", ",", "we", "set", "the", "number", "of", "clusters", "to", "be", "the", "same", "as", "the", "number", "of", "class", "labels", "and", "after", "executing", "the", "reconstruction", "and", "the", "adversarial", "phases", "on", "an", "unlabeled", "mini", "-", "batch", ",", "the", "semi", "-", "supervised", "phase", "is", "executed", "on", "a", "labeled", "mini", "-", "batch", ",", "by", "updating", "the", "weights", "of", "the", "encoder", "to", "minimize", "the", "cross", "-", "entropy", "cost", ".", "The", "semi", "-", "supervised", "cost", "also", "reduces", "the", "mode", "-", "missing", "behavior", "of", "the", "GAN", "training", "by", "enforcing", "the", "encoder", "to", "learn", "all", "the", "modes", "of", "the", "categorical", "distribution", ".", "In", ":", "experiments", ":", "semi", "]", "Section", "[", "reference", "]", ",", "we", "will", "evaluate", "the", "performance", "of", "the", "PixelGAN", "autoencoders", "on", "the", "semi", "-", "supervised", "classification", "tasks", ".", "section", ":", "Experiments", "In", "this", "paper", ",", "we", "presented", "the", "PixelGAN", "autoencoder", "as", "a", "generative", "model", ",", "but", "the", "currently", "available", "metrics", "for", "evaluating", "the", "likelihood", "of", "GAN", "-", "based", "generative", "models", "such", "as", "Parzen", "window", "estimate", "are", "fundamentally", "flawed", "theis", ".", "So", "in", "this", "section", ",", "we", "only", "present", "the", "performance", "of", "the", "PixelGAN", "autoencoder", "on", "downstream", "tasks", "such", "as", "unsupervised", "clustering", "and", "semi", "-", "supervised", "classification", ".", "The", "details", "of", "all", "the", "experiments", "can", "be", "found", "in", "endix", ":", "experiment", "]", "Appendix", "[", "reference", "]", ".", "subsection", ":", "Unsupervised", "Clustering", "We", "trained", "a", "PixelGAN", "autoencoder", "in", "an", "unsupervised", "fashion", "on", "the", "MNIST", "dataset", "(", ":", "pixelgan_cluster", "]", "Figure", "[", "reference", "]", ")", ".", "We", "chose", "the", "number", "of", "clusters", "to", "be", "30", "and", "used", "the", "following", "evaluation", "protocol", ":", "once", "the", "training", "is", "done", ",", "for", "each", "cluster", ",", "we", "found", "the", "validation", "example", "that", "maximizes", ",", "and", "assigned", "the", "label", "of", "to", "all", "the", "points", "in", "the", "cluster", ".", "We", "then", "computed", "the", "test", "error", "based", "on", "the", "assigned", "class", "labels", "to", "each", "cluster", ".", "As", "shown", "in", "the", "first", "column", "of", "le", ":", "semi", "]", "Table", "[", "reference", "]", ",", "the", "performance", "of", "PixelGAN", "autoencoders", "is", "on", "par", "with", "other", "GAN", "-", "based", "clustering", "algorithms", "such", "as", "CatGAN", "catgan", ",", "InfoGAN", "infogan", "and", "adversarial", "autoencoders", "aae", ".", "subsection", ":", "Semi", "-", "supervised", "Classification", "le", ":", "semi", "]", "Table", "[", "reference", "]", "and", ":", "plot", "]", "Figure", "[", "reference", "]", "report", "the", "results", "of", "semi", "-", "supervised", "classification", "experiments", "on", "the", "MNIST", ",", "SVHN", "and", "NORB", "datasets", ".", "On", "the", "MNIST", "dataset", "with", "20", ",", "50", "and", "100", "labels", ",", "our", "classification", "results", "are", "highly", "competitive", ".", "Note", "that", "the", "classification", "rate", "of", "unsupervised", "clustering", "of", "MNIST", "is", "better", "than", "semi", "-", "supervised", "MNIST", "with", "20", "labels", ".", "This", "is", "because", "in", "the", "unsupervised", "case", ",", "the", "number", "of", "clusters", "is", "30", ",", "but", "in", "the", "semi", "-", "supervised", "case", ",", "there", "are", "only", "10", "class", "labels", "which", "makes", "it", "more", "likely", "to", "confuse", "two", "digits", ".", "On", "the", "SVHN", "dataset", "with", "500", "and", "1000", "labels", ",", "the", "PixelGAN", "autoencoder", "outperforms", "all", "the", "other", "methods", "except", "the", "recently", "proposed", "temporal", "ensembling", "work", "temporal", "-", "ensembling", "which", "is", "not", "a", "generative", "model", ".", "On", "the", "NORB", "dataset", "with", "1000", "labels", ",", "the", "PixelGAN", "autoencoder", "outperforms", "all", "the", "other", "reported", "results", ".", ":", "disentangle", "]", "Figure", "[", "reference", "]", "shows", "the", "conditional", "samples", "of", "the", "semi", "-", "supervised", "PixelGAN", "autoencoder", "on", "the", "MNIST", ",", "SVHN", "and", "NORB", "datasets", ".", "Each", "column", "of", "this", "figure", "presents", "sampled", "images", "conditioned", "on", "a", "fixed", "one", "-", "hot", "latent", "code", ".", "We", "can", "see", "from", "this", "figure", "that", "the", "PixelGAN", "autoencoder", "can", "achieve", "a", "rather", "clean", "separation", "of", "style", "and", "content", "on", "these", "datasets", "with", "very", "few", "labeled", "data", ".", "section", ":", "Learning", "Cross", "-", "Domain", "Relations", "with", "PixelGAN", "Autoencoders", "In", "this", "section", ",", "we", "discuss", "how", "the", "PixelGAN", "autoencoder", "can", "be", "viewed", "in", "the", "context", "of", "learning", "cross", "-", "domain", "relations", "between", "two", "different", "domains", ".", "We", "also", "describe", "how", "the", "problem", "of", "clustering", "or", "semi", "-", "supervised", "learning", "can", "be", "cast", "as", "the", "problem", "of", "finding", "a", "smooth", "cross", "-", "domain", "mapping", "from", "the", "data", "distribution", "to", "the", "categorical", "distribution", ".", "Recently", "several", "GAN", "-", "based", "methods", "have", "been", "developed", "to", "learn", "a", "cross", "-", "domain", "mapping", "between", "two", "different", "domains", "discogan", ",", "cyclegan", ",", "cross", "-", "domain", "-", "ilya", ",", "aae", ",", "cross", "-", "domain", "-", "nlp", ".", "In", "cross", "-", "domain", "-", "ilya", ",", "an", "unsupervised", "cost", "function", "called", "the", "output", "distribution", "matching", "(", "ODM", ")", "is", "proposed", "to", "find", "a", "cross", "-", "domain", "mapping", "between", "two", "domains", "and", "by", "imposing", "the", "following", "unsupervised", "constraint", "on", "the", "uncorrelated", "samples", "from", "and", ":", "where", "denotes", "the", "distribution", "of", "the", "random", "variable", ".", "The", "adversarial", "training", "is", "proposed", "as", "one", "of", "the", "methods", "for", "matching", "these", "distributions", ".", "If", "we", "have", "access", "to", "a", "few", "labeled", "pairs", ",", "then", "can", "be", "further", "trained", "on", "them", "in", "a", "supervised", "fashion", "to", "satisfy", ".", "For", "example", ",", "in", "speech", "recognition", ",", "we", "want", "to", "find", "a", "cross", "-", "domain", "mapping", "from", "a", "sequence", "of", "phonemes", "to", "a", "sequence", "of", "characters", ".", "By", "optimizing", "the", "ODM", "cost", "function", "in", "odm", "]", "Equation", "[", "reference", "]", ",", "we", "can", "find", "a", "smooth", "function", "that", "takes", "phonemes", "at", "its", "input", "and", "outputs", "a", "sequence", "of", "characters", "that", "respects", "the", "language", "model", ".", "However", ",", "the", "main", "problem", "with", "this", "method", "is", "that", "the", "network", "can", "learn", "to", "ignore", "part", "of", "the", "input", "distribution", "and", "still", "satisfy", "the", "ODM", "cost", "function", "by", "its", "output", "distribution", ".", "This", "problem", "has", "also", "been", "observed", "in", "other", "works", "such", "as", "discogan", ".", "One", "way", "to", "avoid", "this", "problem", "is", "to", "add", "a", "reconstruction", "term", "to", "the", "ODM", "cost", "function", "by", "introducing", "a", "reverse", "mapping", "from", "the", "output", "of", "the", "encoder", "to", "the", "input", "domain", ".", "The", "is", "essentially", "the", "idea", "of", "the", "adversarial", "autoencoder", "aae", "which", "learns", "a", "generative", "model", "by", "finding", "a", "cross", "-", "domain", "mapping", "between", "a", "Gaussian", "distribution", "and", "the", "data", "distribution", ".", "Using", "the", "ODM", "cost", "function", "along", "with", "a", "reconstruction", "term", "to", "learn", "cross", "-", "domain", "relations", "have", "been", "explored", "in", "several", "previous", "works", ".", "For", "example", ",", "InfoGAN", "infogan", "adds", "a", "mutual", "information", "term", "to", "the", "ODM", "cost", "function", "and", "optimizes", "a", "variational", "lower", "bound", "on", "this", "term", ".", "It", "can", "be", "shown", "that", "maximizing", "this", "variational", "bound", "is", "indeed", "minimizing", "the", "reconstruction", "cost", "of", "an", "autoencoder", "i", "m", ".", "Similarly", ",", "in", "cross", "-", "domain", "-", "nlp", ",", "zhangadversarial", ",", "an", "adversarial", "autoencoder", "is", "used", "to", "learn", "the", "cross", "-", "domain", "relations", "of", "the", "vector", "representations", "of", "words", "from", "two", "different", "languages", ".", "The", "architecture", "of", "the", "recent", "works", "of", "DiscoGAN", "discogan", "and", "CycleGAN", "cyclegan", "are", "also", "similar", "to", "an", "adversarial", "autoencoder", "in", "which", "the", "latent", "representation", "is", "enforced", "to", "have", "the", "distribution", "of", "the", "other", "domain", ".", "Here", "we", "describe", "how", "our", "proposed", "PixelGAN", "autoencoder", "can", "be", "potentially", "used", "in", "all", "these", "application", "areas", "to", "learn", "better", "cross", "-", "domain", "relations", ".", "Suppose", "we", "want", "to", "learn", "a", "mapping", "from", "domain", "to", ".", "In", "the", "architecture", "of", ":", "pixelgan_gaussian", "]", "Figure", "[", "reference", "]", ",", "we", "can", "use", "independent", "samples", "of", "at", "the", "input", "and", "instead", "of", "imposing", "a", "Gaussian", "distribution", "on", "the", "latent", "code", ",", "we", "can", "impose", "the", "distribution", "of", "the", "second", "domain", "using", "its", "independent", "samples", ".", "Unlike", "adversarial", "autoencoders", ",", "the", "encoder", "of", "PixelGAN", "autoencoders", "does", "not", "have", "to", "retain", "all", "the", "input", "information", "in", "order", "to", "have", "a", "lossless", "reconstruction", ".", "So", "the", "encoder", "can", "use", "all", "its", "capacity", "to", "learn", "the", "most", "relevant", "mapping", "from", "to", "and", "at", "the", "same", "time", ",", "the", "PixelCNN", "decoder", "can", "capture", "the", "remaining", "information", "that", "has", "been", "lost", "by", "the", "encoder", ".", "We", "can", "adopt", "the", "ODM", "idea", "for", "semi", "-", "supervised", "learning", "by", "assuming", "is", "the", "image", "domain", "and", "is", "the", "label", "domain", "(", ":", "related", "]", "Figure", "[", "reference", "]", "a", ")", ".", "Independent", "samples", "of", "and", "correspond", "to", "samples", "from", "the", "data", "distribution", "and", "the", "categorical", "distribution", ".", "The", "function", "can", "be", "parametrized", "by", "a", "neural", "network", "that", "is", "trained", "to", "satisfy", "the", "ODM", "cost", "function", "by", "matching", "the", "aggregated", "distribution", "to", "the", "categorical", "distribution", "using", "adversarial", "training", ".", "The", "few", "labeled", "examples", "are", "used", "to", "further", "train", "to", "satisfy", ".", "However", ",", "as", "explained", "above", ",", "the", "problem", "with", "this", "method", "is", "that", "the", "network", "can", "learn", "to", "generate", "the", "categorical", "distribution", "by", "ignoring", "some", "part", "of", "the", "input", "distribution", ".", "The", "adversarial", "autoencoder", "(", ":", "related", "]", "Figure", "[", "reference", "]", "b", ")", "solves", "this", "problem", "by", "adding", "an", "inverse", "mapping", "from", "the", "categorical", "distribution", "to", "the", "data", "distribution", ".", "However", ",", "the", "main", "drawback", "of", "the", "adversarial", "autoencoder", "architecture", "is", "that", "due", "to", "the", "reconstruction", "term", ",", "the", "latent", "representation", "now", "has", "to", "model", "all", "the", "underlying", "factors", "of", "variation", "in", "the", "image", ".", "For", "example", ",", "in", "the", "architecture", "of", ":", "related", "]", "Figure", "[", "reference", "]", "b", ",", "while", "we", "are", "only", "interested", "in", "the", "one", "-", "hot", "label", "representation", "to", "do", "semi", "-", "supervised", "learning", ",", "we", "also", "need", "to", "infer", "the", "style", "of", "the", "image", "so", "that", "we", "can", "have", "a", "lossless", "reconstruction", "of", "the", "image", ".", "The", "PixelGAN", "autoencoder", "solves", "this", "problem", "by", "enabling", "the", "encoder", "to", "only", "infer", "the", "factor", "of", "variation", "that", "we", "are", "interested", "in", "(", "i.e.", ",", "label", "information", ")", ",", "while", "the", "remaining", "structure", "of", "the", "input", "(", "i.e.", ",", "style", "information", ")", "is", "automatically", "captured", "by", "the", "autoregressive", "decoder", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "proposed", "the", "PixelGAN", "autoencoder", ",", "which", "is", "a", "generative", "autoencoder", "that", "combines", "a", "generative", "PixelCNN", "with", "a", "GAN", "inference", "network", "that", "can", "impose", "arbitrary", "priors", "on", "the", "latent", "code", ".", "We", "showed", "that", "imposing", "different", "distributions", "as", "the", "prior", "enables", "us", "to", "learn", "a", "latent", "representation", "that", "captures", "the", "type", "of", "statistics", "that", "we", "care", "about", ",", "while", "the", "remaining", "structure", "of", "the", "image", "is", "captured", "by", "the", "PixelCNN", "decoder", ".", "Specifically", ",", "by", "imposing", "a", "Gaussian", "prior", ",", "we", "were", "able", "to", "disentangle", "the", "low", "-", "frequency", "and", "high", "-", "frequency", "statistics", "of", "the", "images", ",", "and", "by", "imposing", "a", "categorical", "prior", "we", "were", "able", "to", "disentangle", "the", "style", "and", "content", "of", "images", "and", "learn", "representations", "that", "are", "specifically", "useful", "for", "clustering", "and", "semi", "-", "supervised", "learning", "tasks", ".", "While", "the", "main", "focus", "of", "this", "paper", "was", "to", "demonstrate", "the", "application", "of", "PixelGAN", "autoencoders", "in", "downstream", "tasks", "such", "as", "semi", "-", "supervised", "learning", ",", "we", "discussed", "how", "these", "architectures", "have", "many", "other", "potentials", "such", "as", "learning", "cross", "-", "domain", "relations", "between", "two", "different", "domains", ".", "section", ":", "Acknowledgments", "We", "would", "like", "to", "thank", "Nathan", "Killoran", "for", "helpful", "discussions", ".", "We", "also", "thank", "NVIDIA", "for", "GPU", "donations", ".", "bibliography", ":", "References", "section", ":", "Implementation", "Details", "In", "this", "section", ",", "we", "describe", "two", "important", "architecture", "design", "choices", "for", "training", "PixelGAN", "autoencoders", ".", "subsection", ":", "Input", "noise", "In", "all", "the", "semi", "-", "supervised", "experiments", ",", "we", "found", "it", "crucial", "to", "use", "the", "universal", "approximator", "posterior", "discussed", "in", ":", "pixelgan", "]", "Section", "[", "reference", "]", ",", "as", "opposed", "to", "a", "deterministic", "posterior", ".", "Specifically", ",", "the", "input", "noise", "that", "we", "use", "is", "an", "additive", "Gaussian", "noise", ",", "which", "results", "in", "a", "posterior", "distribution", "that", "is", "more", "expressive", "than", "that", "of", "a", "model", "without", "the", "input", "corruption", ".", "This", "is", "similar", "to", "the", "denoising", "criterion", "idea", "proposed", "in", "denoising", "-", "vae", ".", "We", "believe", "this", "additive", "noise", "is", "also", "playing", "an", "important", "role", "in", "preventing", "the", "mode", "-", "missing", "behavior", "of", "the", "GAN", "when", "imposing", "a", "degenerate", "distribution", "such", "as", "the", "categorical", "distribution", ".", "Similar", "related", "ideas", "have", "been", "used", "to", "stabilize", "GAN", "training", "such", "as", "instance", "noise", "instance", "or", "one", "-", "sided", "label", "noise", "improved", "-", "gan", ".", "subsection", ":", "Conditioning", "of", "PixelCNN", "There", "are", "three", "methods", "to", "implement", "how", "the", "PixelCNN", "conditions", "on", "the", "latent", "vector", ".", "Location", "-", "Invariant", "Bias", ".", "This", "is", "the", "method", "that", "was", "proposed", "in", "the", "conditional", "PixelCNN", "model", "pixelcnn", ".", "Suppose", "the", "size", "of", "the", "convolutional", "layer", "of", "the", "decoder", "is", "(", "batch", ",", "width", ",", "height", ",", "channels", ")", ".", "Then", "the", "PixelCNN", "can", "use", "a", "linear", "mapping", "to", "convert", "the", "conditioning", "tensor", "of", "size", "(", "batch", ",", "condition_size", ")", "to", "generate", "a", "tensor", "of", "size", "(", "batch", ",", "channels", ")", "that", "is", "then", "broadcasted", "and", "added", "to", "the", "feature", "maps", "of", "all", "the", "layers", "of", "the", "PixelCNN", "decoder", "as", "an", "adaptive", "bias", ".", "In", "this", "method", ",", "the", "hidden", "code", "is", "encouraged", "to", "learn", "the", "global", "information", "that", "is", "location", "-", "invariant", "(", "the", "what", "information", "and", "not", "the", "where", "information", ")", "such", "as", "the", "class", "label", "information", ".", "We", "use", "this", "method", "in", "all", "the", "clustering", "and", "semi", "-", "supervised", "learning", "experiments", ".", "Location", "-", "Dependent", "Bias", ".", "Suppose", "the", "size", "of", "the", "convolutional", "layer", "of", "the", "PixelCNN", "decoder", "is", "(", "batch", ",", "width", ",", "height", ",", "channels", ")", ".", "Then", "the", "PixelCNN", "can", "use", "a", "one", "layer", "neural", "network", "to", "convert", "the", "conditioning", "tensor", "of", "size", "(", "batch", ",", "condition_size", ")", "to", "generate", "a", "spatial", "tensor", "of", "size", "(", "batch", ",", "width", ",", "height", ",", "k", ")", "followed", "by", "a", "convolutional", "layer", "to", "construct", "a", "tensor", "of", "size", "(", "batch", ",", "width", ",", "height", ",", "channels", ")", "that", "is", "then", "added", "only", "to", "the", "feature", "maps", "of", "the", "first", "layer", "of", "the", "decoder", "as", "an", "adaptive", "bias", "(", "similar", "to", "the", "VPN", "model", "vpn", ")", ".", "When", ",", "we", "can", "simply", "broadcast", "the", "tensor", "of", "size", "(", "batch", ",", "width", ",", "height", ",", "k=1", ")", "to", "get", "a", "tensor", "of", "size", "(", "batch", ",", "width", ",", "height", ",", "channels", ")", "instead", "of", "using", "the", "convolution", ".", "In", "this", "method", ",", "the", "latent", "vector", "has", "spatial", "and", "location", "-", "dependent", "information", "within", "the", "feature", "map", ".", "This", "is", "the", "method", "that", "we", "used", "in", "experiments", "of", ":", "mnist", "]", "Figure", "[", "reference", "]", "a.", "Input", "Channel", ".", "Another", "method", "for", "conditioning", "is", "proposed", "in", "the", "PixelVAE", "pixelvae", "and", "the", "variational", "lossy", "autoencoder", "(", "VLAE", ")", "vlae", ".", "In", "this", "method", ",", "first", "a", "tensor", "of", "size", "(", "batch", ",", "width", ",", "height", ",", "k", ")", "is", "constructed", "using", "the", "conditioning", "tensor", "similar", "to", "the", "location", "-", "dependent", "bias", ".", "This", "tensor", "is", "then", "concatenated", "to", "the", "input", "of", "the", "PixelCNN", ".", "The", "performance", "and", "computational", "complexity", "of", "this", "method", "is", "very", "similar", "to", "that", "of", "the", "location", "-", "dependent", "bias", "method", ".", "section", ":", "Experiment", "Details", "We", "used", "TensorFlow", "tensorflow2015", "-", "whitepaper", "in", "all", "of", "our", "experiments", ".", "As", "suggested", "in", "gan", ",", "in", "order", "to", "improve", "the", "stability", "of", "GAN", "training", ",", "the", "generator", "of", "the", "GAN", "in", "all", "our", "experiments", "is", "trained", "to", "maximize", "rather", "than", "minimizing", ".", "subsection", ":", "MNIST", "Dataset", "The", "MNIST", "dataset", "has", "50", "K", "training", "points", ",", "10", "K", "validation", "points", "and", "10", "K", "test", "points", ".", "We", "perform", "experiments", "on", "both", "the", "binary", "MNIST", "and", "the", "real", "-", "valued", "MNIST", ".", "In", "the", "real", "valued", "MNIST", "experiments", ",", "we", "subtract", "127.5", "from", "the", "data", "points", "and", "then", "divide", "them", "by", "127.5", "and", "use", "the", "discretized", "logistic", "mixture", "likelihood", "pixelcnn", "++", "as", "the", "cost", "function", "for", "the", "PixelCNN", ".", "In", "the", "case", "of", "binary", "MNIST", ",", "the", "data", "points", "are", "binarized", "by", "setting", "pixel", "values", "larger", "than", "0.5", "to", "1", ",", "and", "values", "smaller", "than", "0.5", "to", "0", ".", "subsubsection", ":", "PixelGAN", "Autoencoders", "with", "Gaussian", "Prior", "on", "MNIST", "Here", "we", "describe", "the", "model", "architecture", "used", "for", "training", "the", "PixelGAN", "autoencoder", "with", "a", "Gaussian", "prior", "on", "the", "binary", "MNIST", "dataset", "in", ":", "mnist", "]", "Figure", "[", "reference", "]", "a.", "The", "PixelCNN", "decoder", "uses", "both", "the", "vertical", "and", "horizontal", "stacks", "similar", "to", "pixelcnn", ".", "The", "cost", "function", "of", "the", "PixelCNN", "is", "the", "cross", "-", "entropy", "cost", "function", ".", "The", "PixelCNN", "uses", "the", "location", "-", "dependent", "bias", "as", "described", "in", "endix", ":", "conditioning_of_pixelcnn", "]", "Appendix", "[", "reference", "]", ".", "Specifically", ",", "a", "tensor", "of", "size", "(", "batch", ",", "width", ",", "height", ",", "1", ")", "is", "constructed", "from", "the", "conditioning", "vector", "by", "using", "a", "one", "-", "layer", "neural", "network", "with", "1000", "hidden", "units", ",", "ReLU", "activation", "and", "linear", "output", ".", "This", "tensor", "is", "then", "broadcasted", "and", "added", "only", "to", "the", "feature", "maps", "of", "the", "first", "layer", "of", "the", "PixelCNN", "decoder", ".", "The", "PixelCNN", "is", "designed", "to", "have", "a", "local", "receptive", "field", "by", "having", "3", "residual", "blocks", "(", "filter", "size", "of", "3x5", ",", "32", "feature", "maps", ",", "ReLU", "non", "-", "linearity", "as", "in", "pixelcnn", ")", ".", "The", "adversarial", "discriminator", "has", "two", "layers", "of", "2000", "hidden", "units", "with", "ReLU", "activation", "function", ".", "The", "encoder", "architecture", "has", "two", "fully", "-", "connected", "layers", "of", "size", "2000", "with", "ReLU", "non", "-", "linearity", ".", "The", "last", "layer", "of", "the", "encoder", "has", "a", "linear", "activation", "function", ".", "On", "the", "latent", "representation", "of", "size", ",", "we", "impose", "a", "Gaussian", "distribution", "with", "standard", "deviation", "of", ".", "We", "used", "the", "gradient", "descent", "with", "momentum", "algorithm", "for", "optimizing", "all", "the", "cost", "functions", "of", "the", "network", ".", "For", "the", "PixelCNN", "reconstruction", "cost", ",", "we", "used", "the", "learning", "rate", "of", "0.001", "and", "the", "momentum", "value", "of", "0.9", ".", "After", "25", "epochs", "we", "reduce", "the", "learning", "rate", "to", "0.0001", ".", "For", "both", "of", "the", "generator", "and", "the", "discriminator", "costs", ",", "the", "learning", "rates", "and", "the", "momentum", "values", "were", "set", "to", "0.1", ".", "subsubsection", ":", "Unsupervised", "Clustering", "of", "MNIST", "Here", "we", "describe", "the", "model", "architecture", "used", "for", "clustering", "the", "binary", "MNIST", "dataset", "in", ":", "pixelgan_cluster", "]", "Figure", "[", "reference", "]", "and", ":", "experiments", ":", "unsup", "]", "Section", "[", "reference", "]", ".", "The", "PixelCNN", "decoder", "uses", "both", "the", "vertical", "and", "horizontal", "stacks", "similar", "to", "pixelcnn", ".", "The", "cost", "function", "of", "the", "PixelCNN", "is", "the", "cross", "-", "entropy", "cost", "function", ".", "The", "PixelCNN", "uses", "the", "location", "-", "invariant", "bias", "as", "described", "in", "endix", ":", "conditioning_of_pixelcnn", "]", "Appendix", "[", "reference", "]", "and", "has", "15", "residual", "blocks", "(", "filter", "size", "of", "3x5", ",", "32", "feature", "maps", ",", "ReLU", "non", "-", "linearity", "as", "in", "pixelcnn", ")", ".", "The", "adversarial", "discriminator", "has", "two", "layers", "of", "3000", "hidden", "units", "with", "ReLU", "activation", "function", ".", "The", "encoder", "architecture", "has", "a", "convolutional", "layer", "(", "filter", "size", "of", "7", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", ",", "followed", "by", "another", "convolutional", "layer", "(", "filter", "size", "of", "7", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", "with", "no", "fully", "-", "connected", "layer", ".", "The", "last", "layer", "of", "the", "encoder", "has", "the", "softmax", "activation", "function", ".", "We", "found", "it", "important", "to", "use", "batch", "-", "normalization", "batch", "for", "all", "the", "layers", "of", "the", "encoder", "including", "the", "softmax", "layer", ".", "The", "number", "of", "clusters", "is", "chosen", "to", "be", ".", "The", "clusters", "are", "represented", "by", "a", "discrete", "one", "-", "hot", "variable", "of", "size", "30", ".", "On", "the", "continuous", "probability", "output", "of", "the", "softmax", ",", "we", "impose", "a", "categorical", "distribution", "with", "uniform", "probabilities", ".", "We", "use", "Adam", "Adam", "optimizer", "with", "learning", "rate", "of", "for", "optimizing", "the", "PixelCNN", "reconstruction", "cost", "function", ",", "but", "we", "found", "it", "important", "to", "use", "the", "gradient", "descent", "with", "momentum", "algorithm", "for", "optimizing", "the", "generator", "and", "the", "discriminator", "costs", "of", "the", "adversarial", "network", ".", "For", "both", "of", "the", "generator", "and", "the", "discriminator", "costs", ",", "the", "momentum", "values", "were", "set", "to", "0.1", "and", "the", "learning", "rates", "were", "set", "to", "0.01", ".", "We", "use", "an", "input", "dropout", "noise", "with", "the", "keep", "probability", "of", "at", "the", "input", "layer", "and", "only", "at", "the", "training", "time", ".", "The", "model", "architecture", "used", "for", ":", "pixelgan_cluster_toy", "]", "Figure", "[", "reference", "]", "is", "the", "same", "as", "this", "architecture", "except", "that", "the", "number", "of", "clusters", "is", "chosen", "to", "be", ".", "subsubsection", ":", "Semi", "-", "Supervised", "MNIST", "We", "performed", "semi", "-", "supervised", "learning", "experiments", "on", "both", "binary", "and", "real", "-", "valued", "MNIST", "dataset", ".", "We", "found", "that", "the", "semi", "-", "supervised", "error", "-", "rate", "of", "the", "real", "-", "valued", "MNIST", "is", "roughly", "the", "same", "as", "the", "binary", "MNIST", "(", "about", "1.10", "%", "with", "100", "labels", ")", ",", "but", "it", "takes", "longer", "to", "train", "due", "to", "the", "logistic", "mixture", "likelihood", "cost", "function", "pixelcnn", "++", ".", "So", "in", "le", ":", "semi", "]", "Table", "[", "reference", "]", ",", "we", "only", "report", "the", "performance", "with", "the", "binary", "MNIST", ",", "but", "in", ":", "disentangle", "]", "Figure", "[", "reference", "]", "b", "we", "are", "showing", "the", "samples", "of", "the", "real", "-", "valued", "MNIST", "with", "100", "labels", ".", "Binary", "MNIST", ".", "Here", "we", "describe", "the", "model", "architecture", "used", "for", "the", "semi", "-", "supervised", "learning", "experiments", "on", "the", "binary", "MNIST", "in", ":", "experiments", ":", "semi", "]", "Section", "[", "reference", "]", "and", "le", ":", "semi", "]", "Table", "[", "reference", "]", ".", "The", "PixelCNN", "decoder", "uses", "both", "the", "vertical", "and", "horizontal", "stacks", "similar", "to", "pixelcnn", "and", "uses", "the", "cross", "-", "entropy", "cost", "function", ".", "The", "PixelCNN", "uses", "the", "location", "-", "invariant", "bias", "as", "described", "in", "endix", ":", "conditioning_of_pixelcnn", "]", "Appendix", "[", "reference", "]", ".", "The", "PixelCNN", "has", "6", "residual", "blocks", "(", "filter", "size", "of", "3x5", ",", "32", "feature", "maps", ",", "ReLU", "non", "-", "linearity", "as", "in", "pixelcnn", ")", ".", "The", "adversarial", "discriminator", "has", "two", "layers", "of", "1000", "hidden", "units", "with", "ReLU", "activation", "function", ".", "The", "encoder", "architecture", "has", "three", "convolutional", "layers", "(", "filter", "size", "of", "5", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", ",", "followed", "by", "another", "three", "convolutional", "layers", "(", "filter", "size", "of", "5", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", "with", "no", "fully", "-", "connected", "layer", ".", "The", "last", "layer", "of", "the", "encoder", "has", "the", "softmax", "activation", "function", ".", "All", "the", "convolutional", "layers", "of", "the", "encoder", "except", "the", "softmax", "layer", "use", "batch", "-", "normalization", "batch", ".", "On", "the", "latent", "representation", ",", "we", "impose", "a", "categorical", "distribution", "with", "uniform", "probabilities", ".", "The", "semi", "-", "supervised", "cost", "is", "the", "cross", "-", "entropy", "cost", "function", "at", "the", "output", "of", ".", "We", "use", "Adam", "Adam", "optimizer", "with", "learning", "rate", "of", "for", "optimizing", "the", "PixelCNN", "cost", "and", "the", "cross", "-", "entropy", "cost", ",", "but", "we", "found", "it", "important", "to", "use", "the", "gradient", "descent", "with", "momentum", "algorithm", "for", "optimizing", "the", "generator", "and", "the", "discriminator", "costs", "of", "the", "adversarial", "network", ".", "For", "both", "of", "the", "generator", "and", "the", "discriminator", "costs", ",", "the", "momentum", "values", "were", "set", "to", "0.1", "and", "the", "learning", "rates", "were", "set", "to", "0.1", ".", "We", "add", "a", "Gaussian", "noise", "with", "standard", "deviation", "of", "to", "the", "input", "layer", "as", "described", "in", "endix", ":", "input_noise", "]", "Appendix", "[", "reference", "]", ".", "The", "labeled", "examples", "were", "chosen", "at", "random", "but", "evenly", "distributed", "across", "the", "classes", ".", "Real", "-", "valued", "MNIST", ".", "Here", "we", "describe", "the", "model", "architecture", "used", "for", "the", "semi", "-", "supervised", "learning", "experiments", "on", "the", "real", "-", "valued", "MNIST", "in", ":", "disentangle", "]", "Figure", "[", "reference", "]", "b.", "The", "PixelCNN", "decoder", "uses", "both", "the", "vertical", "and", "horizontal", "stacks", "similar", "to", "pixelcnn", "and", "uses", "a", "discretized", "logistic", "mixture", "likelihood", "cost", "function", "with", "10", "logistic", "distribution", "as", "proposed", "in", "pixelcnn", "++", ".", "The", "PixelCNN", "uses", "the", "location", "-", "invariant", "bias", "as", "described", "in", "endix", ":", "conditioning_of_pixelcnn", "]", "Appendix", "[", "reference", "]", ".", "The", "PixelCNN", "has", "20", "residual", "blocks", "(", "filter", "size", "of", "2x3", ",", "64", "feature", "maps", ",", "gated", "sigmoid", "-", "tanh", "non", "-", "linearity", "as", "in", "pixelcnn", ")", ".", "The", "adversarial", "discriminator", "has", "two", "layers", "of", "1000", "hidden", "units", "with", "ReLU", "activation", "function", ".", "The", "encoder", "architecture", "has", "three", "convolutional", "layers", "(", "filter", "size", "of", "5", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", ",", "followed", "by", "another", "three", "convolutional", "layers", "(", "filter", "size", "of", "5", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", "with", "no", "fully", "-", "connected", "layer", ".", "The", "last", "layer", "of", "the", "encoder", "has", "the", "softmax", "activation", "function", ".", "All", "the", "convolutional", "layers", "of", "the", "encoder", "except", "the", "softmax", "layer", "use", "batch", "-", "normalization", "batch", ".", "On", "the", "latent", "representation", ",", "we", "impose", "a", "categorical", "distribution", "with", "uniform", "probabilities", ".", "The", "semi", "-", "supervised", "cost", "is", "the", "cross", "-", "entropy", "cost", "function", "at", "the", "output", "of", ".", "We", "use", "Adam", "Adam", "optimizer", "with", "learning", "rate", "of", "for", "optimizing", "the", "PixelCNN", "cost", "and", "the", "cross", "-", "entropy", "cost", ",", "but", "we", "found", "it", "important", "to", "use", "the", "gradient", "descent", "with", "momentum", "algorithm", "for", "optimizing", "the", "generator", "and", "the", "discriminator", "costs", "of", "the", "adversarial", "network", ".", "For", "both", "of", "the", "generator", "and", "the", "discriminator", "costs", ",", "the", "momentum", "values", "were", "set", "to", "0.1", "and", "the", "learning", "rates", "were", "set", "to", "0.1", ".", "After", "150", "epochs", ",", "we", "divide", "all", "the", "learning", "rates", "by", "10", ".", "We", "add", "a", "Gaussian", "noise", "with", "standard", "deviation", "of", "to", "the", "input", "layer", "as", "described", "in", "endix", ":", "input_noise", "]", "Appendix", "[", "reference", "]", ".", "The", "labeled", "examples", "were", "chosen", "at", "random", "but", "evenly", "distributed", "across", "the", "classes", ".", "subsection", ":", "SVHN", "Dataset", "The", "SVHN", "dataset", "has", "about", "530", "K", "training", "points", "and", "26", "K", "test", "points", ".", "We", "use", "10", "K", "points", "for", "the", "validation", "set", ".", "Similar", "to", "vat", ",", "we", "downsample", "the", "images", "from", "to", "and", "then", "subtracte", "127.5", "from", "the", "data", "points", "and", "then", "divide", "them", "by", "127.5", ".", "subsubsection", ":", "Semi", "-", "Supervised", "SVHN", "Here", "we", "describe", "the", "model", "architecture", "used", "for", "the", "semi", "-", "supervised", "learning", "experiments", "on", "the", "SVHN", "dataset", "in", ":", "experiments", ":", "semi", "]", "Section", "[", "reference", "]", ".", "The", "PixelCNN", "decoder", "uses", "both", "the", "vertical", "and", "horizontal", "stacks", "similar", "to", "pixelcnn", ".", "The", "cost", "function", "of", "the", "PixelCNN", "is", "a", "discretized", "logistic", "mixture", "likelihood", "cost", "function", "with", "10", "logistic", "distribution", "as", "proposed", "in", "pixelcnn", "++", ".", "The", "PixelCNN", "uses", "the", "location", "-", "invariant", "bias", "as", "described", "in", "endix", ":", "conditioning_of_pixelcnn", "]", "Appendix", "[", "reference", "]", "and", "has", "20", "residual", "blocks", "(", "filter", "size", "of", "3x5", ",", "32", "feature", "maps", ",", "gated", "sigmoid", "-", "tanh", "non", "-", "linearity", "as", "in", "pixelcnn", ")", ".", "The", "adversarial", "discriminator", "has", "two", "layers", "of", "1000", "hidden", "units", "with", "ReLU", "activation", "function", ".", "The", "encoder", "architecture", "has", "two", "convolutional", "layers", "(", "filter", "size", "of", "5", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", ",", "followed", "by", "another", "two", "convolutional", "layers", "(", "filter", "size", "of", "5", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", "with", "no", "fully", "-", "connected", "layer", ".", "The", "last", "layer", "of", "the", "encoder", "has", "the", "softmax", "activation", "function", ".", "All", "the", "convolutional", "layers", "of", "the", "encoder", "except", "the", "softmax", "layer", "use", "batch", "-", "normalization", "batch", ".", "On", "the", "latent", "representation", ",", "we", "impose", "a", "categorical", "distribution", "with", "uniform", "probabilities", ".", "The", "semi", "-", "supervised", "cost", "is", "the", "cross", "-", "entropy", "cost", "function", "at", "the", "output", "of", ".", "We", "use", "Adam", "Adam", "optimizer", "for", "optimizing", "all", "the", "cost", "function", ".", "For", "the", "PixelCNN", "cost", "and", "the", "cross", "-", "entropy", "cost", "we", "use", "the", "learning", "rate", "of", "and", "for", "the", "generator", "and", "the", "discriminator", "costs", "of", "the", "adversarial", "network", "we", "use", "the", "learning", "rate", "of", ".", "We", "add", "a", "Gaussian", "noise", "with", "standard", "deviation", "of", "to", "the", "input", "layer", "as", "described", "in", "endix", ":", "input_noise", "]", "Appendix", "[", "reference", "]", ".", "subsection", ":", "NORB", "Dataset", "The", "NORB", "dataset", "has", "about", "24", "K", "training", "points", "and", "24", "K", "test", "points", ".", "We", "use", "4", "K", "points", "for", "the", "validation", "set", ".", "This", "dataset", "has", "5", "object", "categories", ":", "animals", ",", "human", "figures", ",", "airplanes", ",", "trucks", "and", "cars", ".", "We", "downsample", "the", "images", "to", "have", "the", "size", "of", ",", "subtract", "127.5", "from", "the", "data", "points", "and", "then", "divide", "them", "by", "127.5", ".", "subsubsection", ":", "Semi", "-", "Supervised", "NORB", "The", "PixelCNN", "decoder", "uses", "both", "the", "vertical", "and", "horizontal", "stacks", "similar", "to", "pixelcnn", ".", "The", "cost", "function", "of", "the", "PixelCNN", "is", "a", "discretized", "logistic", "mixture", "likelihood", "cost", "function", "with", "10", "logistic", "distribution", "as", "proposed", "in", "pixelcnn", "++", ".", "The", "PixelCNN", "uses", "the", "location", "-", "invariant", "bias", "as", "described", "in", "endix", ":", "conditioning_of_pixelcnn", "]", "Appendix", "[", "reference", "]", "and", "has", "15", "residual", "blocks", "(", "filter", "size", "of", "3x5", ",", "32", "feature", "maps", ",", "gated", "sigmoid", "-", "tanh", "non", "-", "linearity", "as", "in", "pixelcnn", ")", ".", "The", "adversarial", "discriminator", "has", "two", "layers", "of", "1000", "hidden", "units", "with", "ReLU", "activation", "function", ".", "The", "encoder", "architecture", "has", "a", "convolutional", "layer", "(", "filter", "size", "of", "7", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", ",", "followed", "by", "another", "convolutional", "layer", "(", "filter", "size", "of", "7", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", ",", "followed", "by", "another", "convolutional", "layer", "(", "filter", "size", "of", "7", ",", "32", "feature", "maps", ",", "ReLU", "activation", ")", "and", "a", "max", "-", "pooling", "layer", "(", "pooling", "size", "2", ")", "with", "no", "fully", "-", "connected", "layer", ".", "The", "last", "layer", "of", "the", "encoder", "has", "the", "softmax", "activation", "function", ".", "All", "the", "convolutional", "layers", "of", "the", "encoder", "except", "the", "softmax", "layer", "use", "batch", "-", "normalization", "batch", ".", "On", "the", "latent", "representation", ",", "we", "impose", "a", "categorical", "distribution", "with", "uniform", "probabilities", ".", "The", "semi", "-", "supervised", "cost", "is", "the", "cross", "-", "entropy", "cost", "function", "at", "the", "output", "of", ".", "We", "use", "Adam", "Adam", "optimizer", "for", "optimizing", "all", "the", "cost", "function", ".", "For", "the", "PixelCNN", "cost", "and", "the", "cross", "-", "entropy", "cost", "we", "use", "the", "learning", "rate", "of", "and", "for", "the", "generator", "and", "the", "discriminator", "costs", "of", "the", "adversarial", "network", "we", "use", "the", "learning", "rate", "of", ".", "We", "add", "a", "Gaussian", "noise", "with", "standard", "deviation", "of", "to", "the", "input", "layer", "as", "described", "in", "endix", ":", "input_noise", "]", "Appendix", "[", "reference", "]", ".", "The", "labeled", "examples", "were", "chosen", "at", "random", "but", "evenly", "distributed", "across", "the", "classes", ".", "In", "the", "case", "of", "NORB", "with", "1000", "labels", ",", "the", "test", "error", "after", "10", "epochs", "is", "12.97", "%", ",", "after", "100", "epochs", "is", "11.63", "%", "and", "after", "500", "epochs", "is", "8.17", "%", "."], "entity": [[2, 4, ["PixelGAN", "Autoencoders"], "Method"], [19, 21, ["generative", "autoencoder"], "Method"], [24, 26, ["generative", "path"], "Method"], [28, 34, ["convolutional", "autoregressive", "neural", "network", "on", "pixels"], "Method"], [35, 36, ["PixelCNN"], "Method"], [47, 49, ["recognition", "path"], "Method"], [51, 54, ["generative", "adversarial", "network"], "Method"], [55, 56, ["GAN"], "Method"], [84, 86, ["autoregressive", "decoder"], "Method"], [93, 95, ["Gaussian", "distribution"], "Method"], [130, 132, ["unsupervised", "fashion"], "Method"], [149, 153, ["semi", "-", "supervised", "settings"], "Task"], [155, 160, ["competitive", "semi", "-", "supervised", "classification"], "Task"], [163, 164, ["MNIST"], "Material"], [180, 182, ["generative", "models"], "Method"], [187, 191, ["direct", "back", "-", "propagation"], "Method"], [208, 211, ["generative", "adversarial", "network"], "Method"], [212, 213, ["GAN"], "Method"], [219, 225, ["two", "player", "min", "-", "max", "game"], "Method"], [227, 229, ["generative", "model"], "Method"], [240, 241, ["discriminator"], "Method"], [262, 264, ["generative", "model"], "Method"], [286, 288, ["cost", "function"], "Method"], [289, 290, ["GAN"], "Method"], [291, 292, ["GANs"], "Method"], [300, 303, ["implicit", "generative", "models"], "Method"], [321, 323, ["likelihood", "function"], "Method"], [341, 343, ["approximate", "inference"], "Task"], [367, 369, ["posterior", "approximations"], "Method"], [384, 386, ["variational", "inference"], "Task"], [390, 392, ["adversarial", "autoencoders"], "Method"], [393, 394, ["AAE"], "Method"], [398, 401, ["universal", "approximator", "posterior"], "Method"], [424, 428, ["Adversarial", "variational", "Bayes", "ference"], "Method"], [429, 430, ["avb"], "Method"], [440, 445, ["maximum", "-", "likelihood", "learning", "setting"], "Method"], [456, 462, ["ALI", "ali", "and", "BiGAN", "bigan", "models"], "Method"], [469, 473, ["maximum", "likelihood", "learning", "ference"], "Method"], [480, 484, ["recognition", "and", "generative", "models"], "Method"], [496, 498, ["Variational", "autoencoders"], "Method"], [499, 500, ["VAE"], "Method"], [501, 502, ["vae"], "Method"], [503, 504, ["rezende"], "Method"], [513, 516, ["image", "modeling", "technique"], "Method"], [518, 520, ["neural", "networks"], "Method"], [530, 535, ["top", "-", "down", "generative", "network"], "Method"], [544, 547, ["variational", "lower", "bound"], "Metric"], [559, 561, ["density", "models"], "Method"], [562, 565, ["autoregressive", "neural", "networks"], "Method"], [576, 578, ["PixelCNN", "pixelcnn"], "Method"], [580, 582, ["variational", "autoencoders"], "Method"], [596, 598, ["autoregressive", "models"], "Method"], [610, 613, ["hierarchical", "latent", "representation"], "Method"], [625, 627, ["generative", "autoencoder"], "Method"], [632, 635, ["latent", "variable", "models"], "Method"], [636, 638, ["autoregressive", "architectures"], "Method"], [644, 646, ["generative", "autoencoder"], "Method"], [653, 654, ["PixelCNN"], "Method"], [687, 690, ["adversarial", "autoencoder", "aae"], "Method"], [720, 722, ["latent", "code"], "Method"], [724, 726, ["autoregressive", "decoder"], "Method"], [748, 750, ["autoregressive", "decoder"], "Method"], [792, 795, ["global", "latent", "code"], "Method"], [837, 839, ["PixelGAN", "autoencoders"], "Method"], [841, 845, ["discrete", "vs.", "continuous", "decomposition"], "Method"], [903, 905, ["autoregressive", "decoder"], "Method"], [918, 920, ["unsupervised", "fashion"], "Method"], [925, 927, ["PixelGAN", "autoencoders"], "Method"], [935, 941, ["clustering", "and", "semi", "-", "supervised", "scenarios"], "Task"], [945, 946, ["classification"], "Task"], [970, 972, ["PixelGAN", "autoencoders"], "Method"], [973, 978, ["learning", "cross", "-", "domain", "relations"], "Task"], [995, 997, ["PixelGAN", "Autoencoders"], "Method"], [1013, 1015, ["recognition", "path"], "Method"], [1021, 1024, ["pixelgan_gaussian", "]", "Figure"], "Method"], [1036, 1039, ["deterministic", "neural", "function"], "Method"], [1077, 1081, ["adversarial", "autoencoder", "work", "aae"], "Method"], [1083, 1086, ["universal", "approximator", "posterior"], "Method"], [1108, 1110, ["posterior", "distribution"], "Method"], [1130, 1132, ["PixelGAN", "autoencoders"], "Method"], [1134, 1136, ["generative", "path"], "Method"], [1138, 1141, ["conditional", "PixelCNN", "pixelcnn"], "Method"], [1152, 1154, ["PixelCNN", "layers"], "Method"], [1156, 1157, ["inference"], "Task"], [1171, 1175, ["adversarial", "autoencoder", "work", "aae"], "Method"], [1194, 1195, ["autoencoder"], "Method"], [1241, 1242, ["autoencoder"], "Method"], [1263, 1265, ["PixelCNN", "decoder"], "Method"], [1267, 1268, ["encoder"], "Method"], [1276, 1278, ["reconstruction", "phase"], "Method"], [1292, 1294, ["reconstruction", "phase"], "Task"], [1307, 1308, ["encoder"], "Method"], [1312, 1314, ["PixelCNN", "decoder"], "Method"], [1351, 1352, ["PixelCNN"], "Method"], [1364, 1366, ["discriminative", "network"], "Method"], [1368, 1370, ["generative", "network"], "Method"], [1372, 1373, ["encoder"], "Method"], [1408, 1410, ["PixelCNN", "decoder"], "Method"], [1422, 1425, ["maximum", "likelihood", "learning"], "Method"], [1430, 1434, ["aggregated", "evidence", "lower", "bound"], "Method"], [1435, 1436, ["ELBO"], "Method"], [1491, 1493, ["regularization", "term"], "Method"], [1557, 1558, ["decoder"], "Method"], [1559, 1561, ["Deterministic", "Decoder"], "Method"], [1564, 1565, ["decoder"], "Method"], [1576, 1578, ["factorized", "decoder"], "Method"], [1580, 1581, ["VAE"], "Method"], [1606, 1608, ["reconstruction", "error"], "Metric"], [1638, 1641, ["variational", "lower", "bound"], "Metric"], [1642, 1644, ["i", "m"], "Metric"], [1655, 1656, ["VAE"], "Method"], [1658, 1659, ["MNIST"], "Material"], [1662, 1663, ["reconstruction"], "Method"], [1667, 1670, ["mutual", "information", "term"], "Metric"], [1681, 1683, ["Stochastic", "Decoder"], "Method"], [1689, 1690, ["decoder"], "Method"], [1693, 1694, ["PixelCNN"], "Method"], [1728, 1729, ["ELBO"], "Task"], [1773, 1775, ["latent", "code"], "Method"], [1795, 1796, ["bowman"], "Method"], [1797, 1798, ["vlae"], "Method"], [1808, 1811, ["KL", "term", "bowman"], "Method"], [1814, 1816, ["decoder", "vlae"], "Method"], [1836, 1839, ["maximum", "likelihood", "objective"], "Metric"], [1847, 1849, ["representation", "learning"], "Task"], [1853, 1854, ["decoder"], "Method"], [1858, 1860, ["PixelGAN", "autoencoders"], "Method"], [1873, 1874, ["ELBO"], "Method"], [1924, 1928, ["reconstruction", "phase", "of", "training"], "Method"], [1930, 1933, ["marginal", "KL", "term"], "Method"], [1980, 1981, ["ELBO"], "Method"], [2007, 2009, ["autoregressive", "decoder"], "Method"], [2018, 2020, ["PixelCNN", "decoder"], "Method"], [2025, 2027, ["architectures", "pixelcnn"], "Method"], [2037, 2039, ["PixelCNN", "layer"], "Method"], [2105, 2106, ["decoder"], "Method"], [2121, 2123, ["learnt", "representation"], "Method"], [2133, 2136, ["pixelgan_gaussian", "]", "Section"], "Method"], [2155, 2157, ["PixelGAN", "Autoencoders"], "Method"], [2158, 2160, ["Gaussian", "Priors"], "Method"], [2165, 2167, ["PixelGAN", "autoencoders"], "Method"], [2168, 2170, ["Gaussian", "priors"], "Method"], [2186, 2188, ["autoregressive", "decoder"], "Method"], [2207, 2211, ["location", "-", "dependent", "bias"], "Method"], [2224, 2227, ["decomposition", "of", "information"], "Task"], [2232, 2238, ["2", "-", "D", "Gaussian", "latent", "code"], "Method"], [2265, 2267, ["PixelCNN", "model"], "Method"], [2290, 2292, ["adversarial", "autoencoder"], "Method"], [2294, 2300, ["2", "-", "D", "Gaussian", "latent", "code"], "Method"], [2302, 2303, ["PixelCNN"], "Method"], [2329, 2331, ["adversarial", "autoencoder"], "Method"], [2333, 2335, ["sample", "quality"], "Metric"], [2342, 2343, ["VAE"], "Method"], [2391, 2393, ["PixelGAN", "autoencoders"], "Method"], [2400, 2402, ["conditioning", "architecture"], "Method"], [2413, 2415, ["autoregressive", "decoder"], "Method"], [2434, 2435, ["MNIST"], "Material"], [2445, 2447, ["visualization", "purpose"], "Task"], [2464, 2466, ["shallow", "decoder"], "Method"], [2495, 2496, ["PixelCNN"], "Method"], [2585, 2586, ["PixelCNN"], "Method"], [2590, 2592, ["conditioning", "architecture"], "Method"], [2595, 2598, ["decomposition", "of", "information"], "Task"], [2714, 2718, ["discrete", "vs.", "continuous", "decomposition"], "Task"], [2724, 2731, ["clustering", "and", "semi", "-", "supervised", "learning", "experiments"], "Task"], [2787, 2788, ["adversarial"], "Method"], [2789, 2791, ["variational", "autoencoder"], "Method"], [2811, 2816, ["global", "vs.", "local", "decomposition", "experiments"], "Task"], [2874, 2877, ["latent", "variable", "models"], "Method"], [2878, 2880, ["autoregressive", "decoders"], "Method"], [2882, 2884, ["PixelVAE", "pixelvae"], "Method"], [2885, 2888, ["variational", "lossy", "autoencoders"], "Method"], [2889, 2890, ["VLAE"], "Method"], [2907, 2909, ["PixelCNN", "decoder"], "Method"], [2912, 2915, ["decomposition", "of", "information"], "Task"], [2921, 2923, ["PixelCNN", "decoder"], "Method"], [2994, 2996, ["unsupervised", "fashion"], "Method"], [3025, 3029, ["and", "deep", "PixelCNN", "decoder"], "Method"], [3075, 3077, ["latent", "representation"], "Method"], [3084, 3088, ["discrete", "vs.", "continuous", "decomposition"], "Method"], [3125, 3129, ["discrete", "vs.", "continuous", "decomposition"], "Task"], [3138, 3140, ["PixelCNN", "decoders"], "Method"], [3160, 3161, ["PixelCNN"], "Method"], [3178, 3180, ["PixelGAN", "Autoencoders"], "Method"], [3181, 3183, ["Categorical", "Priors"], "Method"], [3233, 3237, ["semi", "-", "supervised", "settings"], "Task"], [3268, 3270, ["Gaussian", "distribution"], "Method"], [3283, 3288, ["location", "-", "independent", "bias", "architecture"], "Method"], [3296, 3298, ["convolutional", "network"], "Method"], [3300, 3302, ["inference", "network"], "Method"], [3305, 3306, ["encoder"], "Method"], [3320, 3322, ["inference", "network"], "Method"], [3366, 3368, ["softmax", "layer"], "Method"], [3370, 3371, ["encoder"], "Method"], [3392, 3393, ["encoder"], "Method"], [3429, 3430, ["autoencoder"], "Method"], [3515, 3517, ["autoregressive", "decoder"], "Method"], [3549, 3550, ["MNIST"], "Material"], [3583, 3585, ["softmax", "layer"], "Method"], [3587, 3589, ["inference", "network"], "Method"], [3739, 3741, ["autoregressive", "decoder"], "Method"], [3748, 3752, ["test", "error", "-", "rate"], "Metric"], [3759, 3760, ["MNIST"], "Material"], [3769, 3771, ["unsupervised", "fashion"], "Method"], [3780, 3781, ["encoder"], "Method"], [3946, 3947, ["encoder"], "Method"], [3957, 3959, ["error", "rate"], "Metric"], [3960, 3962, ["classifying", "digits"], "Task"], [3991, 3995, ["semi", "-", "supervised", "setting"], "Task"], [4007, 4012, ["semi", "-", "supervised", "training", "phase"], "Method"], [4048, 4052, ["semi", "-", "supervised", "phase"], "Method"], [4067, 4068, ["encoder"], "Method"], [4071, 4075, ["cross", "-", "entropy", "cost"], "Metric"], [4077, 4081, ["semi", "-", "supervised", "cost"], "Method"], [4095, 4096, ["encoder"], "Method"], [4124, 4126, ["PixelGAN", "autoencoders"], "Method"], [4148, 4150, ["generative", "model"], "Method"], [4168, 4171, ["Parzen", "window", "estimate"], "Method"], [4191, 4193, ["downstream", "tasks"], "Task"], [4195, 4197, ["unsupervised", "clustering"], "Task"], [4198, 4202, ["semi", "-", "supervised", "classification"], "Task"], [4224, 4226, ["Unsupervised", "Clustering"], "Method"], [4233, 4235, ["unsupervised", "fashion"], "Method"], [4300, 4302, ["test", "error"], "Metric"], [4331, 4333, ["PixelGAN", "autoencoders"], "Method"], [4345, 4347, ["CatGAN", "catgan"], "Method"], [4348, 4350, ["InfoGAN", "infogan"], "Method"], [4351, 4354, ["adversarial", "autoencoders", "aae"], "Method"], [4357, 4361, ["Semi", "-", "supervised", "Classification"], "Task"], [4381, 4385, ["semi", "-", "supervised", "classification"], "Task"], [4388, 4389, ["MNIST"], "Material"], [4408, 4409, ["classification"], "Task"], [4417, 4419, ["classification", "rate"], "Metric"], [4440, 4442, ["unsupervised", "case"], "Task"], [4453, 4457, ["semi", "-", "supervised", "case"], "Task"], [4496, 4498, ["temporal", "ensembling"], "Method"], [4499, 4502, ["temporal", "-", "ensembling"], "Method"], [4506, 4508, ["generative", "model"], "Method"], [4547, 4548, ["MNIST"], "Material"], [4566, 4571, ["one", "-", "hot", "latent", "code"], "Method"], [4603, 4608, ["Learning", "Cross", "-", "Domain", "Relations"], "Task"], [4609, 4611, ["PixelGAN", "Autoencoders"], "Method"], [4628, 4633, ["learning", "cross", "-", "domain", "relations"], "Task"], [4645, 4651, ["clustering", "or", "semi", "-", "supervised", "learning"], "Task"], [4660, 4665, ["smooth", "cross", "-", "domain", "mapping"], "Task"], [4686, 4690, ["cross", "-", "domain", "mapping"], "Task"], [4698, 4703, ["cross", "-", "domain", "-", "ilya"], "Method"], [4704, 4705, ["aae"], "Method"], [4706, 4711, ["cross", "-", "domain", "-", "nlp"], "Method"], [4713, 4718, ["cross", "-", "domain", "-", "ilya"], "Task"], [4720, 4723, ["unsupervised", "cost", "function"], "Method"], [4725, 4728, ["output", "distribution", "matching"], "Method"], [4729, 4730, ["ODM"], "Method"], [4736, 4740, ["cross", "-", "domain", "mapping"], "Task"], [4800, 4802, ["supervised", "fashion"], "Method"], [4809, 4811, ["speech", "recognition"], "Task"], [4817, 4821, ["cross", "-", "domain", "mapping"], "Task"], [4835, 4838, ["ODM", "cost", "function"], "Method"], [4839, 4842, ["odm", "]", "Equation"], "Method"], [4850, 4852, ["smooth", "function"], "Method"], [4867, 4869, ["language", "model"], "Method"], [4914, 4915, ["discogan"], "Method"], [4930, 4933, ["ODM", "cost", "function"], "Method"], [4956, 4959, ["adversarial", "autoencoder", "aae"], "Method"], [4962, 4964, ["generative", "model"], "Method"], [4967, 4971, ["cross", "-", "domain", "mapping"], "Method"], [4973, 4975, ["Gaussian", "distribution"], "Method"], [4982, 4985, ["ODM", "cost", "function"], "Method"], [4988, 4990, ["reconstruction", "term"], "Method"], [5016, 5019, ["ODM", "cost", "function"], "Method"], [5022, 5025, ["variational", "lower", "bound"], "Metric"], [5042, 5044, ["reconstruction", "cost"], "Metric"], [5053, 5058, ["cross", "-", "domain", "-", "nlp"], "Task"], [5059, 5060, ["zhangadversarial"], "Task"], [5062, 5064, ["adversarial", "autoencoder"], "Method"], [5075, 5077, ["vector", "representations"], "Method"], [5091, 5093, ["DiscoGAN", "discogan"], "Method"], [5094, 5096, ["CycleGAN", "cyclegan"], "Method"], [5101, 5103, ["adversarial", "autoencoder"], "Method"], [5106, 5108, ["latent", "representation"], "Method"], [5160, 5163, ["pixelgan_gaussian", "]", "Figure"], "Method"], [5203, 5205, ["adversarial", "autoencoders"], "Method"], [5225, 5227, ["lossless", "reconstruction"], "Task"], [5251, 5253, ["PixelCNN", "decoder"], "Method"], [5270, 5272, ["ODM", "idea"], "Method"], [5273, 5277, ["semi", "-", "supervised", "learning"], "Task"], [5322, 5324, ["neural", "network"], "Method"], [5390, 5392, ["adversarial", "autoencoder"], "Method"], [5434, 5436, ["reconstruction", "term"], "Method"], [5438, 5440, ["latent", "representation"], "Method"], [5477, 5482, ["one", "-", "hot", "label", "representation"], "Method"], [5484, 5488, ["semi", "-", "supervised", "learning"], "Task"], [5505, 5510, ["lossless", "reconstruction", "of", "the", "image"], "Task"], [5558, 5560, ["autoregressive", "decoder"], "Method"], [5577, 5579, ["generative", "autoencoder"], "Method"], [5582, 5584, ["generative", "PixelCNN"], "Method"], [5613, 5615, ["latent", "representation"], "Method"], [5637, 5639, ["PixelCNN", "decoder"], "Method"], [5691, 5698, ["clustering", "and", "semi", "-", "supervised", "learning", "tasks"], "Task"], [5712, 5714, ["PixelGAN", "autoencoders"], "Method"], [5715, 5717, ["downstream", "tasks"], "Task"], [5719, 5723, ["semi", "-", "supervised", "learning"], "Task"], [5735, 5740, ["learning", "cross", "-", "domain", "relations"], "Task"], [5787, 5789, ["PixelGAN", "autoencoders"], "Method"], [5797, 5801, ["semi", "-", "supervised", "experiments"], "Task"], [5809, 5812, ["universal", "approximator", "posterior"], "Method"], [5826, 5828, ["deterministic", "posterior"], "Method"], [5868, 5871, ["denoising", "criterion", "idea"], "Method"], [5873, 5876, ["denoising", "-", "vae"], "Method"], [5891, 5895, ["mode", "-", "missing", "behavior"], "Task"], [5897, 5898, ["GAN"], "Method"], [5921, 5924, ["instance", "noise", "instance"], "Method"], [5938, 5939, ["PixelCNN"], "Method"], [5968, 5972, ["conditional", "PixelCNN", "model", "pixelcnn"], "Method"], [5978, 5980, ["convolutional", "layer"], "Method"], [5982, 5983, ["decoder"], "Method"], [5996, 5997, ["PixelCNN"], "Method"], [6000, 6002, ["linear", "mapping"], "Method"], [6041, 6043, ["PixelCNN", "decoder"], "Method"], [6091, 6098, ["clustering", "and", "semi", "-", "supervised", "learning", "experiments"], "Task"], [6109, 6111, ["convolutional", "layer"], "Method"], [6113, 6115, ["PixelCNN", "decoder"], "Method"], [6128, 6129, ["PixelCNN"], "Method"], [6132, 6136, ["one", "layer", "neural", "network"], "Method"], [6167, 6169, ["convolutional", "layer"], "Method"], [6208, 6211, ["VPN", "model", "vpn"], "Method"], [6251, 6252, ["convolution"], "Method"], [6296, 6297, ["conditioning"], "Task"], [6301, 6303, ["PixelVAE", "pixelvae"], "Task"], [6305, 6308, ["variational", "lossy", "autoencoder"], "Method"], [6309, 6310, ["VLAE"], "Method"], [6335, 6337, ["conditioning", "tensor"], "Method"], [6355, 6356, ["PixelCNN"], "Method"], [6360, 6362, ["computational", "complexity"], "Metric"], [6372, 6377, ["location", "-", "dependent", "bias", "method"], "Method"], [6384, 6385, ["TensorFlow"], "Method"], [6397, 6398, ["gan"], "Method"], [6410, 6411, ["generator"], "Method"], [6413, 6414, ["GAN"], "Method"], [6455, 6457, ["binary", "MNIST"], "Material"], [6459, 6463, ["real", "-", "valued", "MNIST"], "Material"], [6488, 6493, ["logistic", "mixture", "likelihood", "pixelcnn", "++"], "Method"], [6495, 6497, ["cost", "function"], "Method"], [6499, 6500, ["PixelCNN"], "Method"], [6505, 6507, ["binary", "MNIST"], "Material"], [6533, 6535, ["PixelGAN", "Autoencoders"], "Method"], [6536, 6538, ["Gaussian", "Prior"], "Method"], [6539, 6540, ["MNIST"], "Material"], [6554, 6556, ["Gaussian", "prior"], "Method"], [6571, 6573, ["PixelCNN", "decoder"], "Method"], [6582, 6583, ["pixelcnn"], "Method"], [6585, 6587, ["cost", "function"], "Metric"], [6589, 6590, ["PixelCNN"], "Method"], [6592, 6597, ["cross", "-", "entropy", "cost", "function"], "Metric"], [6599, 6600, ["PixelCNN"], "Method"], [6642, 6647, ["one", "-", "layer", "neural", "network"], "Method"], [6672, 6674, ["first", "layer"], "Method"], [6676, 6678, ["PixelCNN", "decoder"], "Method"], [6680, 6681, ["PixelCNN"], "Method"], [6710, 6711, ["pixelcnn"], "Method"], [6724, 6727, ["ReLU", "activation", "function"], "Method"], [6733, 6737, ["fully", "-", "connected", "layers"], "Method"], [6741, 6745, ["ReLU", "non", "-", "linearity"], "Method"], [6751, 6752, ["encoder"], "Method"], [6754, 6757, ["linear", "activation", "function"], "Method"], [6760, 6764, ["latent", "representation", "of", "size"], "Task"], [6768, 6770, ["Gaussian", "distribution"], "Method"], [6778, 6780, ["gradient", "descent"], "Method"], [6781, 6783, ["momentum", "algorithm"], "Method"], [6795, 6798, ["PixelCNN", "reconstruction", "cost"], "Task"], [6802, 6806, ["learning", "rate", "of", "0.001"], "Metric"], [6819, 6821, ["learning", "rate"], "Metric"], [6828, 6829, ["generator"], "Method"], [6835, 6837, ["learning", "rates"], "Metric"], [6848, 6852, ["Unsupervised", "Clustering", "of", "MNIST"], "Task"], [6860, 6861, ["clustering"], "Task"], [6885, 6887, ["PixelCNN", "decoder"], "Method"], [6896, 6897, ["pixelcnn"], "Method"], [6899, 6901, ["cost", "function"], "Metric"], [6903, 6904, ["PixelCNN"], "Method"], [6906, 6911, ["cross", "-", "entropy", "cost", "function"], "Metric"], [6913, 6914, ["PixelCNN"], "Method"], [6952, 6953, ["pixelcnn"], "Method"], [6966, 6969, ["ReLU", "activation", "function"], "Method"], [6975, 6977, ["convolutional", "layer"], "Method"], [6987, 6989, ["ReLU", "activation"], "Method"], [6992, 6996, ["max", "-", "pooling", "layer"], "Method"], [7005, 7007, ["convolutional", "layer"], "Method"], [7017, 7019, ["ReLU", "activation"], "Method"], [7022, 7026, ["max", "-", "pooling", "layer"], "Method"], [7033, 7037, ["fully", "-", "connected", "layer"], "Method"], [7043, 7044, ["encoder"], "Method"], [7066, 7067, ["encoder"], "Method"], [7069, 7071, ["softmax", "layer"], "Method"], [7103, 7104, ["softmax"], "Method"], [7116, 7119, ["Adam", "Adam", "optimizer"], "Method"], [7120, 7122, ["learning", "rate"], "Metric"], [7126, 7130, ["PixelCNN", "reconstruction", "cost", "function"], "Task"], [7139, 7141, ["gradient", "descent"], "Method"], [7142, 7144, ["momentum", "algorithm"], "Method"], [7161, 7162, ["generator"], "Method"], [7176, 7178, ["learning", "rates"], "Metric"], [7236, 7240, ["Semi", "-", "Supervised", "MNIST"], "Task"], [7242, 7246, ["semi", "-", "supervised", "learning"], "Task"], [7251, 7256, ["real", "-", "valued", "MNIST", "dataset"], "Material"], [7261, 7267, ["semi", "-", "supervised", "error", "-", "rate"], "Metric"], [7269, 7273, ["real", "-", "valued", "MNIST"], "Material"], [7279, 7281, ["binary", "MNIST"], "Material"], [7299, 7306, ["logistic", "mixture", "likelihood", "cost", "function", "pixelcnn", "++"], "Method"], [7325, 7327, ["binary", "MNIST"], "Material"], [7345, 7349, ["real", "-", "valued", "MNIST"], "Material"], [7353, 7355, ["Binary", "MNIST"], "Material"], [7365, 7370, ["semi", "-", "supervised", "learning", "experiments"], "Task"], [7372, 7374, ["binary", "MNIST"], "Material"], [7395, 7397, ["PixelCNN", "decoder"], "Method"], [7406, 7407, ["pixelcnn"], "Method"], [7410, 7415, ["cross", "-", "entropy", "cost", "function"], "Method"], [7417, 7418, ["PixelCNN"], "Method"], [7437, 7438, ["PixelCNN"], "Method"], [7458, 7459, ["pixelcnn"], "Method"], [7472, 7475, ["ReLU", "activation", "function"], "Method"], [7481, 7483, ["convolutional", "layers"], "Method"], [7493, 7495, ["ReLU", "activation"], "Method"], [7498, 7502, ["max", "-", "pooling", "layer"], "Method"], [7512, 7514, ["convolutional", "layers"], "Method"], [7524, 7526, ["ReLU", "activation"], "Method"], [7529, 7533, ["max", "-", "pooling", "layer"], "Method"], [7540, 7544, ["fully", "-", "connected", "layer"], "Method"], [7550, 7551, ["encoder"], "Method"], [7559, 7561, ["convolutional", "layers"], "Method"], [7563, 7564, ["encoder"], "Method"], [7566, 7568, ["softmax", "layer"], "Method"], [7569, 7573, ["batch", "-", "normalization", "batch"], "Method"], [7576, 7578, ["latent", "representation"], "Method"], [7589, 7593, ["semi", "-", "supervised", "cost"], "Metric"], [7595, 7600, ["cross", "-", "entropy", "cost", "function"], "Metric"], [7607, 7610, ["Adam", "Adam", "optimizer"], "Method"], [7611, 7613, ["learning", "rate"], "Metric"], [7617, 7619, ["PixelCNN", "cost"], "Metric"], [7621, 7625, ["cross", "-", "entropy", "cost"], "Metric"], [7634, 7636, ["gradient", "descent"], "Method"], [7637, 7639, ["momentum", "algorithm"], "Method"], [7642, 7643, ["generator"], "Method"], [7656, 7657, ["generator"], "Method"], [7671, 7673, ["learning", "rates"], "Metric"], [7717, 7721, ["Real", "-", "valued", "MNIST"], "Material"], [7731, 7735, ["semi", "-", "supervised", "learning"], "Task"], [7738, 7742, ["real", "-", "valued", "MNIST"], "Material"], [7752, 7754, ["PixelCNN", "decoder"], "Method"], [7763, 7764, ["pixelcnn"], "Method"], [7767, 7773, ["discretized", "logistic", "mixture", "likelihood", "cost", "function"], "Method"], [7780, 7782, ["pixelcnn", "++"], "Method"], [7784, 7785, ["PixelCNN"], "Method"], [7804, 7805, ["PixelCNN"], "Method"], [7819, 7826, ["gated", "sigmoid", "-", "tanh", "non", "-", "linearity"], "Method"], [7828, 7829, ["pixelcnn"], "Method"], [7842, 7845, ["ReLU", "activation", "function"], "Method"], [7851, 7853, ["convolutional", "layers"], "Method"], [7863, 7865, ["ReLU", "activation"], "Method"], [7868, 7872, ["max", "-", "pooling", "layer"], "Method"], [7882, 7884, ["convolutional", "layers"], "Method"], [7894, 7896, ["ReLU", "activation"], "Method"], [7899, 7903, ["max", "-", "pooling", "layer"], "Method"], [7910, 7914, ["fully", "-", "connected", "layer"], "Method"], [7920, 7921, ["encoder"], "Method"], [7929, 7931, ["convolutional", "layers"], "Method"], [7933, 7934, ["encoder"], "Method"], [7936, 7938, ["softmax", "layer"], "Method"], [7939, 7943, ["batch", "-", "normalization", "batch"], "Method"], [7946, 7948, ["latent", "representation"], "Method"], [7959, 7963, ["semi", "-", "supervised", "cost"], "Metric"], [7965, 7970, ["cross", "-", "entropy", "cost", "function"], "Metric"], [7977, 7980, ["Adam", "Adam", "optimizer"], "Method"], [7981, 7983, ["learning", "rate"], "Metric"], [7987, 7989, ["PixelCNN", "cost"], "Metric"], [7991, 7995, ["cross", "-", "entropy", "cost"], "Metric"], [8004, 8006, ["gradient", "descent"], "Method"], [8007, 8009, ["momentum", "algorithm"], "Method"], [8012, 8013, ["generator"], "Method"], [8026, 8027, ["generator"], "Method"], [8041, 8043, ["learning", "rates"], "Metric"], [8056, 8058, ["learning", "rates"], "Metric"], [8131, 8132, ["vat"], "Method"], [8156, 8160, ["Semi", "-", "Supervised", "SVHN"], "Task"], [8164, 8166, ["model", "architecture"], "Method"], [8169, 8174, ["semi", "-", "supervised", "learning", "experiments"], "Task"], [8190, 8192, ["PixelCNN", "decoder"], "Method"], [8201, 8202, ["pixelcnn"], "Method"], [8204, 8206, ["cost", "function"], "Metric"], [8208, 8209, ["PixelCNN"], "Method"], [8211, 8217, ["discretized", "logistic", "mixture", "likelihood", "cost", "function"], "Method"], [8219, 8221, ["logistic", "distribution"], "Method"], [8224, 8226, ["pixelcnn", "++"], "Method"], [8228, 8229, ["PixelCNN"], "Method"], [8261, 8268, ["gated", "sigmoid", "-", "tanh", "non", "-", "linearity"], "Method"], [8270, 8271, ["pixelcnn"], "Method"], [8284, 8287, ["ReLU", "activation", "function"], "Method"], [8293, 8295, ["convolutional", "layers"], "Method"], [8305, 8307, ["ReLU", "activation"], "Method"], [8310, 8314, ["max", "-", "pooling", "layer"], "Method"], [8324, 8326, ["convolutional", "layers"], "Method"], [8336, 8338, ["ReLU", "activation"], "Method"], [8341, 8345, ["max", "-", "pooling", "layer"], "Method"], [8352, 8356, ["fully", "-", "connected", "layer"], "Method"], [8362, 8363, ["encoder"], "Method"], [8371, 8373, ["convolutional", "layers"], "Method"], [8375, 8376, ["encoder"], "Method"], [8378, 8380, ["softmax", "layer"], "Method"], [8381, 8385, ["batch", "-", "normalization", "batch"], "Method"], [8388, 8390, ["latent", "representation"], "Method"], [8401, 8405, ["semi", "-", "supervised", "cost"], "Metric"], [8407, 8412, ["cross", "-", "entropy", "cost", "function"], "Metric"], [8419, 8422, ["Adam", "Adam", "optimizer"], "Method"], [8426, 8428, ["cost", "function"], "Metric"], [8431, 8433, ["PixelCNN", "cost"], "Metric"], [8435, 8439, ["cross", "-", "entropy", "cost"], "Metric"], [8442, 8444, ["learning", "rate"], "Metric"], [8448, 8449, ["generator"], "Method"], [8451, 8453, ["discriminator", "costs"], "Metric"], [8460, 8462, ["learning", "rate"], "Metric"], [8561, 8565, ["Semi", "-", "Supervised", "NORB"], "Method"], [8566, 8568, ["PixelCNN", "decoder"], "Method"], [8577, 8578, ["pixelcnn"], "Method"], [8580, 8582, ["cost", "function"], "Metric"], [8584, 8585, ["PixelCNN"], "Method"], [8587, 8593, ["discretized", "logistic", "mixture", "likelihood", "cost", "function"], "Method"], [8595, 8597, ["logistic", "distribution"], "Method"], [8600, 8602, ["pixelcnn", "++"], "Method"], [8604, 8605, ["PixelCNN"], "Method"], [8637, 8644, ["gated", "sigmoid", "-", "tanh", "non", "-", "linearity"], "Method"], [8646, 8647, ["pixelcnn"], "Method"], [8660, 8663, ["ReLU", "activation", "function"], "Method"], [8669, 8671, ["convolutional", "layer"], "Method"], [8681, 8683, ["ReLU", "activation"], "Method"], [8686, 8690, ["max", "-", "pooling", "layer"], "Method"], [8699, 8701, ["convolutional", "layer"], "Method"], [8711, 8713, ["ReLU", "activation"], "Method"], [8716, 8720, ["max", "-", "pooling", "layer"], "Method"], [8729, 8731, ["convolutional", "layer"], "Method"], [8746, 8750, ["max", "-", "pooling", "layer"], "Method"], [8757, 8761, ["fully", "-", "connected", "layer"], "Method"], [8767, 8768, ["encoder"], "Method"], [8776, 8778, ["convolutional", "layers"], "Method"], [8780, 8781, ["encoder"], "Method"], [8783, 8785, ["softmax", "layer"], "Method"], [8786, 8790, ["batch", "-", "normalization", "batch"], "Method"], [8793, 8795, ["latent", "representation"], "Method"], [8806, 8810, ["semi", "-", "supervised", "cost"], "Metric"], [8812, 8817, ["cross", "-", "entropy", "cost", "function"], "Metric"], [8824, 8827, ["Adam", "Adam", "optimizer"], "Method"], [8831, 8833, ["cost", "function"], "Metric"], [8836, 8838, ["PixelCNN", "cost"], "Metric"], [8840, 8844, ["cross", "-", "entropy", "cost"], "Metric"], [8847, 8849, ["learning", "rate"], "Metric"], [8853, 8854, ["generator"], "Method"], [8856, 8858, ["discriminator", "costs"], "Metric"], [8865, 8867, ["learning", "rate"], "Metric"], [8918, 8920, ["test", "error"], "Metric"], [14, 15, ["autoencoder"], "Method"], [139, 140, ["autoencoder"], "Method"], [214, 215, ["gan"], "Method"], [336, 337, ["GAN"], "Method"], [352, 353, ["aae"], "Method"], [395, 396, ["aae"], "Method"], [408, 409, ["adversarial"], "Method"], [435, 436, ["GAN"], "Method"], [449, 450, ["GAN"], "Method"], [492, 493, ["adversarial"], "Method"], [622, 623, ["autoencoder"], "Method"], [641, 642, ["autoencoder"], "Method"], [679, 680, ["adversarial"], "Method"], [695, 696, ["adversarial"], "Method"], [713, 714, ["autoencoder"], "Method"], [873, 874, ["adversarial"], "Method"], [1018, 1019, ["autoencoder"], "Method"], [1162, 1163, ["GAN"], "Method"], [1181, 1182, ["adversarial"], "Method"], [1216, 1217, ["adversarial"], "Method"], [1231, 1232, ["adversarial"], "Method"], [1238, 1239, ["encoder"], "Method"], [1255, 1256, ["adversarial"], "Method"], [1259, 1260, ["adversarial"], "Method"], [1280, 1281, ["adversarial"], "Method"], [1332, 1333, ["encoder"], "Method"], [1355, 1356, ["adversarial"], "Method"], [1359, 1360, ["adversarial"], "Method"], [1419, 1420, ["autoencoder"], "Method"], [1938, 1939, ["adversarial"], "Method"], [2203, 2204, ["autoencoder"], "Method"], [2214, 2215, ["MNIST"], "Material"], [2250, 2251, ["autoencoder"], "Method"], [2365, 2366, ["autoencoder"], "Method"], [2432, 2433, ["autoencoder"], "Method"], [2897, 2898, ["VLAE"], "Method"], [2960, 2961, ["autoencoder"], "Method"], [3046, 3050, ["semi", "-", "supervised", "classification"], "Task"], [3078, 3079, ["adversarial"], "Method"], [3194, 3195, ["autoencoder"], "Method"], [3354, 3355, ["adversarial"], "Method"], [3381, 3382, ["encoder"], "Method"], [3407, 3408, ["adversarial"], "Method"], [3414, 3415, ["encoder"], "Method"], [3448, 3449, ["adversarial"], "Method"], [3472, 3473, ["encoder"], "Method"], [3529, 3530, ["adversarial"], "Method"], [3542, 3543, ["autoencoder"], "Method"], [3696, 3697, ["adversarial"], "Method"], [3775, 3776, ["autoencoder"], "Method"], [3811, 3812, ["autoencoder"], "Method"], [3816, 3817, ["MNIST"], "Material"], [3980, 3982, ["PixelGAN", "Autoencoders"], "Method"], [3985, 3986, ["autoencoder"], "Method"], [4038, 4039, ["adversarial"], "Method"], [4090, 4091, ["GAN"], "Method"], [4128, 4132, ["semi", "-", "supervised", "classification"], "Task"], [4145, 4146, ["autoencoder"], "Method"], [4161, 4162, ["GAN"], "Method"], [4189, 4190, ["autoencoder"], "Method"], [4230, 4231, ["autoencoder"], "Method"], [4237, 4238, ["MNIST"], "Material"], [4338, 4339, ["GAN"], "Method"], [4397, 4398, ["MNIST"], "Material"], [4423, 4424, ["MNIST"], "Material"], [4430, 4431, ["MNIST"], "Material"], [4486, 4487, ["autoencoder"], "Method"], [4519, 4520, ["autoencoder"], "Method"], [4544, 4545, ["autoencoder"], "Method"], [4581, 4582, ["autoencoder"], "Method"], [4620, 4621, ["autoencoder"], "Method"], [4676, 4677, ["GAN"], "Method"], [4767, 4768, ["adversarial"], "Method"], [4943, 4944, ["encoder"], "Method"], [5046, 5047, ["autoencoder"], "Method"], [5126, 5127, ["autoencoder"], "Method"], [5207, 5208, ["encoder"], "Method"], [5209, 5211, ["PixelGAN", "autoencoders"], "Method"], [5230, 5231, ["encoder"], "Method"], [5264, 5265, ["encoder"], "Method"], [5343, 5344, ["adversarial"], "Method"], [5426, 5427, ["adversarial"], "Method"], [5427, 5428, ["autoencoder"], "Method"], [5513, 5514, ["autoencoder"], "Method"], [5520, 5521, ["encoder"], "Method"], [5572, 5573, ["autoencoder"], "Method"], [5586, 5587, ["GAN"], "Method"], [5917, 5918, ["GAN"], "Method"], [5932, 5933, ["gan"], "Method"], [6406, 6407, ["GAN"], "Method"], [6428, 6429, ["MNIST"], "Material"], [6431, 6432, ["MNIST"], "Material"], [6468, 6469, ["MNIST"], "Material"], [6551, 6552, ["autoencoder"], "Method"], [6559, 6560, ["MNIST"], "Material"], [6714, 6715, ["adversarial"], "Method"], [6729, 6730, ["encoder"], "Method"], [6863, 6864, ["MNIST"], "Material"], [6956, 6957, ["adversarial"], "Method"], [6971, 6972, ["encoder"], "Method"], [7154, 7155, ["adversarial"], "Method"], [7462, 7463, ["adversarial"], "Method"], [7477, 7478, ["encoder"], "Method"], [7649, 7650, ["adversarial"], "Method"], [7832, 7833, ["adversarial"], "Method"], [7847, 7848, ["encoder"], "Method"], [8019, 8020, ["adversarial"], "Method"], [8274, 8275, ["adversarial"], "Method"], [8289, 8290, ["encoder"], "Method"], [8455, 8456, ["adversarial"], "Method"], [8650, 8651, ["adversarial"], "Method"], [8665, 8666, ["encoder"], "Method"], [8860, 8861, ["adversarial"], "Method"]]} {"doc_id": "11356cd6bb0f2776a88cd584ff108470414c6594", "text": ["document", ":", "Submanifold", "Sparse", "Convolutional", "Networks", "Convolutional", "network", "are", "the", "de", "-", "facto", "standard", "for", "analysing", "spatio", "-", "temporal", "data", "such", "as", "images", ",", "videos", ",", "3D", "shapes", ",", "etc", ".", "Whilst", "some", "of", "this", "data", "is", "naturally", "dense", "(", "for", "instance", ",", "photos", ")", ",", "many", "other", "data", "sources", "are", "inherently", "sparse", ".", "Examples", "include", "pen", "-", "strokes", "forming", "on", "a", "piece", "of", "paper", ",", "or", "(", "colored", ")", "3D", "point", "clouds", "that", "were", "obtained", "using", "a", "LiDAR", "scanner", "or", "RGB", "-", "D", "camera", ".", "Standard", "‘", "‘", "dense", "’", "’", "implementations", "of", "convolutional", "networks", "are", "very", "inefficient", "when", "applied", "on", "such", "sparse", "data", ".", "We", "introduce", "a", "sparse", "convolutional", "operation", "tailored", "to", "processing", "sparse", "data", "that", "differs", "from", "prior", "work", "on", "sparse", "convolutional", "networks", "in", "that", "it", "operates", "strictly", "on", "submanifolds", ",", "rather", "than", "‘", "‘", "dilating", "’", "’", "the", "observation", "with", "every", "layer", "in", "the", "network", ".", "Our", "empirical", "analysis", "of", "the", "resulting", "submanifold", "sparse", "convolutional", "networks", "shows", "that", "they", "perform", "on", "par", "with", "state", "-", "of", "-", "the", "-", "art", "methods", "whilst", "requiring", "substantially", "less", "computation", ".", "section", ":", "Introduction", "Convolutional", "networks", "constitute", "the", "state", "-", "of", "-", "the", "art", "method", "for", "a", "wide", "range", "of", "tasks", "that", "involve", "the", "analysis", "of", "data", "with", "spatial", "and", "/", "or", "temporal", "structure", ",", "such", "as", "photographs", ",", "videos", ",", "or", "three", "-", "dimensional", "surface", "models", ".", "While", "such", "data", "frequently", "comprises", "a", "densely", "filled", "(", "2D", "or", "3D", ")", "grid", ",", "other", "spatio", "-", "temporal", "datasets", "are", "naturally", "sparse", ".", "For", "instance", ",", "handwriting", "is", "made", "up", "of", "one", "-", "dimensional", "lines", "in", "two", "-", "dimensional", "space", ",", "pictures", "made", "by", "RGB", "-", "D", "cameras", "are", "three", "-", "dimensional", "point", "clouds", ",", "and", "OFF", "models", "form", "two", "-", "dimensional", "surfaces", "in", "3D", "space", ".", "The", "curse", "of", "dimensionality", "applies", ",", "in", "particular", ",", "on", "data", "that", "lives", "on", "grids", "that", "have", "three", "or", "more", "dimensions", ":", "the", "number", "of", "points", "on", "the", "grid", "grows", "exponentially", "with", "its", "dimensionality", ".", "In", "such", "scenarios", ",", "it", "becomes", "increasingly", "important", "to", "exploit", "data", "sparsity", "whenever", "possible", "in", "order", "to", "reduce", "the", "computational", "resources", "needed", "for", "data", "processing", ".", "Indeed", ",", "exploiting", "sparsity", "is", "paramount", "when", "analyzing", ",", "for", "instance", ",", "RGB", "-", "D", "videos", "which", "are", "sparsely", "populated", "4D", "structures", ".", "Traditional", "convolutional", "network", "implementations", "are", "optimized", "for", "data", "that", "lives", "on", "densely", "populated", "grids", ",", "and", "can", "not", "process", "sparse", "data", "efficiently", ".", "More", "recently", ",", "a", "number", "of", "convolutional", "network", "implementations", "have", "been", "presented", "that", "are", "tailored", "to", "work", "efficiently", "on", "sparse", "data", ".", "Mathematically", ",", "some", "of", "these", "implementations", "are", "identical", "to", "a", "regular", "convolutional", "network", ",", "but", "they", "require", "fewer", "computational", "resources", "in", "terms", "of", "FLOPs", "and", "/", "or", "in", "terms", "of", "memory", ".", "OctNets", "slightly", "modify", "the", "convolution", "operator", "to", "produce", "‘", "‘", "averaged", "’", "’", "hidden", "states", "in", "parts", "of", "the", "grid", "that", "are", "away", "from", "regions", "of", "interest", ".", "One", "of", "the", "downsides", "of", "prior", "sparse", "implementations", "of", "convolutional", "networks", "is", "that", "they", "‘", "‘", "dilate", "’", "’", "the", "sparse", "data", "in", "every", "layer", ",", "because", "they", "implement", "a", "‘", "‘", "full", "’", "’", "convolution", ".", "In", "this", "work", ",", "we", "show", "that", "it", "is", "possible", "to", "successfully", "train", "convolutional", "networks", "that", "keep", "the", "same", "sparsity", "pattern", "throughout", "the", "layers", "of", "the", "network", ",", "without", "dilating", "the", "feature", "maps", ".", "To", "this", "end", ",", "we", "explore", "two", "novel", "convolution", "operators", ":", "sparse", "convolution", "(", "SC", ")", "and", "valid", "sparse", "convolution", "(", "VSC", ")", ".", "In", "our", "experiments", "with", "recognizing", "handwritten", "digits", "and", "3D", "shapes", ",", "networks", "using", "SC", "and", "VSC", "achieve", "state", "-", "of", "-", "the", "-", "art", "performance", "whilst", "reducing", "the", "computation", "and", "memory", "requirements", "by", ".", "section", ":", "Motivation", "We", "define", "a", "-", "dimensional", "convolutional", "network", "as", "a", "network", "that", "takes", "as", "input", "that", "is", "a", "-", "dimensional", "tensor", ":", "the", "input", "tensor", "contains", "spatiotemporal", "dimensions", "(", "such", "as", "length", ",", "width", ",", "height", ",", "time", ",", "etc", ".", ")", "and", "one", "additional", "feature", "space", "dimension", "(", "for", "instance", ",", "RGB", "color", "channels", ",", "surface", "normal", "vectors", ",", "etc", ".", ")", ".", "A", "sparse", "input", "corresponds", "to", "a", "-", "dimensional", "grid", "of", "sites", "that", "is", "associated", "with", "a", "feature", "vector", ".", "We", "define", "a", "site", "in", "the", "input", "to", "be", "active", "if", "any", "element", "in", "the", "feature", "vector", "is", "not", "in", "its", "ground", "state", ",", "for", "instance", ",", "if", "it", "is", "non", "-", "zero", ".", "In", "many", "practical", "problems", ",", "thresholding", "may", "be", "used", "to", "eliminate", "sites", "at", "which", "the", "feature", "vector", "is", "within", "a", "very", "small", "distance", "from", "the", "ground", "state", ".", "Note", "that", "even", "though", "the", "input", "tensor", "is", "-", "dimensional", ",", "activity", "is", "a", "-", "dimensional", "phenomenon", ":", "entire", "planes", "along", "the", "feature", "dimension", "are", "either", "active", "or", "not", ".", "The", "hidden", "layers", "of", "a", "convolutional", "network", "are", "also", "represented", "by", "-", "dimensional", "grids", "of", "feature", "-", "space", "vectors", ".", "When", "propagating", "the", "input", "data", "through", "the", "network", ",", "a", "site", "in", "a", "hidden", "layer", "is", "active", "if", "any", "of", "the", "sites", "in", "the", "layer", "that", "it", "takes", "as", "input", "is", "active", ".", "(", "Note", "that", "when", "using", "convolutions", ",", "each", "site", "is", "connected", "to", "sites", "in", "the", "hidden", "layer", "below", ".", ")", "Activity", "in", "a", "hidden", "layer", "thus", "follows", "an", "inductive", "definition", "in", "which", "each", "layer", "determines", "the", "set", "of", "active", "states", "in", "the", "next", ".", "In", "each", "hidden", "layer", ",", "inactive", "sites", "all", "have", "the", "same", "feature", "vector", ":", "the", "one", "corresponding", "to", "the", "ground", "state", ".", "Note", "that", "the", "ground", "state", "in", "a", "hidden", "layer", "is", "often", "not", "equal", "to", "zero", ",", "in", "particular", ",", "when", "convolutions", "with", "a", "bias", "term", "are", "used", ".", "However", ",", "irrespective", "of", "the", "value", "of", "the", "ground", "state", ",", "the", "ground", "-", "state", "value", "only", "needs", "to", "be", "calculated", "once", "per", "forward", "pass", "during", "training", "(", "and", "only", "once", "for", "all", "forward", "passes", "at", "test", "time", ")", ".", "This", "allows", "for", "substantial", "savings", "in", "computational", "and", "memory", "requirements", ";", "the", "exact", "savings", "depend", "on", "the", "data", "sparsity", "and", "the", "network", "depth", ".", "In", "this", "paper", ",", "we", "argue", "that", "the", "framework", "described", "above", "is", "unduly", "restrictive", ",", "in", "particular", ",", "because", "the", "convolution", "operation", "has", "not", "been", "modified", "to", "accommodate", "the", "sparsity", "of", "the", "input", "data", ".", "If", "the", "input", "data", "contains", "a", "single", "active", "site", ",", "then", "after", "applying", "a", "convolution", ",", "there", "will", "be", "active", "sites", ".", "Applying", "a", "second", "convolution", "of", "the", "same", "size", "will", "yield", "active", "sites", ",", "and", "so", "on", ".", "This", "rapid", "growth", "of", "the", "number", "of", "active", "sites", "is", "a", "poor", "prospect", "when", "implementing", "modern", "convolutional", "network", "architectures", "that", "comprise", "tens", "or", "even", "hundreds", "of", "convolutions", ",", "such", "as", "VGG", "networks", ",", "ResNets", ",", "and", "DenseNets", ".", "Of", "course", ",", "convolutional", "networks", "are", "not", "often", "applied", "to", "inputs", "that", "only", "have", "a", "single", "active", "site", ",", "but", "the", "aforementioned", "‘", "‘", "dilation", "’", "’", "problems", "are", "equally", "problematic", "when", "the", "input", "data", "comprises", "one", "-", "dimensional", "curves", "in", "spaces", "with", "two", "or", "more", "dimensions", ",", "or", "two", "-", "dimensional", "surfaces", "in", "three", "or", "more", "dimensions", ".", "To", "address", "the", "problems", "with", "dilation", "of", "active", "sites", ",", "we", "propose", "two", "slightly", "different", "convolution", "operations", "for", "use", "in", "convolutional", "networks", ".", "What", "the", "two", "operations", "have", "in", "common", "is", "that", "they", "both", "ignore", "the", "ground", "state", ":", "they", "replace", "the", "ground", "state", "with", "a", "zero", "vector", "to", "simplify", "the", "convolution", "operations", ".", "The", "difference", "between", "both", "operations", "is", "in", "how", "they", "handle", "active", "sites", ":", "instead", "of", "automatically", "making", "a", "site", "active", "if", "any", "of", "the", "inputs", "to", "its", "receptive", "field", "is", "active", "(", "thereby", "dilating", "the", "set", "of", "active", "sites", ")", ",", "our", "most", "efficient", "convolutional", "operation", "only", "considers", "the", "central", "input", ".", "As", "a", "result", ",", "the", "output", "set", "of", "active", "sites", "exactly", "mirrors", "that", "of", "the", "input", "set", ".", "We", "empirically", "demonstrate", "that", "use", "of", "our", "adapted", "convolutional", "operators", "allows", "us", "to", "build", "much", "deeper", "networks", "that", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", "whilst", "requiring", "much", "fewer", "resources", "by", "preserving", "sparsity", ".", "subsection", ":", "Submanifold", "Dilation", "In", "Figure", "[", "reference", "]", ",", "we", "show", "an", "example", "of", "a", "one", "-", "dimensional", "curve", "that", "is", "embedded", "on", "a", "two", "-", "dimensional", "grid", ".", "The", "figure", "shows", "that", "even", "when", "we", "apply", "small", "convolutions", "on", "this", "grid", ",", "the", "sparsity", "on", "the", "grid", "rapidly", "disappears", ".", "At", "the", "same", "time", ",", "if", "we", "restrict", "the", "output", "of", "the", "convolution", "only", "to", "the", "set", "of", "active", "input", "points", ",", "hidden", "layers", "in", "the", "network", "can", "not", "capture", "a", "lot", "of", "information", "that", "may", "relevant", "to", "the", "classification", "of", "the", "curve", ".", "In", "particular", ",", "two", "neighboring", "connected", "components", "will", "be", "treated", "completely", "independently", ".", "Luckily", ",", "nearly", "all", "convolutional", "networks", "incorporate", "some", "form", "of", "pooling", ",", "or", "use", "strided", "convolutions", ".", "These", "operations", "are", "essential", "in", "the", "sparse", "convolutional", "networks", "we", "investigate", ",", "as", "they", "allow", "neighboring", "components", "to", "merge", ".", "In", "particular", ",", "the", "closer", "the", "components", "are", ",", "the", "smaller", "the", "number", "of", "poolings", "/", "strided", "convolutions", "is", "that", "is", "necessary", "for", "the", "components", "to", "merge", "in", "the", "hidden", "-", "layer", "representations", ".", "subsection", ":", "Very", "Deep", "Convolutional", "Networks", "In", "image", "classification", ",", "very", "deep", "convolutional", "networks", "with", "small", "filters", ",", "often", "of", "size", "pixels", "and", "a", "padding", "of", "pixel", "(", "to", "preserve", "the", "size", "of", "the", "feature", "maps", ")", ",", "have", "proven", "to", "be", "very", "effective", ".", "Such", "small", "filters", "were", "used", "successfully", "in", "VGG", "networks", ",", "which", "have", "relatively", "wide", "layers", ".", "The", "introduction", "of", "residual", "networks", "(", "ResNets", ")", "showed", "that", "deeper", "but", "narrow", "networks", "with", "small", "filters", "are", "more", "efficient", ".", "The", "success", "of", "very", "deep", "ResNets", ",", "ResNeXt", "models", ",", "and", "DenseNets", "with", "bottleneck", "connections", "shows", "that", "it", "can", "be", "useful", "to", "calculate", "a", "relatively", "small", "number", "of", "features", "at", "a", "time", "and", "amalgamate", "these", "features", "into", "a", "larger", "state", "variable", ",", "either", "by", "vector", "-", "addition", "or", "feature", "-", "vector", "concatenation", ".", "Unfortunately", ",", "these", "techniques", "are", "impractical", "using", "existing", "sparse", "convolutional", "network", "implementations", ".", "One", "problem", "is", "that", "networks", "with", "multiple", "paths", "will", "tend", "to", "generate", "different", "sets", "of", "active", "paths", ",", "which", "would", "have", "to", "be", "merged", "to", "reconnect", "the", "outputs", ".", "It", "seems", "that", "this", "would", "be", "difficult", "to", "perform", "this", "merging", "efficiently", ".", "More", "importantly", ",", "ResNets", "and", "DenseNets", "generate", "such", "large", "receptive", "fields", "that", "sparsity", "would", "almost", "immediately", "be", "destroyed", "by", "the", "explosion", "in", "the", "number", "of", "active", "sites", ".", "section", ":", "(", "Valid", ")", "Sparse", "Convolutions", ":", "SC", "and", "VSC", "We", "define", "a", "sparse", "convolution", "SC", "(", "with", "input", "feature", "planes", ",", "output", "feature", "planes", ",", "a", "filter", "size", "of", ",", "and", "stride", ".", "We", "assume", "and", "to", "be", "odd", "integers", ",", "but", "we", "can", "allow", "generalization", "to", "non", "-", "square", "filters", ",", "e.g.", ",", "or", ",", "if", "we", "want", "to", "implement", "Inception", "-", "style", "factorised", "convolutions", ".", "An", "SC", "convolution", "computes", "the", "set", "of", "active", "sites", "in", "the", "same", "way", "as", "a", "regular", "convolution", ":", "it", "looks", "for", "the", "presence", "of", "any", "active", "sites", "in", "its", "receptive", "field", "of", "size", ".", "If", "the", "input", "has", "size", "then", "the", "output", "will", "have", "size", ".", "An", "SC", "convolution", "differs", "from", "a", "regular", "convolution", "in", "that", "it", "discards", "the", "ground", "state", "for", "non", "-", "active", "sites", "by", "assuming", "that", "the", "input", "from", "those", "sites", "is", "exactly", "zero", ".", "Whereas", "this", "is", "a", "seemingly", "small", "change", "to", "the", "convolution", "operation", ",", "it", "may", "bring", "computational", "benefits", "in", "practice", ".", "Next", ",", "we", "define", "a", "second", "type", "of", "sparse", "convolution", ",", "which", "forms", "the", "main", "contribution", "of", "this", "paper", ".", "Again", ",", "let", "denote", "an", "odd", "number", ",", "or", "collection", "of", "odd", "numbers", ",", "e.g.", ",", "or", ".", "We", "define", "a", "valid", "sparse", "convolution", "VSC", "as", "a", "modified", "SC", "convolution", ".", "First", ",", "we", "pad", "the", "input", "with", "on", "each", "side", ",", "so", "that", "the", "output", "will", "have", "the", "same", "size", "as", "the", "input", ".", "Next", ",", "we", "restrict", "an", "output", "site", "to", "be", "active", "if", "and", "only", "if", "the", "site", "at", "the", "corresponding", "site", "in", "the", "input", "is", "active", "(", "i.e.", ",", "if", "the", "central", "site", "in", "the", "receptive", "field", "is", "active", ")", ".", "Whenever", "an", "output", "site", "is", "determined", "to", "be", "active", ",", "its", "output", "feature", "vector", "is", "calculated", "by", "the", "SC", "operation", ".", "Table", "[", "reference", "]", "presents", "the", "computational", "and", "memory", "requirements", "of", "a", "regular", "convolution", "(", "C", ")", "and", "of", "our", "SC", "and", "VSC", "convolutions", ".", "To", "construct", "convolutional", "networks", "using", "SC", "and", "VSC", ",", "we", "also", "need", "activation", "functions", ",", "batch", "normalization", ",", "and", "pooling", ".", "Activation", "functions", "are", "defined", "as", "usual", ",", "but", "are", "restricted", "to", "the", "set", "of", "active", "sites", ".", "Similarly", ",", "we", "define", "batch", "normalization", "in", "terms", "of", "regular", "batch", "-", "normalization", "applied", "over", "the", "set", "of", "active", "sites", ".", "Max", "-", "pooling", "MP", "and", "average", "-", "pooling", "AP", "operations", "are", "defined", "as", "a", "variant", "of", "SC", ".", "MP", "takes", "the", "maximum", "of", "the", "zero", "vector", "and", "the", "input", "feature", "vectors", "in", "the", "receptive", "field", ".", "AP", "calculates", "times", "the", "sum", "of", "the", "active", "input", "vectors", ".", "We", "also", "define", "a", "deconvolution", "operation", "DC", "as", "an", "inverse", "of", "the", "SC", "convolution", ".", "The", "set", "of", "active", "output", "sites", "from", "a", "DC", "convolution", "is", "exactly", "the", "set", "of", "input", "active", "sites", "to", "the", "matching", "SC", "convolution", ".", "The", "set", "of", "connections", "between", "input", "-", "output", "sites", "is", "simply", "inverted", ".", "subsection", ":", "Submanifold", "Convolutional", "Networks", "We", "use", "a", "combination", "of", "VSC", "convolutions", ",", "strided", "SC", "convolutions", ",", "and", "sparse", "pooling", "operations", "to", "build", "sparse", "versions", "of", "the", "popular", "VGG", ",", "ResNet", ",", "and", "DenseNet", "convolutional", "networks", ".", "The", "blocks", "we", "use", "in", "our", "networks", "are", "presented", "in", "Figure", "[", "reference", "]", ".", "We", "refer", "to", "our", "networks", "as", "submanifold", "convolutional", "networks", ",", "because", "they", "are", "optimised", "to", "process", "low", "-", "dimensional", "data", "living", "in", "a", "space", "of", "higher", "dimensionality", ".", "We", "use", "the", "name", "VGG", "to", "refer", "to", "networks", "that", "contain", "a", "number", "of", "VSC", "(", ",", ",", "3", ",", "1", ")", "convolutions", ",", "separated", "by", "max", "-", "pooling", ".", "Each", "convolution", "is", "followed", "by", "batch", "normalization", "and", "a", "ReLU", "non", "-", "linearity", ".", "Similarly", ",", "we", "define", "‘", "‘", "pre", "-", "activated", "ResNets", "’", "’", "in", "which", "most", "data", "processing", "is", "performed", "by", "pairs", "of", "VSC", "(", ",", ",", "3", ",", "1", ")", "convolutions", ",", "and", "in", "which", "the", "residual", "connections", "are", "identity", "functions", ".", "Whenever", "the", "number", "of", "input", "/", "output", "features", "is", "different", ",", "we", "use", "a", "VSC", "(", ",", ",", "1", ",", "1", ")", "instead", ".", "Whenever", "there", "is", "change", "of", "scale", ",", "we", "replace", "the", "first", "convolution", "and", "the", "residual", "connection", "by", "a", "SC", "(", ",", ",", "3", ",", "2", ")", "convolution", ".", "This", "ensures", "that", "two", "branches", "can", "use", "the", "same", "hash", "table", "of", "active", "sites", ",", "and", "reduces", "additions", "to", "a", "simple", "sum", "of", "two", "equally", "sized", "matrices", ".", "The", "increased", "size", "of", "the", "residual", "connection", "’s", "receptive", "field", "also", "prevents", "excessive", "information", "loss", ".", "We", "also", "experiment", "with", "submanifold", "DenseNets", ".", "Herein", ",", "the", "word", "dense", "does", "not", "refer", "to", "a", "lack", "of", "spatial", "sparsity", "but", "rather", "to", "the", "pattern", "of", "connections", "between", "convolution", "operations", ".", "A", "simple", "DenseNet", "module", "is", "a", "sequence", "of", "convolutions", "in", "which", "each", "convolution", "takes", "as", "input", "the", "concatenated", "output", "of", "all", "the", "previous", "convolution", "operations", ".", "The", "bottleneck", "layers", "in", "our", "submanifold", "DenseNets", "are", "implemented", "in", "the", "same", "way", "as", "for", "ResNets", ".", "section", ":", "Implementation", "To", "implement", "(", "V", ")", "SC", "convolutions", "efficiently", ",", "we", "store", "the", "state", "of", "a", "input", "/", "hidden", "layer", "in", "two", "parts", ":", "a", "hash", "table", "and", "a", "matrix", ".", "The", "matrix", "has", "size", "and", "contains", "one", "row", "for", "each", "of", "the", "active", "sites", ".", "The", "hash", "table", "contains", "(", "location", ",", "row", ")", "pairs", "for", "all", "active", "sites", ":", "the", "location", "is", "a", "tuple", "of", "integer", "coordinates", ",", "and", "the", "row", "number", "indicates", "the", "corresponding", "row", "in", "the", "feature", "matrix", ".", "Given", "a", "convolution", "with", "filter", "size", ",", "we", "define", "a", "rule", "book", "to", "be", "a", "collection", "of", "integer", "matrices", "of", "size", ".", "To", "implement", "an", "SC", "(", "convolution", ",", "we", ":", "Iterate", "once", "through", "the", "the", "input", "hash", "-", "table", ".", "We", "build", "the", "output", "hash", "table", "and", "rule", "book", "on", "-", "the", "-", "fly", "by", "iterating", "over", "points", "in", "the", "output", "layer", "that", "receive", "input", "from", "a", "given", "point", "in", "the", "input", "layer", ".", "When", "an", "output", "site", "is", "visited", "for", "the", "first", "time", ",", "a", "new", "entry", "is", "created", "in", "the", "output", "hash", "table", ".", "Based", "on", "the", "spatial", "offset", "between", "the", "input", "and", "output", "points", ",", "a", "(", "input", "index", ",", "output", "index", ")", "pair", "is", "added", "to", "the", "rule", "book", ".", "Initialize", "the", "output", "matrix", "to", "all", "zeros", ".", "For", "each", ",", "there", "is", "a", "parameter", "matrix", "with", "size", ".", "For", "each", ",", "multiply", "the", "-", "th", "row", "of", "the", "input", "feature", "matrix", "by", "and", "add", "it", "to", "the", "-", "th", "row", "of", "the", "the", "output", "feature", "matrix", ".", "This", "can", "be", "implemented", "very", "efficiently", "on", "GPUs", "because", "it", "is", "a", "matrix", "-", "matrix", "multiply", "-", "add", "operation", ".", "To", "implement", "a", "VSC", "convolution", ",", "we", "re", "-", "use", "the", "input", "hash", "table", "for", "the", "output", ",", "and", "construct", "an", "appropriate", "rule", "book", ".", "Note", "that", "because", "the", "sparsity", "pattern", "does", "not", "change", ",", "the", "same", "rule", "book", "can", "be", "re", "-", "used", "in", "VGG", "/", "ResNet", "/", "DenseNet", "networks", "until", "a", "pooling", "or", "subsampling", "layer", "is", "encountered", ".", "If", "there", "are", "active", "points", "in", "the", "input", "layer", ",", "the", "cost", "of", "building", "the", "input", "hash", "-", "table", "is", ".", "For", "VGG", "/", "ResNet", "/", "DenseNet", "networks", ",", "assuming", "the", "number", "of", "active", "sites", "reduces", "by", "a", "multiplicative", "factor", "with", "each", "pooling", "operation", ",", "the", "cost", "of", "building", "all", "the", "hash", "-", "tables", "and", "rule", "-", "books", "is", "also", ",", "regardless", "of", "the", "depth", "of", "the", "network", ".", "section", ":", "Experiments", "We", "perform", "experiments", "on", "a", "2D", "and", "a", "3D", "dataset", "with", "sparse", "images", ".", "The", "CASIA", "dataset", "contains", "samples", "of", "3755", "GBK", "level", "-", "1", "characters", "with", "approximately", "240", "train", "and", "60", "test", "images", "per", "class", ".", "CJVK", "characters", "are", "good", "test", "cases", "for", "our", "models", "because", "they", "are", "a", "worst", "-", "case", "scenario", "for", "sparse", "convolutional", "networks", ":", "when", "drawn", "at", "scale", ",", "about", "8", "%", "of", "the", "pixels", "are", "active", ",", "but", "this", "percentage", "rapidly", "decreases", "after", "pooling", "due", "to", "the", "small", "density", "of", "the", "pen", "strokes", ".", "This", "makes", "them", "a", "good", "test", "case", "for", "our", "models", ".", "The", "ModelNet", "-", "40", "datasethttp:", "//", "3dshapenets.cs.princeton.edu", "/", "contain", "2468", "CAD", "models", "that", "contain", "shapes", "corresponding", "to", "40", "classes", ".", "We", "follow", "the", "preprocessing", "of", "before", "feeding", "the", "models", "into", "our", "convolutional", "networks", ".", "All", "CAD", "models", "were", "rendered", "as", "surfaces", "at", "size", ".", "subsection", ":", "Results", "on", "CASIA", "We", "first", "experiment", "with", "two", "VGG", "architectures", "on", "the", "CASIA", "dataset", ".", "We", "trained", "all", "models", "for", "100", "epochs", "using", "batches", "of", "size", "100", ",", "SGD", "with", "momentum", "0.9", ",", "a", "weight", "decay", "of", ",", "and", "a", "learning", "rate", "decay", "of", "5", "%", "per", "epoch", ".", "For", "simplicity", ",", "we", "do", "not", "employ", "any", "data", "augmentation", ".", "The", "architectures", "of", "our", "VGG", "networks", "and", "their", "performances", "are", "presented", "in", "Table", "[", "reference", "]", ".", "We", "observe", "that", "‘", "‘", "regular", "’", "’", "C", "convolutions", "and", "‘", "‘", "sparse", "’", "’", "SC", "convolutions", "achieve", "the", "same", "error", ":", "this", "result", "suggests", "that", "discarding", "the", "ground", "state", "has", "essentially", "no", "negative", "impact", "on", "performance", ".", "This", "is", "an", "argument", "for", "always", "discarding", "ground", "states", ",", "as", "it", "makes", "things", "easier", "computationally", "and", "algorithmically", ".", "Comparing", "SC", "with", "VSC", "convolutions", ",", "we", "observe", "a", "minimal", "loss", "in", "performance", "by", "considering", "only", "the", "valid", "part", "of", "the", "convolution", ".", "This", "minimal", "loss", "in", "accuracy", "does", "facilitate", "great", "computational", "improvements", ":", "networks", "using", "VSC", "use", "2", "to", "3", "less", "computation", "and", "memory", ".", "Next", ",", "we", "performed", "experiments", "on", "CASIA", "with", "submanifold", "ResNets", ".", "The", "key", "difference", "between", "our", "implementation", "of", "ResNets", "and", "regular", "ResNets", "is", "that", "stride", "-", "2", "ResNet", "modules", "use", "SC", "(", ",", ",", "3", ",", "2", ")", "convolutions", "for", "the", "strided", "convolution", ",", "rather", "than", "SC", "(", ",", ",", "1", ",", "2", ")", ".", "This", "change", "is", "necessary", "to", "ensure", "the", "two", "branches", "produce", "the", "same", "set", "of", "active", "sites", ",", "which", "simplifies", "bookkeeping", "and", "turns", "the", "addition", "operation", "into", "a", "simple", "matrix", "-", "matrix", "addition", ".", "Unlike", "the", "VSC", "convolutions", "that", "are", "used", "in", "most", "layers", ",", "the", "SC", "(", ",", ",", "3", ",", "2", ")", "we", "use", "after", "downsampling", "leads", "sites", "to", "be", "active", "if", "any", "of", "its", "inputs", "are", "active", ",", "which", "avoids", "information", "loss", "in", "the", "transition", ".", "The", "architectures", "of", "our", "ResNet", "networks", "and", "their", "performances", "are", "presented", "in", "Table", "[", "reference", "]", ".", "The", "results", "with", "ResNets", "are", "in", "line", "with", "those", "obtained", "using", "VGG", "networks", ":", "we", "obtain", "reductions", "in", "computational", "and", "memory", "requirements", "by", "at", "least", "a", "factor", "of", "2", "at", "a", "minimal", "loss", "in", "accuracy", ".", "We", "also", "performed", "experiments", "with", "DenseNets", ";", "please", "see", "Table", "[", "reference", "]", ".", "Next", "we", "experimented", "with", "adding", "extra", "connections", "to", "VGG", "networks", "to", "increase", "the", "effective", "receptive", "fields", "of", "the", "hidden", "states", ";", "see", "Table", "[", "reference", "]", "for", "results", ".", "In", "the", "table", ",", "denotes", "a", "VSC", "convolution", "performed", "in", "parallel", "with", "a", "chain", "of", "SC", "-", "VSC", "-", "DC", "operations", ";", "outputs", "are", "concatenated", "to", "produce", "output", "feature", "planes", ".", "To", "simplify", "the", "network", "design", ",", "we", "switched", "to", "size", "-", "3", "stride", "-", "2", "max", "-", "pooling", ",", "matching", "the", "SC", "convolutions", "in", "the", "SC", "-", "VSC", "-", "DC", "branches", ",", "and", "reduce", "the", "input", "size", "from", "64", "64", "to", "63", "63", ".", "Figure", "[", "reference", "]", "presents", "an", "overview", "of", "all", "our", "results", "on", "the", "CASIA", "dataset", ".", "subsection", ":", "Results", "on", "ModelNet", "In", "a", "second", "set", "of", "experiments", ",", "we", "compare", "two", "submanifold", "VGG", "networks", "with", "a", "state", "-", "of", "-", "the", "-", "art", "dense", "convolutional", "network", "on", "the", "ModelNet", "-", "40", "dataset", ".", "The", "results", "of", "these", "experiments", "are", "shown", "in", "Table", "[", "reference", "]", ":", "the", "left", "part", "of", "the", "table", "shows", "the", "architecture", "and", "performance", "of", "our", "submanifold", "VGG", "networks", ",", "whereas", "the", "right", "part", "of", "the", "table", "shows", "that", "of", "the", "dense", "3DNiN", "network", ".", "The", "results", "clearly", "demonstrate", "that", "submanifold", "have", "the", "potential", "for", "designing", "convolutional", "networks", "for", "sparse", "data", "that", "obtain", "state", "-", "of", "-", "the", "-", "art", "performance", "with", "limited", "computational", "requirements", ":", "in", "particular", ",", "our", "VGG", "-", "A", "network", "makes", "2", "%", "more", "errors", "at", "13", "fewer", "computations", ",", "and", "our", "VGG", "-", "B", "performs", "roughly", "on", "par", "with", "the", "dense", "3DNiN", "whilst", "performing", "fewer", "computations", ".", "section", ":", "Related", "Work", "This", "paper", "is", "not", "the", "first", "to", "study", "sparse", "convolutional", "networks", ".", "Most", "prior", "networks", "for", "sparse", "data", "implements", "a", "standard", "convolutional", "operator", "that", "increases", "the", "number", "of", "active", "sites", "with", "each", "layer", ".", "By", "contrast", ",", "our", "submanifold", "convolutional", "networks", "allows", "sparse", "data", "to", "be", "processed", "whilst", "retraining", "a", "much", "greater", "degree", "of", "sparsity", ".", "We", "have", "shown", "that", "this", "makes", "it", "practical", "to", "train", "deep", "and", "efficient", "VGG", "and", "ResNet", "models", ".", "Submanifold", "convolutional", "networks", "are", "also", "much", "sparser", "than", "OctNets", ".", "OctNet", "stores", "data", "in", "oct", "-", "trees", ":", "a", "data", "structure", "in", "which", "the", "grid", "cube", "is", "progressively", "subdivided", "into", "smaller", "sub", "-", "cubes", "until", "the", "sub", "-", "cubes", "are", "either", "empty", "or", "contain", "a", "single", "active", "site", ".", "To", "compare", "the", "efficiency", "of", "OctNets", "with", "that", "of", "submanifold", "convolutional", "networks", ",", "we", "picked", "a", "random", "sample", "from", "the", "ModelNet", "-", "40", "dataset", "and", "rendered", "it", "in", "a", "cube", "with", "grid", "points", ".", "The", "resulting", "grid", "had", "423", "active", "sites", ",", "which", "corresponds", "to", "1.3", "%", "of", "the", "total", "number", "of", "sites", ".", "Each", "active", "site", "had", "on", "average", "12.4", "active", "neighbors", "(", "the", "maximum", "possible", "number", "of", "neighbors", "is", "27", ")", ".", "VSC", "convolutions", ",", "therefore", ",", "require", "only", "0.6", "%", "of", "the", "work", "of", "a", "dense", "(", "C", ")", "convolution", ".", "However", ",", "in", "the", "OctTree", ",", "80", "%", ",", "13", "%", ",", "4", "%", ",", "and", "3", "%", "of", "the", "volume", "of", "the", "cube", "is", "covered", "by", "sub", "-", "cubes", "of", "size", ",", ",", "and", ",", "respectively", ".", "As", "a", "result", ",", "an", "OctNet", "convolution", ",", "which", "operates", "over", "the", "surfaces", "of", "the", "smaller", "cubes", ",", "requires", "about", "35", "%", "of", "the", "computations", "that", "a", "dense", "(", "C", ")", "convolution", "requires", ".", "In", "this", "particular", "example", ",", "an", "OctNet", "convolution", "thus", "has", "a", "computational", "cost", "that", "is", "60", "times", "higher", "than", "that", "of", "a", "VSC", "convolution", ".", "Submanifold", "convolutional", "networks", "also", "have", "advantages", "in", "terms", "of", "memory", "requirements", ".", "In", "particular", ",", "a", "submanifold", "network", "stores", "a", "single", "feature", "vector", "for", "each", "of", "the", "active", "sites", ".", "By", "contrast", ",", "OctTrees", "have", "about", "twice", "as", "many", "empty", "child", "nodes", "as", "active", "nodes", ",", "which", "implies", "they", "have", "to", "store", "roughly", "three", "times", "as", "many", "features", "as", "a", "submanifold", "convolutional", "network", ".", "Having", "said", "that", ",", "some", "of", "the", "ideas", "of", "may", "be", "combined", "with", "VSC", "convolutions", ".", "In", "particular", ",", "it", "is", "possible", "to", "use", "oct", "-", "trees", "as", "a", "specialized", "hash", "function", "in", "VSC", "convolutions", ".", "Such", "an", "oct", "-", "tree", "-", "based", "hash", "function", "has", "the", "potential", "to", "be", "faster", "than", "a", "standard", "universal", "hash", "function", "that", "operates", "on", "integer", "tuple", "keys", ",", "like", "in", "our", "implementation", "of", "VSC", ".", "section", ":", "Conclusion", "We", "introduced", "a", "new", "sparse", "convolutional", "operator", ",", "called", "valid", "sparse", "convolution", "(", "VSC", ")", ",", "that", "facilitates", "the", "design", "of", "efficient", ",", "deep", "convolutional", "networks", "for", "sparse", "data", ".", "We", "have", "shown", "that", "VSC", "convolutions", "lead", "to", "substantial", "computational", "savings", "whilst", "maintain", "state", "-", "of", "-", "the", "-", "art", "accuracies", "on", "two", "datasets", ":", "a", "dataset", "comprising", "one", "-", "dimensional", "manifolds", "embedded", "in", "two", "-", "dimensional", "space", ",", "and", "a", "dataset", "comprising", "two", "-", "dimensional", "surfaces", "embedded", "in", "three", "-", "dimensional", "space", ".", "As", "part", "of", "this", "paper", ",", "we", "are", "releasing", "easy", "-", "to", "-", "use", "implementations", "of", "VSC", "and", "the", "other", "sparse", "operations", "we", "used", "in", "the", "networks", "described", "in", "this", "paper", ".", "We", "will", "also", "release", "code", "to", "reproduce", "the", "results", "of", "our", "experiments", ".", "bibliography", ":", "References"], "entity": [[2, 6, ["Submanifold", "Sparse", "Convolutional", "Networks"], "Method"], [6, 8, ["Convolutional", "network"], "Method"], [78, 80, ["LiDAR", "scanner"], "Method"], [81, 85, ["RGB", "-", "D", "camera"], "Method"], [89, 93, ["dense", "’", "’", "implementations"], "Method"], [94, 96, ["convolutional", "networks"], "Method"], [109, 112, ["sparse", "convolutional", "operation"], "Method"], [123, 126, ["sparse", "convolutional", "networks"], "Method"], [156, 160, ["submanifold", "sparse", "convolutional", "networks"], "Method"], [184, 186, ["Convolutional", "networks"], "Method"], [204, 214, ["analysis", "of", "data", "with", "spatial", "and", "/", "or", "temporal", "structure"], "Task"], [222, 227, ["three", "-", "dimensional", "surface", "models"], "Method"], [273, 277, ["RGB", "-", "D", "cameras"], "Method"], [285, 287, ["OFF", "models"], "Method"], [354, 356, ["data", "processing"], "Task"], [381, 384, ["convolutional", "network", "implementations"], "Method"], [409, 412, ["convolutional", "network", "implementations"], "Method"], [435, 438, ["regular", "convolutional", "network"], "Method"], [457, 458, ["OctNets"], "Method"], [461, 463, ["convolution", "operator"], "Method"], [491, 496, ["sparse", "implementations", "of", "convolutional", "networks"], "Method"], [517, 521, ["full", "’", "’", "convolution"], "Method"], [535, 537, ["convolutional", "networks"], "Method"], [564, 566, ["convolution", "operators"], "Method"], [567, 569, ["sparse", "convolution"], "Method"], [570, 571, ["SC"], "Method"], [573, 576, ["valid", "sparse", "convolution"], "Method"], [577, 578, ["VSC"], "Method"], [584, 590, ["recognizing", "handwritten", "digits", "and", "3D", "shapes"], "Task"], [593, 594, ["SC"], "Method"], [595, 596, ["VSC"], "Method"], [620, 624, ["-", "dimensional", "convolutional", "network"], "Method"], [738, 739, ["thresholding"], "Task"], [796, 798, ["convolutional", "network"], "Method"], [824, 826, ["hidden", "layer"], "Method"], [849, 850, ["convolutions"], "Method"], [872, 874, ["inductive", "definition"], "Method"], [930, 931, ["convolutions"], "Method"], [1022, 1024, ["convolution", "operation"], "Method"], [1051, 1052, ["convolution"], "Method"], [1092, 1095, ["convolutional", "network", "architectures"], "Method"], [1109, 1110, ["ResNets"], "Method"], [1112, 1113, ["DenseNets"], "Method"], [1117, 1119, ["convolutional", "networks"], "Method"], [1178, 1182, ["dilation", "of", "active", "sites"], "Task"], [1188, 1190, ["convolution", "operations"], "Method"], [1193, 1195, ["convolutional", "networks"], "Method"], [1271, 1273, ["convolutional", "operation"], "Method"], [1305, 1307, ["convolutional", "operators"], "Method"], [1312, 1314, ["deeper", "networks"], "Method"], [1335, 1337, ["Submanifold", "Dilation"], "Method"], [1372, 1373, ["convolutions"], "Method"], [1397, 1398, ["convolution"], "Method"], [1424, 1428, ["classification", "of", "the", "curve"], "Task"], [1446, 1448, ["convolutional", "networks"], "Method"], [1452, 1453, ["pooling"], "Method"], [1456, 1458, ["strided", "convolutions"], "Method"], [1465, 1468, ["sparse", "convolutional", "networks"], "Method"], [1495, 1497, ["strided", "convolutions"], "Method"], [1508, 1512, ["hidden", "-", "layer", "representations"], "Method"], [1516, 1519, ["Deep", "Convolutional", "Networks"], "Method"], [1520, 1522, ["image", "classification"], "Task"], [1524, 1527, ["deep", "convolutional", "networks"], "Method"], [1528, 1530, ["small", "filters"], "Method"], [1577, 1579, ["residual", "networks"], "Method"], [1580, 1581, ["ResNets"], "Method"], [1586, 1588, ["narrow", "networks"], "Method"], [1589, 1591, ["small", "filters"], "Method"], [1599, 1601, ["deep", "ResNets"], "Task"], [1602, 1604, ["ResNeXt", "models"], "Method"], [1606, 1610, ["DenseNets", "with", "bottleneck", "connections"], "Method"], [1639, 1642, ["vector", "-", "addition"], "Method"], [1643, 1647, ["feature", "-", "vector", "concatenation"], "Method"], [1656, 1660, ["sparse", "convolutional", "network", "implementations"], "Method"], [1706, 1707, ["ResNets"], "Method"], [1708, 1709, ["DenseNets"], "Method"], [1736, 1738, ["Sparse", "Convolutions"], "Method"], [1739, 1740, ["SC"], "Method"], [1741, 1742, ["VSC"], "Method"], [1745, 1748, ["sparse", "convolution", "SC"], "Method"], [1794, 1799, ["Inception", "-", "style", "factorised", "convolutions"], "Method"], [1815, 1817, ["regular", "convolution"], "Method"], [1852, 1854, ["regular", "convolution"], "Method"], [1887, 1889, ["convolution", "operation"], "Method"], [1906, 1908, ["sparse", "convolution"], "Method"], [1940, 1943, ["sparse", "convolution", "VSC"], "Method"], [2046, 2048, ["regular", "convolution"], "Method"], [2054, 2055, ["SC"], "Method"], [2056, 2058, ["VSC", "convolutions"], "Method"], [2061, 2063, ["convolutional", "networks"], "Method"], [2064, 2065, ["SC"], "Method"], [2066, 2067, ["VSC"], "Method"], [2074, 2076, ["batch", "normalization"], "Method"], [2078, 2079, ["pooling"], "Method"], [2101, 2103, ["batch", "normalization"], "Method"], [2106, 2110, ["regular", "batch", "-", "normalization"], "Method"], [2118, 2122, ["Max", "-", "pooling", "MP"], "Method"], [2123, 2128, ["average", "-", "pooling", "AP", "operations"], "Method"], [2134, 2135, ["SC"], "Method"], [2136, 2137, ["MP"], "Method"], [2154, 2155, ["AP"], "Method"], [2169, 2172, ["deconvolution", "operation", "DC"], "Method"], [2188, 2190, ["DC", "convolution"], "Method"], [2219, 2222, ["Submanifold", "Convolutional", "Networks"], "Method"], [2227, 2229, ["VSC", "convolutions"], "Method"], [2230, 2233, ["strided", "SC", "convolutions"], "Method"], [2235, 2238, ["sparse", "pooling", "operations"], "Method"], [2240, 2242, ["sparse", "versions"], "Method"], [2245, 2246, ["VGG"], "Method"], [2247, 2248, ["ResNet"], "Method"], [2250, 2253, ["DenseNet", "convolutional", "networks"], "Method"], [2275, 2278, ["submanifold", "convolutional", "networks"], "Method"], [2301, 2302, ["VGG"], "Method"], [2323, 2326, ["max", "-", "pooling"], "Method"], [2328, 2329, ["convolution"], "Method"], [2332, 2334, ["batch", "normalization"], "Method"], [2336, 2340, ["ReLU", "non", "-", "linearity"], "Method"], [2356, 2358, ["data", "processing"], "Task"], [2363, 2364, ["VSC"], "Method"], [2397, 2398, ["VSC"], "Method"], [2417, 2419, ["first", "convolution"], "Method"], [2513, 2515, ["DenseNet", "module"], "Method"], [2519, 2520, ["convolutions"], "Method"], [2523, 2524, ["convolution"], "Method"], [2534, 2536, ["convolution", "operations"], "Method"], [2542, 2544, ["submanifold", "DenseNets"], "Method"], [2552, 2553, ["ResNets"], "Method"], [2641, 2645, ["convolution", "with", "filter", "size"], "Method"], [2819, 2820, ["GPUs"], "Method"], [2824, 2831, ["matrix", "-", "matrix", "multiply", "-", "add", "operation"], "Method"], [2877, 2878, ["VGG"], "Method"], [2879, 2880, ["ResNet"], "Method"], [2881, 2883, ["DenseNet", "networks"], "Method"], [2885, 2889, ["pooling", "or", "subsampling", "layer"], "Method"], [2914, 2915, ["VGG"], "Method"], [2916, 2917, ["ResNet"], "Method"], [2918, 2920, ["DenseNet", "networks"], "Method"], [2934, 2936, ["pooling", "operation"], "Method"], [3019, 3022, ["sparse", "convolutional", "networks"], "Method"], [3043, 3044, ["pooling"], "Method"], [3075, 3077, ["CAD", "models"], "Method"], [3096, 3098, ["convolutional", "networks"], "Method"], [3100, 3102, ["CAD", "models"], "Method"], [3139, 3140, ["SGD"], "Method"], [3151, 3154, ["learning", "rate", "decay"], "Metric"], [3168, 3170, ["data", "augmentation"], "Method"], [3196, 3198, ["C", "convolutions"], "Method"], [3247, 3248, ["SC"], "Method"], [3249, 3251, ["VSC", "convolutions"], "Method"], [3273, 3274, ["accuracy"], "Metric"], [3280, 3281, ["networks"], "Method"], [3282, 3283, ["VSC"], "Method"], [3298, 3299, ["CASIA"], "Task"], [3300, 3302, ["submanifold", "ResNets"], "Task"], [3310, 3311, ["ResNets"], "Method"], [3312, 3314, ["regular", "ResNets"], "Method"], [3333, 3335, ["strided", "convolution"], "Method"], [3370, 3372, ["addition", "operation"], "Method"], [3375, 3379, ["matrix", "-", "matrix", "addition"], "Method"], [3382, 3384, ["VSC", "convolutions"], "Method"], [3445, 3446, ["ResNets"], "Method"], [3460, 3461, ["computational"], "Metric"], [3462, 3464, ["memory", "requirements"], "Metric"], [3476, 3477, ["accuracy"], "Metric"], [3536, 3542, ["SC", "-", "VSC", "-", "DC", "operations"], "Method"], [3555, 3557, ["network", "design"], "Method"], [3564, 3570, ["stride", "-", "2", "max", "-", "pooling"], "Method"], [3577, 3583, ["SC", "-", "VSC", "-", "DC", "branches"], "Method"], [3639, 3642, ["dense", "convolutional", "network"], "Method"], [3690, 3693, ["dense", "3DNiN", "network"], "Method"], [3699, 3700, ["submanifold"], "Method"], [3705, 3707, ["convolutional", "networks"], "Method"], [3754, 3756, ["dense", "3DNiN"], "Method"], [3773, 3776, ["sparse", "convolutional", "networks"], "Method"], [3778, 3780, ["prior", "networks"], "Method"], [3786, 3788, ["convolutional", "operator"], "Method"], [3803, 3806, ["submanifold", "convolutional", "networks"], "Method"], [3831, 3835, ["deep", "and", "efficient", "VGG"], "Method"], [3836, 3838, ["ResNet", "models"], "Method"], [3839, 3842, ["Submanifold", "convolutional", "networks"], "Method"], [3847, 3848, ["OctNets"], "Method"], [3849, 3850, ["OctNet"], "Method"], [3853, 3856, ["oct", "-", "trees"], "Method"], [3858, 3860, ["data", "structure"], "Method"], [3893, 3894, ["OctNets"], "Method"], [3897, 3900, ["submanifold", "convolutional", "networks"], "Method"], [3962, 3964, ["VSC", "convolutions"], "Method"], [3976, 3981, ["dense", "(", "C", ")", "convolution"], "Method"], [3986, 3987, ["OctTree"], "Method"], [4025, 4027, ["OctNet", "convolution"], "Method"], [4047, 4052, ["dense", "(", "C", ")", "convolution"], "Method"], [4060, 4062, ["OctNet", "convolution"], "Method"], [4079, 4082, ["Submanifold", "convolutional", "networks"], "Method"], [4088, 4090, ["memory", "requirements"], "Metric"], [4095, 4097, ["submanifold", "network"], "Method"], [4112, 4113, ["OctTrees"], "Method"], [4139, 4142, ["submanifold", "convolutional", "network"], "Method"], [4156, 4158, ["VSC", "convolutions"], "Method"], [4167, 4170, ["oct", "-", "trees"], "Method"], [4173, 4175, ["hash", "function"], "Method"], [4176, 4178, ["VSC", "convolutions"], "Method"], [4181, 4188, ["oct", "-", "tree", "-", "based", "hash", "function"], "Method"], [4197, 4200, ["universal", "hash", "function"], "Method"], [4212, 4213, ["VSC"], "Method"], [4221, 4224, ["sparse", "convolutional", "operator"], "Method"], [4226, 4229, ["valid", "sparse", "convolution"], "Method"], [4230, 4231, ["VSC"], "Method"], [4240, 4243, ["deep", "convolutional", "networks"], "Method"], [4251, 4253, ["VSC", "convolutions"], "Method"], [4267, 4268, ["accuracies"], "Metric"], [4317, 4318, ["VSC"], "Method"], [4321, 4323, ["sparse", "operations"], "Method"], [350, 351, ["computational"], "Metric"], [443, 444, ["computational"], "Metric"], [610, 612, ["memory", "requirements"], "Metric"], [984, 985, ["computational"], "Metric"], [986, 988, ["memory", "requirements"], "Metric"], [1106, 1107, ["VGG"], "Method"], [1565, 1566, ["VGG"], "Method"], [1801, 1802, ["SC"], "Method"], [1847, 1848, ["SC"], "Method"], [1893, 1894, ["computational"], "Metric"], [1946, 1947, ["SC"], "Method"], [2031, 2032, ["SC"], "Method"], [2040, 2041, ["computational"], "Metric"], [2042, 2044, ["memory", "requirements"], "Metric"], [2177, 2178, ["SC"], "Method"], [2201, 2202, ["SC"], "Method"], [2311, 2312, ["VSC"], "Method"], [2425, 2426, ["SC"], "Method"], [2562, 2563, ["SC"], "Method"], [2664, 2665, ["SC"], "Method"], [2835, 2836, ["VSC"], "Method"], [3119, 3120, ["VGG"], "Method"], [3175, 3176, ["VGG"], "Method"], [3204, 3205, ["SC"], "Method"], [3277, 3278, ["computational"], "Metric"], [3319, 3320, ["ResNet"], "Method"], [3322, 3323, ["SC"], "Method"], [3338, 3339, ["SC"], "Method"], [3392, 3393, ["SC"], "Method"], [3429, 3430, ["ResNet"], "Method"], [3453, 3454, ["VGG"], "Method"], [3500, 3501, ["VGG"], "Method"], [3527, 3528, ["VSC"], "Method"], [3573, 3574, ["SC"], "Method"], [3628, 3629, ["VGG"], "Method"], [3676, 3677, ["VGG"], "Method"], [3722, 3723, ["computational"], "Metric"], [3729, 3730, ["VGG"], "Method"], [3745, 3746, ["VGG"], "Method"], [4065, 4066, ["computational"], "Metric"], [4076, 4077, ["VSC"], "Method"], [4256, 4257, ["computational"], "Metric"]]} {"doc_id": "11da0c54ba904a1cb31a09d10da55f73e8825c61", "text": ["document", ":", "Natural", "Language", "Inference", "by", "Tree", "-", "Based", "Convolution", "and", "Heuristic", "Matching", "In", "this", "paper", ",", "we", "propose", "the", "TBCNN", "-", "pair", "model", "to", "recognize", "entailment", "and", "contradiction", "between", "two", "sentences", ".", "In", "our", "model", ",", "a", "tree", "-", "based", "convolutional", "neural", "network", "(", "TBCNN", ")", "captures", "sentence", "-", "level", "semantics", ";", "then", "heuristic", "matching", "layers", "like", "concatenation", ",", "element", "-", "wise", "product", "/", "difference", "combine", "the", "information", "in", "individual", "sentences", ".", "Experimental", "results", "show", "that", "our", "model", "outperforms", "existing", "sentence", "encoding", "-", "based", "approaches", "by", "a", "large", "margin", ".", "section", ":", "Introduction", "Recognizing", "entailment", "and", "contradiction", "between", "two", "sentences", "(", "called", "a", "premise", "and", "a", "hypothesis", ")", "is", "known", "as", "natural", "language", "inference", "(", "NLI", ")", "in", "inference", ".", "Provided", "with", "a", "premise", "sentence", ",", "the", "task", "is", "to", "judge", "whether", "the", "hypothesis", "can", "be", "inferred", "(", "entailment", ")", ",", "or", "the", "hypothesis", "can", "not", "be", "true", "(", "contradiction", ")", ".", "Several", "examples", "are", "illustrated", "in", "Table", "[", "reference", "]", ".", "NLI", "is", "in", "the", "core", "of", "natural", "language", "understanding", "and", "has", "wide", "applications", "in", "NLP", ",", "e.g.", ",", "question", "answering", "and", "automatic", "summarization", ".", "Moreover", ",", "NLI", "is", "also", "related", "to", "other", "tasks", "of", "sentence", "pair", "modeling", ",", "including", "paraphrase", "detection", ",", "relation", "recognition", "of", "discourse", "units", ",", "etc", ".", "Traditional", "approaches", "to", "NLI", "mainly", "fall", "into", "two", "groups", ":", "feature", "-", "rich", "models", "and", "formal", "reasoning", "methods", ".", "Feature", "-", "based", "approaches", "typically", "leverage", "machine", "learning", "models", ",", "but", "require", "intensive", "human", "engineering", "to", "represent", "lexical", "and", "syntactic", "information", "in", "two", "sentences", ".", "Formal", "reasoning", ",", "on", "the", "other", "hand", ",", "converts", "a", "sentence", "into", "a", "formal", "logical", "representation", "and", "uses", "interpreters", "to", "search", "for", "a", "proof", ".", "However", ",", "such", "approaches", "are", "limited", "in", "terms", "of", "scope", "and", "accuracy", ".", "The", "renewed", "prosperity", "of", "neural", "networks", "has", "made", "significant", "achievements", "in", "various", "NLP", "applications", ",", "including", "individual", "sentence", "modeling", "as", "well", "as", "sentence", "matching", ".", "A", "typical", "neural", "architecture", "to", "model", "sentence", "pairs", "is", "the", "“", "Siamese", "”", "structure", ",", "which", "involves", "an", "underlying", "sentence", "model", "and", "a", "matching", "layer", "to", "determine", "the", "relationship", "between", "two", "sentences", ".", "Prevailing", "sentence", "models", "include", "convolutional", "networks", "and", "recurrent", "/", "recursive", "networks", ".", "Although", "they", "have", "achieved", "high", "performance", ",", "they", "may", "either", "fail", "to", "fully", "make", "use", "of", "the", "syntactical", "information", "in", "sentences", "or", "be", "difficult", "to", "train", "due", "to", "the", "long", "propagation", "path", ".", "Recently", ",", "we", "propose", "a", "novel", "tree", "-", "based", "convolutional", "neural", "network", "(", "TBCNN", ")", "to", "alleviate", "the", "aforementioned", "problems", "and", "have", "achieved", "higher", "performance", "in", "two", "sentence", "classification", "tasks", ".", "However", ",", "it", "is", "less", "clear", "whether", "TBCNN", "can", "be", "harnessed", "to", "model", "sentence", "pairs", "for", "implicit", "logical", "inference", ",", "as", "is", "in", "the", "NLI", "task", ".", "In", "this", "paper", ",", "we", "propose", "the", "TBCNN", "-", "pair", "neural", "model", "to", "recognize", "entailment", "and", "contradiction", "between", "two", "sentences", ".", "We", "leverage", "our", "newly", "proposed", "TBCNN", "model", "to", "capture", "structural", "information", "in", "sentences", ",", "which", "is", "important", "to", "NLI", ".", "For", "example", ",", "the", "phrase", "“", "riding", "bicycles", "on", "the", "streets", "”", "in", "Table", "[", "reference", "]", "can", "be", "well", "recognized", "by", "TBCNN", "via", "the", "dependency", "relations", "dobj", "(", "riding", ",", "bicycles", ")", "and", "prep_on", "(", "riding", ",", "street", ")", ".", "As", "we", "can", "see", ",", "TBCNN", "is", "more", "robust", "than", "sequential", "convolution", "in", "terms", "of", "word", "order", "distortion", ",", "which", "may", "be", "introduced", "by", "determinators", ",", "modifiers", ",", "etc", ".", "A", "pooling", "layer", "then", "aggregates", "information", "along", "the", "tree", ",", "serving", "as", "a", "way", "of", "semantic", "compositonality", ".", "Finally", ",", "two", "sentences", "’", "information", "is", "combined", "by", "several", "heuristic", "matching", "layers", ",", "including", "concatenation", ",", "element", "-", "wise", "product", "and", "difference", ";", "they", "are", "effective", "in", "capturing", "relationships", "between", "two", "sentences", ",", "but", "remain", "low", "complexity", ".", "To", "sum", "up", ",", "the", "main", "contributions", "of", "this", "paper", "are", "two", "-", "fold", ":", "(", "1", ")", "We", "are", "the", "first", "to", "introduce", "tree", "-", "based", "convolution", "to", "sentence", "pair", "modeling", "tasks", "like", "NLI", ";", "(", "2", ")", "Leveraging", "additional", "heuristics", "further", "improves", "the", "accuracy", "while", "remaining", "low", "complexity", ",", "outperforming", "existing", "sentence", "encoding", "-", "based", "approaches", "to", "a", "large", "extent", ",", "including", "feature", "-", "rich", "methods", "and", "long", "short", "term", "memory", "(", "LSTM", ")-", "based", "recurrent", "networks", ".", "section", ":", "Related", "Work", "Entailment", "recognition", "can", "be", "viewed", "as", "a", "task", "of", "sentence", "pair", "modeling", ".", "Most", "neural", "networks", "in", "this", "field", "involve", "a", "sentence", "-", "level", "model", ",", "followed", "by", "one", "or", "a", "few", "matching", "layers", ".", "They", "are", "sometimes", "called", "“", "Siamese", "”", "architectures", ".", "CNN", ":", "NIPS", "and", "CNN", ":", "NAACL", "apply", "convolutional", "neural", "networks", "(", "CNNs", ")", "as", "the", "individual", "sentence", "model", ",", "where", "a", "set", "of", "feature", "detectors", "over", "successive", "words", "are", "designed", "to", "extract", "local", "features", ".", "LSTM", ":", "AAAI", "build", "sentence", "pair", "models", "upon", "recurrent", "neural", "networks", "(", "RNNs", ")", "to", "iteratively", "integrate", "information", "along", "a", "sentence", ".", "recurparaphrase", "dynamically", "construct", "tree", "structures", "(", "analogous", "to", "parse", "trees", ")", "by", "recursive", "autoencoders", "to", "detect", "paraphrase", "between", "two", "sentences", ".", "As", "shown", ",", "inherent", "structural", "information", "in", "sentences", "is", "oftentimes", "important", "to", "natural", "language", "understanding", ".", "The", "simplest", "approach", "to", "match", "two", "sentences", ",", "perhaps", ",", "is", "to", "concatenate", "their", "vector", "representations", ".", "Concatenation", "is", "also", "applied", "in", "our", "previous", "work", "of", "matching", "the", "subject", "and", "object", "in", "relation", "classification", ".", "CNN", ":", "EMNLP", "apply", "additional", "heuristics", ",", "namely", "Euclidean", "distance", ",", "cosine", "measure", ",", "and", "element", "-", "wise", "absolute", "difference", ".", "The", "above", "methods", "operate", "on", "a", "fixed", "-", "size", "vector", "representation", "of", "a", "sentence", ",", "categorized", "as", "sentence", "encoding", "-", "based", "approaches", ".", "Thus", "the", "matching", "complexity", "is", ",", "i.e.", ",", "independent", "of", "the", "sentence", "length", ".", "Word", "-", "by", "-", "word", "similarity", "matrices", "are", "introduced", "to", "enhance", "interaction", ".", "To", "obtain", "the", "similarity", "matrix", ",", "CNN", ":", "NIPS", "(", "Arc", "-", "II", ")", "concatenate", "two", "words", "’", "vectors", "(", "after", "convolution", ")", ",", "recurparaphrase", "compute", "Euclidean", "distance", ",", "and", "LSTM", ":", "AAAI", "apply", "tensor", "product", ".", "In", "this", "way", ",", "the", "complexity", "is", "of", ",", "where", "is", "the", "length", "of", "a", "sentence", ";", "hence", "similarity", "matrices", "are", "difficult", "to", "scale", "and", "less", "efficient", "for", "large", "datasets", ".", "Recently", ",", "attention", "introduce", "several", "context", "-", "aware", "methods", "for", "sentence", "matching", ".", "They", "report", "that", "RNNs", "over", "a", "single", "chain", "of", "two", "sentences", "are", "more", "informative", "than", "separate", "RNNs", ";", "a", "static", "attention", "over", "the", "first", "sentence", "is", "also", "useful", "when", "modeling", "the", "second", "one", ".", "Such", "context", "-", "awareness", "interweaves", "the", "sentence", "modeling", "and", "matching", "steps", ".", "In", "some", "scenarios", "like", "sentence", "pair", "re", "-", "ranking", ",", "it", "is", "not", "feasible", "to", "pre", "-", "calculate", "the", "vector", "representations", "of", "sentences", ",", "so", "the", "matching", "complexity", "is", "of", ".", "attention", "further", "develop", "a", "word", "-", "by", "-", "word", "attention", "mechanism", "and", "obtain", "a", "higher", "accuracy", "with", "a", "complexity", "order", "of", ".", "section", ":", "Our", "Approach", "We", "follow", "the", "“", "Siamese", "”", "architecture", "(", "like", "most", "work", "in", "Section", "[", "reference", "]", ")", "and", "adopt", "a", "two", "-", "step", "strategy", "to", "classify", "the", "relation", "between", "two", "sentences", ".", "Concretely", ",", "our", "model", "comprises", "two", "parts", ":", "A", "tree", "-", "based", "convolutional", "neural", "network", "models", "each", "individual", "sentence", "(", "Figure", "[", "reference", "]", "a", ")", ".", "Notice", "that", ",", "the", "two", "sentences", ",", "premise", "and", "hypothesis", ",", "share", "a", "same", "TBCNN", "model", "(", "with", "same", "parameters", ")", ",", "because", "this", "part", "aims", "to", "capture", "general", "semantics", "of", "sentences", ".", "A", "matching", "layer", "combines", "two", "sentences", "’", "information", "by", "heuristics", "(", "Figure", "[", "reference", "]", "b", ")", ".", "After", "individual", "sentence", "models", ",", "we", "design", "a", "sentence", "matching", "layer", "to", "aggregate", "information", ".", "We", "use", "simple", "heuristics", ",", "including", "concatenation", ",", "element", "-", "wise", "product", "and", "difference", ",", "which", "are", "effective", "and", "efficient", ".", "Finally", ",", "we", "add", "a", "softmax", "layer", "for", "output", ".", "The", "training", "objective", "is", "cross", "-", "entropy", "loss", ",", "and", "we", "adopt", "mini", "-", "batch", "stochastic", "gradient", "descent", ",", "computed", "by", "back", "-", "propagation", ".", "subsection", ":", "Tree", "-", "Based", "Convolution", "The", "tree", "-", "based", "convolutoinal", "neural", "network", "(", "TBCNN", ")", "is", "first", "proposed", "in", "our", "previous", "work", "to", "classify", "program", "source", "code", ";", "later", ",", "we", "further", "propose", "TBCNN", "variants", "to", "model", "sentences", ".", "This", "subsection", "details", "the", "tree", "-", "based", "convolution", "process", ".", "The", "basic", "idea", "of", "TBCNN", "is", "to", "design", "a", "set", "of", "subtree", "feature", "detectors", "sliding", "over", "the", "parse", "tree", "of", "a", "sentence", ";", "either", "a", "constituency", "tree", "or", "a", "dependency", "tree", "applies", ".", "In", "this", "paper", ",", "we", "prefer", "the", "dependency", "tree", "-", "based", "convolution", "for", "its", "efficiency", "and", "compact", "expressiveness", ".", "Concretely", ",", "a", "sentence", "is", "first", "converted", "to", "a", "dependency", "parse", "tree", ".", "Each", "node", "in", "the", "dependency", "tree", "corresponds", "to", "a", "word", "in", "the", "sentence", ";", "an", "edge", "indicates", "is", "governed", "by", ".", "Edges", "are", "labeled", "with", "grammatical", "relations", "(", "e.g.", ",", "nsubj", ")", "between", "the", "parent", "node", "and", "its", "children", ".", "Words", "are", "represented", "by", "pretrained", "vector", "representations", ",", "also", "known", "as", "word", "embeddings", ".", "Now", ",", "we", "consider", "a", "set", "of", "two", "-", "layer", "subtree", "feature", "detectors", "sliding", "over", "the", "dependency", "tree", ".", "At", "a", "position", "where", "the", "parent", "node", "is", "with", "child", "nodes", ",", "the", "output", "of", "the", "feature", "detector", ",", ",", "is", "Let", "us", "assume", "word", "embeddings", "(", "and", ")", "are", "of", "dimensions", ";", "that", "the", "convolutional", "layer", "is", "-", "dimensional", ".", "is", "the", "weight", "matrix", ";", "is", "the", "bias", "vector", ".", "denotes", "the", "dependency", "relation", "between", "and", ".", "is", "the", "non", "-", "linear", "activation", "function", ",", "and", "we", "apply", "ReLU", "in", "our", "experiments", ".", "After", "tree", "-", "based", "convolution", ",", "we", "obtain", "a", "set", "of", "feature", "maps", ",", "which", "are", "one", "-", "one", "corresponding", "to", "original", "words", "in", "the", "sentence", ".", "Therefore", ",", "they", "may", "vary", "in", "size", "and", "length", ".", "A", "dynamic", "pooling", "layer", "is", "applied", "to", "aggregate", "information", "along", "different", "parts", "of", "the", "tree", ",", "serving", "as", "a", "way", "of", "semantic", "compositionality", ".", "We", "use", "the", "pooling", "operation", ",", "which", "takes", "the", "maximum", "value", "in", "each", "dimension", ".", "Then", "we", "add", "a", "fully", "-", "connected", "hidden", "layer", "to", "further", "mix", "the", "information", "in", "a", "sentence", ".", "The", "obtained", "vector", "representation", "of", "a", "sentence", "is", "denoted", "as", "(", "also", "called", "a", "sentence", "embedding", ")", ".", "Notice", "that", "the", "same", "tree", "-", "based", "convolution", "applies", "to", "both", "the", "premise", "and", "hypothesis", ".", "Tree", "-", "based", "convolution", "along", "with", "pooling", "enables", "structural", "features", "to", "reach", "the", "output", "layer", "with", "short", "propagation", "paths", ",", "as", "opposed", "to", "the", "recursive", "network", ",", "which", "is", "also", "structure", "-", "sensitive", "but", "may", "suffer", "from", "the", "problem", "of", "long", "propagation", "path", ".", "By", "contrast", ",", "TBCNN", "is", "effective", "and", "efficient", "in", "learning", "such", "structural", "information", ".", "subsection", ":", "Matching", "Heuristics", "In", "this", "part", ",", "we", "introduce", "how", "vector", "representations", "of", "individual", "sentences", "are", "combined", "to", "capture", "the", "relation", "between", "the", "premise", "and", "hypothesis", ".", "As", "the", "dataset", "is", "large", ",", "we", "prefer", "matching", "operations", "because", "of", "efficiency", "concerns", ".", "Concretely", ",", "we", "have", "three", "matching", "heuristics", ":", "Concatenation", "of", "the", "two", "sentence", "vectors", ",", "Element", "-", "wise", "product", ",", "and", "Element", "-", "wise", "difference", ".", "The", "first", "heuristic", "follows", "the", "most", "standard", "procedure", "of", "the", "“", "Siamese", "”", "architectures", ",", "while", "the", "latter", "two", "are", "certain", "measures", "of", "“", "similarity", "”", "or", "“", "closeness", ".", "”", "These", "matching", "layers", "are", "further", "concatenated", "(", "Figure", "[", "reference", "]", "b", ")", ",", "given", "by", "where", "and", "are", "the", "sentence", "vectors", "of", "the", "premise", "and", "hypothesis", ",", "respectively", ";", "“", "”", "denotes", "element", "-", "wise", "product", ";", "semicolons", "refer", "to", "column", "vector", "concatenation", ".", "is", "the", "output", "of", "the", "matching", "layer", ".", "We", "would", "like", "to", "point", "out", "that", ",", "with", "subsequent", "linear", "transformation", ",", "element", "-", "wise", "difference", "is", "a", "special", "case", "of", "concatenation", ".", "If", "we", "assume", "the", "subsequent", "transformation", "takes", "the", "form", "of", ",", "where", "is", "the", "weights", "for", "concatenated", "sentence", "representations", ",", "then", "element", "-", "wise", "difference", "can", "be", "viewed", "as", "such", "that", ".", "(", "is", "the", "weights", "corresponding", "to", "element", "-", "wise", "difference", ".", ")", "Thus", ",", "our", "third", "heuristic", "can", "be", "absorbed", "into", "the", "first", "one", "in", "terms", "of", "model", "capacity", ".", "However", ",", "as", "will", "be", "shown", "in", "the", "experiment", ",", "explicitly", "specifying", "this", "heuristic", "significantly", "improves", "the", "performance", ",", "indicating", "that", "optimization", "differs", ",", "despite", "the", "same", "model", "capacity", ".", "Moreover", ",", "word", "embedding", "studies", "show", "that", "linear", "offset", "of", "vectors", "can", "capture", "relationships", "between", "two", "words", ",", "but", "it", "has", "not", "been", "exploited", "in", "sentence", "-", "pair", "relation", "recognition", ".", "Although", "element", "-", "wise", "distance", "is", "used", "to", "detect", "paraphrase", "in", "CNN", ":", "EMNLP", ",", "it", "mainly", "reflects", "“", "similarity", "”", "information", ".", "Our", "study", "verifies", "that", "vector", "offset", "is", "useful", "in", "capturing", "generic", "sentence", "relationships", ",", "akin", "to", "the", "word", "analogy", "task", ".", "section", ":", "Evaluation", "subsection", ":", "Dataset", "To", "evaluate", "our", "TBCNN", "-", "pair", "model", ",", "we", "used", "the", "newly", "published", "Stanford", "Natural", "Language", "Inference", "(", "SNLI", ")", "dataset", ".", "The", "dataset", "is", "constructed", "by", "crowdsourced", "efforts", ",", "each", "sentence", "written", "by", "humans", ".", "Moreover", ",", "the", "SNLI", "dataset", "is", "magnitudes", "of", "larger", "than", "previous", "resources", ",", "and", "hence", "is", "particularly", "suitable", "for", "comparing", "neural", "models", ".", "The", "target", "labels", "comprise", "three", "classes", ":", "Entailment", ",", "Contradiction", ",", "and", "Neutral", "(", "two", "irrelevant", "sentences", ")", ".", "We", "applied", "the", "standard", "train", "/", "validation", "/", "test", "split", ",", "contraining", "550k", ",", "10k", ",", "and", "10k", "samples", ",", "respectively", ".", "Figure", "[", "reference", "]", "presents", "additional", "dataset", "statistics", ",", "especially", "those", "relevant", "to", "dependency", "parse", "trees", ".", "subsection", ":", "Hyperparameter", "Settings", "All", "our", "neural", "layers", ",", "including", "embeddings", ",", "were", "set", "to", "300", "dimensions", ".", "The", "model", "is", "mostly", "robust", "when", "the", "dimension", "is", "large", ",", "e.g.", ",", "several", "hundred", ".", "Word", "embeddings", "were", "pretrained", "ourselves", "by", "word2vec", "on", "the", "English", "Wikipedia", "corpus", "and", "fined", "tuned", "during", "training", "as", "a", "part", "of", "model", "parameters", ".", "We", "applied", "penalty", "of", ";", "dropout", "was", "chosen", "by", "validation", "with", "a", "granularity", "of", "0.1", "(", "Figure", "[", "reference", "]", ")", ".", "We", "see", "that", "a", "large", "dropout", "rate", "(", "0.3", ")", "hurts", "the", "performance", "(", "and", "also", "makes", "training", "slow", ")", "for", "such", "a", "large", "dataset", "as", "opposed", "to", "small", "datasets", "in", "other", "tasks", ".", "Initial", "learning", "rate", "was", "set", "to", "1", ",", "and", "a", "power", "decay", "was", "applied", ".", "We", "used", "stochastic", "gradient", "descent", "with", "a", "batch", "size", "of", "50", ".", "subsection", ":", "Performance", "Table", "[", "reference", "]", "compares", "our", "model", "with", "previous", "results", ".", "As", "seen", ",", "the", "TBCNN", "sentence", "pair", "model", ",", "followed", "by", "simple", "concatenation", "alone", ",", "outperforms", "existing", "sentence", "encoding", "-", "based", "approaches", "(", "without", "pretraining", ")", ",", "including", "a", "feature", "-", "rich", "method", "using", "6", "groups", "of", "human", "-", "engineered", "features", ",", "long", "short", "term", "memory", "(", "LSTM", ")-", "based", "RNNs", ",", "and", "traditional", "CNNs", ".", "This", "verifies", "the", "rationale", "for", "using", "tree", "-", "based", "convolution", "as", "the", "sentence", "-", "level", "neural", "model", "for", "NLI", ".", "Table", "[", "reference", "]", "compares", "different", "heuristics", "of", "matching", ".", "We", "first", "analyze", "each", "heuristic", "separately", ":", "using", "element", "-", "wise", "product", "alone", "is", "significantly", "worse", "than", "concatenation", "or", "element", "-", "wise", "difference", ";", "the", "latter", "two", "are", "comparable", "to", "each", "other", ".", "Combining", "different", "matching", "heuristics", "improves", "the", "result", ":", "the", "TBCNN", "-", "pair", "model", "with", "concatenation", ",", "element", "-", "wise", "product", "and", "difference", "yields", "the", "highest", "performance", "of", "82.1", "%", ".", "As", "analyzed", "in", "Section", "[", "reference", "]", ",", "the", "element", "-", "wise", "difference", "matching", "layer", "does", "not", "add", "to", "model", "complexity", "and", "can", "be", "absorbed", "as", "a", "special", "case", "into", "simple", "concatenation", ".", "However", ",", "explicitly", "using", "such", "heuristic", "yields", "an", "accuracy", "boost", "of", "1–2", "%", ".", "Further", "applying", "element", "-", "wise", "product", "improves", "the", "accuracy", "by", "another", "0.5", "%", ".", "The", "full", "TBCNN", "-", "pair", "model", "outperforms", "all", "existing", "sentence", "encoding", "-", "based", "approaches", ",", "including", "a", "1024d", "gated", "recurrent", "unit", "(", "GRU", ")-", "based", "RNN", "with", "“", "skip", "-", "thought", "”", "pretraining", ".", "The", "results", "obtained", "by", "our", "model", "are", "also", "comparable", "to", "several", "attention", "-", "based", "LSTMs", ",", "which", "are", "more", "computationally", "intensive", "than", "ours", "in", "terms", "of", "complexity", "order", ".", "subsection", ":", "Complexity", "Concerns", "For", "most", "sentence", "models", "including", "TBCNN", ",", "the", "overall", "complexity", "is", "at", "least", ".", "However", ",", "an", "efficient", "matching", "approach", "is", "still", "important", ",", "especially", "to", "retrieval", "-", "and", "-", "reranking", "systems", ".", "For", "example", ",", "in", "a", "retrieval", "-", "based", "question", "-", "answering", "or", "conversation", "system", ",", "we", "can", "largely", "reduce", "response", "time", "by", "performing", "sentence", "matching", "based", "on", "precomputed", "candidates", "’", "embeddings", ".", "By", "contrast", ",", "context", "-", "aware", "matching", "approaches", "as", "described", "in", "Section", "[", "reference", "]", "involve", "processing", "each", "candidate", "given", "a", "new", "user", "-", "issued", "query", ",", "which", "is", "time", "-", "consuming", "in", "terms", "of", "most", "industrial", "products", ".", "In", "our", "experiments", ",", "the", "matching", "part", "(", "Figure", "[", "reference", "]", "b", ")", "counts", "1.71", "%", "of", "the", "total", "time", "during", "prediction", "(", "single", "-", "CPU", ",", "C", "++", "implementation", ")", ",", "showing", "the", "potential", "applications", "of", "our", "approach", "in", "efficient", "retrieval", "of", "semantically", "related", "sentences", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "proposed", "the", "TBCNN", "-", "pair", "model", "for", "natural", "language", "inference", ".", "Our", "model", "relies", "on", "the", "tree", "-", "based", "convolutional", "neural", "network", "(", "TBCNN", ")", "to", "capture", "sentence", "-", "level", "semantics", ";", "then", "two", "sentences", "’", "information", "is", "combined", "by", "several", "heuristics", "including", "concatenation", ",", "element", "-", "wise", "product", "and", "difference", ".", "Experimental", "results", "on", "a", "large", "dataset", "show", "a", "high", "performance", "of", "our", "TBCNN", "-", "pair", "model", "while", "remaining", "a", "low", "complexity", "order", ".", "section", ":", "Acknowledgments", "We", "thank", "all", "anonymous", "reviewers", "for", "their", "constructive", "comments", ",", "especially", "those", "on", "complexity", "issues", ".", "We", "also", "thank", "Sam", "Bowman", ",", "Edward", "Grefenstette", ",", "and", "Tim", "Rocktäschel", "for", "their", "discussion", ".", "This", "research", "was", "supported", "by", "the", "National", "Basic", "Research", "Program", "of", "China", "(", "the", "973", "Program", ")", "under", "Grant", "No", ".", "2015CB352201", "and", "the", "National", "Natural", "Science", "Foundation", "of", "China", "under", "Grant", "Nos", ".", "61232015", ",", "61421091", ",", "and", "61502014", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["Natural", "Language", "Inference"], "Task"], [6, 10, ["Tree", "-", "Based", "Convolution"], "Method"], [11, 13, ["Heuristic", "Matching"], "Method"], [20, 24, ["TBCNN", "-", "pair", "model"], "Method"], [25, 29, ["recognize", "entailment", "and", "contradiction"], "Task"], [38, 44, ["tree", "-", "based", "convolutional", "neural", "network"], "Method"], [45, 46, ["TBCNN"], "Method"], [54, 57, ["heuristic", "matching", "layers"], "Method"], [58, 59, ["concatenation"], "Method"], [81, 86, ["sentence", "encoding", "-", "based", "approaches"], "Method"], [94, 96, ["Recognizing", "entailment"], "Task"], [116, 117, ["NLI"], "Task"], [119, 120, ["inference"], "Task"], [163, 164, ["NLI"], "Task"], [169, 172, ["natural", "language", "understanding"], "Task"], [177, 178, ["NLP"], "Task"], [181, 183, ["question", "answering"], "Task"], [184, 186, ["automatic", "summarization"], "Task"], [189, 190, ["NLI"], "Task"], [197, 200, ["sentence", "pair", "modeling"], "Task"], [202, 204, ["paraphrase", "detection"], "Task"], [205, 210, ["relation", "recognition", "of", "discourse", "units"], "Task"], [216, 217, ["NLI"], "Task"], [223, 227, ["feature", "-", "rich", "models"], "Method"], [228, 231, ["formal", "reasoning", "methods"], "Method"], [232, 236, ["Feature", "-", "based", "approaches"], "Method"], [238, 241, ["machine", "learning", "models"], "Method"], [257, 259, ["Formal", "reasoning"], "Task"], [270, 273, ["formal", "logical", "representation"], "Method"], [275, 276, ["interpreters"], "Method"], [291, 292, ["scope"], "Metric"], [293, 294, ["accuracy"], "Metric"], [299, 301, ["neural", "networks"], "Method"], [307, 309, ["NLP", "applications"], "Task"], [311, 314, ["individual", "sentence", "modeling"], "Task"], [317, 319, ["sentence", "matching"], "Task"], [322, 324, ["neural", "architecture"], "Method"], [326, 328, ["sentence", "pairs"], "Task"], [339, 341, ["sentence", "model"], "Method"], [343, 345, ["matching", "layer"], "Method"], [354, 356, ["sentence", "models"], "Method"], [357, 359, ["convolutional", "networks"], "Method"], [360, 364, ["recurrent", "/", "recursive", "networks"], "Method"], [404, 410, ["tree", "-", "based", "convolutional", "neural", "network"], "Method"], [411, 413, ["TBCNN", ")"], "Method"], [425, 428, ["sentence", "classification", "tasks"], "Task"], [436, 437, ["TBCNN"], "Method"], [453, 455, ["NLI", "task"], "Task"], [463, 468, ["TBCNN", "-", "pair", "neural", "model"], "Method"], [482, 484, ["TBCNN", "model"], "Method"], [495, 496, ["NLI"], "Task"], [519, 520, ["TBCNN"], "Method"], [543, 544, ["TBCNN"], "Method"], [548, 550, ["sequential", "convolution"], "Method"], [569, 571, ["pooling", "layer"], "Method"], [583, 585, ["semantic", "compositonality"], "Method"], [596, 599, ["heuristic", "matching", "layers"], "Method"], [601, 602, ["concatenation"], "Method"], [603, 607, ["element", "-", "wise", "product"], "Method"], [608, 609, ["difference"], "Method"], [622, 624, ["low", "complexity"], "Metric"], [649, 653, ["tree", "-", "based", "convolution"], "Method"], [654, 658, ["sentence", "pair", "modeling", "tasks"], "Task"], [659, 660, ["NLI"], "Task"], [670, 671, ["accuracy"], "Metric"], [673, 675, ["low", "complexity"], "Metric"], [678, 683, ["sentence", "encoding", "-", "based", "approaches"], "Method"], [689, 693, ["feature", "-", "rich", "methods"], "Method"], [694, 698, ["long", "short", "term", "memory"], "Method"], [699, 700, ["LSTM"], "Method"], [709, 711, ["Entailment", "recognition"], "Task"], [718, 721, ["sentence", "pair", "modeling"], "Task"], [723, 725, ["neural", "networks"], "Method"], [730, 734, ["sentence", "-", "level", "model"], "Method"], [741, 743, ["matching", "layers"], "Method"], [749, 752, ["Siamese", "”", "architectures"], "Method"], [753, 754, ["CNN"], "Method"], [755, 756, ["NIPS"], "Method"], [757, 758, ["CNN"], "Method"], [759, 760, ["NAACL"], "Method"], [761, 764, ["convolutional", "neural", "networks"], "Method"], [765, 766, ["CNNs"], "Method"], [770, 772, ["sentence", "model"], "Method"], [777, 779, ["feature", "detectors"], "Method"], [789, 790, ["LSTM"], "Method"], [791, 792, ["AAAI"], "Method"], [793, 796, ["sentence", "pair", "models"], "Method"], [797, 800, ["recurrent", "neural", "networks"], "Method"], [801, 802, ["RNNs"], "Method"], [811, 812, ["recurparaphrase"], "Method"], [823, 825, ["recursive", "autoencoders"], "Method"], [844, 847, ["natural", "language", "understanding"], "Task"], [862, 864, ["vector", "representations"], "Method"], [865, 866, ["Concatenation"], "Method"], [876, 882, ["subject", "and", "object", "in", "relation", "classification"], "Task"], [883, 884, ["CNN"], "Method"], [885, 886, ["EMNLP"], "Method"], [894, 896, ["cosine", "measure"], "Method"], [898, 903, ["element", "-", "wise", "absolute", "difference"], "Method"], [910, 915, ["fixed", "-", "size", "vector", "representation"], "Method"], [921, 926, ["sentence", "encoding", "-", "based", "approaches"], "Method"], [929, 931, ["matching", "complexity"], "Metric"], [952, 953, ["interaction"], "Task"], [960, 961, ["CNN"], "Method"], [962, 963, ["NIPS"], "Method"], [975, 976, ["convolution"], "Method"], [978, 979, ["recurparaphrase"], "Method"], [980, 982, ["Euclidean", "distance"], "Method"], [984, 985, ["LSTM"], "Method"], [986, 987, ["AAAI"], "Method"], [988, 990, ["tensor", "product"], "Method"], [996, 997, ["complexity"], "Metric"], [1027, 1031, ["context", "-", "aware", "methods"], "Method"], [1032, 1034, ["sentence", "matching"], "Task"], [1038, 1039, ["RNNs"], "Method"], [1051, 1052, ["RNNs"], "Method"], [1070, 1073, ["context", "-", "awareness"], "Method"], [1075, 1080, ["sentence", "modeling", "and", "matching", "steps"], "Method"], [1085, 1090, ["sentence", "pair", "re", "-", "ranking"], "Task"], [1100, 1104, ["vector", "representations", "of", "sentences"], "Method"], [1107, 1109, ["matching", "complexity"], "Metric"], [1116, 1123, ["word", "-", "by", "-", "word", "attention", "mechanism"], "Method"], [1127, 1128, ["accuracy"], "Metric"], [1142, 1145, ["Siamese", "”", "architecture"], "Method"], [1179, 1185, ["tree", "-", "based", "convolutional", "neural", "network"], "Method"], [1211, 1213, ["TBCNN", "model"], "Method"], [1231, 1233, ["matching", "layer"], "Method"], [1250, 1252, ["sentence", "models"], "Method"], [1256, 1259, ["sentence", "matching", "layer"], "Method"], [1266, 1267, ["heuristics"], "Method"], [1271, 1275, ["element", "-", "wise", "product"], "Method"], [1276, 1277, ["difference"], "Method"], [1289, 1291, ["softmax", "layer"], "Method"], [1295, 1297, ["training", "objective"], "Metric"], [1298, 1302, ["cross", "-", "entropy", "loss"], "Metric"], [1306, 1312, ["mini", "-", "batch", "stochastic", "gradient", "descent"], "Method"], [1315, 1318, ["back", "-", "propagation"], "Method"], [1321, 1325, ["Tree", "-", "Based", "Convolution"], "Method"], [1326, 1332, ["tree", "-", "based", "convolutoinal", "neural", "network"], "Method"], [1333, 1334, ["TBCNN"], "Method"], [1353, 1355, ["TBCNN", "variants"], "Method"], [1363, 1368, ["tree", "-", "based", "convolution", "process"], "Method"], [1373, 1374, ["TBCNN"], "Method"], [1380, 1383, ["subtree", "feature", "detectors"], "Method"], [1409, 1414, ["dependency", "tree", "-", "based", "convolution"], "Method"], [1418, 1420, ["compact", "expressiveness"], "Metric"], [1478, 1481, ["pretrained", "vector", "representations"], "Method"], [1485, 1487, ["word", "embeddings"], "Method"], [1495, 1501, ["two", "-", "layer", "subtree", "feature", "detectors"], "Method"], [1523, 1525, ["feature", "detector"], "Method"], [1542, 1544, ["convolutional", "layer"], "Method"], [1576, 1577, ["ReLU"], "Method"], [1582, 1586, ["tree", "-", "based", "convolution"], "Method"], [1619, 1622, ["dynamic", "pooling", "layer"], "Method"], [1639, 1641, ["semantic", "compositionality"], "Task"], [1645, 1647, ["pooling", "operation"], "Method"], [1661, 1666, ["fully", "-", "connected", "hidden", "layer"], "Method"], [1677, 1679, ["vector", "representation"], "Method"], [1689, 1691, ["sentence", "embedding"], "Task"], [1697, 1701, ["tree", "-", "based", "convolution"], "Method"], [1709, 1713, ["Tree", "-", "based", "convolution"], "Method"], [1715, 1716, ["pooling"], "Method"], [1733, 1735, ["recursive", "network"], "Method"], [1756, 1757, ["TBCNN"], "Method"], [1769, 1771, ["Matching", "Heuristics"], "Method"], [1778, 1780, ["vector", "representations"], "Method"], [1803, 1805, ["matching", "operations"], "Method"], [1815, 1817, ["matching", "heuristics"], "Method"], [1825, 1829, ["Element", "-", "wise", "product"], "Method"], [1831, 1835, ["Element", "-", "wise", "difference"], "Method"], [1847, 1850, ["Siamese", "”", "architectures"], "Method"], [1859, 1861, ["“", "similarity"], "Metric"], [1868, 1870, ["matching", "layers"], "Method"], [1908, 1911, ["column", "vector", "concatenation"], "Method"], [1917, 1919, ["matching", "layer"], "Method"], [1930, 1932, ["linear", "transformation"], "Method"], [1933, 1937, ["element", "-", "wise", "difference"], "Method"], [2003, 2005, ["model", "capacity"], "Metric"], [2027, 2028, ["optimization"], "Task"], [2038, 2040, ["word", "embedding"], "Task"], [2061, 2066, ["sentence", "-", "pair", "relation", "recognition"], "Task"], [2075, 2077, ["detect", "paraphrase"], "Task"], [2078, 2079, ["CNN"], "Method"], [2080, 2081, ["EMNLP"], "Method"], [2094, 2096, ["vector", "offset"], "Method"], [2100, 2103, ["generic", "sentence", "relationships"], "Task"], [2107, 2110, ["word", "analogy", "task"], "Task"], [2120, 2124, ["TBCNN", "-", "pair", "model"], "Method"], [2130, 2134, ["Stanford", "Natural", "Language", "Inference"], "Material"], [2135, 2136, ["SNLI"], "Material"], [2173, 2175, ["neural", "models"], "Method"], [2199, 2205, ["train", "/", "validation", "/", "test", "split"], "Method"], [2240, 2242, ["neural", "layers"], "Method"], [2244, 2245, ["embeddings"], "Method"], [2319, 2321, ["dropout", "rate"], "Metric"], [2331, 2332, ["training"], "Task"], [2349, 2351, ["learning", "rate"], "Metric"], [2365, 2368, ["stochastic", "gradient", "descent"], "Method"], [2393, 2397, ["TBCNN", "sentence", "pair", "model"], "Method"], [2401, 2403, ["concatenation", "alone"], "Method"], [2406, 2411, ["sentence", "encoding", "-", "based", "approaches"], "Method"], [2413, 2414, ["pretraining"], "Method"], [2418, 2422, ["feature", "-", "rich", "method"], "Method"], [2443, 2444, ["CNNs"], "Method"], [2451, 2455, ["tree", "-", "based", "convolution"], "Method"], [2457, 2462, ["sentence", "-", "level", "neural", "model"], "Method"], [2463, 2464, ["NLI"], "Task"], [2473, 2474, ["matching"], "Task"], [2483, 2488, ["element", "-", "wise", "product", "alone"], "Method"], [2510, 2512, ["matching", "heuristics"], "Method"], [2517, 2521, ["TBCNN", "-", "pair", "model"], "Method"], [2522, 2523, ["concatenation"], "Method"], [2524, 2528, ["element", "-", "wise", "product"], "Method"], [2529, 2530, ["difference"], "Method"], [2547, 2553, ["element", "-", "wise", "difference", "matching", "layer"], "Method"], [2557, 2559, ["model", "complexity"], "Metric"], [2569, 2570, ["concatenation"], "Method"], [2587, 2591, ["element", "-", "wise", "product"], "Method"], [2593, 2594, ["accuracy"], "Metric"], [2608, 2613, ["sentence", "encoding", "-", "based", "approaches"], "Method"], [2617, 2620, ["gated", "recurrent", "unit"], "Method"], [2621, 2622, ["GRU"], "Method"], [2644, 2648, ["attention", "-", "based", "LSTMs"], "Method"], [2659, 2661, ["complexity", "order"], "Metric"], [2664, 2666, ["Complexity", "Concerns"], "Metric"], [2668, 2670, ["sentence", "models"], "Method"], [2671, 2672, ["TBCNN"], "Method"], [2674, 2676, ["overall", "complexity"], "Metric"], [2684, 2686, ["matching", "approach"], "Method"], [2692, 2698, ["retrieval", "-", "and", "-", "reranking", "systems"], "Task"], [2704, 2713, ["retrieval", "-", "based", "question", "-", "answering", "or", "conversation", "system"], "Task"], [2718, 2720, ["response", "time"], "Metric"], [2722, 2724, ["sentence", "matching"], "Method"], [2734, 2739, ["context", "-", "aware", "matching", "approaches"], "Method"], [2767, 2769, ["industrial", "products"], "Task"], [2792, 2793, ["prediction"], "Task"], [2812, 2817, ["retrieval", "of", "semantically", "related", "sentences"], "Task"], [2828, 2832, ["TBCNN", "-", "pair", "model"], "Method"], [2842, 2848, ["tree", "-", "based", "convolutional", "neural", "network"], "Method"], [2849, 2850, ["TBCNN"], "Method"], [2869, 2870, ["concatenation"], "Method"], [2871, 2875, ["element", "-", "wise", "product"], "Method"], [2890, 2894, ["TBCNN", "-", "pair", "model"], "Method"], [2898, 2900, ["complexity", "order"], "Metric"], [114, 115, ["inference"], "Task"], [447, 448, ["inference"], "Task"], [2156, 2157, ["SNLI"], "Material"], [2431, 2435, ["long", "short", "term", "memory"], "Method"], [2436, 2437, ["LSTM"], "Method"], [2439, 2440, ["RNNs"], "Method"], [2579, 2580, ["accuracy"], "Metric"], [2601, 2605, ["TBCNN", "-", "pair", "model"], "Method"], [2835, 2836, ["inference"], "Task"]]} {"doc_id": "1235dd37312cb20aced0e97d953f6379d8a0c7d4", "text": ["document", ":", "Grounded", "Textual", "Entailment", "Capturing", "semantic", "relations", "between", "sentences", ",", "such", "as", "entailment", ",", "is", "a", "long", "-", "standing", "challenge", "for", "computational", "semantics", ".", "Logic", "-", "based", "models", "analyse", "entailment", "in", "terms", "of", "possible", "worlds", "(", "interpretations", ",", "or", "situations", ")", "where", "a", "premise", "P", "entails", "a", "hypothesis", "H", "iff", "in", "all", "worlds", "where", "P", "is", "true", ",", "H", "is", "also", "true", ".", "Statistical", "models", "view", "this", "relationship", "probabilistically", ",", "addressing", "it", "in", "terms", "of", "whether", "a", "human", "would", "likely", "infer", "H", "from", "P.", "In", "this", "paper", ",", "we", "wish", "to", "bridge", "these", "two", "perspectives", ",", "by", "arguing", "for", "a", "visually", "-", "grounded", "version", "of", "the", "Textual", "Entailment", "task", ".", "Specifically", ",", "we", "ask", "whether", "models", "can", "perform", "better", "if", ",", "in", "addition", "to", "P", "and", "H", ",", "there", "is", "also", "an", "image", "(", "corresponding", "to", "the", "relevant", "“", "world", "”", "or", "“", "situation", "”", ")", ".", "We", "use", "a", "multimodal", "version", "of", "the", "SNLI", "dataset", "and", "we", "compare", "“", "blind", "”", "and", "visually", "-", "augmented", "models", "of", "textual", "entailment", ".", "We", "show", "that", "visual", "information", "is", "beneficial", ",", "but", "we", "also", "conduct", "an", "in", "-", "depth", "error", "analysis", "that", "reveals", "that", "current", "multimodal", "models", "are", "not", "performing", "“", "grounding", "”", "in", "an", "optimal", "fashion", ".", "section", ":", "Introduction", "Correspondence", "should", "be", "addressed", "to", "Raffaella", "Bernardi", "(", "raffaella.bernardi@unitn.it", ")", "and", "Albert", "Gatt", "(", "albert.gatt@um.edu.mt", ")", ".", "This", "work", "is", "licensed", "under", "a", "Creative", "Commons", "Attribution", "4.0", "International", "License", ".", "License", "details", ":", ".", "The", "dataset", ",", "annotation", "and", "code", "is", "available", "from", ".", "Evaluating", "the", "ability", "to", "infer", "information", "from", "a", "text", "is", "a", "crucial", "test", "of", "the", "capability", "of", "models", "to", "grasp", "meaning", ".", "As", "a", "result", ",", "the", "computational", "linguistics", "community", "has", "invested", "huge", "efforts", "into", "developing", "textual", "entailment", "(", "TE", ")", "datasets", ".", "After", "formal", "semanticists", "developed", "FraCas", "in", "the", "mid", "’", "90", ",", "an", "increase", "in", "statistical", "approaches", "to", "computational", "semantics", "gave", "rise", "to", "the", "need", "for", "suitable", "evaluation", "datasets", ".", "Hence", ",", "Recognizing", "Textual", "Entailment", "(", "RTE", ")", "shared", "tasks", "have", "been", "organized", "regularly", ".", "Recent", "work", "on", "compositional", "distributional", "models", "has", "motivated", "the", "development", "of", "the", "SICK", "dataset", "of", "sentence", "pairs", "in", "entailment", "relations", "for", "evaluating", "such", "models", ".", "Further", "advances", "with", "Neural", "Networks", "(", "NNs", ")", "have", "once", "more", "motivated", "efforts", "to", "develop", "a", "large", "natural", "language", "inference", "dataset", ",", "SNLI", ",", "since", "NNs", "need", "to", "be", "trained", "on", "big", "data", ".", "However", ",", "meaning", "is", "not", "something", "we", "obtain", "just", "from", "text", "and", "the", "ability", "to", "reason", "is", "not", "unimodal", "either", ".", "The", "importance", "of", "enriching", "meaning", "representations", "with", "other", "modalities", "has", "been", "advocated", "by", "cognitive", "scientists", ",", "(", "e.g.", ",", ")", "and", "computational", "linguists", "(", "e.g.", ",", ")", ".", "While", "efforts", "have", "been", "put", "into", "developing", "multimodal", "datasets", "for", "the", "task", "of", "checking", "Semantic", "Text", "Similarity", "Text", ",", "we", "are", "not", "aware", "of", "any", "available", "datasets", "to", "tackle", "the", "problem", "of", "Grounded", "Textual", "Entailment", "(", "GTE", ")", ".", "Our", "paper", "is", "a", "first", "effort", "in", "this", "direction", ".", "[", "b", "]", "0.45", "[", "scale=0.5", "]", "figures", "/", "warm", "-", "entailment", "[", "b", "]", "0.45", "[", "scale=0.5", "]", "figures", "/", "neutral", "-", "mis", "Textual", "Entailment", "is", "defined", "in", "terms", "of", "the", "likelihood", "of", "two", "sentences", "(", "a", "premise", "P", "and", "an", "hypothesis", "H", ")", "to", "be", "in", "a", "certain", "relation", ":", "P", "entails", ",", "contradicts", "or", "is", "unrelated", "to", "H.", "For", "instance", ",", "the", "premise", "“", "People", "trying", "to", "get", "warm", "in", "front", "of", "a", "chimney", "”", "and", "the", "hypothesis", "“", "People", "trying", "to", "get", "warm", "at", "home", "”", "are", "highly", "likely", "to", "be", "in", "an", "entailment", "relation", ".", "Our", "question", "is", "whether", "having", "an", "image", "that", "illustrates", "the", "event", "(", "e.g.", ",", "Figure", "[", "reference", "]", ")", "can", "help", "a", "model", "to", "capture", "the", "relation", ".", "In", "order", "to", "answer", "this", "question", ",", "we", "augment", "the", "largest", "available", "TE", "dataset", "with", "images", ",", "we", "enhance", "a", "state", "of", "the", "art", "model", "of", "textual", "entailment", "to", "take", "images", "into", "account", "and", "we", "evaluate", "it", "against", "the", "GTE", "task", ".", "The", "inclusion", "of", "images", "can", "also", "alter", "relations", "which", ",", "based", "on", "text", "alone", ",", "would", "seem", "likely", ".", "For", "example", ",", "to", "a", "“", "blind", "”", "model", "the", "sentences", "of", "the", "sentence", "pair", "in", "Figure", "[", "reference", "]", "would", "seem", "to", "be", "unrelated", ",", "but", "when", "the", "two", "sentences", "are", "viewed", "in", "the", "context", "of", "the", "image", ",", "they", "do", "become", "related", ".", "A", "suitable", "GTE", "model", "therefore", "has", "to", "perform", "two", "sub", "-", "tasks", ":", "(", "a", ")", "it", "needs", "to", "ground", "its", "linguistic", "representations", "of", "P", ",", "H", "or", "both", "in", "non", "-", "linguistic", "(", "visual", ")", "data", ";", "(", "b", ")", "it", "needs", "to", "reason", "about", "the", "possible", "relationship", "between", "P", "and", "H", "(", "modulo", "the", "visual", "information", ")", ".", "section", ":", "Related", "Work", "Grounding", "language", "through", "vision", "has", "recently", "become", "the", "focus", "of", "several", "tasks", ",", "including", "Image", "Captioning", "(", "IC", ",", "e.g.", ")", "and", "Visual", "Question", "Answering", "(", "VQA", ",", "eg", ".", ")", ",", "and", "even", "more", "recently", ",", "Visual", "Reasoning", "and", "Visual", "Dialog", ".", "Our", "focus", "is", "on", "Grounded", "Textual", "Entailment", "(", "GTE", ")", ".", "While", "the", "literature", "on", "TE", "is", "rather", "vast", ",", "GTE", "is", "still", "rather", "unexplored", "territory", ".", "paragraph", ":", "Textual", "Entailment", "Throughout", "the", "history", "of", "Computational", "Linguistics", "various", "datasets", "have", "been", "built", "to", "evaluate", "Computational", "Semantics", "models", "on", "the", "TE", "task", ".", "Usually", "they", "contain", "data", "divided", "into", "entailment", ",", "contradiction", "or", "unknown", "classes", ".", "The", "“", "unknown", "”", "label", "has", "sometimes", "been", "replaced", "with", "the", "“", "unrelated", "”", "or", "“", "neutral", "”", "label", ",", "capturing", "slightly", "different", "types", "of", "phenomena", ".", "Interestingly", ",", "the", "“", "entailment", "”", "and", "“", "contradiction", "”", "classes", "also", "differ", "across", "datasets", ".", "In", "the", "mid", "-", "’90s", "a", "group", "of", "formal", "semanticists", "developed", "FraCaS", "(", "Framework", "for", "Computational", "Semantics", ")", ".", "The", "dataset", "contains", "logical", "entailment", "problems", "in", "which", "a", "conclusion", "has", "to", "be", "derived", "from", "one", "or", "more", "premises", "(", "but", "not", "necessarily", "all", "premises", "are", "needed", "to", "verify", "the", "entailment", ")", ".", "The", "entailments", "are", "driven", "by", "logical", "properties", "of", "linguistic", "expressions", ",", "like", "the", "monotonicity", "of", "quantifiers", ",", "or", "their", "conservativity", "property", "etc", ".", "Hence", ",", "the", "set", "of", "premises", "entails", "the", "conclusion", "iff", "in", "all", "the", "interpretations", "(", "worlds", ")", "in", "which", "the", "premises", "are", "true", "the", "conclusion", "is", "also", "true", ";", "otherwise", "the", "conclusion", "contradicts", "the", "premises", ".", "In", "2005", ",", "the", "PASCAL", "RTE", "(", "Recognizing", "Textual", "Entailment", ")", "challenge", "was", "launched", ",", "to", "become", "a", "task", "organized", "annually", ".", "In", "2008", ",", "the", "RTE", "-", "4", "committee", "made", "the", "task", "more", "fine", "-", "grained", "by", "requiring", "the", "classification", "of", "the", "pairs", "as", "“", "entailment", "”", ",", "“", "contradiction", "”", "and", "“", "unknown", "”", ".", "The", "RTE", "datasets", ",", "unlike", "FraCaS", ",", "contain", "real", "-", "life", "natural", "language", "sentences", "and", "the", "sort", "of", "entailment", "problems", "which", "occur", "in", "corpora", "collected", "from", "the", "web", ".", "Importantly", ",", "the", "sentence", "pair", "relations", "are", "annotated", "as", "entailment", ",", "contradiction", "or", "neutral", "based", "on", "a", "likelihood", "condition", ":", "if", "a", "human", "reading", "the", "premise", "would", "typically", "infer", "that", "the", "conclusion", "(", "called", "the", "hypothesis", ")", "is", "most", "likely", "true", "(", "entailment", ")", ",", "its", "negation", "is", "most", "likely", "true", "(", "contradiction", ")", "or", "the", "conclusion", "can", "be", "either", "true", "or", "false", "(", "neutral", ")", ".", "At", "SemEval", "2014", ",", "in", "order", "to", "evaluate", "Compositional", "Distributional", "Semantics", "Models", "focusing", "on", "the", "compositionality", "ability", "of", "those", "models", ",", "the", "SICK", "dataset", "(", "Sentences", "Involving", "Compositional", "Knowledge", ")", "was", "used", "in", "a", "shared", "entailment", "task", ".", "Sentence", "pairs", "were", "obtained", "through", "re", "-", "writing", "rules", "and", "annotated", "with", "the", "three", "RTE", "labels", "via", "a", "crowdsourcing", "platform", ".", "Both", "in", "RTE", "and", "SICK", "the", "label", "assigned", "to", "the", "sentence", "pairs", "captures", "the", "relation", "holding", "between", "the", "two", "sentences", ".", "A", "different", "approach", "has", "been", "used", "to", "build", "the", "much", "larger", "SNLI", "(", "Stanford", "Natural", "Language", "Inference", ")", "dataset", ":", "Premises", "are", "taken", "from", "a", "dataset", "of", "images", "annotated", "with", "descriptive", "captions", ";", "the", "corresponding", "hypotheses", "were", "produced", "through", "crowdsourcing", ",", "where", "for", "a", "given", "premise", ",", "annotators", "provided", "a", "sentence", "which", "is", "true", "or", "not", "true", "with", "respect", "to", "a", "possible", "image", "which", "the", "premise", "could", "describe", ".", "A", "consequence", "of", "this", "choice", "is", "that", "the", "contradiction", "relation", "can", "be", "assigned", "to", "pairs", "which", "are", "rather", "unrelated", "(", "“", "A", "person", "in", "a", "black", "wetsuit", "is", "surfing", "a", "small", "wave", "”", "and", "“", "A", "woman", "is", "trying", "to", "sleep", "on", "her", "bed", "”", ")", ",", "differently", "from", "what", "happens", "in", "RTE", "and", "SICK", ".", "Since", "the", "inception", "of", "RTE", "shared", "tasks", ",", "there", "has", "been", "an", "increasing", "emphasis", "on", "data", "-", "driven", "approaches", "which", ",", "given", "the", "hypothesis", "H", "and", "premise", "P", ",", "seek", "to", "classify", "the", "semantic", "relation", "(", "see", "for", "a", "review", ")", ".", "More", "recently", ",", "neural", "approaches", "have", "come", "to", "dominate", "the", "scene", ",", "as", "shown", "by", "the", "recent", "RepEval", "2017", "task", ",", "where", "all", "submissions", "relied", "on", "bidirectional", "LSTM", "models", ",", "with", "or", "without", "pretrained", "embeddings", ".", "RTE", "also", "intersects", "with", "a", "number", "of", "related", "inference", "problems", ",", "including", "semantic", "text", "similarity", "and", "Question", "Answering", ",", "and", "some", "models", "have", "been", "proposed", "to", "address", "several", "such", "problems", ".", "In", "one", "popular", "approach", ",", "both", "P", "and", "H", "are", "encoded", "within", "the", "same", "embedding", "space", ",", "using", "a", "single", "RNN", ",", "with", "a", "decision", "made", "based", "on", "the", "encodings", "of", "the", "two", "sentences", ".", "This", "is", "the", "approach", "we", "adopt", "for", "our", "baseline", "LSTM", "in", "Section", "[", "reference", "]", ",", "based", "on", "the", "model", "proposed", "by", "snli", ":", "emnlp2015", ",", "albeit", "with", "some", "modifications", "(", "see", "also", ")", ".", "A", "second", "promising", "approach", ",", "based", "on", "which", "we", "adapt", "our", "state", "of", "the", "art", "model", ",", "relies", "on", "matching", "and", "aggregation", ".", "Here", ",", "the", "decision", "concerning", "the", "relationship", "between", "P", "and", "H", "is", "based", "on", "an", "aggregate", "representation", "achieved", "after", "the", "two", "sentences", "are", "matched", ".", "Yet", "another", "area", "where", "neural", "approaches", "are", "being", "applied", "to", "sentence", "pairs", "in", "an", "entailment", "relationship", "is", "generation", ",", "where", "an", "RNN", "generates", "an", "entailed", "hypothesis", "(", "or", "a", "chain", "of", "such", "hypotheses", ")", "given", "an", "encoding", "of", "the", "premise", ".", "paragraph", ":", "Vision", "and", "Textual", "Entailment", "In", "recent", "years", ",", "several", "models", "have", "been", "proposed", "to", "integrate", "the", "language", "and", "vision", "modalities", ";", "usually", "the", "integration", "is", "operationalized", "by", "element", "-", "wise", "multiplication", "between", "linguistic", "and", "visual", "vectors", ".", "Though", "the", "interest", "in", "these", "modalities", "has", "spread", "in", "an", "astonishing", "way", "thanks", "to", "various", "multimodal", "tasks", "proposed", ",", "including", "the", "IC", ",", "VQA", ",", "Visual", "Reasoning", "and", "Visual", "Dialogue", "tasks", "mentioned", "above", ",", "very", "little", "work", "has", "been", "done", "on", "grounding", "entailment", ".", "Interestingly", ",", "youn", ":", "from14", "has", "proposed", "the", "idea", "of", "considering", "images", "as", "the", "“", "possible", "worlds", "”", "on", "which", "sentences", "find", "their", "denotation", ".", "Hence", ",", "they", "released", "a", "“", "visual", "denotation", "graph", "”", "which", "associates", "sentences", "with", "their", "denotation", "(", "sets", "of", "images", ")", ".", "The", "idea", "has", "been", "further", "exploited", "by", "lai", ":", "lear17", "and", "han", ":", "visd17", ".", "vend", ":", "orde16", "look", "at", "hypernymy", ",", "textual", "entailment", "and", "image", "captioning", "as", "special", "cases", "of", "a", "single", "visual", "-", "semantic", "hierarchy", "over", "words", ",", "sentences", "and", "images", ",", "and", "they", "claim", "that", "modelling", "the", "partial", "order", "structure", "of", "this", "hierarchy", "in", "visual", "and", "linguistic", "semantic", "spaces", "improves", "model", "performance", "on", "those", "three", "tasks", ".", "We", "share", "with", "this", "work", "the", "idea", "that", "the", "image", "can", "be", "taken", "as", "a", "possible", "world", ".", "However", ",", "we", "do", "n’t", "use", "sets", "of", "images", "to", "obtain", "the", "visual", "denotation", "of", "text", "in", "order", "to", "check", "whether", "entailment", "is", "logically", "valid", "/", "highly", "likely", ".", "Rather", ",", "we", "take", "the", "image", "to", "be", "the", "world", "/", "situation", "in", "which", "the", "text", "finds", "its", "interpretation", ".", "The", "only", "work", "that", "is", "close", "to", "ours", "is", "an", "unpublished", "student", "report", ",", "which", "however", "lacks", "the", "in", "-", "depth", "analysis", "presented", "here", ".", "section", ":", "Annotated", "dataset", "of", "images", "and", "sentence", "pairs", "We", "took", "as", "our", "starting", "point", "the", "Stanford", "Natural", "Language", "Inference", "(", "SNLI", ")", "dataset", ",", "the", "largest", "natural", "language", "inference", "dataset", "available", "with", "sentence", "pairs", "labelled", "with", "entailment", ",", "contradiction", "and", "neutral", "relations", ".", "We", "augmented", "this", "dataset", "with", "images", ".", "It", "has", "been", "shown", "very", "recently", "that", "SNLI", "contains", "language", "bias", ",", "such", "that", "a", "simple", "classifier", "can", "achieve", "high", "accuracy", "in", "predicting", "the", "three", "classes", "just", "by", "having", "as", "input", "the", "hypothesis", "sentence", ".", "A", "subset", "of", "the", "SNLI", "test", "set", "with", "‘", "hard", "’", "cases", ",", "where", "such", "a", "simplistic", "classifier", "fails", "(", "hereafter", "SNLI", ")", "has", "been", "released", ".", "Hence", ",", "in", "this", "paper", "we", "will", "report", "our", "results", "on", "both", "the", "full", "dataset", "and", "the", "hard", "test", "set", ",", "but", "then", "zoom", "in", "on", "SNLI", "to", "understand", "the", "models", "’", "behaviour", ".", "We", "briefly", "introduce", "SNLI", "and", "the", "new", "test", "set", "and", "compare", "them", "through", "our", "annotation", "of", "linguistic", "phenomena", ".", "subsection", ":", "Dataset", "construction", "paragraph", ":", "SNLI", "and", "SNLI", "test", "set", "The", "SNLI", "dataset", "was", "built", "through", "Amazon", "Mechanical", "Turk", ".", "Workers", "were", "shown", "captions", "of", "photographs", "without", "the", "photo", "and", "were", "asked", "to", "write", "a", "new", "caption", "that", "(", "a", ")", "is", "definitely", "a", "true", "description", "of", "the", "photo", "(", "entailment", ")", ";", "(", "b", ")", "might", "be", "a", "true", "description", "of", "the", "photo", "(", "neutral", ")", ";", "(", "c", ")", "is", "definitely", "a", "false", "description", "of", "the", "photo", "(", "contradiction", ")", ".", "Examples", "were", "provided", "for", "each", "of", "the", "three", "cases", ".", "The", "premises", "are", "captions", "which", "come", "mostly", "from", "Flickr30", "K", ";", "only", "4", "K", "captions", "are", "from", "VisualGenome", ".", "In", "total", ",", "the", "dataset", "contains", "570", ",", "152", "sentence", "pairs", ",", "balanced", "with", "respect", "to", "the", "three", "labels", ".", "Around", "10", "%", "of", "these", "data", "have", "been", "validated", "(", "4", "annotators", "for", "each", "example", "plus", "the", "label", "assigned", "through", "the", "previous", "data", "collection", "phase", ")", ".", "The", "development", "and", "test", "datasets", "contain", "10", "K", "examples", "each", ".", "Moreover", ",", "each", "Image", "/", "Flickr", "caption", "occurs", "in", "only", "one", "of", "the", "three", "sets", ",", "and", "all", "the", "examples", "in", "the", "development", "and", "test", "sets", "have", "been", "validated", ".", "paragraph", ":", "V", "-", "SNLI", "and", "V", "-", "SNLI", "test", "set", "Our", "grounded", "version", "of", "SNLI", ",", "V", "-", "SNLI", ",", "has", "been", "built", "by", "matching", "each", "sentence", "pair", "in", "SNLI", "with", "the", "corresponding", "image", "coming", "from", "the", "Flickr30", "K", "dataset", ";", "thus", "the", "V", "-", "SNLI", "dataset", "is", "slightly", "smaller", "than", "the", "original", ",", "which", "also", "contains", "captions", "from", "VisualGenome", ".", "V", "-", "SNLI", "consists", "of", "565", ",", "286", "pairs", "(", "187", ",", "969", "neutral", ",", "188", ",", "453", "contradiction", ",", "and", "188", ",", "864", "entailment", ")", ".", "Training", ",", "test", ",", "and", "development", "splits", "have", "been", "built", "according", "to", "the", "splits", "in", "SNLI", ".", "The", "main", "statistics", "of", "the", "splits", "of", "the", "dataset", "are", "reported", "in", "Table", "[", "reference", "]", "together", "with", "statistics", "for", "the", "visual", "counterpart", "of", "Hard", "SNLI", ",", "namely", "V", "-", "SNLI", ".", "By", "construction", ",", "V", "-", "SNLI", "contains", "datapoints", "such", "that", "the", "premise", "is", "always", "true", "with", "respect", "to", "the", "image", ",", "whereas", "the", "hypothesis", "can", "be", "either", "true", "(", "entailment", "or", "neutral", "cases", ")", "or", "false", "(", "contradiction", "or", "neutral", "cases", ".", ")", "subsection", ":", "Dataset", "annotation", "For", "deeper", "analysis", "and", "comparison", "of", "the", "contents", "of", "SNLI", "and", "SNLI", ",", "we", "have", "annotated", "the", "SNLI", "dataset", "by", "both", "automatically", "detecting", "some", "surface", "linguistic", "cues", "and", "manually", "labelling", "less", "trivial", "phenomena", ".", "Using", "an", "in", "-", "house", "annotation", "interface", ",", "we", "collected", "human", "judgments", "aiming", "to", "(", "a", ")", "filter", "out", "those", "cases", "for", "which", "the", "gold", "-", "standard", "annotation", "was", "considered", "to", "be", "wrong", ";", "(", "b", ")", "connect", "the", "three", "ungrounded", "relations", "to", "various", "linguistic", "phenomena", ".", "To", "achieve", "this", ",", "we", "annotated", "a", "random", "sample", "of", "the", "SNLI", "test", "set", "containing", "527", "sentence", "pairs", "(", "185", "entailment", ",", "171", "contradiction", ",", "171", "neutral", ")", ",", "out", "of", "which", "176", "were", "from", "the", "hard", "test", "set", "(", "56", "entailment", ",", "62", "contradiction", ",", "58", "neutral", ")", ".", "All", "the", "pairs", "were", "annotated", "by", "at", "least", "two", "annotators", ",", "as", "follows", ":", "(", "a", ")", "We", "filtered", "out", "all", "the", "pairs", "which", "had", "a", "wrong", "gold", "label", "(", "see", "Table", "[", "reference", "]", "for", "details", ")", ".", "When", "our", "annotators", "did", "not", "agree", "whether", "a", "given", "relation", "holds", "for", "a", "specific", "pair", ",", "we", "appealed", "to", "the", "corresponding", "five", "judgments", "coming", "from", "the", "validation", "stage", "of", "the", "SNLI", "dataset", "to", "reach", "a", "consensus", "based", "on", "the", "majority", "of", "labels", ".", "(", "b", ")", "We", "considered", "as", "valid", "any", "linguistic", "tag", "assigned", "by", "at", "least", "one", "annotator", ".", "Since", "the", "annotation", "for", "(", "a", ")", "is", "binary", "whereas", "for", "(", "b", ")", "it", "is", "multi", "-", "class", ",", "we", "used", "Cohen", "’s", "for", "the", "former", "and", "also", "Scott", "’s", "and", "Krippendorff", "’s", "for", "the", "latter", "as", "suggested", "by", "Passonneau", "pass", ":", "inte06", ".", "The", "inter", "-", "annotator", "agreement", "for", "the", "relation", "type", "(", "a", ")", "was", ";", "for", "(", "b", ")", "linguistic", "tags", "it", "was", ",", ",", "and", ".", "paragraph", ":", "Linguistic", "phenomena", "Following", "the", "error", "analysis", "approach", "described", "in", "recent", "work", ",", "we", "compiled", "a", "new", "list", "of", "linguistic", "features", "that", "can", "be", "of", "interest", "when", "contrasting", "SNLI", "and", "SNLI", ",", "as", "well", "as", "for", "evaluating", "RTE", "models", ".", "Some", "of", "these", "were", "detected", "automatically", ",", "while", "others", "were", "assigned", "manually", ".", "Automatic", "tags", "included", "Synonym", "and", "Antonym", ",", "which", "were", "detected", "using", "WordNet", ".", "Quantifier", ",", "Pronoun", ",", "Diff", "Tense", ",", "Superlative", "and", "Bare", "NP", "were", "identified", "using", "Penn", "treebank", "labels", ",", "while", "labels", "such", "as", "Negation", "were", "found", "with", "a", "straightforward", "keyword", "search", ".", "The", "tag", "Long", "has", "been", "assigned", "to", "sentence", "pairs", "with", "a", "premise", "containing", "more", "than", "30", "tokens", ",", "or", "a", "hypothesis", "with", "more", "than", "16", "tokens", ".", "Details", "about", "the", "tags", "used", "in", "the", "manual", "annotation", "are", "presented", "in", "Table", "[", "reference", "]", ".", "We", "examined", "the", "differences", "in", "the", "tags", "distributions", "between", "the", "SNLI", "and", "SNLI", "test", "sets", "(", "Table", "[", "reference", "]", ")", ".", "Interestingly", ",", "the", "hard", "sentence", "pairs", "from", "our", "random", "sample", "include", "proportionately", "more", "antonyms", "but", "fewer", "pronouns", ",", "as", "well", "as", "examples", "with", "different", "verb", "tenses", "in", "the", "premise", "and", "hypothesis", ",", "compared", "to", "the", "full", "test", "set", ".", "Furthermore", ",", "SNLI", "contains", "a", "significantly", "larger", "proportion", "of", "gold", "-", "standard", "labels", "which", "become", "wrong", "when", "the", "image", "is", "factored", "in", "(", "-", "test", "with", ")", ".", "section", ":", "Models", "In", "this", "section", ",", "we", "describe", "a", "variety", "of", "models", "that", "were", "compared", "on", "both", "V", "-", "SNLI", "and", "V", "-", "SNLI", ",", "ranging", "from", "baseline", "models", "based", "on", "snli", ":", "emnlp2015", "to", "a", "state", "of", "the", "art", "model", "by", "Wang2017", ".", "We", "compare", "the", "original", "‘", "blind", "’", "version", "of", "a", "model", "with", "a", "visually", "-", "augmented", "counterpart", ".", "In", "what", "follows", ",", "we", "use", "P", "and", "H", "to", "refer", "to", "a", "premise", "and", "hypothesis", ",", "respectively", ".", "paragraph", ":", "LSTM", "baseline", "(", "Blind", ")", "This", "model", "exploits", "a", "Recurrent", "Neural", "Network", "with", "Long", "Short", "-", "Term", "Memory", "units", "to", "encode", "both", "P", "and", "H", "in", "512D", "vectors", ".", "The", "two", "vectors", "are", "then", "concatenated", "in", "a", "stack", "of", "three", "512D", "layers", "having", "a", "ReLU", "activation", "function", ",", "with", "a", "final", "softmax", "layer", "to", "classify", "the", "relation", "between", "the", "two", "sentences", "as", "entailment", ",", "contradiction", "or", "neutral", ".", "The", "model", "is", "inspired", "by", "the", "LSTM", "baseline", "proposed", "by", "snli", ":", "emnlp2015", ".", "The", "model", "exploits", "the", "300", ",", "000", "most", "frequent", "pretrained", "GloVe", "embeddings", "and", "improves", "them", "during", "the", "training", "process", ".", "To", "regularize", "the", "model", ",", "Dropout", "is", "applied", "to", "the", "inputs", "and", "outputs", "of", "the", "recurrent", "layers", "and", "to", "the", "ReLU", "fully", "connected", "layers", "with", "a", "keeping", "probability", "of", "0.5", ".", "The", "model", "is", "trained", "using", "the", "Adam", "optimizer", "with", "a", "learning", "rate", "of", "0.001", "until", "its", "accuracy", "on", "the", "development", "set", "drops", "for", "three", "successive", "iterations", ".", "paragraph", ":", "V", "-", "LSTM", "baseline", "The", "LSTM", "model", "described", "above", "is", "augmented", "with", "a", "visual", "component", "following", "a", "standard", "Visual", "Question", "Answering", "baseline", "model", ".", "Following", "initial", "representation", "of", "P", "and", "H", "in", "512D", "vectors", "through", "an", "LSTM", ",", "a", "fully", "-", "connected", "layer", "projects", "the", "L2", "-", "normalized", "4096D", "image", "vector", "coming", "from", "the", "penultimate", "layer", "of", "a", "VGGnet16", "Convolutional", "Neural", "Network", "to", "a", "reduced", "512D", "vector", ".", "A", "fully", "-", "connected", "layer", "with", "a", "ReLU", "activation", "function", "is", "also", "applied", "to", "P", "and", "H", "to", "obtain", "two", "512D", "vectors", ".", "The", "multimodal", "fusion", "between", "the", "text", "and", "the", "image", "is", "obtained", "by", "performing", "an", "element", "-", "wise", "multiplication", "between", "the", "vector", "of", "the", "text", "representation", "and", "the", "reduced", "vector", "of", "the", "image", ".", "The", "multimodal", "fusion", "is", "performed", "between", "the", "image", "and", "both", "the", "premise", "and", "the", "hypothesis", ",", "resulting", "in", "two", "multimodal", "representations", ".", "The", "relation", "between", "them", "is", "captured", "as", "in", "the", "model", "described", "above", ".", "This", "model", "uses", "GloVe", "embeddings", "and", "the", "same", "optimization", "and", "procedure", "described", "above", ".", "We", "have", "also", "adapted", "a", "state", "-", "of", "-", "the", "-", "art", "attention", "-", "based", "model", "for", "IC", "and", "VQA", "to", "the", "GTE", "task", ".", "It", "obtains", "results", "comparable", "to", "the", "V", "-", "LSTM", ".", "This", "lack", "of", "improvement", "might", "be", "due", "to", "the", "need", "of", "further", "parameter", "tuning", ".", "We", "report", "the", "details", "of", "our", "implementation", "and", "its", "results", "in", "the", "Supplementary", "Material", ".", "paragraph", ":", "BiMPM", "The", "Bilateral", "Multi", "-", "Perspective", "Matching", "(", "BiMPM", ")", "model", "obtains", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "the", "SNLI", "dataset", ",", "achieving", "a", "maximum", "accuracy", "of", "86.9", "%", ",", "and", "going", "up", "to", "88.8", "%", "in", "an", "ensemble", "setup", ".", "An", "initial", "embedding", "layer", "vectorises", "words", "in", "P", "and", "H", "using", "pretrained", "GLoVe", "embeddings", ",", "and", "passing", "them", "to", "a", "context", "representation", "layer", ",", "which", "uses", "bidirectional", "LSTMs", "(", "BiLSTMs", ")", "to", "encode", "context", "vectors", "for", "each", "time", "-", "step", ".", "The", "core", "part", "of", "the", "model", "is", "the", "subsequent", "matching", "layer", ",", "where", "each", "contextual", "embedding", "or", "time", "-", "step", "of", "P", "is", "matched", "against", "all", "the", "embeddings", "of", "H", ",", "and", "vice", "versa", ".", "The", "output", "of", "this", "layer", "is", "composed", "of", "two", "sequences", "of", "matching", "vectors", ",", "which", "constitute", "the", "input", "to", "another", "BiLSTM", "at", "the", "aggregation", "layer", ".", "The", "vectors", "from", "the", "last", "time", "-", "step", "of", "the", "BiLSTM", "are", "concatenated", "into", "a", "fixed", "-", "length", "vector", ",", "which", "is", "passed", "to", "the", "final", "prediction", "tier", ",", "a", "two", "-", "layer", "feed", "-", "forward", "network", "which", "classifies", "the", "relation", "between", "P", "and", "H", "via", "softmax", ".", "Matching", "is", "performed", "via", "a", "cosine", "operation", ",", "which", "yields", "an", "-", "dimensional", "vector", ",", "where", "is", "the", "number", "of", "perspectives", ".", "Wang2017", "experiment", "with", "four", "different", "matching", "strategies", ".", "In", "their", "results", ",", "the", "best", "-", "performing", "version", "of", "the", "BiMPM", "model", "used", "all", "four", "matching", "strategies", ".", "We", "adopt", "this", "version", "of", "the", "model", "in", "what", "follows", ".", "paragraph", ":", "V", "-", "BiMPM", "model", "We", "enhanced", "BiMPM", "to", "account", "for", "the", "image", ",", "too", ".", "Our", "version", "of", "this", "model", "is", "referred", "to", "as", "the", "V", "-", "BiMPM", ".", "Here", ",", "the", "feature", "vector", "for", "an", "image", "is", "obtained", "from", "the", "layer", "before", "the", "fully", "-", "connected", "layer", "of", "a", "VGGnet", "-", "16", ".", "This", "results", "in", "a", "tensor", ",", "which", "we", "consider", "as", "49", "512", "-", "dimensional", "vectors", ".", "The", "same", "matching", "operations", "are", "performed", ",", "except", "that", "matching", "occurs", "between", "P", ",", "H", ",", "and", "the", "image", ".", "Since", "the", "textual", "and", "visual", "vectors", "have", "different", "dimensionality", "and", "belong", "to", "different", "spaces", ",", "we", "first", "map", "them", "to", "a", "mutual", "space", "using", "an", "affine", "transformation", ".", "We", "match", "textual", "and", "image", "vectors", "using", "a", "cosine", "operation", ",", "as", "before", ".", "Full", "details", "of", "the", "model", "are", "reported", "in", "the", "Supplementary", "Materials", "for", "this", "paper", ".", "section", ":", "Experiments", "and", "Results", "The", "models", "described", "in", "the", "previous", "sections", "were", "evaluated", "on", "both", "(", "V", "-)", "SNLI", "and", "(", "V", "-)", "SNLI", ".", "For", "the", "visually", "-", "augmented", "models", ",", "we", "experimented", "with", "configurations", "where", "image", "vectors", "were", "combined", "with", "both", "P", "and", "H", "(", "namely", "P", "+", "I", "and", "H", "+", "I", ")", ",", "or", "only", "with", "H", "(", "P", "and", "H", "+", "I", ")", ".", "The", "best", "setting", "was", "invariably", "the", "one", "where", "only", "H", "was", "grounded", ";", "hence", ",", "we", "focus", "on", "these", "results", "in", "what", "follows", ",", "comparing", "them", "to", "“", "blind", "”", "models", ".", "In", "view", "of", "recent", "results", "suggesting", "that", "biases", "in", "SNLI", "afford", "a", "high", "accuracy", "in", "the", "prediction", "task", "with", "only", "the", "hypothesis", "sentence", "as", "input", ",", "we", "also", "include", "results", "for", "the", "blind", "models", "without", "the", "premise", "(", "denoted", "with", "[", "H", "]", "in", "what", "follows", ")", ".", "Table", "[", "reference", "]", "shows", "the", "results", "of", "the", "various", "models", "on", "the", "full", "V", "-", "SNLI", "dataset", ".", "The", "same", "models", "are", "compared", "in", "Table", "[", "reference", "]", "on", "V", "-", "SNLI", ".", "First", ",", "note", "that", "the", "LSTM", "[", "H", "]", "model", "evinces", "a", "drop", "in", "performance", "compared", "to", "LSTM", "(", "from", "81.49", "%", "to", "68.49", "%", ")", ",", "though", "the", "drop", "is", "much", "greater", "on", "the", "unbiased", "SNLI", "subset", "(", "from", "60.99", "to", "25.57", "%", ")", ".", "This", "confirms", "the", "results", "reported", "by", "guru", ":", "anno18", "and", "justifies", "our", "additional", "focus", "on", "this", "subset", "of", "the", "data", ".", "The", "effect", "of", "grounding", "in", "these", "models", "is", "less", "clear", ".", "The", "LSTM", "baseline", "performs", "worse", "when", "it", "is", "visually", "augmented", ";", "this", "is", "the", "case", "of", "V", "-", "SNLI", "and", ",", "even", "more", "drastically", ",", "V", "-", "SNLI", ".", "It", "is", "also", "true", "irrespective", "of", "the", "relationship", "type", ".", "On", "the", "other", "hand", ",", "the", "V", "-", "BiMPM", "model", "improves", "marginally", "across", "the", "board", ",", "compared", "to", "BiMPM", ",", "on", "the", "V", "-", "SNLI", "data", ".", "On", "the", "hard", "subset", ",", "the", "images", "appear", "to", "hurt", "performance", "somewhat", "in", "the", "case", "of", "contradiction", "(", "from", "77.62", "%", "to", "76.12", "%", ")", ",", "but", "improve", "it", "by", "a", "substantial", "margin", "on", "neutral", "cases", "(", "from", "59.36", "%", "to", "63.67", "%", ")", ".", "The", "neutral", "case", "is", "the", "hardest", "for", "all", "models", ",", "with", "the", "possible", "exception", "of", "LSTM", "[", "H", "]", "on", "the", "full", "dataset", ".", "Overall", ",", "the", "results", "suggest", "that", "factoring", "in", "images", "either", "hinders", "performance", "(", "as", "in", "the", "case", "of", "the", "V", "-", "LSTM", "baseline", ")", ",", "or", "helps", "only", "marginally", "(", "as", "in", "the", "case", "of", "V", "-", "BiMPM", ")", ".", "In", "the", "latter", "case", ",", "we", "also", "observe", "instances", "where", "factoring", "in", "images", "hurts", "performance", ".", "In", "an", "effort", "to", "understand", "the", "results", ",", "we", "turn", "to", "a", "more", "detailed", "error", "analysis", "of", "the", "V", "-", "BiMPM", "model", ",", "first", "in", "relation", "to", "the", "dataset", "annotations", ",", "and", "then", "by", "zooming", "in", "somewhat", "closer", "on", "V", "-", "SNLI", ".", "subsection", ":", "Error", "analysis", "by", "linguistic", "annotation", "label", "In", "Table", "[", "reference", "]", ",", "accuracies", "for", "the", "blind", "and", "grounded", "version", "of", "BiMPM", "are", "broken", "down", "by", "the", "labels", "given", "to", "the", "sentence", "pairs", "in", "the", "annotated", "subset", "of", "SNLI", "described", "in", "Section", "[", "reference", "]", ".", "We", "only", "observe", "a", "significant", "difference", "in", "the", "Entity", "case", ",", "that", "is", ",", "where", "the", "referents", "in", "P", "and", "H", "are", "inconsistent", ".", "Here", ",", "the", "blind", "model", "outperforms", "the", "grounded", "one", ",", "an", "unexpected", "result", ",", "since", "one", "would", "assume", "a", "grounded", "model", "to", "be", "better", "equipped", "to", "identify", "mismatched", "referents", ".", "Hence", ",", "in", "the", "following", "we", "aim", "to", "understand", "whether", "the", "models", "properly", "deal", "with", "the", "grounding", "sub", "-", "task", ".", "subsection", ":", "Error", "analysis", "on", "grounding", "in", "the", "SNLI", "We", "next", "turn", "to", "the", "“", "hard", "”", "subset", "of", "the", "data", ",", "where", "V", "-", "BiMPM", "showed", "some", "improvement", "over", "the", "blind", "case", ",", "but", "suffered", "on", "contradiction", "cases", "(", "Table", "[", "reference", "]", ")", ".", "We", "analysed", "the", "207", "cases", "in", "SNLI", "where", "the", "V", "-", "BiMPM", "made", "incorrect", "predictions", "compared", "to", "the", "blind", "model", ",", "that", "is", ",", "where", "the", "image", "hurt", "performance", ".", "These", "were", "annotated", "independently", "by", "two", "of", "the", "authors", "(", "raw", "inter", "-", "annotator", "agreement", ":", "96", "%", ")", "who", "(", "a", ")", "read", "the", "two", "sentences", ",", "P", "and", "H", ";", "(", "b", ")", "checked", "whether", "the", "relation", "annotated", "in", "the", "dataset", "actually", "held", "or", "whether", "it", "was", "an", "annotation", "error", ";", "(", "c", ")", "in", "those", "cases", "where", "it", "held", ",", "checked", "whether", "including", "the", "image", "actually", "resulted", "in", "a", "change", "in", "the", "relation", ".", "Table", "[", "reference", "]", "displays", "the", "proportions", "of", "image", "mismatch", "and", "incorrect", "annotations", ".", "As", "the", "table", "suggests", ",", "in", "the", "cases", "where", "images", "hinder", "performance", "in", "the", "V", "-", "BiMPM", ",", "it", "is", "usually", "because", "the", "image", "changes", "the", "relation", "(", "thus", ",", "these", "are", "cases", "of", "image", "mismatch", ";", "see", "Section", "[", "reference", "]", "for", "an", "example", ")", ";", "this", "occurs", "in", "a", "large", "proportion", "of", "cases", "labelled", "as", "neutral", "in", "the", "dataset", ".", "Inspired", "by", "the", "work", "in", ",", "we", "further", "explored", "the", "impact", "of", "visual", "grounding", "in", "both", "the", "V", "-", "LSTM", "and", "V", "-", "BiMPM", "by", "comparing", "their", "performance", "on", "SNLI", ",", "with", "the", "same", "subset", "incorporating", "image", "“", "foils", "”", ".", "Vectors", "for", "the", "images", "in", "the", "V", "-", "SNLI", "test", "set", "were", "compared", "pairwise", "using", "cosine", ",", "and", "for", "each", "test", "case", "in", "V", "-", "SNLI", ",", "the", "actual", "image", "was", "replaced", "with", "the", "most", "dissimilar", "image", "in", "the", "full", "test", "set", ".", "The", "rationale", "is", "that", ",", "if", "visual", "grounding", "is", "really", "helpful", "in", "recognising", "the", "semantic", "relationship", "between", "P", "and", "H", ",", "we", "should", "observe", "a", "drop", "in", "performance", "when", "the", "images", "are", "unrelated", "to", "the", "scenario", "described", "by", "the", "sentences", ".", "The", "results", "are", "displayed", "in", "Table", "[", "reference", "]", ",", "which", "also", "reproduces", "the", "original", "results", "on", "V", "-", "SNLI", "from", "Table", "[", "reference", "]", "for", "ease", "of", "reference", ".", "As", "the", "results", "show", ",", "models", "are", "not", "hurt", "by", "the", "foil", "image", ",", "contrary", "to", "our", "expectations", ".", "V", "-", "BiMPM", "overall", "drops", "just", "by", "0.67", "%", "whereas", "V", "-", "LSTM", "drop", "is", "somewhat", "higher", "(", "-", "2.11", "%", ")", "showing", "it", "might", "be", "doing", "a", "better", "job", "on", "the", "grounding", "sub", "-", "task", ".", "As", "a", "final", "check", ",", "we", "sought", "to", "isolate", "the", "grounding", "from", "the", "reasoning", "sub", "-", "task", ",", "focusing", "only", "on", "the", "former", ".", "We", "compared", "the", "models", "when", "grounding", "only", "the", "hypothesis", "[", "H", "+", "I", "]", ",", "while", "leaving", "out", "the", "premise", ".", "Note", "that", "this", "test", "is", "different", "from", "the", "evaluation", "of", "the", "model", "using", "only", "the", "hypothesis", "[", "H", "]", ":", "Whereas", "in", "that", "case", "the", "input", "is", "not", "expected", "to", "provide", "any", "useful", "information", "to", "perform", "the", "task", ",", "here", "it", "is", ".", "As", "we", "noted", "in", "Section", "[", "reference", "]", ",", "by", "construction", "the", "premise", "is", "always", "true", "with", "respect", "to", "the", "image", "while", "the", "hypothesis", "can", "be", "either", "true", "(", "entailment", "or", "neutral", "cases", ")", "or", "false", "(", "contradiction", "or", "neutral", "cases", ")", ".", "A", "model", "that", "is", "grounding", "the", "text", "adequately", "would", "be", "expected", "to", "confuse", "both", "entailment", "and", "contradiction", "cases", "with", "neutral", "ones", ";", "on", "the", "other", "hand", ",", "neutral", "cases", "should", "be", "confused", "with", "entailments", "or", "contradictions", ".", "Confusing", "contradictions", "with", "entailments", "would", "be", "a", "sign", "that", "a", "model", "is", "grounding", "inadequately", ",", "since", "it", "is", "not", "recognising", "that", "H", "is", "false", "with", "respect", "to", "the", "image", ".", "As", "the", "left", "panel", "of", "Table", "[", "reference", "]", "shows", ",", "V", "-", "BiMPM", "outperforms", "V", "-", "LSTM", "by", "a", "substantial", "margin", ",", "though", "the", "performance", "of", "both", "models", "drops", "substantially", "with", "this", "setup", ".", "The", "right", "panel", "in", "the", "table", "shows", "that", "neither", "model", "is", "free", "of", "implausible", "errors", "(", "confusing", "entailments", "and", "contradictions", ")", ",", "though", "V", "-", "BiMPM", "makes", "substantially", "fewer", "of", "these", ".", "section", ":", "Conclusion", "This", "paper", "has", "investigated", "the", "potential", "of", "grounding", "the", "textual", "entailment", "task", "in", "visual", "data", ".", "We", "argued", "that", "a", "Grounded", "Textual", "Entailment", "model", "needs", "to", "perform", "two", "tasks", ":", "(", "a", ")", "the", "grounding", "itself", ",", "and", "(", "b", ")", "reasoning", "about", "the", "relation", "between", "the", "sentences", ",", "against", "the", "visual", "information", ".", "Our", "results", "suggest", "that", "a", "model", "based", "on", "matching", "and", "aggregation", "like", "the", "BiMPM", "model", "can", "perform", "very", "well", "at", "the", "reasoning", "task", ",", "classifying", "entailment", "relations", "correctly", "much", "more", "frequently", "than", "a", "baseline", "V", "-", "LSTM", ".", "On", "the", "other", "hand", ",", "it", "is", "not", "clear", "that", "grounding", "is", "being", "performed", "adequately", "in", "this", "model", ".", "It", "is", "primarily", "in", "the", "case", "of", "contradictions", "that", "the", "image", "seems", "to", "play", "a", "direct", "role", "in", "biasing", "the", "classification", "towards", "the", "right", "or", "wrong", "class", ",", "depending", "on", "whether", "the", "image", "is", "correct", ".", "In", "summary", ",", "two", "conclusions", "can", "be", "drawn", "from", "these", "results", ".", "First", ",", "in", "those", "cases", "where", "the", "inclusion", "of", "visual", "information", "results", "in", "a", "loss", "of", "accuracy", ",", "this", "is", "often", "due", "to", "the", "image", "resulting", "in", "a", "change", "in", "the", "original", "relation", "annotated", "in", "the", "dataset", ".", "A", "related", "observation", "is", "that", "using", "foil", "images", "results", "in", "a", "greater", "drop", "in", "performance", "on", "contradiction", "cases", ",", "possibly", "because", "in", "such", "cases", ",", "grounding", "serves", "to", "identify", "a", "mismatch", "between", "the", "hypothesis", "and", "the", "scene", "described", "by", "the", "premise", ",", "a", "situation", "which", "is", "rendered", "opaque", "by", "the", "introduction", "of", "foils", ".", "Second", ",", "in", "those", "cases", "where", "improvements", "are", "observed", "in", "the", "state", "of", "the", "art", "V", "-", "BiMPM", ",", "the", "precise", "role", "played", "by", "the", "image", "is", "not", "straightforward", ".", "Indeed", ",", "we", "find", "that", "this", "model", "still", "marginally", "outperforms", "the", "‘", "blind", "’", ",", "text", "-", "only", "model", "overall", ",", "when", "the", "images", "involved", "are", "foils", "rather", "than", "actual", "images", ".", "We", "believe", "that", "further", "research", "on", "grounded", "TE", "is", "worthy", "of", "the", "NLP", "community", "’s", "attention", ".", "While", "linking", "language", "with", "perception", "is", "currently", "a", "topical", "issue", ",", "there", "has", "been", "relatively", "little", "work", "on", "linking", "grounding", "directly", "with", "inference", ".", "By", "drawing", "closer", "to", "a", "joint", "solution", "to", "the", "grounding", "and", "inference", "tasks", ",", "models", "will", "also", "be", "better", "able", "to", "address", "language", "understanding", "in", "the", "real", "world", ".", "The", "present", "paper", "presented", "a", "first", "step", "in", "this", "direction", "using", "a", "version", "of", "an", "existing", "TE", "dataset", "which", "was", "augmented", "with", "images", "that", "could", "be", "paired", "directly", "with", "the", "premises", ",", "since", "these", "were", "originally", "captions", "for", "those", "images", ".", "However", ",", "it", "is", "important", "to", "note", "that", "in", "this", "dataset", "premise", "-", "hypotheses", "pairs", "were", "not", "generated", "directly", "with", "reference", "to", "the", "images", "themselves", ".", "An", "important", "issue", "to", "consider", "in", "future", "work", "on", "GTE", ",", "besides", "the", "development", "of", "better", "models", ",", "is", "the", "development", "of", "datasets", "in", "which", "the", "role", "of", "perceptual", "information", "is", "controlled", ",", "ensuring", "that", "the", "data", "on", "which", "models", "are", "trained", "represents", "truly", "grounded", "inferences", ".", "section", ":", "Acknowledgements", "We", "kindly", "acknowledge", "the", "European", "Network", "on", "Integrating", "Vision", "and", "Language", "(", "iV", "&", "L", "Net", ")", "ICT", "COST", "Action", "IC1307", ".", "Moreover", ",", "we", "thank", "the", "Erasmus", "Mundus", "European", "Program", "in", "Language", "and", "Communication", "Technology", ".", "Marc", "Tanti", "’s", "work", "is", "partially", "funded", "by", "the", "Endeavour", "Scholarship", "Scheme", "(", "Malta", ")", ",", "part", "-", "financed", "by", "the", "European", "Union", "’s", "European", "Social", "Fund", "(", "ESF", ")", ".", "Finally", ",", "we", "gratefully", "acknowledge", "the", "support", "of", "NVIDIA", "Corporation", "with", "the", "donations", "to", "the", "University", "of", "Trento", "of", "the", "GPUs", "used", "in", "our", "research", ".", "section", ":", "Appendix", "A", ":", "Bottom", "-", "up", "top", "-", "down", "attention", "(", "VQA", ")", "We", "adapted", "the", "Visual", "Question", "Answering", "model", "proposed", "in", "to", "the", "Grounded", "Textual", "Entailment", "task", ".", "The", "model", "presents", "a", "more", "fine", "-", "grained", "attention", "mechanism", "which", "allows", "to", "identify", "the", "most", "important", "regions", "discovered", "in", "the", "image", "and", "to", "perform", "attention", "over", "each", "of", "them", ".", "The", "model", "uses", "a", "a", "Recurrent", "Neural", "Network", "with", "Long", "Short", "-", "Term", "Memory", "units", "to", "encode", "the", "premise", "P", "and", "hypothesis", "H", "in", "512D", "vectors", ".", "A", "bottom", "-", "up", "attention", "mechanism", "exploits", "a", "Fast", "R", "-", "CNN", "based", "on", "a", "ResNet", "-", "101", "convolutional", "neural", "network", "to", "obtain", "region", "proposals", "corresponding", "to", "the", "36", "most", "informative", "regions", "of", "the", "image", ".", "A", "top", "-", "down", "attention", "mechanism", "is", "used", "between", "the", "premise", "(", "resp", ".", "hypothesis", ")", "and", "each", "of", "the", "L2", "-", "normalized", "2048D", "image", "vectors", "corresponding", "to", "the", "region", "proposals", "to", "obtain", "an", "attention", "score", "for", "each", "of", "them", ".", "Then", ",", "a", "2048D", "image", "vector", "encoding", "the", "most", "interesting", "visual", "features", "for", "the", "premise", "(", "hypothesis", ")", "is", "obtained", "as", "a", "sum", "of", "the", "36", "image", "vectors", "weighted", "by", "the", "corresponding", "attention", "scores", "for", "the", "premise", "(", "hypothesis", ")", ".", "A", "fully", "-", "connected", "layer", "with", "a", "gated", "tanh", "activation", "function", "is", "applied", "to", "the", "image", "vector", "of", "the", "most", "interesting", "visual", "features", "for", "the", "premise", "and", "for", "the", "hypothesis", "to", "obtain", "a", "reduced", "512D", "vector", "for", "each", "of", "them", ".", "A", "fully", "-", "connected", "layer", "with", "a", "gated", "tanh", "activation", "function", "is", "also", "applied", "to", "the", "premise", "and", "to", "the", "hypothesis", "in", "order", "to", "obtain", "a", "reduced", "512D", "vector", "for", "each", "of", "them", ".", "The", "multimodal", "fusion", "between", "the", "premise", "(", "hypothesis", ")", "and", "the", "image", "vector", "of", "the", "most", "interesting", "visual", "features", "for", "the", "premise", "(", "hypothesis", ")", "is", "obtained", "by", "performing", "an", "element", "-", "wise", "multiplication", "between", "the", "reduced", "vector", "of", "the", "premise", "(", "hypothesis", ")", "and", "the", "reduced", "vector", "of", "the", "most", "interesting", "visual", "features", "for", "the", "premise", "(", "hypothesis", ")", ".", "After", "that", ",", "the", "model", "feeds", "the", "concatenation", "of", "the", "two", "resulting", "multimodal", "representations", "to", "a", "stack", "of", "three", "512D", "layers", "having", "a", "gated", "tanh", "activation", "function", ",", "with", "a", "final", "softmax", "layer", "to", "classify", "the", "relation", "between", "the", "two", "sentences", "as", "entailment", ",", "contradiction", "or", "neutral", ".", "This", "model", "uses", "GloVe", "embeddings", "and", "the", "same", "optimization", "tricks", "and", "procedure", "of", "the", "LSTM", "and", "V", "-", "LSTM", "models", ".", "We", "report", "the", "accuracies", "of", "the", "VQA", "models", "against", "the", "various", "tests", "reported", "in", "the", "paper", ".", "For", "ease", "of", "comparison", "we", "reproduce", "the", "full", "table", "from", "the", "main", "paper", ",", "with", "the", "addition", "of", "the", "VQA", "results", ".", "section", ":", "Appendix", "B", ":", "V", "-", "biMPM", "Model", "details", "arrows", ",", "calc", ",", "fit", "positioning", ",", "shapes.multipart", ",", "shapes.callouts", "decorations.pathreplacing", "hbox=", "[", "rectangle", ",", "minimum", "width=7pt", ",", "minimum", "height=25pt", "]", "vbox=", "[", "rectangle", ",", "fill", "=", "black!25", ",", "minimum", "width=35pt", ",", "minimum", "height=10pt", "]", "blue", "hbox=", "[", "hbox", ",", "fill", "=", "blue", "]", "dots", "hbox=", "[", "hbox", "]", "green", "hbox=", "[", "hbox", ",", "fill", "=", "green", "]", "green", "vbox=", "[", "vbox", ",", "fill", "=", "green", "]", "tbox=", "[", "rectangle", ",", "minimum", "width=45pt", ",", "minimum", "height=20pt", ",", "draw", "=", "black", ",", "text", "centered", ",", "text", "=", "black", "]", "red", "arrow=", "[-", "¿", ",", "red", "]", "match", "arrow=", "[-", "¿", ",", "black", "]", "red", "matching=", "[", "text", "=", "red", ",", "inner", "sep=0", "]", "black", "matching=", "[", "text", "=", "black", ",", "inner", "sep=0", "]", "context", "container", "/", ".style", "=", "draw", "=", "orange", ",", "rectangle", "callout", ",", "inner", "sep=0.6em", "0.95", "!", "[", "scale=0.2", "]", "[", "]", "(", "prem_text", ")", "at", "(", "10", ",-", "7", ")", "Premise", ";", "[", "]", "(", "hypo_text", ")", "at", "(", "35", ",-", "7", ")", "Hypothesis", ";", "[", "blue", "hbox", ",", "label", "=", "below", ":", "]", "(", "w", "-", "1", ")", "at", "(", "0", ",", "0", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "w", "-", "1", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "]", "(", "w", "-", "2", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "w", "-", "2", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "…", "]", "(", "w", "-", "3", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "w", "-", "3", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "]", "(", "w", "-", "4", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "w", "-", "4", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "…", "]", "(", "w", "-", "5", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "w", "-", "5", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "]", "(", "w", "-", "6", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "w", "-", "6", ",", "node", "distance=1.5cm*1.2", ",", "label", "=", "below", ":", "]", "(", "w", "-", "7", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "w", "-", "7", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "]", "(", "w", "-", "8", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "w", "-", "8", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "…", "]", "(", "w", "-", "9", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "w", "-", "9", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "]", "(", "w", "-", "10", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "w", "-", "10", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "…", "]", "(", "w", "-", "11", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "w", "-", "11", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "]", "(", "w", "-", "12", ")", ";", "[", "inner", "sep=0pt", ",", "right", "of", "=", "w", "-", "12", ",", "node", "distance=1.5cm*2.15", "]", "(", "image", ")", "[", "width=.10", "]", "figures", "/", "cat.jpg", ";", "[", "blue", "hbox", "]", "(", "c", "-", "1", ")", "at", "(", "0", ",", "1.5cm*7", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "1", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "2", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "c", "-", "2", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "3", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "3", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "4", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "c", "-", "4", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "5", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "5", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "6", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "1", ")", "–", "(", "c", "-", "2", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "2", ")", "–", "(", "c", "-", "3", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "3", ")", "–", "(", "c", "-", "4", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "4", ")", "–", "(", "c", "-", "5", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "5", ")", "–", "(", "c", "-", "6", ")", ";", "[", "blue", "hbox", ",", "above", "of", "=", "c", "-", "1", ",", "node", "distance=1.5cm", "/", "1.2", "]", "(", "c", "-", "1u", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "1u", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "2u", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "c", "-", "2u", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "3u", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "3u", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "4u", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "c", "-", "4u", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "5u", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "5u", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "6u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "2u", ")", "–", "(", "c", "-", "1u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "3u", ")", "–", "(", "c", "-", "2u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "4u", ")", "–", "(", "c", "-", "3u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "5u", ")", "–", "(", "c", "-", "4u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "6u", ")", "–", "(", "c", "-", "5u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "10pt", ",", "shorten", "¡", "=", "-", "3pt", "]", "(", "w", "-", "3", ")", "–", "(", "c", "-", "3", ")", ";", "[", "inner", "sep=0", "]", "(", "p", "-", "callout", "-", "pointer", ")", "at", "(", ")", ";", "[", "context", "container", ",", "callout", "absolute", "pointer=", "(", "p", "-", "callout", "-", "pointer", ")", ",", "fit=", "(", "c", "-", "1", ")", "(", "c", "-", "6u", ")", "]", "(", "premise", "-", "callout", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "6", ",", "node", "distance=1.5cm*1.2", "]", "(", "c", "-", "7", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "7", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "8", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "c", "-", "8", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "9", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "9", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "10", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "c", "-", "10", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "11", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "11", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "12", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "7", ")", "–", "(", "c", "-", "8", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "8", ")", "–", "(", "c", "-", "9", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "9", ")", "–", "(", "c", "-", "10", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "10", ")", "–", "(", "c", "-", "11", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "11", ")", "–", "(", "c", "-", "12", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "6u", ",", "node", "distance=1.5cm*1.2", "]", "(", "c", "-", "7u", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "7u", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "8u", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "c", "-", "8u", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "9u", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "9u", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "10u", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "c", "-", "10u", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "11u", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "11u", ",", "node", "distance=0.7", "cm", "]", "(", "c", "-", "12u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "8u", ")", "–", "(", "c", "-", "7u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "9u", ")", "–", "(", "c", "-", "8u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "10u", ")", "–", "(", "c", "-", "9u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "11u", ")", "–", "(", "c", "-", "10u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "c", "-", "12u", ")", "–", "(", "c", "-", "11u", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "10pt", ",", "shorten", "¡", "=", "-", "3pt", "]", "(", "w", "-", "9", ")", "–", "(", "c", "-", "9", ")", ";", "[", "inner", "sep=0", "]", "(", "h", "-", "callout", "-", "pointer", ")", "at", "(", ")", ";", "[", "context", "container", ",", "callout", "absolute", "pointer=", "(", "h", "-", "callout", "-", "pointer", ")", ",", "fit=", "(", "c", "-", "7", ")(", "c", "-", "12u", ")", "]", "(", "left", "-", "callout", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "c", "-", "12", ",", "node", "distance=1.5cm*1.2", ",", "yshift=1.2", "cm", ",", "label", "=", "below", ":", "]", "(", "ci", "-", "1", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "ci", "-", "1", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "]", "(", "ci", "-", "2", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "ci", "-", "2", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "…", "]", "(", "ci", "-", "3", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "ci", "-", "3", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "]", "(", "ci", "-", "4", ")", ";", "[", "dots", "hbox", ",", "right", "of", "=", "ci", "-", "4", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "…", "]", "(", "ci", "-", "5", ")", "…", ";", "[", "blue", "hbox", ",", "right", "of", "=", "ci", "-", "5", ",", "node", "distance=0.7", "cm", ",", "label", "=", "below", ":", "]", "(", "ci", "-", "6", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "9pt", ",", "shorten", "¡", "=", "5pt", "]", "(", "image", ")", "–", "(", "ci", "-", "3", ")", "node", "[", "pos=0.35", ",", "right", ",", "font=", "]", "VGGnet", ";", "[", "inner", "sep=0em", "]", "(", "i", "-", "callout", "-", "pointer", ")", "at", "(", "-(-", "⁢ci1", ")(", "⁢3cm", ",-", "⁢1", "cm", ")", ")", ";", "container", ",", "callout", "absolute", "pointer=", "(", "i", "-", "callout", "-", "pointer", ")", ",", "fit=", "(", "ci", "-", "1", ")(", "ci", "-", "6", ")", ",", "inner", "sep=0.26em", "]", "(", "left", "-", "callout", ")", ";", "[", "blue", "hbox", ",", "above", "of", "=", "c", "-", "1u", ",", "node", "distance=1.5cm*2.9", "]", "(", "m", "-", "1", ")", ";", "hbox", ",", "right", "of", "=", "m", "-", "1", ",", "node", "distance=0.5", "cm", "]", "(", "m", "-", "2", ")", ";", "hbox", ",", "right", "of", "=", "m", "-", "2", ",", "node", "distance=0.5", "cm", ",", "]", "(", "m", "-", "3", ")", "…", ";", "hbox", ",", "right", "of", "=", "m", "-", "3", ",", "node", "distance=0.5", "cm", "]", "(", "m", "-", "4", ")", ";", "hbox", ",", "right", "of", "=", "m", "-", "4", ",", "node", "distance=0.5", "cm", "]", "(", "m", "-", "5", ")", "…", ";", "hbox", ",", "right", "of", "=", "m", "-", "5", ",", "node", "distance=0.5", "cm", ",", "]", "(", "m", "-", "6", ")", ";", "(", "[", "xshift=0pt", ",", "yshift=", "-", "0.2cm", "]", "m", "-", "6.south", "east", ")", "–", "(", "[", "xshift=0pt", ",", "yshift=", "-", "0.2cm", "]", "m", "-", "1.south", "west", ")", "node", "[", "black", ",", "midway", ",", "yshift=", "-", "0.35", "cm", "]", "P", "vs", "H", ";", "[", "blue", "hbox", ",", "right", "of", "=", "m", "-", "6", ",", "node", "distance=0.9cm*1.5", "]", "(", "m", "-", "7", ")", ";", "hbox", ",", "right", "of", "=", "m", "-", "7", ",", "node", "distance=0.5", "cm", "]", "(", "m", "-", "8", ")", ";", "hbox", ",", "right", "of", "=", "m", "-", "8", ",", "node", "distance=0.5", "cm", "]", "(", "m", "-", "9", ")", "…", ";", "hbox", ",", "right", "of", "=", "m", "-", "9", ",", "node", "distance=0.5", "cm", "]", "(", "m", "-", "10", ")", ";", "hbox", ",", "right", "of", "=", "m", "-", "10", ",", "node", "distance=0.5", "cm", "]", "(", "m", "-", "11", ")", "…", ";", "hbox", ",", "right", "of", "=", "m", "-", "11", ",", "node", "distance=0.5", "cm", "]", "(", "m", "-", "12", ")", ";", "(", "[", "xshift=0pt", ",", "yshift=", "-", "0.2cm", "]", "m", "-", "12.south", "east", ")", "–", "(", "[", "xshift=0pt", ",", "yshift=", "-", "0.2cm", "]", "m", "-", "7.south", "west", ")", "node", "[", "black", ",", "midway", ",", "yshift=", "-", "0.35", "cm", "]", "H", "vs", "P", ";", "[", "blue", "hbox", ",", "right", "of", "=", "m", "-", "12", ",", "node", "distance=0.9cm*1.5", "]", "(", "mi", "-", "13", ")", ";", "hbox", ",", "right", "of", "=", "mi", "-", "13", ",", "node", "distance=0.5", "cm", "]", "(", "mi", "-", "14", ")", ";", "hbox", ",", "right", "of", "=", "mi", "-", "14", ",", "node", "distance=0.5", "cm", "]", "(", "mi", "-", "15", ")", "…", ";", "hbox", ",", "right", "of", "=", "mi", "-", "15", ",", "node", "distance=0.5", "cm", "]", "(", "mi", "-", "16", ")", ";", "hbox", ",", "right", "of", "=", "mi", "-", "16", ",", "node", "distance=0.5", "cm", "]", "(", "mi", "-", "17", ")", "…", ";", "hbox", ",", "right", "of", "=", "mi", "-", "17", ",", "node", "distance=0.5", "cm", "]", "(", "mi", "-", "18", ")", ";", "(", "[", "xshift=0pt", ",", "yshift=", "-", "0.2cm", "]", "mi", "-", "18.south", "east", ")", "–", "(", "[", "xshift=0pt", ",", "yshift=", "-", "0.2cm", "]", "mi", "-", "13.south", "west", ")", "node", "[", "black", ",", "midway", ",", "yshift=", "-", "0.35", "cm", "]", "H", "vs", "Image", ";", "[", "blue", "hbox", ",", "right", "of", "=", "mi", "-", "18", ",", "node", "distance=0.9cm*1.5", "]", "(", "mi", "-", "19", ")", ";", "hbox", ",", "right", "of", "=", "mi", "-", "19", ",", "node", "distance=0.5", "cm", "]", "(", "mi", "-", "20", ")", ";", "hbox", ",", "right", "of", "=", "mi", "-", "20", ",", "node", "distance=0.5", "cm", "]", "(", "mi", "-", "21", ")", "…", ";", "hbox", ",", "right", "of", "=", "mi", "-", "21", ",", "node", "distance=0.5", "cm", "]", "(", "mi", "-", "22", ")", ";", "hbox", ",", "right", "of", "=", "mi", "-", "22", ",", "node", "distance=0.5", "cm", "]", "(", "mi", "-", "23", ")", "…", ";", "hbox", ",", "right", "of", "=", "mi", "-", "23", ",", "node", "distance=0.5", "cm", "]", "(", "mi", "-", "24", ")", ";", "(", "[", "xshift=0pt", ",", "yshift=", "-", "0.2cm", "]", "mi", "-", "24.south", "east", ")", "–", "(", "[", "xshift=0pt", ",", "yshift=", "-", "0.2cm", "]", "mi", "-", "19.south", "west", ")", "node", "[", "black", ",", "midway", ",", "yshift=", "-", "0.35", "cm", "]", "Image", "vs", "H", ";", "[", "red", "matching", ",", "below", "of", "=", "m", "-", "1", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "1", ")", "⊗", ";", "matching", ",", "below", "of", "=", "m", "-", "2", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "2", ")", "⊗", ";", "matching", ",", "below", "of", "=", "m", "-", "4", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "3", ")", "⊗", ";", "matching", ",", "below", "of", "=", "m", "-", "6", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "4", ")", "⊗", ";", "[", "red", "matching", ",", "below", "of", "=", "m", "-", "7", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "5", ")", "⊗", ";", "matching", ",", "below", "of", "=", "m", "-", "8", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "6", ")", "⊗", ";", "matching", ",", "below", "of", "=", "m", "-", "10", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "7", ")", "⊗", ";", "matching", ",", "below", "of", "=", "m", "-", "12", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "8", ")", "⊗", ";", "[", "black", "matching", ",", "below", "of", "=", "mi", "-", "13", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "9", ")", "⊗", ";", "matching", ",", "below", "of", "=", "mi", "-", "14", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "10", ")", "⊗", ";", "matching", ",", "below", "of", "=", "mi", "-", "16", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "11", ")", "⊗", ";", "matching", ",", "below", "of", "=", "mi", "-", "18", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "12", ")", "⊗", ";", "[", "black", "matching", ",", "below", "of", "=", "mi", "-", "19", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "13", ")", "⊗", ";", "matching", ",", "below", "of", "=", "mi", "-", "20", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "14", ")", "⊗", ";", "matching", ",", "below", "of", "=", "mi", "-", "22", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "15", ")", "⊗", ";", "matching", ",", "below", "of", "=", "mi", "-", "24", ",", "node", "distance=1.5cm*1.5", "]", "(", "mp", "-", "16", ")", "⊗", ";", "mp", "-", "1", ")", ";", "h", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=165", ",", "␣", "in=", "-", "70", "]", "mp", "-", "2", ")", ";", "h", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=160", ",", "␣", "in=", "-", "75", "]", "mp", "-", "3", ")", ";", "h", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=155", ",", "␣", "in=", "-", "80", "]", "mp", "-", "4", ")", ";", "mp", "-", "5", ")", ";", "p", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=110", ",", "␣", "in=", "-", "100", "]", "mp", "-", "6", ")", ";", "p", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=65", ",", "␣", "in=", "-", "90", "]", "mp", "-", "7", ")", ";", "p", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=60", ",", "␣", "in=", "-", "90", "]", "mp", "-", "8", ")", ";", "mp", "-", "13", ")", ";", "h", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=20", ",", "␣", "in=", "-", "100", "]", "mp", "-", "14", ")", ";", "h", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=15", ",", "␣", "in=", "-", "105", "]", "mp", "-", "15", ")", ";", "h", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=10", ",", "␣", "in=", "-", "110", "]", "mp", "-", "16", ")", ";", "mp", "-", "9", ")", ";", "i", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=145", ",", "␣", "in=", "-", "60", "]", "mp", "-", "10", ")", ";", "i", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=140", ",", "␣", "in=", "-", "105", "]", "mp", "-", "11", ")", ";", "i", "-", "callout", "-", "pointer", ")", "␣", "edge", "[-", ">", ",", "␣", "out=65", ",", "␣", "in=", "-", "60", "]", "mp", "-", "12", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "1", ")", "–", "(", "m", "-", "1", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "2", ")", "–", "(", "m", "-", "2", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "3", ")", "–", "(", "m", "-", "4", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "4", ")", "–", "(", "m", "-", "6", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "5", ")", "–", "(", "m", "-", "7", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "6", ")", "–", "(", "m", "-", "8", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "7", ")", "–", "(", "m", "-", "10", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "8", ")", "–", "(", "m", "-", "12", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "9", ")", "–", "(", "mi", "-", "13", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "10", ")", "–", "(", "mi", "-", "14", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "11", ")", "–", "(", "mi", "-", "16", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "12", ")", "–", "(", "mi", "-", "18", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "13", ")", "–", "(", "mi", "-", "19", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "14", ")", "–", "(", "mi", "-", "20", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "15", ")", "–", "(", "mi", "-", "22", ")", ";", "arrow", ",", "shorten", "¿", "=", "18pt", "]", "(", "mp", "-", "16", ")", "–", "(", "mi", "-", "24", ")", ";", "[", "match", "arrow", ",", "shorten", "¡", "=", "6pt", "]", "(", "c", "-", "1u", ")", "–", "(", "mp", "-", "1", ")", ";", "arrow", ",", "shorten", "¡", "=", "6pt", "]", "(", "c", "-", "2u", ")", "–", "(", "mp", "-", "2", ")", ";", "arrow", ",", "shorten", "¡", "=", "7pt", "]", "(", "c", "-", "4u", ")", "to", "[", "out=80", ",", "in=", "-", "45", "]", "(", "mp", "-", "3", ")", ";", "arrow", ",", "shorten", "¡", "=", "6pt", "]", "(", "c", "-", "6u", ")", "to", "[", "out=100", ",", "in=", "-", "45", "]", "(", "mp", "-", "4", ")", ";", "[", "match", "arrow", ",", "shorten", "¡", "=", "9pt", "]", "(", "c", "-", "7u", ")", "to", "[", "out=110", ",", "in=", "-", "40", "]", "(", "mp", "-", "5", ")", ";", "arrow", ",", "shorten", "¡", "=", "7pt", "]", "(", "c", "-", "8u", ")", "to", "[", "out=100", ",", "in=", "-", "35", "]", "(", "mp", "-", "6", ")", ";", "arrow", ",", "shorten", "¡", "=", "20pt", "]", "(", "c", "-", "10u", ")", "–", "(", "mp", "-", "7", ")", ";", "arrow", ",", "shorten", "¡", "=", "22pt", "]", "(", "c", "-", "12u", ")", "–", "(", "mp", "-", "8", ")", ";", "[", "match", "arrow", ",", "shorten", "¡", "=", "22pt", "]", "(", "c", "-", "7u", ")", "–", "(", "mp", "-", "9", ")", ";", "arrow", ",", "shorten", "¡", "=", "20pt", "]", "(", "c", "-", "8u", ")", "–", "(", "mp", "-", "10", ")", ";", "arrow", ",", "shorten", "¡", "=", "7pt", "]", "(", "c", "-", "10u", ")", "to", "[", "out=80", ",", "in=220", "]", "(", "mp", "-", "11", ")", ";", "arrow", ",", "shorten", "¡", "=", "7pt", "]", "(", "c", "-", "12u", ")", "to", "[", "out=80", ",", "in=230", "]", "(", "mp", "-", "12", ")", ";", "[", "match", "arrow", ",", "shorten", "¡", "=", "3pt", "]", "(", "ci", "-", "1", ")", "to", "[", "out=80", ",", "in=220", "]", "(", "mp", "-", "13", ")", ";", "arrow", ",", "shorten", "¡", "=", "3pt", "]", "(", "ci", "-", "2", ")", "to", "[", "out=90", ",", "in=220", "]", "(", "mp", "-", "14", ")", ";", "arrow", ",", "shorten", "¡", "=", "10pt", "]", "(", "ci", "-", "4", ")", "to", "[", "out=60", ",", "in=", "-", "70", "]", "(", "mp", "-", "15", ")", ";", "arrow", ",", "shorten", "¡", "=", "3pt", "]", "(", "ci", "-", "6", ")", "–", "(", "mp", "-", "16", ")", ";", "[", "blue", "hbox", ",", "above", "of", "=", "m", "-", "1", ",", "node", "distance=1.5cm*1.2", "]", "(", "a", "-", "1", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "1", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "2", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "2", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "3", ")", "…", ";", "hbox", ",", "right", "of", "=", "a", "-", "3", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "4", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "4", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "5", ")", "…", ";", "hbox", ",", "right", "of", "=", "a", "-", "5", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "6", ")", ";", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "a", "-", "1", ")", "–", "(", "a", "-", "2", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "a", "-", "2", ")", "–", "(", "a", "-", "3", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "a", "-", "3", ")", "–", "(", "a", "-", "4", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "a", "-", "4", ")", "–", "(", "a", "-", "5", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "a", "-", "5", ")", "–", "(", "a", "-", "6", ")", ";", "[", "green", "hbox", ",", "above", "of", "=", "a", "-", "1", ",", "node", "distance=1.5cm", "/", "1.2", "]", "(", "a", "-", "1u", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "1u", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "2u", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "2u", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "3u", ")", "…", ";", "hbox", ",", "right", "of", "=", "a", "-", "3u", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "4u", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "4u", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "5u", ")", "…", ";", "hbox", ",", "right", "of", "=", "a", "-", "5u", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "6u", ")", ";", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "a", "-", "2u", ")", "–", "(", "a", "-", "1u", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "a", "-", "3u", ")", "–", "(", "a", "-", "2u", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "a", "-", "4u", ")", "–", "(", "a", "-", "3u", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "a", "-", "5u", ")", "–", "(", "a", "-", "4u", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "a", "-", "6u", ")", "–", "(", "a", "-", "5u", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "a", "-", "6", ",", "node", "distance=0.9cm*1.5", "]", "(", "a", "-", "7", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "7", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "8", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "8", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "9", ")", "…", ";", "hbox", ",", "right", "of", "=", "a", "-", "9", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "10", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "10", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "11", ")", "…", ";", "hbox", ",", "right", "of", "=", "a", "-", "11", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "12", ")", ";", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "a", "-", "7", ")", "–", "(", "a", "-", "8", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "a", "-", "8", ")", "–", "(", "a", "-", "9", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "a", "-", "9", ")", "–", "(", "a", "-", "10", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "a", "-", "10", ")", "–", "(", "a", "-", "11", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "a", "-", "11", ")", "–", "(", "a", "-", "12", ")", ";", "[", "green", "hbox", ",", "right", "of", "=", "a", "-", "6u", ",", "node", "distance=0.9cm*1.5", "]", "(", "a", "-", "7u", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "7u", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "8u", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "8u", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "9u", ")", "…", ";", "hbox", ",", "right", "of", "=", "a", "-", "9u", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "10u", ")", ";", "hbox", ",", "right", "of", "=", "a", "-", "10u", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "11u", ")", "…", ";", "hbox", ",", "right", "of", "=", "a", "-", "11u", ",", "node", "distance=0.5", "cm", "]", "(", "a", "-", "12u", ")", ";", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "a", "-", "8u", ")", "–", "(", "a", "-", "7u", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "a", "-", "9u", ")", "–", "(", "a", "-", "8u", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "a", "-", "10u", ")", "–", "(", "a", "-", "9u", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "a", "-", "11u", ")", "–", "(", "a", "-", "10u", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "a", "-", "12u", ")", "–", "(", "a", "-", "11u", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "a", "-", "12", ",", "node", "distance=0.9cm*1.5", "]", "(", "ai", "-", "13", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "13", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "14", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "14", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "15", ")", "…", ";", "hbox", ",", "right", "of", "=", "ai", "-", "15", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "16", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "16", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "17", ")", "…", ";", "hbox", ",", "right", "of", "=", "ai", "-", "17", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "18", ")", ";", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "ai", "-", "13", ")", "–", "(", "ai", "-", "14", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "ai", "-", "14", ")", "–", "(", "ai", "-", "15", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "ai", "-", "15", ")", "–", "(", "ai", "-", "16", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "ai", "-", "16", ")", "–", "(", "ai", "-", "17", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "ai", "-", "17", ")", "–", "(", "ai", "-", "18", ")", ";", "[", "green", "hbox", ",", "right", "of", "=", "a", "-", "12u", ",", "node", "distance=0.9cm*1.5", "]", "(", "ai", "-", "13u", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "13u", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "14u", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "14u", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "15u", ")", "…", ";", "hbox", ",", "right", "of", "=", "ai", "-", "15u", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "16u", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "16u", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "17u", ")", "…", ";", "hbox", ",", "right", "of", "=", "ai", "-", "17u", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "18u", ")", ";", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "ai", "-", "14u", ")", "–", "(", "ai", "-", "13u", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "ai", "-", "15u", ")", "–", "(", "ai", "-", "14u", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "ai", "-", "16u", ")", "–", "(", "ai", "-", "15u", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "ai", "-", "17u", ")", "–", "(", "ai", "-", "16u", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "ai", "-", "18u", ")", "–", "(", "ai", "-", "17u", ")", ";", "[", "blue", "hbox", ",", "right", "of", "=", "ai", "-", "18", ",", "node", "distance=0.9cm*1.5", "]", "(", "ai", "-", "19", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "19", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "20", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "20", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "21", ")", "…", ";", "hbox", ",", "right", "of", "=", "ai", "-", "21", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "22", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "22", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "23", ")", "…", ";", "hbox", ",", "right", "of", "=", "ai", "-", "23", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "24", ")", ";", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "ai", "-", "19", ")", "–", "(", "ai", "-", "20", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "ai", "-", "20", ")", "–", "(", "ai", "-", "21", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "ai", "-", "21", ")", "–", "(", "ai", "-", "22", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "ai", "-", "22", ")", "–", "(", "ai", "-", "23", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "ai", "-", "23", ")", "–", "(", "ai", "-", "24", ")", ";", "[", "green", "hbox", ",", "right", "of", "=", "ai", "-", "18u", ",", "node", "distance=0.9cm*1.5", "]", "(", "ai", "-", "19u", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "19u", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "20u", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "20u", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "21u", ")", "…", ";", "hbox", ",", "right", "of", "=", "ai", "-", "21u", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "22u", ")", ";", "hbox", ",", "right", "of", "=", "ai", "-", "22u", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "23u", ")", "…", ";", "hbox", ",", "right", "of", "=", "ai", "-", "23u", ",", "node", "distance=0.5", "cm", "]", "(", "ai", "-", "24u", ")", ";", "arrow", ",", "shorten", "¿", "=", "1pt", ",", "shorten", "¡", "=", "1pt", "]", "(", "ai", "-", "20u", ")", "–", "(", "ai", "-", "19u", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "ai", "-", "21u", ")", "–", "(", "ai", "-", "20u", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "ai", "-", "22u", ")", "–", "(", "ai", "-", "21u", ")", ";", "arrow", ",", "shorten", "¿", "=", "0pt", ",", "shorten", "¡", "=", "-", "2pt", "]", "(", "ai", "-", "23u", ")", "–", "(", "ai", "-", "22u", ")", ";", "arrow", ",", "shorten", "¿", "=", "-", "2pt", ",", "shorten", "¡", "=", "0pt", "]", "(", "ai", "-", "24u", ")", "–", "(", "ai", "-", "23u", ")", ";", "arrow", ",", "shorten", "¿", "=", "6pt", ",", "shorten", "¡", "=", "-", "5pt", "]", "(", "m", "-", "3", ")", "–", "(", "a", "-", "3", ")", ";", "arrow", ",", "shorten", "¿", "=", "6pt", ",", "shorten", "¡", "=", "-", "5pt", "]", "(", "m", "-", "9", ")", "–", "(", "a", "-", "9", ")", ";", "arrow", ",", "shorten", "¿", "=", "6pt", ",", "shorten", "¡", "=", "-", "5pt", "]", "(", "mi", "-", "15", ")", "–", "(", "ai", "-", "15", ")", ";", "arrow", ",", "shorten", "¿", "=", "6pt", ",", "shorten", "¡", "=", "-", "5pt", "]", "(", "mi", "-", "21", ")", "–", "(", "ai", "-", "21", ")", ";", "[", "green", "vbox", ",", "above", "of", "=", "a", "-", "3u", ",", "node", "distance=1.5cm*1.2", ",", "xshift=", "-", "0.3", "cm", "]", "(", "ap", "-", "1", ")", ";", "vbox", ",", "right", "of", "=", "ap", "-", "1", ",", "node", "distance=0.9cm*2", "]", "(", "ap", "-", "2", ")", ";", "vbox", ",", "right", "of", "=", "ap", "-", "2", ",", "node", "distance=0.9cm*2", "]", "(", "ap", "-", "3", ")", ";", "vbox", ",", "right", "of", "=", "ap", "-", "3", ",", "node", "distance=0.9cm*2", "]", "(", "ap", "-", "4", ")", ";", "vbox", ",", "right", "of", "=", "ap", "-", "4", ",", "node", "distance=0.9cm*2", "]", "(", "ap", "-", "5", ")", ";", "vbox", ",", "right", "of", "=", "ap", "-", "5", ",", "node", "distance=0.9cm*2", "]", "(", "ap", "-", "6", ")", ";", "vbox", ",", "right", "of", "=", "ap", "-", "6", ",", "node", "distance=0.9cm*2", "]", "(", "ap", "-", "7", ")", ";", "vbox", ",", "right", "of", "=", "ap", "-", "7", ",", "node", "distance=0.9cm*2", "]", "(", "ap", "-", "8", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "3pt", ",", "shorten", "¡", "=", "3pt", "]", "(", "a", "-", "1u", ")", "–", "(", "ap", "-", "1", ")", ";", "arrow", ",", "shorten", "¿", "=", "3pt", ",", "shorten", "¡", "=", "3pt", "]", "(", "a", "-", "6.east", ")", "–", "(", "[", "xshift=", "-", "1", "cm", "]", "ap", "-", "2.south", "east", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "3pt", ",", "shorten", "¡", "=", "3pt", "]", "(", "a", "-", "7u", ")", "–", "(", "ap", "-", "3", ")", ";", "arrow", ",", "shorten", "¿", "=", "3pt", ",", "shorten", "¡", "=", "3pt", "]", "(", "a", "-", "12.west", ")", "–", "(", "[", "xshift=", "-", "0.5cm", "]", "ap", "-", "4.south", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "3pt", ",", "shorten", "¡", "=", "3pt", "]", "(", "ai", "-", "13u", ")", "–", "(", "ap", "-", "5", ")", ";", "arrow", ",", "shorten", "¿", "=", "3pt", ",", "shorten", "¡", "=", "3pt", "]", "(", "ai", "-", "18.west", ")", "–", "(", "[", "xshift=", "-", "1cm", "]", "ap", "-", "6", ")", ";", "[", "red", "arrow", ",", "shorten", "¿", "=", "3pt", ",", "shorten", "¡", "=", "3pt", "]", "(", "ai", "-", "19u", ")", "–", "(", "ap", "-", "7", ")", ";", "arrow", ",", "shorten", "¿", "=", "3pt", ",", "shorten", "¡", "=", "3pt", "]", "(", "ai", "-", "24.west", ")", "–", "(", "ap", "-", "8", ")", ";", "[", "tbox", ",", "above", "of", "=", "ap", "-", "4", ",", "node", "distance=0.9cm*1.6", ",", "xshift=1", "cm", "]", "(", "softmax", ")", "softmax", ";", "of", "=", "softmax", ",", "node", "distance=0.9", "cm", "]", "(", "output", ")", "P", "(", "y|premise", ",", "image", ",", "hypothesis", ")", ";", "arrow", ",", "shorten", "¿", "=", "3pt", ",", "shorten", "¡", "=", "3pt", "]", "(", "[", "xshift=1.5cm", "]", "ap", "-", "4.north", "east", ")", "–", "(", "softmax", ")", ";", "[", "decorate", ",", "thick", ",", "decoration", "=", "brace", ",", "amplitude=6pt", ",", "yshift=0pt", "]", "(", "[", "yshift=", "-", "65pt", ",", "xshift=", "-", "50pt", "]", "w", "-", "1.south", "west", ")", "–", "(", "[", "xshift=", "-", "50pt", ",", "yshift=", "-", "125pt", "]", "c", "-", "1.south", "west", ")", "node", "[", "black", ",", "midway", ",", "xshift=", "-", "1", "cm", ",", "yshift=", "-", "0.2", "cm", ",", "rotate=90", ",", "anchor", "=", "north", "]", "Embedding", "layer", ";", "[", "decorate", ",", "thick", ",", "decoration", "=", "brace", ",", "amplitude=6pt", ",", "yshift=0pt", "]", "(", "[", "yshift=", "-", "5pt", ",", "xshift=", "-", "50pt", "]", "c", "-", "1.south", "west", ")", "–", "(", "[", "xshift=", "-", "50pt", ",", "yshift=25pt", "]", "c", "-", "1u.north", "west", ")", "node", "[", "black", ",", "midway", ",", "xshift=", "-", "1", "cm", ",", "rotate=90", ",", "anchor", "=", "north", "]", "Context", "layer", ";", "[", "decorate", ",", "thick", ",", "decoration", "=", "brace", ",", "amplitude=6pt", ",", "yshift=0pt", "]", "(", "[", "yshift=", "-", "95pt", ",", "xshift=", "-", "50pt", "]", "mp", "-", "1.south", "west", ")", "–", "(", "[", "xshift=", "-", "50pt", ",", "yshift=", "-", "45pt", "]", "m", "-", "1.north", "west", ")", "node", "[", "black", ",", "midway", ",", "xshift=", "-", "1", "cm", ",", "rotate=90", ",", "anchor", "=", "north", "]", "Matching", "layer", ";", "[", "decorate", ",", "thick", ",", "decoration", "=", "brace", ",", "amplitude=6pt", ",", "yshift=0pt", "]", "(", "[", "xshift=", "-", "50pt", "]", "a", "-", "1.south", "west", ")", "–", "(", "[", "xshift=", "-", "50pt", ",", "yshift=1.5cm*3", "]", "a", "-", "1u.north", "west", ")", "node", "[", "black", ",", "midway", ",", "xshift=", "-", "1", "cm", ",", "rotate=90", ",", "anchor", "=", "north", "]", "Aggregation", "layer", ";", "[", "decorate", ",", "thick", ",", "decoration", "=", "brace", ",", "amplitude=6pt", ",", "yshift=0pt", "]", "(", "[", "xshift=", "-", "50pt", ",", "yshift=1.5cm*5.5", "]", "a", "-", "1u.north", "west", ")", "–", "(", "[", "xshift=", "-", "50pt", ",", "yshift=1.5cm*12.5", "]", "a", "-", "1u.north", "west", ")", "node", "[", "black", ",", "midway", ",", "xshift=", "-", "1", "cm", ",", "rotate=90", ",", "anchor", "=", "north", "]", "Prediction", "layer", ";", "Here", ",", "we", "report", "some", "further", "details", "of", "our", "implementation", "of", "the", "V", "-", "BiMPM", "model", "described", "in", "Section", "4", "of", "the", "main", "paper", ",", "based", "on", "the", "work", "of", "Wang2017", ".", "Our", "model", "is", "displayed", "in", "Figure", "[", "reference", "]", ".", "The", "core", "part", "of", "the", "original", "BiMPM", "is", "the", "matching", "layer", ".", "Given", "two", "-", "dimensional", "vectors", "and", ",", "each", "replicated", "times", "(", "is", "the", "number", "of", "‘", "perspectives", "’", ")", "and", "a", "trainable", "weight", "matrix", ",", "matching", "involves", "a", "cosine", "similarity", "computation", "that", "yields", "an", "-", "dimensional", "matching", "vector", ",", "whose", "elements", "are", "defined", "as", "follows", ":", "The", "matching", "operations", "included", "are", "the", "following", ":", "full", "-", "matching", ",", "where", "each", "forward", "or", "backward", "contextual", "embedding", "of", "the", "premise", "P", "(", "resp", ".", "the", "hypothesis", "H", ")", "is", "matched", "to", "the", "last", "time", "-", "step", "of", "H", "(", "resp", ".", "P", ")", ";", "max", "-", "pooling", ",", "where", "each", "forward", "/", "backward", "contextual", "embedding", "of", "one", "sentence", "is", "compared", "to", "the", "embeddings", "of", "the", "other", ",", "retaining", "the", "maximum", "value", "for", "each", "dimension", ";", "attentive", "matching", ",", "where", "first", ",", "the", "pairwise", "cosine", "similarity", "between", "forward", "/", "backward", "embeddings", "of", "P", "and", "H", "is", "estimated", ",", "before", "calculating", "an", "attentive", "vector", "over", "the", "weighted", "sum", "of", "contextual", "embeddings", "for", "H", "and", "matching", "each", "forward", "/", "backward", "embedding", "of", "P", "against", "the", "attentive", "vector", ";", "max", "-", "attentive", "matching", ",", "a", "version", "of", "attentive", "matching", "where", "the", "contextual", "embedding", "with", "the", "highest", "cosine", "is", "used", "as", "the", "attentive", "vector", ",", "instead", "of", "the", "weighted", "sum", ".", "The", "visually", "-", "augmented", "version", "of", "the", "original", "model", ",", "V", "-", "BiMPM", ",", "is", "displayed", "in", "Figure", "[", "reference", "]", ".", "To", "perform", "multimodal", "matching", ",", "the", "visual", "and", "textual", "vectors", "are", "mapped", "to", "a", "mutual", "space", "using", "the", "following", "affine", "transformation", ":", "where", ",", ",", ",", "and", "are", "the", "weight", "matrix", ",", "the", "bias", ",", "the", "input", "features", "and", "output", "features", ",", "respectively", ",", "and", "is", "any", "text", "(", "P", "or", "H", ")", ".", "Given", "weight", "matrices", "for", "text", "and", "for", "images", ",", "we", "compute", "the", "matching", "vector", "between", "a", "textual", "vector", "and", "image", "vector", "as", ":", "bibliography", ":", "References"], "entity": [[2, 5, ["Grounded", "Textual", "Entailment"], "Task"], [5, 10, ["Capturing", "semantic", "relations", "between", "sentences"], "Task"], [22, 24, ["computational", "semantics"], "Task"], [25, 29, ["Logic", "-", "based", "models"], "Method"], [64, 66, ["Statistical", "models"], "Method"], [101, 105, ["visually", "-", "grounded", "version"], "Method"], [107, 110, ["Textual", "Entailment", "task"], "Task"], [151, 153, ["multimodal", "version"], "Method"], [164, 168, ["visually", "-", "augmented", "models"], "Method"], [169, 171, ["textual", "entailment"], "Task"], [187, 190, ["depth", "error", "analysis"], "Task"], [194, 196, ["multimodal", "models"], "Method"], [281, 284, ["computational", "linguistics", "community"], "Task"], [311, 313, ["statistical", "approaches"], "Method"], [314, 316, ["computational", "semantics"], "Task"], [328, 336, ["Recognizing", "Textual", "Entailment", "(", "RTE", ")", "shared", "tasks"], "Task"], [344, 347, ["compositional", "distributional", "models"], "Method"], [369, 371, ["Neural", "Networks"], "Method"], [372, 373, ["NNs"], "Method"], [383, 386, ["natural", "language", "inference"], "Task"], [391, 392, ["NNs"], "Method"], [425, 427, ["meaning", "representations"], "Task"], [442, 444, ["computational", "linguists"], "Task"], [462, 467, ["checking", "Semantic", "Text", "Similarity", "Text"], "Task"], [481, 484, ["Grounded", "Textual", "Entailment"], "Task"], [485, 486, ["GTE"], "Task"], [652, 654, ["textual", "entailment"], "Task"], [665, 667, ["GTE", "task"], "Task"], [734, 736, ["GTE", "model"], "Method"], [741, 744, ["sub", "-", "tasks"], "Task"], [753, 755, ["linguistic", "representations"], "Method"], [796, 798, ["Grounding", "language"], "Task"], [810, 812, ["Image", "Captioning"], "Task"], [813, 814, ["IC"], "Task"], [818, 821, ["Visual", "Question", "Answering"], "Task"], [822, 823, ["VQA"], "Task"], [833, 835, ["Visual", "Reasoning"], "Task"], [836, 838, ["Visual", "Dialog"], "Task"], [843, 846, ["Grounded", "Textual", "Entailment"], "Task"], [847, 848, ["GTE"], "Task"], [854, 855, ["TE"], "Task"], [859, 860, ["GTE"], "Method"], [868, 870, ["Textual", "Entailment"], "Task"], [883, 886, ["Computational", "Semantics", "models"], "Method"], [955, 957, ["formal", "semanticists"], "Method"], [958, 959, ["FraCaS"], "Method"], [962, 964, ["Computational", "Semantics"], "Task"], [969, 972, ["logical", "entailment", "problems"], "Task"], [1062, 1064, ["PASCAL", "RTE"], "Task"], [1065, 1068, ["Recognizing", "Textual", "Entailment"], "Task"], [1084, 1088, ["RTE", "-", "4", "committee"], "Method"], [1098, 1099, ["classification"], "Task"], [1120, 1121, ["FraCaS"], "Method"], [1133, 1135, ["entailment", "problems"], "Task"], [1219, 1223, ["Compositional", "Distributional", "Semantics", "Models"], "Method"], [1245, 1248, ["shared", "entailment", "task"], "Task"], [1267, 1269, ["crowdsourcing", "platform"], "Method"], [1330, 1331, ["crowdsourcing"], "Method"], [1420, 1423, ["RTE", "shared", "tasks"], "Task"], [1431, 1435, ["data", "-", "driven", "approaches"], "Method"], [1461, 1463, ["neural", "approaches"], "Method"], [1484, 1487, ["bidirectional", "LSTM", "models"], "Method"], [1494, 1495, ["RTE"], "Method"], [1502, 1504, ["inference", "problems"], "Task"], [1506, 1509, ["semantic", "text", "similarity"], "Task"], [1510, 1512, ["Question", "Answering"], "Task"], [1545, 1546, ["RNN"], "Method"], [1568, 1570, ["baseline", "LSTM"], "Method"], [1582, 1583, ["snli"], "Method"], [1614, 1617, ["matching", "and", "aggregation"], "Method"], [1633, 1635, ["aggregate", "representation"], "Method"], [1647, 1649, ["neural", "approaches"], "Method"], [1653, 1655, ["sentence", "pairs"], "Task"], [1657, 1659, ["entailment", "relationship"], "Task"], [1660, 1661, ["generation"], "Task"], [1664, 1665, ["RNN"], "Method"], [1686, 1687, ["Vision"], "Task"], [1688, 1690, ["Textual", "Entailment"], "Task"], [1713, 1717, ["element", "-", "wise", "multiplication"], "Method"], [1738, 1740, ["multimodal", "tasks"], "Task"], [1744, 1745, ["IC"], "Task"], [1746, 1747, ["VQA"], "Task"], [1748, 1750, ["Visual", "Reasoning"], "Task"], [1751, 1754, ["Visual", "Dialogue", "tasks"], "Task"], [1764, 1766, ["grounding", "entailment"], "Task"], [1798, 1801, ["visual", "denotation", "graph"], "Method"], [1836, 1838, ["textual", "entailment"], "Task"], [1839, 1841, ["image", "captioning"], "Task"], [2003, 2006, ["natural", "language", "inference"], "Task"], [2043, 2044, ["classifier"], "Method"], [2047, 2048, ["accuracy"], "Metric"], [2078, 2080, ["simplistic", "classifier"], "Method"], [2137, 2141, ["annotation", "of", "linguistic", "phenomena"], "Task"], [2345, 2348, ["V", "-", "SNLI"], "Material"], [2349, 2353, ["V", "-", "SNLI", "test"], "Material"], [2360, 2363, ["V", "-", "SNLI"], "Material"], [2387, 2391, ["V", "-", "SNLI", "dataset"], "Material"], [2405, 2408, ["V", "-", "SNLI"], "Material"], [2470, 2472, ["visual", "counterpart"], "Method"], [2477, 2480, ["V", "-", "SNLI"], "Material"], [2484, 2487, ["V", "-", "SNLI"], "Material"], [2835, 2838, ["error", "analysis", "approach"], "Method"], [2867, 2869, ["RTE", "models"], "Method"], [2924, 2926, ["keyword", "search"], "Method"], [2961, 2963, ["manual", "annotation"], "Task"], [3078, 3081, ["V", "-", "SNLI"], "Material"], [3082, 3085, ["V", "-", "SNLI"], "Material"], [3088, 3090, ["baseline", "models"], "Method"], [3118, 3122, ["visually", "-", "augmented", "counterpart"], "Method"], [3144, 3146, ["LSTM", "baseline"], "Method"], [3153, 3156, ["Recurrent", "Neural", "Network"], "Method"], [3157, 3163, ["Long", "Short", "-", "Term", "Memory", "units"], "Method"], [3184, 3186, ["512D", "layers"], "Method"], [3188, 3191, ["ReLU", "activation", "function"], "Method"], [3195, 3197, ["softmax", "layer"], "Method"], [3218, 3220, ["LSTM", "baseline"], "Method"], [3235, 3238, ["pretrained", "GloVe", "embeddings"], "Method"], [3251, 3252, ["Dropout"], "Method"], [3266, 3270, ["ReLU", "fully", "connected", "layers"], "Method"], [3283, 3285, ["Adam", "optimizer"], "Method"], [3287, 3289, ["learning", "rate"], "Metric"], [3293, 3294, ["accuracy"], "Metric"], [3306, 3310, ["V", "-", "LSTM", "baseline"], "Material"], [3311, 3313, ["LSTM", "model"], "Method"], [3319, 3321, ["visual", "component"], "Method"], [3324, 3329, ["Visual", "Question", "Answering", "baseline", "model"], "Method"], [3342, 3343, ["LSTM"], "Method"], [3345, 3349, ["fully", "-", "connected", "layer"], "Method"], [3360, 3362, ["penultimate", "layer"], "Method"], [3364, 3368, ["VGGnet16", "Convolutional", "Neural", "Network"], "Method"], [3375, 3379, ["fully", "-", "connected", "layer"], "Method"], [3398, 3400, ["multimodal", "fusion"], "Task"], [3411, 3415, ["element", "-", "wise", "multiplication"], "Method"], [3417, 3422, ["vector", "of", "the", "text", "representation"], "Method"], [3431, 3433, ["multimodal", "fusion"], "Method"], [3449, 3451, ["multimodal", "representations"], "Method"], [3468, 3470, ["GloVe", "embeddings"], "Method"], [3491, 3495, ["attention", "-", "based", "model"], "Method"], [3496, 3497, ["IC"], "Task"], [3498, 3499, ["VQA"], "Task"], [3501, 3503, ["GTE", "task"], "Task"], [3510, 3513, ["V", "-", "LSTM"], "Material"], [3526, 3528, ["parameter", "tuning"], "Method"], [3546, 3547, ["BiMPM"], "Method"], [3548, 3553, ["Bilateral", "Multi", "-", "Perspective", "Matching"], "Method"], [3554, 3555, ["BiMPM"], "Method"], [3574, 3575, ["accuracy"], "Metric"], [3587, 3589, ["ensemble", "setup"], "Method"], [3592, 3594, ["embedding", "layer"], "Method"], [3602, 3604, ["GLoVe", "embeddings"], "Method"], [3610, 3613, ["context", "representation", "layer"], "Method"], [3616, 3618, ["bidirectional", "LSTMs"], "Method"], [3619, 3620, ["BiLSTMs"], "Method"], [3640, 3642, ["matching", "layer"], "Method"], [3686, 3687, ["BiLSTM"], "Method"], [3689, 3691, ["aggregation", "layer"], "Method"], [3702, 3703, ["BiLSTM"], "Method"], [3718, 3720, ["prediction", "tier"], "Method"], [3722, 3729, ["two", "-", "layer", "feed", "-", "forward", "network"], "Method"], [3738, 3739, ["softmax"], "Method"], [3740, 3741, ["Matching"], "Task"], [3745, 3747, ["cosine", "operation"], "Method"], [3767, 3769, ["matching", "strategies"], "Method"], [3786, 3788, ["matching", "strategies"], "Method"], [3808, 3809, ["BiMPM"], "Method"], [3827, 3830, ["V", "-", "BiMPM"], "Method"], [3846, 3850, ["fully", "-", "connected", "layer"], "Method"], [3852, 3855, ["VGGnet", "-", "16"], "Method"], [3874, 3876, ["matching", "operations"], "Method"], [3881, 3882, ["matching"], "Task"], [3917, 3919, ["affine", "transformation"], "Method"], [3928, 3930, ["cosine", "operation"], "Method"], [3966, 3969, ["V", "-)", "SNLI"], "Material"], [3971, 3974, ["V", "-)", "SNLI"], "Material"], [3977, 3981, ["visually", "-", "augmented", "models"], "Method"], [4047, 4050, ["blind", "”", "models"], "Method"], [4064, 4065, ["accuracy"], "Metric"], [4067, 4069, ["prediction", "task"], "Task"], [4083, 4085, ["blind", "models"], "Method"], [4129, 4132, ["V", "-", "SNLI"], "Material"], [4138, 4143, ["LSTM", "[", "H", "]", "model"], "Method"], [4150, 4151, ["LSTM"], "Method"], [4168, 4171, ["unbiased", "SNLI", "subset"], "Metric"], [4212, 4214, ["LSTM", "baseline"], "Method"], [4227, 4230, ["V", "-", "SNLI"], "Material"], [4236, 4239, ["V", "-", "SNLI"], "Material"], [4268, 4269, ["BiMPM"], "Method"], [4337, 4338, ["LSTM"], "Method"], [4365, 4369, ["V", "-", "LSTM", "baseline"], "Material"], [4381, 4384, ["V", "-", "BiMPM"], "Method"], [4416, 4418, ["error", "analysis"], "Method"], [4441, 4444, ["V", "-", "SNLI"], "Material"], [4447, 4449, ["Error", "analysis"], "Task"], [4459, 4460, ["accuracies"], "Metric"], [4462, 4466, ["blind", "and", "grounded", "version"], "Method"], [4467, 4468, ["BiMPM"], "Method"], [4500, 4502, ["Entity", "case"], "Task"], [4519, 4521, ["blind", "model"], "Method"], [4535, 4537, ["grounded", "model"], "Method"], [4562, 4566, ["grounding", "sub", "-", "task"], "Task"], [4569, 4571, ["Error", "analysis"], "Task"], [4572, 4573, ["grounding"], "Task"], [4590, 4593, ["V", "-", "BiMPM"], "Method"], [4598, 4600, ["blind", "case"], "Task"], [4622, 4625, ["V", "-", "BiMPM"], "Method"], [4631, 4633, ["blind", "model"], "Method"], [4748, 4751, ["V", "-", "BiMPM"], "Method"], [4808, 4810, ["visual", "grounding"], "Task"], [4817, 4820, ["V", "-", "BiMPM"], "Method"], [4852, 4853, ["cosine"], "Method"], [4860, 4863, ["V", "-", "SNLI"], "Material"], [4886, 4888, ["visual", "grounding"], "Task"], [4970, 4973, ["V", "-", "BiMPM"], "Method"], [5002, 5006, ["grounding", "sub", "-", "task"], "Task"], [5020, 5024, ["reasoning", "sub", "-", "task"], "Task"], [5216, 5219, ["V", "-", "BiMPM"], "Method"], [5220, 5223, ["V", "-", "LSTM"], "Material"], [5263, 5266, ["V", "-", "BiMPM"], "Method"], [5284, 5287, ["textual", "entailment", "task"], "Task"], [5295, 5299, ["Grounded", "Textual", "Entailment", "model"], "Method"], [5309, 5310, ["grounding"], "Task"], [5337, 5340, ["matching", "and", "aggregation"], "Method"], [5350, 5352, ["reasoning", "task"], "Task"], [5353, 5356, ["classifying", "entailment", "relations"], "Task"], [5406, 5407, ["classification"], "Task"], [5450, 5451, ["accuracy"], "Metric"], [5541, 5544, ["V", "-", "BiMPM"], "Method"], [5568, 5575, ["blind", "’", ",", "text", "-", "only", "model"], "Method"], [5606, 5608, ["linking", "language"], "Task"], [5609, 5610, ["perception"], "Task"], [5623, 5625, ["linking", "grounding"], "Task"], [5627, 5628, ["inference"], "Task"], [5638, 5642, ["grounding", "and", "inference", "tasks"], "Task"], [5651, 5653, ["language", "understanding"], "Task"], [5734, 5735, ["GTE"], "Task"], [5792, 5796, ["ICT", "COST", "Action", "IC1307"], "Task"], [5807, 5811, ["Language", "and", "Communication", "Technology"], "Task"], [5877, 5881, ["top", "-", "down", "attention"], "Task"], [5882, 5883, ["VQA"], "Method"], [5887, 5891, ["Visual", "Question", "Answering", "model"], "Method"], [5895, 5899, ["Grounded", "Textual", "Entailment", "task"], "Task"], [5907, 5910, ["grained", "attention", "mechanism"], "Method"], [5936, 5939, ["Recurrent", "Neural", "Network"], "Method"], [5940, 5946, ["Long", "Short", "-", "Term", "Memory", "units"], "Method"], [5959, 5964, ["bottom", "-", "up", "attention", "mechanism"], "Method"], [5966, 5970, ["Fast", "R", "-", "CNN"], "Method"], [5973, 5979, ["ResNet", "-", "101", "convolutional", "neural", "network"], "Method"], [5995, 6000, ["top", "-", "down", "attention", "mechanism"], "Method"], [6028, 6030, ["attention", "score"], "Metric"], [6077, 6081, ["fully", "-", "connected", "layer"], "Method"], [6083, 6087, ["gated", "tanh", "activation", "function"], "Method"], [6118, 6122, ["fully", "-", "connected", "layer"], "Method"], [6124, 6128, ["gated", "tanh", "activation", "function"], "Method"], [6152, 6154, ["multimodal", "fusion"], "Task"], [6181, 6185, ["element", "-", "wise", "multiplication"], "Method"], [6224, 6226, ["multimodal", "representations"], "Method"], [6231, 6233, ["512D", "layers"], "Method"], [6235, 6239, ["gated", "tanh", "activation", "function"], "Method"], [6243, 6245, ["softmax", "layer"], "Method"], [6263, 6265, ["GloVe", "embeddings"], "Method"], [6268, 6270, ["optimization", "tricks"], "Method"], [6284, 6285, ["accuracies"], "Metric"], [6287, 6289, ["VQA", "models"], "Method"], [6325, 6329, ["V", "-", "biMPM", "Model"], "Method"], [8958, 8959, ["matching"], "Task"], [9076, 9077, ["matching"], "Task"], [9095, 9096, ["matching"], "Task"], [10758, 10759, ["hbox"], "Method"], [10778, 10779, ["hbox"], "Method"], [10980, 10981, ["hbox"], "Method"], [10999, 11000, ["hbox"], "Method"], [11963, 11964, ["hbox"], "Method"], [12090, 12093, ["ai", "-", "23u"], "Method"], [12611, 12612, ["softmax"], "Method"], [12971, 12973, ["Prediction", "layer"], "Method"], [13022, 13023, ["BiMPM"], "Method"], [13025, 13027, ["matching", "layer"], "Method"], [13050, 13052, ["weight", "matrix"], "Method"], [13053, 13054, ["matching"], "Task"], [13056, 13059, ["cosine", "similarity", "computation"], "Method"], [13075, 13077, ["matching", "operations"], "Method"], [13082, 13085, ["full", "-", "matching"], "Task"], [13120, 13123, ["max", "-", "pooling"], "Method"], [13126, 13131, ["forward", "/", "backward", "contextual", "embedding"], "Method"], [13151, 13153, ["attentive", "matching"], "Method"], [13198, 13200, ["attentive", "vector"], "Method"], [13201, 13205, ["max", "-", "attentive", "matching"], "Method"], [13209, 13211, ["attentive", "matching"], "Method"], [13229, 13231, ["weighted", "sum"], "Method"], [13242, 13245, ["V", "-", "BiMPM"], "Method"], [13256, 13258, ["multimodal", "matching"], "Task"], [293, 294, ["TE"], "Task"], [638, 639, ["TE"], "Task"], [888, 889, ["TE"], "Task"], [3781, 3782, ["BiMPM"], "Method"], [3804, 3805, ["BiMPM"], "Method"], [4113, 4116, ["V", "-", "SNLI"], "Material"], [4258, 4259, ["BiMPM"], "Method"], [4272, 4275, ["V", "-", "SNLI"], "Material"], [4422, 4423, ["BiMPM"], "Method"], [4813, 4816, ["V", "-", "LSTM"], "Material"], [4843, 4846, ["V", "-", "SNLI"], "Material"], [4938, 4941, ["V", "-", "SNLI"], "Material"], [4980, 4983, ["V", "-", "LSTM"], "Material"], [5342, 5343, ["BiMPM"], "Method"], [5363, 5366, ["V", "-", "LSTM"], "Material"], [5595, 5596, ["TE"], "Task"], [5674, 5675, ["TE"], "Task"], [6276, 6279, ["V", "-", "LSTM"], "Material"], [12988, 12989, ["BiMPM"], "Method"]]} {"doc_id": "14318685b5959b51d0f1e3db34643eb2855dc6d9", "text": ["document", ":", "Going", "deeper", "with", "convolutions", "We", "propose", "a", "deep", "convolutional", "neural", "network", "architecture", "codenamed", "“Inception”", ",", "which", "was", "responsible", "for", "setting", "the", "new", "state", "of", "the", "art", "for", "classification", "and", "detection", "in", "the", "ImageNet", "Large", "-", "Scale", "Visual", "Recognition", "Challenge", "2014", "(", "ILSVRC’14", ")", ".", "The", "main", "hallmark", "of", "this", "architecture", "is", "the", "improved", "utilization", "of", "the", "computing", "resources", "inside", "the", "network", ".", "This", "was", "achieved", "by", "a", "carefully", "crafted", "design", "that", "allows", "for", "increasing", "the", "depth", "and", "width", "of", "the", "network", "while", "keeping", "the", "computational", "budget", "constant", ".", "To", "optimize", "quality", ",", "the", "architectural", "decisions", "were", "based", "on", "the", "Hebbian", "principle", "and", "the", "intuition", "of", "multi", "-", "scale", "processing", ".", "One", "particular", "incarnation", "used", "in", "our", "submission", "for", "ILSVRC’14", "is", "called", "GoogLeNet", ",", "a", "22", "layers", "deep", "network", ",", "the", "quality", "of", "which", "is", "assessed", "in", "the", "context", "of", "classification", "and", "detection", ".", "parskip", "[", "]", "section", ":", "Introduction", "In", "the", "last", "three", "years", ",", "mainly", "due", "to", "the", "advances", "of", "deep", "learning", ",", "more", "concretely", "convolutional", "networks", ",", "the", "quality", "of", "image", "recognition", "and", "object", "detection", "has", "been", "progressing", "at", "a", "dramatic", "pace", ".", "One", "encouraging", "news", "is", "that", "most", "of", "this", "progress", "is", "not", "just", "the", "result", "of", "more", "powerful", "hardware", ",", "larger", "datasets", "and", "bigger", "models", ",", "but", "mainly", "a", "consequence", "of", "new", "ideas", ",", "algorithms", "and", "improved", "network", "architectures", ".", "No", "new", "data", "sources", "were", "used", ",", "for", "example", ",", "by", "the", "top", "entries", "in", "the", "ILSVRC", "2014", "competition", "besides", "the", "classification", "dataset", "of", "the", "same", "competition", "for", "detection", "purposes", ".", "Our", "GoogLeNet", "submission", "to", "ILSVRC", "2014", "actually", "uses", "fewer", "parameters", "than", "the", "winning", "architecture", "of", "Krizhevsky", "et", "al", "from", "two", "years", "ago", ",", "while", "being", "significantly", "more", "accurate", ".", "The", "biggest", "gains", "in", "object", "-", "detection", "have", "not", "come", "from", "the", "utilization", "of", "deep", "networks", "alone", "or", "bigger", "models", ",", "but", "from", "the", "synergy", "of", "deep", "architectures", "and", "classical", "computer", "vision", ",", "like", "the", "R", "-", "CNN", "algorithm", "by", "Girshick", "et", "al", ".", "Another", "notable", "factor", "is", "that", "with", "the", "ongoing", "traction", "of", "mobile", "and", "embedded", "computing", ",", "the", "efficiency", "of", "our", "algorithms", "–", "especially", "their", "power", "and", "memory", "use", "–", "gains", "importance", ".", "It", "is", "noteworthy", "that", "the", "considerations", "leading", "to", "the", "design", "of", "the", "deep", "architecture", "presented", "in", "this", "paper", "included", "this", "factor", "rather", "than", "having", "a", "sheer", "fixation", "on", "accuracy", "numbers", ".", "For", "most", "of", "the", "experiments", ",", "the", "models", "were", "designed", "to", "keep", "a", "computational", "budget", "of", "billion", "multiply", "-", "adds", "at", "inference", "time", ",", "so", "that", "the", "they", "do", "not", "end", "up", "to", "be", "a", "purely", "academic", "curiosity", ",", "but", "could", "be", "put", "to", "real", "world", "use", ",", "even", "on", "large", "datasets", ",", "at", "a", "reasonable", "cost", ".", "In", "this", "paper", ",", "we", "will", "focus", "on", "an", "efficient", "deep", "neural", "network", "architecture", "for", "computer", "vision", ",", "codenamed", "“Inception”", ",", "which", "derives", "its", "name", "from", "the", "“Network", "in", "network”", "paper", "by", "Lin", "et", "al", "in", "conjunction", "with", "the", "famous", "“", "we", "need", "to", "go", "deeper", "”", "internet", "meme", ".", "In", "our", "case", ",", "the", "word", "“", "deep", "”", "is", "used", "in", "two", "different", "meanings", ":", "first", "of", "all", ",", "in", "the", "sense", "that", "we", "introduce", "a", "new", "level", "of", "organization", "in", "the", "form", "of", "the", "“", "Inception", "module", "”", "and", "also", "in", "the", "more", "direct", "sense", "of", "increased", "network", "depth", ".", "In", "general", ",", "one", "can", "view", "the", "Inception", "model", "as", "a", "logical", "culmination", "of", "while", "taking", "inspiration", "and", "guidance", "from", "the", "theoretical", "work", "by", "Arora", "et", "al", ".", "The", "benefits", "of", "the", "architecture", "are", "experimentally", "verified", "on", "the", "ILSVRC", "2014", "classification", "and", "detection", "challenges", ",", "on", "which", "it", "significantly", "outperforms", "the", "current", "state", "of", "the", "art", ".", "section", ":", "Related", "Work", "Starting", "with", "LeNet", "-", "5", ",", "convolutional", "neural", "networks", "(", "CNN", ")", "have", "typically", "had", "a", "standard", "structure", "–", "stacked", "convolutional", "layers", "(", "optionally", "followed", "by", "contrast", "normalization", "and", "max", "-", "pooling", ")", "are", "followed", "by", "one", "or", "more", "fully", "-", "connected", "layers", ".", "Variants", "of", "this", "basic", "design", "are", "prevalent", "in", "the", "image", "classification", "literature", "and", "have", "yielded", "the", "best", "results", "to", "-", "date", "on", "MNIST", ",", "CIFAR", "and", "most", "notably", "on", "the", "ImageNet", "classification", "challenge", ".", "Â", "For", "larger", "datasets", "such", "as", "Imagenet", ",", "the", "recent", "trend", "has", "been", "to", "increase", "the", "number", "of", "layers", "and", "layer", "size", ",", "while", "using", "dropout", "to", "address", "the", "problem", "of", "overfitting", ".", "Despite", "concerns", "that", "max", "-", "pooling", "layers", "result", "in", "loss", "of", "accurate", "spatial", "information", ",", "the", "same", "convolutional", "network", "architecture", "as", "has", "also", "been", "successfully", "employed", "for", "localization", ",", "object", "detection", "and", "human", "pose", "estimation", ".", "Inspired", "by", "a", "neuroscience", "model", "of", "the", "primate", "visual", "cortex", ",", "Serre", "et", "al", ".", "use", "a", "series", "of", "fixed", "Gabor", "filters", "of", "different", "sizes", "in", "order", "to", "handle", "multiple", "scales", ",", "similarly", "to", "the", "Inception", "model", ".", "However", ",", "contrary", "to", "the", "fixed", "2", "-", "layer", "deep", "model", "of", ",", "all", "filters", "in", "the", "Inception", "model", "are", "learned", ".", "Furthermore", ",", "Inception", "layers", "are", "repeated", "many", "times", ",", "leading", "to", "a", "22", "-", "layer", "deep", "model", "in", "the", "case", "of", "the", "GoogLeNet", "model", ".", "Network", "-", "in", "-", "Network", "is", "an", "approach", "proposed", "by", "Lin", "et", "al", ".", "in", "order", "to", "increase", "the", "representational", "power", "of", "neural", "networks", ".", "When", "applied", "to", "convolutional", "layers", ",", "the", "method", "could", "be", "viewed", "as", "additional", "convolutional", "layers", "followed", "typically", "by", "the", "rectified", "linear", "activation", ".", "This", "enables", "it", "to", "be", "easily", "integrated", "in", "the", "current", "CNN", "pipelines", ".", "We", "use", "this", "approach", "heavily", "in", "our", "architecture", ".", "However", ",", "in", "our", "setting", ",", "convolutions", "have", "dual", "purpose", ":", "most", "critically", ",", "they", "are", "used", "mainly", "as", "dimension", "reduction", "modules", "to", "remove", "computational", "bottlenecks", ",", "that", "would", "otherwise", "limit", "the", "size", "of", "our", "networks", ".", "Â", "This", "allows", "for", "not", "just", "increasing", "the", "depth", ",", "but", "also", "the", "width", "of", "our", "networks", "without", "significant", "performance", "penalty", ".", "The", "current", "leading", "approach", "for", "object", "detection", "is", "the", "Regions", "with", "Convolutional", "Neural", "Networks", "(", "R", "-", "CNN", ")", "proposed", "by", "Girshick", "et", "al", ".", ".", "R", "-", "CNN", "decomposes", "the", "overall", "detection", "problem", "into", "two", "subproblems", ":", "to", "first", "utilize", "low", "-", "level", "cues", "such", "as", "color", "and", "superpixel", "consistency", "for", "potential", "object", "proposals", "in", "a", "category", "-", "agnostic", "fashion", ",", "and", "to", "then", "use", "CNN", "classifiers", "to", "identify", "object", "categories", "at", "those", "locations", ".", "Such", "a", "two", "stage", "approach", "leverages", "the", "accuracy", "of", "bounding", "box", "segmentation", "with", "low", "-", "level", "cues", ",", "as", "well", "as", "the", "highly", "powerful", "classification", "power", "of", "state", "-", "of", "-", "the", "-", "art", "CNNs", ".", "We", "adopted", "a", "similar", "pipeline", "in", "our", "detection", "submissions", ",", "but", "have", "explored", "enhancements", "in", "both", "stages", ",", "such", "as", "multi", "-", "box", "prediction", "for", "higher", "object", "bounding", "box", "recall", ",", "and", "ensemble", "approaches", "for", "better", "categorization", "of", "bounding", "box", "proposals", ".", "section", ":", "Motivation", "and", "High", "Level", "Considerations", "The", "most", "straightforward", "way", "of", "improving", "the", "performance", "of", "deep", "neural", "networks", "is", "by", "increasing", "their", "size", ".", "This", "includes", "both", "increasing", "the", "depth", "–", "the", "number", "of", "levels", "–", "of", "the", "network", "and", "its", "width", ":", "the", "number", "of", "units", "at", "each", "level", ".", "This", "is", "as", "an", "easy", "and", "safe", "way", "of", "training", "higher", "quality", "models", ",", "especially", "given", "the", "availability", "of", "a", "large", "amount", "of", "labeled", "training", "data", ".", "However", "this", "simple", "solution", "comes", "with", "two", "major", "drawbacks", ".", "Bigger", "size", "typically", "means", "a", "larger", "number", "of", "parameters", ",", "which", "makes", "the", "enlarged", "network", "more", "prone", "to", "overfitting", ",", "especially", "if", "the", "number", "of", "labeled", "examples", "in", "the", "training", "set", "is", "limited", ".", "This", "can", "become", "a", "major", "bottleneck", ",", "since", "the", "creation", "of", "high", "quality", "training", "sets", "can", "be", "tricky", "and", "expensive", ",", "especially", "if", "expert", "human", "raters", "are", "necessary", "to", "distinguish", "between", "fine", "-", "grained", "visual", "categories", "like", "those", "in", "ImageNet", "(", "even", "in", "the", "1000", "-", "class", "ILSVRC", "subset", ")", "as", "demonstrated", "by", "Figure", "[", "reference", "]", ".", "Another", "drawback", "of", "uniformly", "increased", "network", "size", "is", "the", "dramatically", "increased", "use", "of", "computational", "resources", ".", "For", "example", ",", "in", "a", "deep", "vision", "network", ",", "if", "two", "convolutional", "layers", "are", "chained", ",", "any", "uniform", "increase", "in", "the", "number", "of", "their", "filters", "results", "in", "a", "quadratic", "increase", "of", "computation", ".", "If", "the", "added", "capacity", "is", "used", "inefficiently", "(", "for", "example", ",", "if", "most", "weights", "end", "up", "to", "be", "close", "to", "zero", ")", ",", "then", "a", "lot", "of", "computation", "is", "wasted", ".", "Since", "in", "practice", "the", "computational", "budget", "is", "always", "finite", ",", "an", "efficient", "distribution", "of", "computing", "resources", "is", "preferred", "to", "an", "indiscriminate", "increase", "of", "size", ",", "even", "when", "the", "main", "objective", "is", "to", "increase", "the", "quality", "of", "results", ".", "The", "fundamental", "way", "of", "solving", "both", "issues", "would", "be", "by", "ultimately", "moving", "from", "fully", "connected", "to", "sparsely", "connected", "architectures", ",", "even", "inside", "the", "convolutions", ".", "Besides", "mimicking", "biological", "systems", ",", "this", "would", "also", "have", "the", "advantage", "of", "firmer", "theoretical", "underpinnings", "due", "to", "the", "groundbreaking", "work", "of", "Arora", "et", "al", ".", ".", "Their", "main", "result", "states", "that", "if", "the", "probability", "distribution", "of", "the", "data", "-", "set", "is", "representable", "by", "a", "large", ",", "very", "sparse", "deep", "neural", "network", ",", "then", "the", "optimal", "network", "topology", "can", "be", "constructed", "layer", "by", "layer", "by", "analyzing", "the", "correlation", "statistics", "of", "the", "activations", "of", "the", "last", "layer", "and", "clustering", "neurons", "with", "highly", "correlated", "outputs", ".", "Although", "the", "strict", "mathematical", "proof", "requires", "very", "strong", "conditions", ",", "the", "fact", "that", "this", "statement", "resonates", "with", "the", "well", "known", "Hebbian", "principle", "–", "neurons", "that", "fire", "together", ",", "wire", "together", "–", "suggests", "that", "the", "underlying", "idea", "is", "applicable", "even", "under", "less", "strict", "conditions", ",", "in", "practice", ".", "On", "the", "downside", ",", "today’s", "computing", "infrastructures", "are", "very", "inefficient", "when", "it", "comes", "to", "numerical", "calculation", "on", "non", "-", "uniform", "sparse", "data", "structures", ".", "Even", "if", "the", "number", "of", "arithmetic", "operations", "is", "reduced", "by", ",", "the", "overhead", "of", "lookups", "and", "cache", "misses", "is", "so", "dominant", "that", "switching", "to", "sparse", "matrices", "would", "not", "pay", "off", ".", "The", "gap", "is", "widened", "even", "further", "by", "the", "use", "of", "steadily", "improving", ",", "highly", "tuned", ",", "numerical", "libraries", "that", "allow", "for", "extremely", "fast", "dense", "matrix", "multiplication", ",", "exploiting", "the", "minute", "details", "of", "the", "underlying", "CPU", "or", "GPU", "hardware", ".", "Also", ",", "non", "-", "uniform", "sparse", "models", "require", "more", "sophisticated", "engineering", "and", "computing", "infrastructure", ".", "Most", "current", "vision", "oriented", "machine", "learning", "systems", "utilize", "sparsity", "in", "the", "spatial", "domain", "just", "by", "the", "virtue", "of", "employing", "convolutions", ".", "However", ",", "convolutions", "are", "implemented", "as", "collections", "of", "dense", "connections", "to", "the", "patches", "in", "the", "earlier", "layer", ".", "ConvNets", "have", "traditionally", "used", "random", "and", "sparse", "connection", "tables", "in", "the", "feature", "dimensions", "since", "in", "order", "to", "break", "the", "symmetry", "and", "improve", "learning", ",", "the", "trend", "changed", "back", "to", "full", "connections", "with", "in", "order", "to", "better", "optimize", "parallel", "computing", ".", "The", "uniformity", "of", "the", "structure", "and", "a", "large", "number", "of", "filters", "and", "greater", "batch", "size", "allow", "for", "utilizing", "efficient", "dense", "computation", ".", "This", "raises", "the", "question", "whether", "there", "is", "any", "hope", "for", "a", "next", ",", "intermediate", "step", ":", "an", "architecture", "that", "makes", "use", "of", "the", "extra", "sparsity", ",", "even", "at", "filter", "level", ",", "as", "suggested", "by", "the", "theory", ",", "but", "exploits", "our", "current", "hardware", "by", "utilizing", "computations", "on", "dense", "matrices", ".", "The", "vast", "literature", "on", "sparse", "matrix", "computations", "(", "e.g.", ")", "suggests", "that", "clustering", "sparse", "matrices", "into", "relatively", "dense", "submatrices", "tends", "to", "give", "state", "of", "the", "art", "practical", "performance", "for", "sparse", "matrix", "multiplication", ".", "It", "does", "not", "seem", "far", "-", "fetched", "to", "think", "that", "similar", "methods", "would", "be", "utilized", "for", "the", "automated", "construction", "of", "non", "-", "uniform", "deep", "-", "learning", "architectures", "in", "the", "near", "future", ".", "The", "Inception", "architecture", "started", "out", "as", "a", "case", "study", "of", "the", "first", "author", "for", "assessing", "the", "hypothetical", "output", "of", "a", "sophisticated", "network", "topology", "construction", "algorithm", "that", "tries", "to", "approximate", "a", "sparse", "structure", "implied", "by", "for", "vision", "networks", "and", "covering", "the", "hypothesized", "outcome", "by", "dense", ",", "readily", "available", "components", ".", "Despite", "being", "a", "highly", "speculative", "undertaking", ",", "only", "after", "two", "iterations", "on", "the", "exact", "choice", "of", "topology", ",", "we", "could", "already", "see", "modest", "gains", "against", "the", "reference", "architecture", "based", "on", ".", "After", "further", "tuning", "of", "learning", "rate", ",", "hyperparameters", "and", "improved", "training", "methodology", ",", "we", "established", "that", "the", "resulting", "Inception", "architecture", "was", "especially", "useful", "in", "the", "context", "of", "localization", "and", "object", "detection", "as", "the", "base", "network", "for", "and", ".", "Interestingly", ",", "while", "most", "of", "the", "original", "architectural", "choices", "have", "been", "questioned", "and", "tested", "thoroughly", ",", "they", "turned", "out", "to", "be", "at", "least", "locally", "optimal", ".", "One", "must", "be", "cautious", "though", ":", "although", "the", "proposed", "architecture", "has", "become", "a", "success", "for", "computer", "vision", ",", "it", "is", "still", "questionable", "whether", "its", "quality", "can", "be", "attributed", "to", "the", "guiding", "principles", "that", "have", "lead", "to", "its", "construction", ".", "Making", "sure", "would", "require", "much", "more", "thorough", "analysis", "and", "verification", ":", "for", "example", ",", "if", "automated", "tools", "based", "on", "the", "principles", "described", "below", "would", "find", "similar", ",", "but", "better", "topology", "for", "the", "vision", "networks", ".", "The", "most", "convincing", "proof", "would", "be", "if", "an", "automated", "system", "would", "create", "network", "topologies", "resulting", "in", "similar", "gains", "in", "other", "domains", "using", "the", "same", "algorithm", "but", "with", "very", "differently", "looking", "global", "architecture", ".", "At", "very", "least", ",", "the", "initial", "success", "of", "the", "Inception", "architecture", "yields", "firm", "motivation", "for", "exciting", "future", "work", "in", "this", "direction", ".", "section", ":", "Architectural", "Details", "The", "main", "idea", "of", "the", "Inception", "architecture", "is", "based", "on", "finding", "out", "how", "an", "optimal", "local", "sparse", "structure", "in", "a", "convolutional", "vision", "network", "can", "be", "approximated", "and", "covered", "by", "readily", "available", "dense", "components", ".", "Note", "that", "assuming", "translation", "invariance", "means", "that", "our", "network", "will", "be", "built", "from", "convolutional", "building", "blocks", ".", "All", "we", "need", "is", "to", "find", "the", "optimal", "local", "construction", "and", "to", "repeat", "it", "spatially", ".", "Arora", "et", "al", ".", "suggests", "a", "layer", "-", "by", "layer", "construction", "in", "which", "one", "should", "analyze", "the", "correlation", "statistics", "of", "the", "last", "layer", "and", "cluster", "them", "into", "groups", "of", "units", "with", "high", "correlation", ".", "These", "clusters", "form", "the", "units", "of", "the", "next", "layer", "and", "are", "connected", "to", "the", "units", "in", "the", "previous", "layer", ".", "We", "assume", "that", "each", "unit", "from", "the", "earlier", "layer", "corresponds", "to", "some", "region", "of", "the", "input", "image", "and", "these", "units", "are", "grouped", "into", "filter", "banks", ".", "In", "the", "lower", "layers", "(", "the", "ones", "close", "to", "the", "input", ")", "correlated", "units", "would", "concentrate", "in", "local", "regions", ".", "This", "means", ",", "we", "would", "end", "up", "with", "a", "lot", "of", "clusters", "concentrated", "in", "a", "single", "region", "and", "they", "can", "be", "covered", "by", "a", "layer", "of", "convolutions", "in", "the", "next", "layer", ",", "as", "suggested", "in", ".", "However", ",", "one", "can", "also", "expect", "that", "there", "will", "be", "a", "smaller", "number", "of", "more", "spatially", "spread", "out", "clusters", "that", "can", "be", "covered", "by", "convolutions", "over", "larger", "patches", ",", "and", "there", "will", "be", "a", "decreasing", "number", "of", "patches", "over", "larger", "and", "larger", "regions", ".", "In", "order", "to", "avoid", "patch", "-", "alignment", "issues", ",", "current", "incarnations", "of", "the", "Inception", "architecture", "are", "restricted", "to", "filter", "sizes", ",", "and", ",", "however", "this", "decision", "was", "based", "more", "on", "convenience", "rather", "than", "necessity", ".", "It", "also", "means", "that", "the", "suggested", "architecture", "is", "a", "combination", "of", "all", "those", "layers", "with", "their", "output", "filter", "banks", "concatenated", "into", "a", "single", "output", "vector", "forming", "the", "input", "of", "the", "next", "stage", ".", "Additionally", ",", "since", "pooling", "operations", "have", "been", "essential", "for", "the", "success", "in", "current", "state", "of", "the", "art", "convolutional", "networks", ",", "it", "suggests", "that", "adding", "an", "alternative", "parallel", "pooling", "path", "in", "each", "such", "stage", "should", "have", "additional", "beneficial", "effect", ",", "too", "(", "see", "Figure", "[", "reference", "]", ")", ".", "As", "these", "“", "Inception", "modules", "”", "are", "stacked", "on", "top", "of", "each", "other", ",", "their", "output", "correlation", "statistics", "are", "bound", "to", "vary", ":", "as", "features", "of", "higher", "abstraction", "are", "captured", "by", "higher", "layers", ",", "their", "spatial", "concentration", "is", "expected", "to", "decrease", "suggesting", "that", "the", "ratio", "of", "and", "convolutions", "should", "increase", "as", "we", "move", "to", "higher", "layers", ".", "One", "big", "problem", "with", "the", "above", "modules", ",", "at", "least", "in", "this", "naïve", "form", ",", "is", "that", "even", "a", "modest", "number", "of", "convolutions", "can", "be", "prohibitively", "expensive", "on", "top", "of", "a", "convolutional", "layer", "with", "a", "large", "number", "of", "filters", ".", "This", "problem", "becomes", "even", "more", "pronounced", "once", "pooling", "units", "are", "added", "to", "the", "mix", ":", "their", "number", "of", "output", "filters", "equals", "to", "the", "number", "of", "filters", "in", "the", "previous", "stage", ".", "The", "merging", "of", "the", "output", "of", "the", "pooling", "layer", "with", "the", "outputs", "of", "convolutional", "layers", "would", "lead", "to", "an", "inevitable", "increase", "in", "the", "number", "of", "outputs", "from", "stage", "to", "stage", ".", "Even", "while", "this", "architecture", "might", "cover", "the", "optimal", "sparse", "structure", ",", "it", "would", "do", "it", "very", "inefficiently", ",", "leading", "to", "a", "computational", "blow", "up", "within", "a", "few", "stages", ".", "This", "leads", "to", "the", "second", "idea", "of", "the", "proposed", "architecture", ":", "judiciously", "applying", "dimension", "reductions", "and", "projections", "wherever", "the", "computational", "requirements", "would", "increase", "too", "much", "otherwise", ".", "This", "is", "based", "on", "the", "success", "of", "embeddings", ":", "even", "low", "dimensional", "embeddings", "might", "contain", "a", "lot", "of", "information", "about", "a", "relatively", "large", "image", "patch", ".", "However", ",", "embeddings", "represent", "information", "in", "a", "dense", ",", "compressed", "form", "and", "compressed", "information", "is", "harder", "to", "model", ".", "We", "would", "like", "to", "keep", "our", "representation", "sparse", "at", "most", "places", "(", "as", "required", "by", "the", "conditions", "of", ")", "and", "compress", "the", "signals", "only", "whenever", "they", "have", "to", "be", "aggregated", "en", "masse", ".", "That", "is", ",", "convolutions", "are", "used", "to", "compute", "reductions", "before", "the", "expensive", "and", "convolutions", ".", "Besides", "being", "used", "as", "reductions", ",", "they", "also", "include", "the", "use", "of", "rectified", "linear", "activation", "which", "makes", "them", "dual", "-", "purpose", ".", "The", "final", "result", "is", "depicted", "in", "Figure", "[", "reference", "]", ".", "In", "general", ",", "an", "Inception", "network", "is", "a", "network", "consisting", "of", "modules", "of", "the", "above", "type", "stacked", "upon", "each", "other", ",", "with", "occasional", "max", "-", "pooling", "layers", "with", "stride", "2", "to", "halve", "the", "resolution", "of", "the", "grid", ".", "For", "technical", "reasons", "(", "memory", "efficiency", "during", "training", ")", ",", "it", "seemed", "beneficial", "to", "start", "using", "Inception", "modules", "only", "at", "higher", "layers", "while", "keeping", "the", "lower", "layers", "in", "traditional", "convolutional", "fashion", ".", "This", "is", "not", "strictly", "necessary", ",", "simply", "reflecting", "some", "infrastructural", "inefficiencies", "in", "our", "current", "implementation", ".", "One", "of", "the", "main", "beneficial", "aspects", "of", "this", "architecture", "is", "that", "it", "allows", "for", "increasing", "the", "number", "of", "units", "at", "each", "stage", "significantly", "without", "an", "uncontrolled", "blow", "-", "up", "in", "computational", "complexity", ".", "The", "ubiquitous", "use", "of", "dimension", "reduction", "allows", "for", "shielding", "the", "large", "number", "of", "input", "filters", "of", "the", "last", "stage", "to", "the", "next", "layer", ",", "first", "reducing", "their", "dimension", "before", "convolving", "over", "them", "with", "a", "large", "patch", "size", ".", "Another", "practically", "useful", "aspect", "of", "this", "design", "is", "that", "it", "aligns", "with", "the", "intuition", "that", "visual", "information", "should", "be", "processed", "at", "various", "scales", "and", "then", "aggregated", "so", "that", "the", "next", "stage", "can", "abstract", "features", "from", "different", "scales", "simultaneously", ".", "The", "improved", "use", "of", "computational", "resources", "allows", "for", "increasing", "both", "the", "width", "of", "each", "stage", "as", "well", "as", "the", "number", "of", "stages", "without", "getting", "into", "computational", "difficulties", ".", "Another", "way", "to", "utilize", "the", "inception", "architecture", "is", "to", "create", "slightly", "inferior", ",", "but", "computationally", "cheaper", "versions", "of", "it", ".", "We", "have", "found", "that", "all", "the", "included", "the", "knobs", "and", "levers", "allow", "for", "a", "controlled", "balancing", "of", "computational", "resources", "that", "can", "result", "in", "networks", "that", "are", "faster", "than", "similarly", "performing", "networks", "with", "non", "-", "Inception", "architecture", ",", "however", "this", "requires", "careful", "manual", "design", "at", "this", "point", ".", "section", ":", "GoogLeNet", "We", "chose", "GoogLeNet", "as", "our", "team", "-", "name", "in", "the", "ILSVRC14", "competition", ".", "This", "name", "is", "an", "homage", "to", "Yann", "LeCun’s", "pioneering", "LeNet", "5", "network", ".", "We", "also", "use", "GoogLeNet", "to", "refer", "to", "the", "particular", "incarnation", "of", "the", "Inception", "architecture", "used", "in", "our", "submission", "for", "the", "competition", ".", "We", "have", "also", "used", "a", "deeper", "and", "wider", "Inception", "network", ",", "the", "quality", "of", "which", "was", "slightly", "inferior", ",", "but", "adding", "it", "to", "the", "ensemble", "seemed", "to", "improve", "the", "results", "marginally", ".", "We", "omit", "the", "details", "of", "that", "network", ",", "since", "our", "experiments", "have", "shown", "that", "the", "influence", "of", "the", "exact", "architectural", "parameters", "is", "relatively", "minor", ".", "Here", ",", "the", "most", "successful", "particular", "instance", "(", "named", "GoogLeNet", ")", "is", "described", "in", "Table", "[", "reference", "]", "for", "demonstrational", "purposes", ".", "The", "exact", "same", "topology", "(", "trained", "with", "different", "sampling", "methods", ")", "was", "used", "for", "6", "out", "of", "the", "7", "models", "in", "our", "ensemble", ".", "All", "the", "convolutions", ",", "including", "those", "inside", "the", "Inception", "modules", ",", "use", "rectified", "linear", "activation", ".", "The", "size", "of", "the", "receptive", "field", "in", "our", "network", "is", "taking", "RGB", "color", "channels", "with", "mean", "subtraction", ".", "“", "reduce", "”", "and", "“", "reduce", "”", "stands", "for", "the", "number", "of", "filters", "in", "the", "reduction", "layer", "used", "before", "the", "and", "convolutions", ".", "One", "can", "see", "the", "number", "of", "filters", "in", "the", "projection", "layer", "after", "the", "built", "-", "in", "max", "-", "pooling", "in", "the", "“pool", "proj”", "column", ".", "All", "these", "reduction", "/", "projection", "layers", "use", "rectified", "linear", "activation", "as", "well", ".", "The", "network", "was", "designed", "with", "computational", "efficiency", "and", "practicality", "in", "mind", ",", "so", "that", "inference", "can", "be", "run", "on", "individual", "devices", "including", "even", "those", "with", "limited", "computational", "resources", ",", "especially", "with", "low", "-", "memory", "footprint", ".", "The", "network", "is", "22", "layers", "deep", "when", "counting", "only", "layers", "with", "parameters", "(", "or", "27", "layers", "if", "we", "also", "count", "pooling", ")", ".", "The", "overall", "number", "of", "“layers”", "(", "independent", "building", "blocks", ")", "used", "for", "the", "construction", "of", "the", "network", "is", "about", "100", ".", "However", "this", "number", "depends", "on", "the", "machine", "learning", "infrastructure", "system", "used", ".", "The", "use", "of", "average", "pooling", "before", "the", "classifier", "is", "based", "on", ",", "although", "our", "implementation", "differs", "in", "that", "we", "use", "an", "extra", "linear", "layer", ".", "This", "enables", "adapting", "and", "fine", "-", "tuning", "our", "networks", "for", "other", "label", "sets", "easily", ",", "but", "it", "is", "mostly", "convenience", "and", "we", "do", "not", "expect", "it", "to", "have", "a", "major", "effect", ".", "It", "was", "found", "that", "a", "move", "from", "fully", "connected", "layers", "to", "average", "pooling", "improved", "the", "top", "-", "1", "accuracy", "by", "about", "0.6", "%", ",", "however", "the", "use", "of", "dropout", "remained", "essential", "even", "after", "removing", "the", "fully", "connected", "layers", ".", "Given", "the", "relatively", "large", "depth", "of", "the", "network", ",", "the", "ability", "to", "propagate", "gradients", "back", "through", "all", "the", "layers", "in", "an", "effective", "manner", "was", "a", "concern", ".", "One", "interesting", "insight", "is", "that", "the", "strong", "performance", "of", "relatively", "shallower", "networks", "on", "this", "task", "suggests", "that", "the", "features", "produced", "by", "the", "layers", "in", "the", "middle", "of", "the", "network", "should", "be", "very", "discriminative", ".", "By", "adding", "auxiliary", "classifiers", "connected", "to", "these", "intermediate", "layers", ",", "we", "would", "expect", "to", "encourage", "discrimination", "in", "the", "lower", "stages", "in", "the", "classifier", ",", "increase", "the", "gradient", "signal", "that", "gets", "propagated", "back", ",", "and", "provide", "additional", "regularization", ".", "These", "classifiers", "take", "the", "form", "of", "smaller", "convolutional", "networks", "put", "on", "top", "of", "the", "output", "of", "the", "Inception", "(", "4a", ")", "and", "(", "4d", ")", "modules", ".", "During", "training", ",", "their", "loss", "gets", "added", "to", "the", "total", "loss", "of", "the", "network", "with", "a", "discount", "weight", "(", "the", "losses", "of", "the", "auxiliary", "classifiers", "were", "weighted", "by", "0.3", ")", ".", "At", "inference", "time", ",", "these", "auxiliary", "networks", "are", "discarded", ".", "The", "exact", "structure", "of", "the", "extra", "network", "on", "the", "side", ",", "including", "the", "auxiliary", "classifier", ",", "is", "as", "follows", ":", "An", "average", "pooling", "layer", "with", "filter", "size", "and", "stride", ",", "resulting", "in", "an", "output", "for", "the", "(", "4a", ")", ",", "and", "for", "the", "(", "4d", ")", "stage", ".", "A", "convolution", "with", "128", "filters", "for", "dimension", "reduction", "and", "rectified", "linear", "activation", ".", "A", "fully", "connected", "layer", "with", "1024", "units", "and", "rectified", "linear", "activation", ".", "A", "dropout", "layer", "with", "70", "%", "ratio", "of", "dropped", "outputs", ".", "A", "linear", "layer", "with", "softmax", "loss", "as", "the", "classifier", "(", "predicting", "the", "same", "1000", "classes", "as", "the", "main", "classifier", ",", "but", "removed", "at", "inference", "time", ")", ".", "A", "schematic", "view", "of", "the", "resulting", "network", "is", "depicted", "in", "Figure", "[", "reference", "]", ".", "section", ":", "Training", "Methodology", "Our", "networks", "were", "trained", "using", "the", "DistBelief", "distributed", "machine", "learning", "system", "using", "modest", "amount", "of", "model", "and", "data", "-", "parallelism", ".", "Although", "we", "used", "CPU", "based", "implementation", "only", ",", "a", "rough", "estimate", "suggests", "that", "the", "GoogLeNet", "network", "could", "be", "trained", "to", "convergence", "using", "few", "high", "-", "end", "GPUs", "within", "a", "week", ",", "the", "main", "limitation", "being", "the", "memory", "usage", ".", "Our", "training", "used", "asynchronous", "stochastic", "gradient", "descent", "with", "0.9", "momentum", ",", "fixed", "learning", "rate", "schedule", "(", "decreasing", "the", "learning", "rate", "by", "4", "%", "every", "8", "epochs", ")", ".", "Polyak", "averaging", "was", "used", "to", "create", "the", "final", "model", "used", "at", "inference", "time", ".", "Our", "image", "sampling", "methods", "have", "changed", "substantially", "over", "the", "months", "leading", "to", "the", "competition", ",", "and", "already", "converged", "models", "were", "trained", "on", "with", "other", "options", ",", "sometimes", "in", "conjunction", "with", "changed", "hyperparameters", ",", "like", "dropout", "and", "learning", "rate", ",", "so", "it", "is", "hard", "to", "give", "a", "definitive", "guidance", "to", "the", "most", "effective", "single", "way", "to", "train", "these", "networks", ".", "To", "complicate", "matters", "further", ",", "some", "of", "the", "models", "were", "mainly", "trained", "on", "smaller", "relative", "crops", ",", "others", "on", "larger", "ones", ",", "inspired", "by", ".", "Still", ",", "one", "prescription", "that", "was", "verified", "to", "work", "very", "well", "after", "the", "competition", "includes", "sampling", "of", "various", "sized", "patches", "of", "the", "image", "whose", "size", "is", "distributed", "evenly", "between", "8", "%", "and", "100", "%", "of", "the", "image", "area", "and", "whose", "aspect", "ratio", "is", "chosen", "randomly", "between", "and", ".", "Also", ",", "we", "found", "that", "the", "photometric", "distortions", "by", "Andrew", "Howard", "were", "useful", "to", "combat", "overfitting", "to", "some", "extent", ".", "In", "addition", ",", "we", "started", "to", "use", "random", "interpolation", "methods", "(", "bilinear", ",", "area", ",", "nearest", "neighbor", "and", "cubic", ",", "with", "equal", "probability", ")", "for", "resizing", "relatively", "late", "and", "in", "conjunction", "with", "other", "hyperparameter", "changes", ",", "so", "we", "could", "not", "tell", "definitely", "whether", "the", "final", "results", "were", "affected", "positively", "by", "their", "use", ".", "section", ":", "ILSVRC", "2014", "Classification", "Challenge", "Setup", "and", "Results", "The", "ILSVRC", "2014", "classification", "challenge", "involves", "the", "task", "of", "classifying", "the", "image", "into", "one", "of", "1000", "leaf", "-", "node", "categories", "in", "the", "Imagenet", "hierarchy", ".", "There", "are", "about", "1.2", "million", "images", "for", "training", ",", "50", ",", "000", "for", "validation", "and", "100", ",", "000", "images", "for", "testing", ".", "Each", "image", "is", "associated", "with", "one", "ground", "truth", "category", ",", "and", "performance", "is", "measured", "based", "on", "the", "highest", "scoring", "classifier", "predictions", ".", "Two", "numbers", "are", "usually", "reported", ":", "the", "top", "-", "1", "accuracy", "rate", ",", "which", "compares", "the", "ground", "truth", "against", "the", "first", "predicted", "class", ",", "and", "the", "top", "-", "5", "error", "rate", ",", "which", "compares", "the", "ground", "truth", "against", "the", "first", "5", "predicted", "classes", ":", "an", "image", "is", "deemed", "correctly", "classified", "if", "the", "ground", "truth", "is", "among", "the", "top", "-", "5", ",", "regardless", "of", "its", "rank", "in", "them", ".", "The", "challenge", "uses", "the", "top", "-", "5", "error", "rate", "for", "ranking", "purposes", ".", "We", "participated", "in", "the", "challenge", "with", "no", "external", "data", "used", "for", "training", ".", "In", "addition", "to", "the", "training", "techniques", "aforementioned", "in", "this", "paper", ",", "we", "adopted", "a", "set", "of", "techniques", "during", "testing", "to", "obtain", "a", "higher", "performance", ",", "which", "we", "elaborate", "below", ".", "We", "independently", "trained", "7", "versions", "of", "the", "same", "GoogLeNet", "model", "(", "including", "one", "wider", "version", ")", ",", "and", "performed", "ensemble", "prediction", "with", "them", ".", "These", "models", "were", "trained", "with", "the", "same", "initialization", "(", "even", "with", "the", "same", "initial", "weights", ",", "mainly", "because", "of", "an", "oversight", ")", "and", "learning", "rate", "policies", ",", "and", "they", "only", "differ", "in", "sampling", "methodologies", "and", "the", "random", "order", "in", "which", "they", "see", "input", "images", ".", "During", "testing", ",", "we", "adopted", "a", "more", "aggressive", "cropping", "approach", "than", "that", "of", "Krizhevsky", "et", "al", ".", ".", "Specifically", ",", "we", "resize", "the", "image", "to", "4", "scales", "where", "the", "shorter", "dimension", "(", "height", "or", "width", ")", "is", "256", ",", "288", ",", "320", "and", "352", "respectively", ",", "take", "the", "left", ",", "center", "and", "right", "square", "of", "these", "resized", "images", "(", "in", "the", "case", "of", "portrait", "images", ",", "we", "take", "the", "top", ",", "center", "and", "bottom", "squares", ")", ".", "For", "each", "square", ",", "we", "then", "take", "the", "4", "corners", "and", "the", "center", "crop", "as", "well", "as", "the", "square", "resized", "to", ",", "and", "their", "mirrored", "versions", ".", "This", "results", "in", "crops", "per", "image", ".", "A", "similar", "approach", "was", "used", "by", "Andrew", "Howard", "in", "the", "previous", "year", "’s", "entry", ",", "which", "we", "empirically", "verified", "to", "perform", "slightly", "worse", "than", "the", "proposed", "scheme", ".", "We", "note", "that", "such", "aggressive", "cropping", "may", "not", "be", "necessary", "in", "real", "applications", ",", "as", "the", "benefit", "of", "more", "crops", "becomes", "marginal", "after", "a", "reasonable", "number", "of", "crops", "are", "present", "(", "as", "we", "will", "show", "later", "on", ")", ".", "The", "softmax", "probabilities", "are", "averaged", "over", "multiple", "crops", "and", "over", "all", "the", "individual", "classifiers", "to", "obtain", "the", "final", "prediction", ".", "In", "our", "experiments", "we", "analyzed", "alternative", "approaches", "on", "the", "validation", "data", ",", "such", "as", "max", "pooling", "over", "crops", "and", "averaging", "over", "classifiers", ",", "but", "they", "lead", "to", "inferior", "performance", "than", "the", "simple", "averaging", ".", "In", "the", "remainder", "of", "this", "paper", ",", "we", "analyze", "the", "multiple", "factors", "that", "contribute", "to", "the", "overall", "performance", "of", "the", "final", "submission", ".", "Our", "final", "submission", "in", "the", "challenge", "obtains", "a", "top", "-", "5", "error", "of", "6.67", "%", "on", "both", "the", "validation", "and", "testing", "data", ",", "ranking", "the", "first", "among", "other", "participants", ".", "This", "is", "a", "56.5", "%", "relative", "reduction", "compared", "to", "the", "SuperVision", "approach", "in", "2012", ",", "and", "about", "40", "%", "relative", "reduction", "compared", "to", "the", "previous", "year", "’s", "best", "approach", "(", "Clarifai", ")", ",", "both", "of", "which", "used", "external", "data", "for", "training", "the", "classifiers", ".", "The", "following", "table", "shows", "the", "statistics", "of", "some", "of", "the", "top", "-", "performing", "approaches", ".", "We", "also", "analyze", "and", "report", "the", "performance", "of", "multiple", "testing", "choices", ",", "by", "varying", "the", "number", "of", "models", "and", "the", "number", "of", "crops", "used", "when", "predicting", "an", "image", "in", "the", "following", "table", ".", "When", "we", "use", "one", "model", ",", "we", "chose", "the", "one", "with", "the", "lowest", "top", "-", "1", "error", "rate", "on", "the", "validation", "data", ".", "All", "numbers", "are", "reported", "on", "the", "validation", "dataset", "in", "order", "to", "not", "overfit", "to", "the", "testing", "data", "statistics", ".", "section", ":", "ILSVRC", "2014", "Detection", "Challenge", "Setup", "and", "Results", "The", "ILSVRC", "detection", "task", "is", "to", "produce", "bounding", "boxes", "around", "objects", "in", "images", "among", "200", "possible", "classes", ".", "Detected", "objects", "count", "as", "correct", "if", "they", "match", "the", "class", "of", "the", "groundtruth", "and", "their", "bounding", "boxes", "overlap", "by", "at", "least", "50", "%", "(", "using", "the", "Jaccard", "index", ")", ".", "Extraneous", "detections", "count", "as", "false", "positives", "and", "are", "penalized", ".", "Contrary", "to", "the", "classification", "task", ",", "each", "image", "may", "contain", "many", "objects", "or", "none", ",", "and", "their", "scale", "may", "vary", "from", "large", "to", "tiny", ".", "Results", "are", "reported", "using", "the", "mean", "average", "precision", "(", "mAP", ")", ".", "The", "approach", "taken", "by", "GoogLeNet", "for", "detection", "is", "similar", "to", "the", "R", "-", "CNN", "by", ",", "but", "is", "augmented", "with", "the", "Inception", "model", "as", "the", "region", "classifier", ".", "Additionally", ",", "the", "region", "proposal", "step", "is", "improved", "by", "combining", "the", "Selective", "Search", "approach", "with", "multi", "-", "box", "predictions", "for", "higher", "object", "bounding", "box", "recall", ".", "In", "order", "to", "cut", "down", "the", "number", "of", "false", "positives", ",", "the", "superpixel", "size", "was", "increased", "by", ".", "This", "halves", "the", "proposals", "coming", "from", "the", "selective", "search", "algorithm", ".", "We", "added", "back", "200", "region", "proposals", "coming", "from", "multi", "-", "box", "resulting", ",", "in", "total", ",", "in", "about", "60", "%", "of", "the", "proposals", "used", "by", ",", "while", "increasing", "the", "coverage", "from", "92", "%", "to", "93", "%", ".", "The", "overall", "effect", "of", "cutting", "the", "number", "of", "proposals", "with", "increased", "coverage", "is", "a", "1", "%", "improvement", "of", "the", "mean", "average", "precision", "for", "the", "single", "model", "case", ".", "Finally", ",", "we", "use", "an", "ensemble", "of", "6", "ConvNets", "when", "classifying", "each", "region", "which", "improves", "results", "from", "40", "%", "to", "43.9", "%", "accuracy", ".", "Note", "that", "contrary", "to", "R", "-", "CNN", ",", "we", "did", "not", "use", "bounding", "box", "regression", "due", "to", "lack", "of", "time", ".", "We", "first", "report", "the", "top", "detection", "results", "and", "show", "the", "progress", "since", "the", "first", "edition", "of", "the", "detection", "task", ".", "Compared", "to", "the", "2013", "result", ",", "the", "accuracy", "has", "almost", "doubled", ".", "The", "top", "performing", "teams", "all", "use", "Convolutional", "Networks", ".", "We", "report", "the", "official", "scores", "in", "Table", "[", "reference", "]", "and", "common", "strategies", "for", "each", "team", ":", "the", "use", "of", "external", "data", ",", "ensemble", "models", "or", "contextual", "models", ".", "The", "external", "data", "is", "typically", "the", "ILSVRC12", "classification", "data", "for", "pre", "-", "training", "a", "model", "that", "is", "later", "refined", "on", "the", "detection", "data", ".", "Some", "teams", "also", "mention", "the", "use", "of", "the", "localization", "data", ".", "Since", "a", "good", "portion", "of", "the", "localization", "task", "bounding", "boxes", "are", "not", "included", "in", "the", "detection", "dataset", ",", "one", "can", "pre", "-", "train", "a", "general", "bounding", "box", "regressor", "with", "this", "data", "the", "same", "way", "classification", "is", "used", "for", "pre", "-", "training", ".", "The", "GoogLeNet", "entry", "did", "not", "use", "the", "localization", "data", "for", "pretraining", ".", "In", "Table", "[", "reference", "]", ",", "we", "compare", "results", "using", "a", "single", "model", "only", ".", "The", "top", "performing", "model", "is", "by", "Deep", "Insight", "and", "surprisingly", "only", "improves", "by", "0.3", "points", "with", "an", "ensemble", "of", "3", "models", "while", "the", "GoogLeNet", "obtains", "significantly", "stronger", "results", "with", "the", "ensemble", ".", "section", ":", "Conclusions", "Our", "results", "seem", "to", "yield", "a", "solid", "evidence", "that", "approximating", "the", "expected", "optimal", "sparse", "structure", "by", "readily", "available", "dense", "building", "blocks", "is", "a", "viable", "method", "for", "improving", "neural", "networks", "for", "computer", "vision", ".", "The", "main", "advantage", "of", "this", "method", "is", "a", "significant", "quality", "gain", "at", "a", "modest", "increase", "of", "computational", "requirements", "compared", "to", "shallower", "and", "less", "wide", "networks", ".", "Also", "note", "that", "our", "detection", "work", "was", "competitive", "despite", "of", "neither", "utilizing", "context", "nor", "performing", "bounding", "box", "regression", "and", "this", "fact", "provides", "further", "evidence", "of", "the", "strength", "of", "the", "Inception", "architecture", ".", "Although", "it", "is", "expected", "that", "similar", "quality", "of", "result", "can", "be", "achieved", "by", "much", "more", "expensive", "networks", "of", "similar", "depth", "and", "width", ",", "our", "approach", "yields", "solid", "evidence", "that", "moving", "to", "sparser", "architectures", "is", "feasible", "and", "useful", "idea", "in", "general", ".", "This", "suggest", "promising", "future", "work", "towards", "creating", "sparser", "and", "more", "refined", "structures", "in", "automated", "ways", "on", "the", "basis", "of", ".", "section", ":", "Acknowledgements", "We", "would", "like", "to", "thank", "Sanjeev", "Arora", "and", "Aditya", "Bhaskara", "for", "fruitful", "discussions", "on", ".", "Also", "we", "are", "indebted", "to", "the", "DistBelief", "team", "for", "their", "support", "especially", "to", "Rajat", "Monga", ",", "Jon", "Shlens", ",", "Alex", "Krizhevsky", ",", "Jeff", "Dean", ",", "Ilya", "Sutskever", "and", "Andrea", "Frome", ".", "We", "would", "also", "like", "to", "thank", "to", "Tom", "Duerig", "and", "Ning", "Ye", "for", "their", "help", "on", "photometric", "distortions", ".", "Also", "our", "work", "would", "not", "have", "been", "possible", "without", "the", "support", "of", "Chuck", "Rosenberg", "and", "Hartwig", "Adam", ".", "bibliography", ":", "References"], "entity": [[5, 6, ["convolutions"], "Method"], [9, 16, ["deep", "convolutional", "neural", "network", "architecture", "codenamed", "“Inception”"], "Method"], [29, 30, ["classification"], "Task"], [31, 32, ["detection"], "Task"], [34, 42, ["ImageNet", "Large", "-", "Scale", "Visual", "Recognition", "Challenge", "2014"], "Task"], [43, 44, ["ILSVRC’14"], "Task"], [92, 93, ["quality"], "Metric"], [101, 103, ["Hebbian", "principle"], "Method"], [105, 111, ["intuition", "of", "multi", "-", "scale", "processing"], "Method"], [120, 121, ["ILSVRC’14"], "Task"], [123, 124, ["GoogLeNet"], "Method"], [126, 130, ["22", "layers", "deep", "network"], "Method"], [141, 142, ["classification"], "Task"], [143, 144, ["detection"], "Task"], [163, 165, ["deep", "learning"], "Method"], [168, 170, ["convolutional", "networks"], "Method"], [174, 176, ["image", "recognition"], "Task"], [177, 179, ["object", "detection"], "Task"], [223, 225, ["network", "architectures"], "Method"], [242, 245, ["ILSVRC", "2014", "competition"], "Task"], [258, 260, ["GoogLeNet", "submission"], "Method"], [261, 263, ["ILSVRC", "2014"], "Task"], [284, 285, ["accurate"], "Metric"], [300, 302, ["deep", "networks"], "Method"], [304, 306, ["bigger", "models"], "Method"], [312, 314, ["deep", "architectures"], "Method"], [316, 318, ["computer", "vision"], "Method"], [321, 325, ["R", "-", "CNN", "algorithm"], "Method"], [340, 344, ["mobile", "and", "embedded", "computing"], "Task"], [355, 357, ["memory", "use"], "Metric"], [373, 375, ["deep", "architecture"], "Method"], [389, 391, ["accuracy", "numbers"], "Metric"], [460, 464, ["deep", "neural", "network", "architecture"], "Method"], [465, 467, ["computer", "vision"], "Task"], [537, 539, ["Inception", "module"], "Method"], [559, 561, ["Inception", "model"], "Method"], [615, 618, ["LeNet", "-", "5"], "Method"], [619, 622, ["convolutional", "neural", "networks"], "Method"], [623, 624, ["CNN"], "Method"], [632, 635, ["stacked", "convolutional", "layers"], "Method"], [639, 641, ["contrast", "normalization"], "Method"], [642, 645, ["max", "-", "pooling"], "Method"], [652, 656, ["fully", "-", "connected", "layers"], "Method"], [697, 698, ["Imagenet"], "Material"], [716, 717, ["dropout"], "Method"], [722, 723, ["overfitting"], "Task"], [727, 731, ["max", "-", "pooling", "layers"], "Method"], [741, 744, ["convolutional", "network", "architecture"], "Method"], [751, 752, ["localization"], "Task"], [753, 755, ["object", "detection"], "Task"], [756, 759, ["human", "pose", "estimation"], "Task"], [763, 765, ["neuroscience", "model"], "Method"], [779, 782, ["fixed", "Gabor", "filters"], "Method"], [795, 797, ["Inception", "model"], "Method"], [803, 809, ["fixed", "2", "-", "layer", "deep", "model"], "Method"], [815, 817, ["Inception", "model"], "Method"], [822, 823, ["Inception"], "Method"], [832, 837, ["22", "-", "layer", "deep", "model"], "Method"], [842, 844, ["GoogLeNet", "model"], "Method"], [845, 850, ["Network", "-", "in", "-", "Network"], "Task"], [864, 866, ["representational", "power"], "Method"], [867, 869, ["neural", "networks"], "Method"], [873, 875, ["convolutional", "layers"], "Method"], [889, 892, ["rectified", "linear", "activation"], "Method"], [903, 905, ["CNN", "pipelines"], "Method"], [921, 922, ["convolutions"], "Method"], [934, 937, ["dimension", "reduction", "modules"], "Method"], [979, 981, ["object", "detection"], "Task"], [983, 988, ["Regions", "with", "Convolutional", "Neural", "Networks"], "Method"], [989, 992, ["R", "-", "CNN"], "Method"], [1000, 1003, ["R", "-", "CNN"], "Method"], [1040, 1042, ["CNN", "classifiers"], "Method"], [1057, 1058, ["accuracy"], "Metric"], [1059, 1062, ["bounding", "box", "segmentation"], "Task"], [1074, 1075, ["classification"], "Task"], [1084, 1085, ["CNNs"], "Method"], [1106, 1110, ["multi", "-", "box", "prediction"], "Method"], [1112, 1116, ["object", "bounding", "box", "recall"], "Task"], [1118, 1120, ["ensemble", "approaches"], "Method"], [1122, 1127, ["categorization", "of", "bounding", "box", "proposals"], "Task"], [1144, 1147, ["deep", "neural", "networks"], "Method"], [1230, 1232, ["enlarged", "network"], "Method"], [1275, 1277, ["human", "raters"], "Metric"], [1290, 1291, ["ImageNet"], "Material"], [1330, 1333, ["deep", "vision", "network"], "Task"], [1336, 1338, ["convolutional", "layers"], "Method"], [1356, 1357, ["computation"], "Task"], [1440, 1442, ["fully", "connected"], "Method"], [1443, 1446, ["sparsely", "connected", "architectures"], "Method"], [1450, 1451, ["convolutions"], "Method"], [1454, 1456, ["biological", "systems"], "Method"], [1499, 1503, ["sparse", "deep", "neural", "network"], "Method"], [1507, 1509, ["network", "topology"], "Method"], [1528, 1530, ["clustering", "neurons"], "Method"], [1555, 1557, ["Hebbian", "principle"], "Method"], [1587, 1589, ["computing", "infrastructures"], "Task"], [1596, 1598, ["numerical", "calculation"], "Task"], [1653, 1655, ["numerical", "libraries"], "Method"], [1660, 1663, ["dense", "matrix", "multiplication"], "Task"], [1678, 1683, ["non", "-", "uniform", "sparse", "models"], "Method"], [1693, 1698, ["vision", "oriented", "machine", "learning", "systems"], "Method"], [1710, 1711, ["convolutions"], "Method"], [1714, 1715, ["convolutions"], "Method"], [1730, 1731, ["ConvNets"], "Method"], [1752, 1753, ["learning"], "Task"], [1767, 1769, ["parallel", "computing"], "Task"], [1789, 1791, ["dense", "computation"], "Task"], [1845, 1848, ["sparse", "matrix", "computations"], "Task"], [1870, 1873, ["sparse", "matrix", "multiplication"], "Task"], [1891, 1901, ["automated", "construction", "of", "non", "-", "uniform", "deep", "-", "learning", "architectures"], "Task"], [1927, 1931, ["network", "topology", "construction", "algorithm"], "Method"], [1941, 1943, ["vision", "networks"], "Method"], [1990, 1992, ["learning", "rate"], "Metric"], [1993, 1994, ["hyperparameters"], "Method"], [2013, 2014, ["localization"], "Task"], [2015, 2017, ["object", "detection"], "Task"], [2065, 2067, ["computer", "vision"], "Task"], [2098, 2099, ["verification"], "Task"], [2104, 2106, ["automated", "tools"], "Method"], [2121, 2123, ["vision", "networks"], "Task"], [2132, 2134, ["automated", "system"], "Method"], [2203, 2206, ["convolutional", "vision", "network"], "Method"], [2214, 2216, ["dense", "components"], "Method"], [2230, 2233, ["convolutional", "building", "blocks"], "Method"], [2242, 2244, ["local", "construction"], "Method"], [2256, 2261, ["layer", "-", "by", "layer", "construction"], "Method"], [2327, 2329, ["filter", "banks"], "Method"], [2376, 2377, ["convolutions"], "Method"], [2410, 2411, ["convolutions"], "Method"], [2434, 2438, ["patch", "-", "alignment", "issues"], "Task"], [2482, 2484, ["filter", "banks"], "Method"], [2501, 2503, ["pooling", "operations"], "Method"], [2515, 2517, ["convolutional", "networks"], "Method"], [2562, 2564, ["correlation", "statistics"], "Metric"], [2625, 2626, ["convolutions"], "Method"], [2634, 2636, ["convolutional", "layer"], "Method"], [2681, 2683, ["pooling", "layer"], "Method"], [2687, 2689, ["convolutional", "layers"], "Method"], [2747, 2749, ["dimension", "reductions"], "Method"], [2753, 2755, ["computational", "requirements"], "Metric"], [2768, 2769, ["embeddings"], "Method"], [2771, 2774, ["low", "dimensional", "embeddings"], "Method"], [2842, 2843, ["convolutions"], "Method"], [2866, 2869, ["rectified", "linear", "activation"], "Method"], [2910, 2914, ["max", "-", "pooling", "layers"], "Method"], [2929, 2931, ["memory", "efficiency"], "Metric"], [2932, 2933, ["training"], "Task"], [2954, 2956, ["convolutional", "fashion"], "Method"], [3003, 3005, ["computational", "complexity"], "Metric"], [3010, 3012, ["dimension", "reduction"], "Method"], [3116, 3118, ["inception", "architecture"], "Method"], [3139, 3140, ["knobs"], "Method"], [3183, 3184, ["GoogLeNet"], "Method"], [3191, 3193, ["ILSVRC14", "competition"], "Task"], [3295, 3296, ["GoogLeNet"], "Method"], [3316, 3318, ["sampling", "methods"], "Method"], [3334, 3335, ["convolutions"], "Method"], [3344, 3347, ["rectified", "linear", "activation"], "Method"], [3363, 3365, ["mean", "subtraction"], "Method"], [3381, 3383, ["reduction", "layer"], "Method"], [3386, 3388, ["and", "convolutions"], "Method"], [3405, 3408, ["max", "-", "pooling"], "Method"], [3416, 3420, ["reduction", "/", "projection", "layers"], "Method"], [3421, 3424, ["rectified", "linear", "activation"], "Method"], [3432, 3434, ["computational", "efficiency"], "Metric"], [3441, 3442, ["inference"], "Task"], [3513, 3517, ["machine", "learning", "infrastructure", "system"], "Method"], [3522, 3524, ["average", "pooling"], "Method"], [3526, 3527, ["classifier"], "Method"], [3541, 3543, ["linear", "layer"], "Method"], [3583, 3586, ["fully", "connected", "layers"], "Method"], [3587, 3589, ["average", "pooling"], "Method"], [3591, 3595, ["top", "-", "1", "accuracy"], "Metric"], [3652, 3654, ["shallower", "networks"], "Method"], [3678, 3680, ["auxiliary", "classifiers"], "Method"], [3698, 3699, ["classifier"], "Method"], [3720, 3723, ["smaller", "convolutional", "networks"], "Method"], [3731, 3732, ["Inception"], "Method"], [3742, 3743, ["training"], "Task"], [3764, 3766, ["auxiliary", "classifiers"], "Method"], [3777, 3779, ["auxiliary", "networks"], "Method"], [3787, 3789, ["extra", "network"], "Method"], [3795, 3797, ["auxiliary", "classifier"], "Method"], [3803, 3806, ["average", "pooling", "layer"], "Method"], [3831, 3832, ["convolution"], "Method"], [3833, 3835, ["128", "filters"], "Method"], [3836, 3838, ["dimension", "reduction"], "Task"], [3839, 3842, ["rectified", "linear", "activation"], "Task"], [3844, 3847, ["fully", "connected", "layer"], "Method"], [3851, 3854, ["rectified", "linear", "activation"], "Method"], [3856, 3858, ["dropout", "layer"], "Method"], [3867, 3869, ["linear", "layer"], "Method"], [3874, 3875, ["classifier"], "Method"], [3884, 3885, ["classifier"], "Method"], [3918, 3923, ["DistBelief", "distributed", "machine", "learning", "system"], "Method"], [3936, 3939, ["CPU", "based", "implementation"], "Method"], [3947, 3949, ["GoogLeNet", "network"], "Method"], [3975, 3979, ["asynchronous", "stochastic", "gradient", "descent"], "Method"], [3990, 3992, ["learning", "rate"], "Metric"], [4000, 4002, ["Polyak", "averaging"], "Method"], [4015, 4018, ["image", "sampling", "methods"], "Method"], [4050, 4052, ["learning", "rate"], "Metric"], [4173, 4176, ["random", "interpolation", "methods"], "Method"], [4223, 4225, ["Classification", "Challenge"], "Task"], [4266, 4267, ["validation"], "Task"], [4294, 4296, ["classifier", "predictions"], "Method"], [4313, 4315, ["ground", "truth"], "Metric"], [4323, 4328, ["top", "-", "5", "error", "rate"], "Metric"], [4332, 4334, ["ground", "truth"], "Metric"], [4369, 4374, ["top", "-", "5", "error", "rate"], "Metric"], [4375, 4377, ["ranking", "purposes"], "Task"], [4395, 4397, ["training", "techniques"], "Method"], [4429, 4431, ["GoogLeNet", "model"], "Method"], [4440, 4442, ["ensemble", "prediction"], "Method"], [4468, 4471, ["learning", "rate", "policies"], "Method"], [4477, 4479, ["sampling", "methodologies"], "Method"], [4498, 4500, ["cropping", "approach"], "Method"], [4633, 4635, ["aggressive", "cropping"], "Method"], [4681, 4682, ["classifiers"], "Method"], [4686, 4687, ["prediction"], "Task"], [4702, 4706, ["max", "pooling", "over", "crops"], "Method"], [4707, 4710, ["averaging", "over", "classifiers"], "Method"], [4720, 4721, ["averaging"], "Method"], [4753, 4757, ["top", "-", "5", "error"], "Metric"], [4785, 4787, ["SuperVision", "approach"], "Method"], [4817, 4818, ["classifiers"], "Method"], [4880, 4885, ["top", "-", "1", "error", "rate"], "Metric"], [4911, 4915, ["ILSVRC", "2014", "Detection", "Challenge"], "Task"], [4962, 4964, ["Jaccard", "index"], "Metric"], [5006, 5009, ["mean", "average", "precision"], "Metric"], [5010, 5011, ["mAP"], "Metric"], [5017, 5018, ["GoogLeNet"], "Method"], [5019, 5020, ["detection"], "Task"], [5024, 5027, ["R", "-", "CNN"], "Method"], [5034, 5036, ["Inception", "model"], "Method"], [5038, 5040, ["region", "classifier"], "Method"], [5044, 5047, ["region", "proposal", "step"], "Method"], [5052, 5055, ["Selective", "Search", "approach"], "Method"], [5056, 5060, ["multi", "-", "box", "predictions"], "Method"], [5062, 5066, ["object", "bounding", "box", "recall"], "Metric"], [5092, 5095, ["selective", "search", "algorithm"], "Method"], [5125, 5126, ["coverage"], "Metric"], [5152, 5155, ["mean", "average", "precision"], "Metric"], [5169, 5170, ["ConvNets"], "Method"], [5183, 5184, ["accuracy"], "Metric"], [5189, 5192, ["R", "-", "CNN"], "Method"], [5197, 5200, ["bounding", "box", "regression"], "Method"], [5211, 5212, ["detection"], "Task"], [5233, 5234, ["accuracy"], "Metric"], [5244, 5246, ["Convolutional", "Networks"], "Method"], [5250, 5252, ["official", "scores"], "Metric"], [5270, 5272, ["ensemble", "models"], "Method"], [5273, 5275, ["contextual", "models"], "Method"], [5282, 5285, ["ILSVRC12", "classification", "data"], "Task"], [5335, 5339, ["general", "bounding", "box", "regressor"], "Method"], [5345, 5346, ["classification"], "Task"], [5349, 5352, ["pre", "-", "training"], "Task"], [5363, 5364, ["pretraining"], "Task"], [5403, 5404, ["GoogLeNet"], "Method"], [5433, 5436, ["dense", "building", "blocks"], "Method"], [5442, 5444, ["neural", "networks"], "Method"], [5445, 5447, ["computer", "vision"], "Task"], [5457, 5458, ["quality"], "Metric"], [5464, 5466, ["computational", "requirements"], "Metric"], [5468, 5473, ["shallower", "and", "less", "wide", "networks"], "Method"], [5478, 5479, ["detection"], "Task"], [5489, 5492, ["bounding", "box", "regression"], "Method"], [5537, 5539, ["sparser", "architectures"], "Method"], [247, 248, ["classification"], "Task"], [254, 255, ["detection"], "Task"], [292, 293, ["detection"], "Task"], [592, 593, ["classification"], "Task"], [594, 595, ["detection"], "Task"], [667, 668, ["classification"], "Task"], [688, 689, ["classification"], "Task"], [735, 736, ["accurate"], "Metric"], [1006, 1007, ["detection"], "Task"], [1093, 1094, ["detection"], "Task"], [1907, 1908, ["Inception"], "Method"], [2004, 2005, ["Inception"], "Method"], [2166, 2167, ["Inception"], "Method"], [2188, 2189, ["Inception"], "Method"], [2443, 2444, ["Inception"], "Method"], [2549, 2550, ["Inception"], "Method"], [2891, 2892, ["Inception"], "Method"], [2941, 2942, ["Inception"], "Method"], [3165, 3166, ["Inception"], "Method"], [3219, 3220, ["Inception"], "Method"], [3237, 3238, ["Inception"], "Method"], [3340, 3341, ["Inception"], "Method"], [4231, 4232, ["classification"], "Task"], [4250, 4251, ["Imagenet"], "Material"], [4307, 4308, ["accuracy"], "Metric"], [4920, 4921, ["detection"], "Task"], [4979, 4980, ["classification"], "Task"], [5223, 5224, ["detection"], "Task"], [5297, 5298, ["detection"], "Task"], [5308, 5309, ["localization"], "Task"], [5317, 5318, ["localization"], "Task"], [5326, 5327, ["detection"], "Task"], [5360, 5361, ["localization"], "Task"], [5503, 5504, ["Inception"], "Method"]]} {"doc_id": "16051bbe3a7f7c77a952ebf76722ea655e8906ca", "text": ["Image", "Super", "-", "resolution", "via", "Feature", "-", "augmented", "Random", "Forest", "section", ":", "Abstract", "-", "Recent", "random", "-", "forest", "(", "RF", ")-", "based", "image", "super", "-", "resolution", "approaches", "inherit", "some", "properties", "from", "dictionary", "-", "learning", "-", "based", "algorithms", ",", "but", "the", "effectiveness", "of", "the", "properties", "in", "RF", "is", "overlooked", "in", "the", "literature", ".", "In", "this", "paper", ",", "we", "present", "a", "novel", "feature", "-", "augmented", "random", "forest", "(", "FARF", ")", "for", "image", "super", "-", "resolution", ",", "where", "the", "conventional", "gradient", "-", "based", "features", "are", "augmented", "with", "gradient", "magnitudes", "and", "different", "feature", "recipes", "are", "formulated", "on", "different", "stages", "in", "an", "RF", ".", "The", "advantages", "of", "our", "method", "are", "that", ",", "firstly", ",", "the", "dictionary", "-", "learning", "-", "based", "features", "are", "enhanced", "by", "adding", "gradient", "magnitudes", ",", "based", "on", "the", "observation", "that", "the", "non", "-", "linear", "gradient", "magnitude", "are", "with", "highly", "discriminative", "property", ".", "Secondly", ",", "generalized", "locality", "-", "sensitive", "hashing", "(", "LSH", ")", "is", "used", "to", "replace", "principal", "component", "analysis", "(", "PCA", ")", "for", "feature", "dimensionality", "reduction", "and", "original", "high", "-", "dimensional", "features", "are", "employed", ",", "instead", "of", "the", "compressed", "ones", ",", "for", "the", "leaf", "-", "nodes", "'", "regressors", ",", "since", "regressors", "can", "benefit", "from", "higher", "dimensional", "features", ".", "This", "original", "-", "compressed", "coupled", "feature", "sets", "scheme", "unifies", "the", "unsupervised", "LSH", "evaluation", "on", "both", "image", "super", "-", "resolution", "and", "content", "-", "based", "image", "retrieval", "(", "CBIR", ")", ".", "Finally", ",", "we", "present", "a", "generalized", "weighted", "ridge", "regression", "(", "GWRR", ")", "model", "for", "the", "leaf", "-", "nodes", "'", "regressors", ".", "Experiment", "results", "on", "several", "public", "benchmark", "datasets", "show", "that", "our", "FARF", "method", "can", "achieve", "an", "average", "gain", "of", "about", "0.3", "dB", ",", "compared", "to", "traditional", "RF", "-", "based", "methods", ".", "Furthermore", ",", "a", "fine", "-", "tuned", "FARF", "model", "can", "compare", "to", "or", "(", "in", "many", "cases", ")", "outperform", "some", "recent", "stateof", "-", "the", "-", "art", "deep", "-", "learning", "-", "based", "algorithms", ".", "section", ":", "INTRODUCTION", "In", "the", "past", "few", "years", ",", "random", "forest", "(", "RF", ")", "[", "reference", "][", "reference", "]", "as", "a", "machine", "-", "learning", "tool", ",", "working", "via", "an", "ensemble", "of", "multiple", "decision", "trees", ",", "has", "been", "employed", "for", "efficient", "classification", "or", "regression", "problems", ",", "and", "applied", "to", "a", "large", "variety", "of", "computer", "-", "vision", "applications", ",", "such", "as", "object", "recognition", "[", "reference", "]", ",", "face", "alignment", "[", "reference", "][", "reference", "][", "reference", "]", ",", "data", "clustering", "[", "reference", "]", ",", "single", "image", "super", "-", "resolution", "(", "SISR", ")", "[", "reference", "][", "reference", "]", ",", "and", "so", "on", ".", "The", "RF", "method", ",", "which", "benefits", "from", "its", "simple", "implementation", "of", "binary", "trees", ",", "has", "been", "widely", "used", ",", "and", "exhibits", "a", "number", "of", "merits", ",", "including", "(", "1", ")", "it", "works", "with", "an", "ensemble", "of", "multiple", "decision", "trees", "to", "express", "the", "principle", "that", "\"", "two", "heads", "are", "better", "than", "one", "\"", ",", "[", "reference", "]", "it", "is", "easy", "to", "be", "sped", "up", "with", "parallel", "processing", "technology", ",", "on", "both", "the", "training", "and", "inference", "stages", ",", "(", "3", ")", "it", "has", "sub", "-", "linear", "search", "complexity", ",", "because", "of", "the", "use", "of", "the", "binary", "tree", "structure", ",", "(", "4", ")", "the", "bagging", "strategy", "for", "feature", "candidates", "on", "splitnodes", "enable", "it", "to", "handle", "high", "-", "dimensional", "features", "and", "avoid", "over", "-", "fitting", "on", "regression", ",", "and", "(", "5", ")", "the", "clustering", "-", "regression", "scheme", "employs", "the", "\"", "divide", "and", "conquer", "\"", "strategy", ",", "which", "can", "tackle", "the", "classification", "and", "regression", "tasks", "with", "more", "stable", "performance", ".", "The", "RF", "-", "based", "image", "super", "-", "resolution", "approach", "can", "be", "considered", "as", "a", "clustering", "/", "classificationbased", "method", ",", "as", "shown", "in", "Fig", ".", "1", ".", "But", "the", "clustering", "and", "regression", "problems", "in", "RF", "require", "with", "different", "discriminative", "features", ",", "which", "have", "not", "been", "systematically", "studied", "in", "existing", "literature", ".", "Feature", "engineering", "has", "been", "a", "research", "hotspot", "for", "decades", ".", "Several", "features", "have", "been", "proposed", "for", "learning", "the", "mapping", "functions", "from", "low", "-", "resolution", "(", "LR", ")", "patches", "to", "high", "-", "resolution", "(", "HR", ")", "patches", "on", "image", "restoration", "problems", ".", "Pioneer", "work", "in", "[", "reference", "]", "used", "a", "simple", "high", "-", "pass", "filter", "as", "simple", "as", "subtracting", "a", "low", "-", "pass", "filtered", "values", "from", "the", "input", "image", "raw", "values", ".", "Meanwhile", ",", "most", "algorithms", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "follow", "the", "approach", "in", "[", "reference", "]", ",", "which", "concatenates", "the", "first", "-", "and", "second", "-", "order", "gradients", "to", "form", "the", "features", ",", "as", "an", "inexpensive", "solution", "to", "approximating", "high", "-", "pass", "filtering", ".", "Since", "RF", "is", "used", "as", "a", "dictionarylearning", "-", "based", "tool", ",", "it", "inherits", "many", "properties", "from", "the", "conventional", "dictionary", "-", "learning", "-", "based", "algorithms", "on", "feature", "extraction", ".", "However", ",", "the", "discriminative", "ability", "of", "those", "gradient", "-", "based", "features", "for", "random", "forest", "has", "been", "overlooked", "in", "the", "literature", ".", "We", "found", ",", "from", "experiments", ",", "that", "augmented", "features", "based", "on", "two", "gradient", "-", "magnitude", "filters", "can", "achieve", "more", "than", "0.1dB", "quality", "improvement", "in", "RF", "based", "SISR", ",", "with", "the", "same", "parameter", "setting", ".", "In", "most", "dictionary", "-", "learning", "-", "based", "algorithms", ",", "principal", "component", "analysis", "(", "PCA", ")", "is", "used", "for", "dimensionality", "reduction", "before", "classification", "and", "regression", "processes", ".", "The", "impact", "of", "using", "PCA", "has", "also", "been", "paid", "less", "attention", "in", "the", "literature", ".", "PCA", "projection", "may", "damage", "the", "structure", "of", "features", ",", "which", "are", "originally", "discriminative", "for", "clustering", "at", "the", "split", "-", "nodes", "and", "regression", "at", "the", "leaf", "-", "nodes", ".", "Motivated", "from", "content", "-", "based", "image", "retrieval", "(", "CBIR", ")", "[", "reference", "][", "reference", "]", ",", "where", "the", "coarse", "-", "level", "search", "uses", "compressed", "features", ",", "while", "the", "fine", "-", "level", "search", "uses", "augmented", "features", ".", "Therefore", ",", "in", "our", "method", ",", "we", "use", "the", "original", "features", "rather", "than", "the", "compressed", "features", "generated", "by", "PCA", "as", "worked", "in", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ",", "so", "that", "more", "accurate", "regression", "and", "higher", "image", "quality", "improvement", "can", "be", "achieved", ".", "Moreover", ",", "the", "unsupervised", "locality", "-", "sensitive", "hashing", "(", "LSH", ")", "model", ",", "instead", "of", "PCA", ",", "is", "employed", "for", "feature", "dimensionality", "reduction", ",", "which", "can", "reduce", "the", "damage", "on", "the", "feature", "structure", "for", "the", "compressed", "features", "used", "on", "clustering", "at", "the", "split", "-", "nodes", "and", "thus", "improve", "the", "final", "image", "quality", ".", "For", "regression", "problems", "at", "the", "leaf", "-", "nodes", ",", "we", "propose", "a", "generalized", "weighted", "ridge", "regression", "(", "GWRR", ")", "as", "an", "extension", "of", "the", "work", "in", "[", "reference", "]", ".", "GWRR", "models", "are", "generated", "based", "on", "the", "data", "distributions", "from", "the", "leaf", "-", "nodes", ".", "The", "main", "contribution", "of", "our", "method", "is", "on", "feature", "augmentation", ",", "so", "we", "call", "our", "method", "featureaugmented", "random", "forest", "(", "FARF", ")", ".", "The", "pipeline", "of", "our", "FARF", "method", ",", "which", "includes", "feature", "extraction", ",", "the", "training", "stage", ",", "and", "inference", "stages", "for", "SISR", ",", "is", "shown", "in", "Fig", ".", "1", ".", "In", "the", "FARF", "-", "based", "image", "SR", "scheme", ",", "higher", "discriminative", "features", "are", "extracted", "by", "using", "the", "first", "-", "and", "second", "-", "order", "gradients", "and", "their", "magnitudes", ".", "Then", ",", "the", "conventional", "PCA", "is", "replaced", "by", "the", "generalized", "LSH", "for", "dimensionality", "reduction", ",", "and", "the", "compressed", "features", "are", "used", "for", "clustering", "in", "the", "split", "-", "nodes", "on", "an", "RF", ".", "Finally", ",", "the", "respective", "regressors", "at", "the", "leaf", "-", "nodes", "are", "learned", "by", "using", "the", "original", "high", "dimensional", "features", "with", "the", "GWRR", "models", ".", "Having", "introduced", "the", "main", "idea", "of", "our", "paper", ",", "the", "remainder", "of", "this", "paper", "is", "organized", "as", "follows", ".", "In", "Section", "2", ",", "we", "review", "the", "related", "works", "on", "SISR", ",", "particularly", "the", "RF", "-", "based", "approaches", "and", "our", "insights", ".", "In", "Section", "3", ",", "we", "introduce", "the", "proposed", "method", "FARF", ",", "including", "the", "discriminative", "feature", "augmented", "by", "the", "gradient", "-", "magnitude", "filters", ",", "the", "generalized", "weighted", "ridge", "regression", "(", "GWRR", ")", "model", ",", "and", "the", "fine", "-", "tuned", "FARF", "version", ".", "In", "Section", "4", ",", "we", "evaluate", "our", "FARF", "scheme", "on", "public", "datasets", ",", "and", "conclusions", "are", "given", "in", "Section", "5", ".", "section", ":", "IMAGE", "SUPER", "-", "RESOLUTION", "VIA", "RANDOM", "FOREST", "section", ":", "Image", "Super", "-", "Resolution", "Image", "SR", "attempts", "to", "achieve", "an", "impressive", "HR", "quality", "image", "from", "one", "or", "a", "set", "of", "LR", "images", "via", "artistic", "skills", ",", "which", "has", "been", "an", "active", "research", "topic", "for", "decades", "in", "the", "image", "restoration", "area", ".", "Generalized", "SR", "includes", "interpolation", "algorithms", ",", "such", "as", "the", "classic", "bicubic", "interpolation", ",", "and", "other", "edge", "-", "preserving", "algorithms", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "The", "traditional", "super", "-", "resolution", "algorithms", "are", "based", "on", "pixel", "operations", ".", "Intuitively", ",", "operating", "on", "a", "\"", "big", "pixel", "\"", ",", "i.e.", "a", "patch", "[", "reference", "]", ",", "is", "more", "effective", ".", "Since", "patch", "-", "based", "algorithms", "can", "preserve", "the", "local", "texture", "structure", "of", "an", "image", ",", "various", "methods", "based", "on", "image", "patches", ",", "such", "as", "non", "-", "local", "means", "[", "reference", "]", ",", "self", "-", "similarity", "[", "reference", "]", ",", "manifold", "learning", "[", "reference", "]", ",", "block", "-", "matching", "and", "3D", "filtering", "(", "BM3D", ")", "[", "reference", "]", ",", "sparse", "representation", "[", "reference", "]", ",", "etc", ".", "have", "been", "proposed", ".", "The", "neighbor", "-", "embedding", "(", "NE", ")", "methods", "[", "reference", "][", "reference", "]", "are", "the", "milestone", "for", "patch", "-", "based", "dictionary", "learning", "methods", ".", "NE", "learns", "the", "mapping", "between", "low", "-", "and", "high", "-", "resolution", "patches", ",", "with", "the", "use", "of", "manifold", "learning", ".", "Based", "on", "the", "locally", "linear", "embedding", "(", "LLE", ")", "theory", ",", "an", "LR", "patch", "can", "be", "represented", "as", "a", "linear", "combination", "of", "its", "nearest", "neighbors", "in", "a", "learned", "dictionary", ",", "and", "its", "HR", "counterpart", "can", "be", "approximated", "as", "a", "linear", "combination", "of", "the", "corresponding", "HR", "patches", "of", "its", "LR", "neighbors", ",", "with", "the", "same", "coefficients", ".", "Although", "the", "NE", "method", "is", "simple", "and", "sounds", "practical", ",", "a", "problem", "with", "the", "method", "is", "how", "to", "build", "a", "feasible", "patch", "dictionary", ".", "For", "example", ",", "for", "a", "patch", "size", "of", "5×5", ",", "with", "256", "gray", "levels", ",", "it", "is", "necessary", "to", "have", "a", "massive", "dataset", ",", "which", "has", "millions", "of", "patches", ",", "in", "order", "to", "achieve", "high", "-", "quality", "reconstructed", "HR", "patches", ",", "if", "the", "patches", "are", "collected", "directly", "from", "natural", "scene", "images", ".", "Because", "of", "the", "large", "dictionary", "size", ",", "it", "is", "time", "consuming", "to", "search", "for", "a", "neighbor", "in", "such", "a", "large", "dataset", ".", "Other", "method", "to", "reduce", "the", "dictionary", "size", "is", "to", "learn", "a", "relatively", "smaller", "dictionary", "with", "discrete", "cosine", "transform", "(", "DCT", ")", "or", "wavelet", "fixed", "basis", ",", "which", "the", "adaptiveness", "is", "sacrificed", ".", "In", "2010", ",", "Yang", "et", "al", ".", "[", "reference", "]", "proposed", "a", "sparse", "prior", "for", "dictionary", "learning", ".", "Using", "sparse", "coding", ",", "image", "representation", "can", "work", "with", "a", "relatively", "smaller", "dictionary", "while", "keep", "the", "adaptiveness", "by", "learning", "the", "basis", "from", "data", "directly", ",", "which", "opens", "the", "era", "for", "sparse", "coding", "in", "the", "image", "inverse", "problems", ".", "With", "the", "sparse", "constraint", "used", "in", "the", "sparse", "-", "coding", "super", "-", "resolution", "(", "ScSR", ")", "framework", ",", "an", "LR", "patch", "and", "its", "corresponding", "HR", "patch", "can", "both", "be", "reconstructed", "through", "two", "learned", "coupled", "dictionaries", ",", "with", "the", "same", "coefficients", "as", "following", ":", "where", "and", "denote", "an", "LR", "patch", "and", "its", "HR", "counterpart", ",", "respectively", ",", "and", "D", "and", "D", "ℎ", "are", "the", "low", "and", "high", "-", "resolution", "coupled", "dictionaries", "trained", "jointly", "from", "LR", "and", "HR", "patch", "samples", ".", "The", "value", "of", "in", "‖", "‖", "ϑ", "is", "the", "sparsity", "factor", "of", "the", "coefficients", ".", "‖", "‖", "0", ",", "called", "the", "0", "-", "norm", ",", "is", "the", "non", "-", "zero", "count", "of", "the", "coefficients", "in", ".", "The", "LR", "and", "HR", "coupled", "dictionaries", "are", "trained", "jointly", "with", "a", "sparsity", "constraint", ",", "as", "following", ":", "an", "LR", "patch", "of", "an", "input", "LR", "image", "Y", "can", "be", "formulated", "in", "terms", "of", "D", "as", "following", ":", "where", "is", "a", "feature", "-", "extraction", "operator", "on", "the", "LR", "patches", ",", "which", "aims", "to", "extract", "discriminative", "features", "from", "LR", "patches", ",", "rather", "than", "using", "the", "raw", "pixel", "intensity", ".", "Although", "the", "0", "-", "norm", "of", "α", "is", "an", "ideal", "regularization", "term", "for", "the", "sparse", "constraint", ",", "this", "strong", "constraint", "leads", "to", "an", "NP", "-", "hard", "problem", "in", "solving", "the", "coefficients", "α", ".", "Yang", "et", "al", ".", "[", "reference", "]", "relaxed", "the", "0", "-", "norm", "to", "1", "-", "norm", ",", "so", "as", "to", "achieve", "a", "feasible", "solution", "as", "following", ":", "and", "an", "equivalent", "formulation", "can", "be", "achieved", "by", "using", "the", "Lagrange", "multiplier", ",", "where", "the", "parameter", "balances", "the", "sparsity", "of", "the", "solution", "and", "the", "fidelity", "of", "the", "approximation", "to", ".", "As", "the", "sparse", "constraint", "in", "[", "reference", "]", "is", "still", "a", "bottleneck", "on", "training", "dictionaries", "considering", "the", "computation", ",", "an", "intuitive", "way", "to", "solve", "it", "is", "to", "relax", "the", "constraint", "again", "to", "2", "-", "norm", ".", "Meanwhile", ",", "the", "effectiveness", "of", "sparsity", "is", "challenged", "[", "reference", "][", "reference", "]", "by", "researchers", "as", "to", "whether", "sparsity", "or", "collaborative", "representation", "really", "helps", "in", "image", "classification", "and", "restoration", ".", "As", "a", "natural", "solution", "to", "that", ",", "Timofte", "et", "al", ".", "proposed", "an", "anchored", "neighborhood", "regression", "(", "ANR", ")", "[", "reference", "]", "framework", ",", "where", "there", "is", "no", "sparse", "constraint", "in", "the", "model", ".", "ANR", "replaces", "the", "sparse", "-", "decomposition", "optimization", "(", "1", "-", "norm", ")", "with", "a", "ridge", "regression", "(", "i.e.", "2", "-", "norm", ")", ",", "where", "the", "coefficients", "can", "be", "computed", "offline", "and", "each", "coefficient", "can", "be", "stored", "as", "an", "atom", "(", "anchor", ")", "in", "the", "dictionary", ".", "This", "offline", "learning", "can", "greatly", "speed", "-", "up", "the", "prediction", "stage", ",", "and", "this", "approach", "has", "subsequently", "led", "to", "several", "variant", "algorithms", ".", "Timofte", "et", "al", ".", "later", "extended", "the", "ANR", "approach", "to", "the", "A", "+", "[", "reference", "]", ".", "In", "A", "+", "[", "reference", "]", ",", "the", "coupled", "dictionaries", "are", "trained", "from", "a", "large", "pool", "of", "training", "samples", "(", "in", "the", "order", "of", "millions", ")", "rather", "than", "only", "from", "the", "anchoring", "atoms", ",", "which", "greatly", "improves", "the", "image", "quality", ".", "After", "that", ",", "more", "extensions", "based", "on", "ANR", "and", "A", "+", "have", "emerged", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "However", ",", "in", "the", "above", "-", "mentioned", "dictionary", "-", "learning", "methods", ",", "the", "complexity", "of", "finding", "those", "similar", "patches", "by", "comparing", "an", "input", "patch", "with", "all", "the", "dictionary", "items", "has", "been", "overlooked", ".", "Recently", ",", "algorithms", "using", "random", "forest", "(", "RF", ")", "[", "reference", "][", "reference", "][", "reference", "]", "have", "achieved", "state", "-", "of", "-", "the", "-", "art", "performances", ",", "in", "terms", "of", "both", "accuracy", "and", "efficiency", "for", "classification", "and", "regression", "tasks", ".", "This", "is", "mainly", "due", "to", "the", "use", "of", "ensemble", "learning", "and", "sublinear", "search", "based", "on", "binary", "trees", ".", "Schulter", "et", "al", ".", "[", "reference", "]", "adopted", "random", "forest", "and", "the", "clustering", "-", "regression", "scheme", "to", "learn", "regressors", "from", "the", "patches", "in", "leaf", "-", "nodes", "for", "SISR", ".", "With", "the", "same", "number", "of", "regressors", ",", "the", "RF", "-", "based", "algorithm", "can", "outperform", "or", "achieve", "comparable", "performance", "with", "A", "+", "and", "its", "variants", ",", "in", "terms", "of", "accuracy", "but", "with", "less", "computational", "complexity", ".", "In", "recent", "years", ",", "deep", "learning", "has", "achieved", "promising", "performances", "on", "image", "super", "-", "resolution", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "In", "[", "reference", "][", "reference", "]", ",", "milestone", "works", "on", "image", "super", "-", "resolution", "based", "on", "deep", "learning", "were", "presented", ",", "where", "a", "convolutional", "neural", "network", "(", "SRCNN", ")", "was", "proposed", "to", "learn", "an", "end", "-", "to", "-", "end", "mapping", "between", "LR", "and", "HR", "images", "for", "image", "super", "-", "resolution", ".", "Later", "a", "scheme", "with", "very", "deep", "networks", "for", "SISR", "was", "proposed", "in", "[", "reference", "]", ",", "where", "the", "convergence", "rate", "of", "the", "deep", "network", "is", "improved", "by", "using", "residual", "learning", "and", "extremely", "high", "learning", "rates", ".", "In", "addition", ",", "Ledig", "et", "al", ".", "[", "reference", "]", "introduced", "a", "generative", "adversarial", "network", "(", "GAN", ")", "based", "image", "super", "-", "resolution", "model", "(", "SRGAN", ")", ",", "where", "the", "image", "perceptual", "loss", "function", "is", "reformulated", "as", "the", "combination", "of", "content", "loss", "and", "adversarial", "loss", ".", "Although", "deeplearning", "-", "based", "approaches", "have", "achieved", "promising", "progress", "on", "SISR", ",", "the", "heavy", "computational", "requirement", "is", "still", "a", "large", "burden", "even", "though", "the", "implementation", "is", "accelerated", "by", "GPU", ".", "This", "may", "limit", "them", "from", "those", "applications", "without", "powerful", "GPU", ",", "such", "as", "smart", "mobile", "terminals", ".", "cluster", "all", "the", "feature", "data", "assigned", "to", "this", "node", ".", "This", "results", "in", "separating", "the", "three", "data", "samples", "(", "quadrangle", ",", "pentagon", "and", "hexagon", ")", "into", "three", "leaf", "nodes", ".", "section", ":", "Image", "Super", "-", "Resolution", "via", "Random", "Forest", "In", "the", "inference", "stage", ",", "each", "decision", "tree", "returns", "a", "class", "probability", "(", "|", ")", "for", "a", "given", "test", "sample", "∈", ",", "and", "the", "final", "class", "label", "*", "is", "then", "obtained", "via", "averaging", ",", "as", "follows", ":", "A", "splitting", "function", "(", ";", "Θ", ")", "is", "typically", "parameterized", "by", "two", "values", ":", "(", "i", ")", "a", "feature", "dimensional", "index", ":", "Θ", "{1", ",", ".", ".", ".", ",", "}", ",", "and", "(", "ii", ")", "a", "threshold", "Θ", "ℝ.", "The", "splitting", "function", "is", "defined", "as", "follows", ":", "where", "the", "outcome", "defines", "to", "which", "child", "node", "is", "routed", ",", "and", "0", "and", "1", "are", "the", "two", "labels", "belonging", "to", "the", "left", "and", "right", "child", "node", ",", "respectively", ".", "Each", "node", "chooses", "the", "best", "splitting", "function", "Θ", "*", "out", "of", "a", "randomly", "sampled", "set", "{", "Θ", "}", ",", "and", "the", "threshold", "Θ", "is", "determined", "by", "optimizing", "the", "following", "function", ":", "where", "and", "are", "the", "sets", "of", "samples", "that", "are", "routed", "to", "the", "left", "and", "right", "child", "nodes", ",", "respectively", ",", "and", "|", "|", "represents", "the", "number", "of", "samples", "in", "the", "set", ".", "During", "the", "training", "of", "an", "RF", ",", "the", "decision", "trees", "are", "provided", "with", "a", "random", "subset", "of", "the", "training", "data", "(", "i.e.", "bagging", ")", ",", "and", "are", "trained", "independently", ".", "Training", "a", "single", "decision", "tree", "involves", "recursively", "splitting", "each", "node", ",", "such", "that", "the", "training", "data", "in", "the", "newly", "created", "child", "node", "is", "clustered", "conforming", "to", "class", "labels", ".", "Each", "tree", "is", "grown", "until", "a", "stopping", "criterion", "is", "reached", "(", "e.g.", "the", "number", "of", "samples", "in", "a", "node", "is", "less", "than", "a", "threshold", "or", "the", "tree", "depth", "reaches", "a", "maximum", "value", ")", "and", "the", "class", "probability", "distributions", "are", "estimated", "in", "the", "leaf", "nodes", ".", "After", "fulfilling", "one", "of", "the", "stopping", "criteria", ",", "the", "density", "model", "(", ")", "in", "each", "leaf", "node", "is", "estimated", "by", "using", "all", "the", "samples", "falling", "into", "the", "leaf", "node", ",", "which", "will", "be", "used", "as", "a", "prediction", "of", "class", "probabilities", "in", "the", "inference", "stage", ".", "A", "simple", "way", "to", "estimate", "the", "probability", "distribution", "function", "(", ")", "is", "by", "averaging", "all", "the", "samples", "in", "the", "leaf", "node", ",", "and", "there", "are", "many", "variants", ",", "such", "as", "fitting", "a", "Gaussian", "distribution", ",", "kernel", "density", "estimation", ",", "etc", ".", "In", "(", "9", ")", ",", "(", ")", "is", "the", "local", "score", "for", "a", "set", "of", "samples", "in", "S", "(", "S", "is", "either", "L", "or", "R", ")", ",", "which", "is", "usually", "calculated", "by", "entropy", ",", "as", "shown", "in", "Eqn", ".", "[", "reference", "]", ",", "and", "it", "can", "be", "replaced", "by", "variance", "[", "reference", "][", "reference", "][", "reference", "]", "or", "by", "the", "Gini", "index", "[", "reference", "]", ".", "where", "is", "the", "number", "of", "classes", ",", "and", "(", "|", ")", "is", "the", "probability", "for", "class", ",", "which", "is", "estimated", "from", "the", "set", ".", "For", "the", "regression", "problem", ",", "the", "differential", "entropy", "is", "used", ",", "and", "is", "defined", "as", ",", "where", "(", "|", ")", "denotes", "the", "conditional", "probability", "of", "a", "target", "variable", "given", "an", "input", "sample", ".", "Assuming", "that", "(", ".", ",", ".", ")", "is", "of", "Gaussian", "distribution", ",", "and", "has", "only", "a", "set", "of", "finite", "samples", ",", "the", "differential", "entropy", "can", "be", "written", "as", ",", "where", "det", "(", "Σ", ")", "is", "the", "determinant", "of", "the", "estimated", "covariance", "matrix", "of", "the", "target", "variables", "in", ".", "RF", "-", "based", "approaches", "hold", "some", "properties", ",", "which", "make", "them", "powerful", "classifiers", "as", "SVM", "(", "support", "vector", "machine", ")", "[", "reference", "]", "and", "AdaBoost", "(", "short", "for", "\"", "Adaptive", "Boosting", "\"", ")", "[", "reference", "]", ".", "Both", "SVM", "and", "AdaBoost", "work", "as", "to", "approximate", "the", "Bayes", "decision", "rule", "-", "known", "to", "be", "the", "optimal", "classifiers", "-", "via", "minimizing", "a", "margin", "-", "based", "global", "loss", "function", ".", "RF", "-", "based", "image", "super", "-", "resolution", "(", "SR", ")", ",", "following", "a", "recent", "emerging", "stream", "[", "reference", "][", "reference", "]", "on", "single", "-", "image", "SR", ",", "formulates", "the", "SR", "problem", "as", "a", "clustering", "-", "regression", "problem", ".", "These", "emerging", "approaches", "attempt", "to", "reconstruct", "an", "HR", "image", "from", "patches", "with", "the", "aid", "of", "an", "external", "database", ".", "These", "methods", "first", "decompose", "an", "image", "into", "patches", ",", "then", "classify", "the", "patches", "into", "different", "clusters", ",", "and", "later", "regressors", "are", "trained", "for", "all", "the", "clusters", "respectively", ",", "which", "generate", "mappings", "from", "an", "input", "LR", "patch", "'s", "features", "to", "its", "corresponding", "HR", "patch", ".", "In", "the", "inference", "stage", ",", "an", "LR", "image", "follows", "the", "same", "procedures", ",", "such", "that", "it", "is", "divided", "into", "patches", "and", "features", "are", "extracted", "from", "each", "patch", ".", "Then", ",", "the", "patches", "are", "classified", "into", "different", "clusters", "using", "K", "-", "NN", "[", "reference", "][", "reference", "]", "or", "RF", "[", "reference", "][", "reference", "][", "reference", "]", ",", "and", "their", "super", "-", "resolved", "HR", "patches", "are", "computed", "through", "regression", "in", "the", "leaf", "nodes", "(", "see", "Fig", ".", "1", ")", ".", "This", "kind", "of", "clustering", "-", "regression", "-", "based", "random", "forest", "[", "reference", "][", "reference", "][", "reference", "]", "methods", "has", "achieved", "state", "-", "of", "-", "the", "-", "art", "performance", "in", "SISR", ",", "both", "in", "terms", "of", "accuracy", "and", "efficiency", ".", "section", ":", "FEATURE", "-", "AUGMENTED", "RANDOM", "FOREST", "Classification", "and", "regression", "can", "be", "regarded", "as", "probability", "problems", "from", "the", "statistical", "theory", ".", "Historical", "frequentist", "probability", "is", "the", "probability", "obtained", "from", "the", "relative", "frequency", "in", "a", "large", "number", "of", "trials", ".", "In", "contrast", ",", "the", "Bayesian", "probability", "is", "an", "interpretation", "of", "the", "concept", "of", "probability", ",", "in", "which", "probability", "is", "interpreted", "as", "an", "expectation", "taking", "the", "knowledge", "and", "personal", "belief", "into", "account", ".", "From", "the", "Bayesian", "theory", ",", "the", "posterior", "probability", "of", "a", "random", "event", "is", "a", "conditional", "probability", ",", "which", "can", "be", "calculated", "if", "the", "relevant", "evidence", "or", "context", "is", "considered", ".", "Therefore", ",", "the", "posterior", "probability", "is", "the", "probability", "(", "|", ")", "of", "the", "parameters", "given", "the", "evidence", ".", "We", "denote", "the", "probability", "distribution", "function", "of", "the", "prior", "for", "parameters", "as", "(", ")", ",", "and", "the", "likelihood", "as", "(", "|", ")", ",", "which", "is", "the", "probability", "of", "given", ".", "Then", ",", "based", "on", "the", "Bayesian", "rule", ",", "the", "posterior", "probability", "can", "be", "defined", "as", "follows", ":", "The", "posterior", "probability", "can", "be", "denoted", "in", "a", "memorable", "form", "as", ":", "Based", "on", "the", "Bayesian", "framework", ",", "the", "likelihood", "term", "and", "the", "prior", "term", "are", "both", "required", "to", "be", "determined", "in", "order", "to", "solve", "the", "inverse", "problems", ",", "and", "the", "extracted", "features", "are", "normally", "worked", "as", "prior", "or", "likelihood", ",", "particularly", "on", "some", "image", "restoration", "problems", ".", "From", "this", "point", "of", "view", ",", "most", "research", "works", ",", "from", "classic", "feature", "extractors", "to", "deep", "-", "learning", "neural", "networks", ",", "are", "essentially", "done", "under", "the", "Bayesian", "inference", "framework", ".", "Since", "SISR", "is", "a", "well", "-", "known", "ill", "-", "posed", "problem", ",", "researchers", "have", "put", "their", "efforts", "into", "the", "priors", "of", "the", "problem", "with", "skills", "from", "mathematics", ",", "computer", "vision", "and", "machine", "learning", ".", "One", "of", "the", "obvious", "and", "most", "studied", "priors", "is", "the", "edge", "prior", ",", "which", "can", "be", "found", "in", "many", "pioneering", "works", ":", "new", "edgedirected", "interpolation", "(", "NEDI", ")", "[", "reference", "]", ",", "soft", "-", "decision", "adaptive", "interpolation", "(", "SAI", ")", "[", "reference", "]", ",", "directional", "filtering", "and", "data", "-", "fusion", "(", "DFDF", ")", "[", "reference", "]", ",", "modified", "edge", "-", "directed", "interpolation", "(", "MEDI", ")", "[", "reference", "]", ",", "and", "so", "on", ".", "The", "edge", "prior", "is", "effective", "on", "image", "processing", ",", "and", "the", "first", "and", "second", "-", "order", "gradients", "are", "studied", "and", "employed", "by", "Yang", "et", "al", ".", "[", "reference", "]", "in", "a", "pioneering", "dictionary", "-", "learning", "-", "based", "algorithm", ".", "However", ",", "the", "effect", "of", "edgebased", "features", "has", "not", "been", "investigated", "in", "depth", ".", "For", "the", "clustering", "and", "classification", "problems", ",", "feature", "engineering", "is", "a", "critical", "research", "point", ",", "and", "in", "some", "cases", ",", "the", "chosen", "feature", "may", "dominate", "the", "performance", ".", "As", "shown", "in", "Eqn", ".", "(", "6", ")", ",", "a", "feature", "filter", ",", "whose", "coefficients", "are", "computed", "to", "fit", "the", "most", "relevant", "parts", "in", "the", "LR", "image", "patches", ",", "is", "employed", ",", "and", "the", "generated", "features", "can", "achieve", "more", "accurate", "predictions", "for", "reconstructing", "their", "counterpart", "HR", "image", "patches", ",", "as", "shown", "in", "Fig", ".", "3", ".", "section", ":", "Augmented", "Features", "via", "Gradient", "Magnitude", "Filters", "Normally", "it", "is", "unstable", "to", "directly", "use", "pixel", "intensities", "as", "features", ",", "which", "are", "susceptible", "to", "the", "environmental", "lighting", "variations", "and", "camera", "noise", ".", "Instead", ",", "the", "differences", "between", "the", "neighboring", "pixels", "'", "intensity", "values", ",", "which", "are", "computationally", "efficient", ",", "and", "are", "immune", "to", "lighting", "changes", "and", "noise", ",", "are", "examined", ".", "This", "type", "of", "features", "can", "be", "implemented", "efficiently", "through", "convolutional", "filters", ".", "Typically", ",", "the", "feature", "filter", "can", "be", "chosen", "as", "a", "high", "-", "pass", "filter", ",", "while", "in", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ",", "the", "first", "and", "second", "-", "order", "gradient", "operators", "are", "used", "to", "generate", "an", "up", "-", "sampled", "version", "from", "a", "low", "-", "resolution", "image", ",", "then", "four", "patches", "are", "extracted", "from", "the", "gradient", "maps", "at", "each", "location", ",", "and", "finally", "the", "patches", "are", "concatenated", "to", "form", "feature", "vectors", ".", "The", "four", "1", "-", "D", "filters", "used", "to", "extract", "the", "derivatives", "are", "described", "in", "Eqn", ".", "(", "14", ")", ",", "These", "features", "can", "work", "well", "on", "dictionary", "-", "learning", "-", "based", "methods", ",", "because", "when", "searching", "a", "matched", "patch", "in", "a", "dictionary", ",", "the", "distance", "is", "calculated", "based", "on", "the", "whole", "feature", "vectors", "with", "the", "Euclidean", "distance", ".", "However", ",", "when", "training", "a", "split", "node", "in", "a", "decision", "tree", "of", "an", "RF", ",", "only", "one", "or", "a", "few", "of", "the", "feature", "dimensions", "are", "chosen", "as", "candidate", "features", "for", "comparison", ".", "Therefore", ",", "more", "discriminative", "features", "are", "required", "for", "RF", ",", "when", "compared", "with", "dictionary", "-", "learning", "-", "based", "methods", ".", "The", "first", "and", "second", "-", "order", "gradients", "of", "an", "image", "can", "provide", "the", "directions", "of", "edges", "in", "a", "perceptual", "manner", "as", "shown", "in", "Fig", ".", "4", "and", "Fig", ".", "5", ",", "which", "can", "be", "calculated", "as", "Eqn", ".", "(", "15", ")", ",", "where", "/", "and", "/", "are", "the", "gradients", "in", "the", "x", "-", "axis", "and", "y", "-", "axis", "directions", ",", "respectively", ",", "at", "a", "given", "pixel", ".", "Meanwhile", ",", "the", "gradient", "magnitude", "image", "can", "provide", "the", "edge", "strength", ",", "as", "described", "in", "Eqn", ".", "[", "reference", "]", ".", "Fig", ".", "4", "shows", "a", "toy", "example", "of", "a", "man", "-", "made", "\"", "circle", "\"", "image", ",", "to", "demonstrate", "its", "discriminative", "property", ".", "With", "a", "natural", "image", "shown", "in", "Fig", ".", "5", ",", "it", "can", "be", "observed", "that", "the", "gradient", "magnitude", "image", "has", "more", "detailed", "textures", "than", "the", "gradient", "images", "(", "/", "and", "/", ")", ",", "as", "well", "as", "the", "sum", "of", "the", "horizontal", "gradient", "and", "vertical", "gradient", "image", ",", "i.e.", "/", "+", "/", ",", "perceptually", ".", "An", "explanation", "for", "this", "phenomenon", "is", "that", "non", "-", "linear", "features", "are", "usually", "more", "discriminative", ".", "Thus", ",", "in", "our", "work", ",", "all", "the", "first", "and", "second", "-", "order", "gradients", ",", "and", "gradient", "magnitude", "are", "employed", ",", "and", "are", "concatenated", "to", "form", "more", "discriminative", ",", "augmented", "features", ".", "On", "the", "other", "hand", ",", "the", "image", "orientation", "(", "gradient", "angle", ")", "is", "defined", "by", "the", "following", "formulation", ",", "where", "atan", "(", ")", "is", "the", "gradient", "orientation", ",", "with", "a", "value", "between", "-", "90", "and", "90.", "As", "shown", "in", "Eqn", ".", "(", "17", ")", ",", "when", "the", "value", "of", "is", "equal", "to", "0", "or", "close", "to", "0", ",", "the", "value", "of", "∠∇", "becomes", "infinitely", "large", "and", "unstable", ",", "i.e.", ",", "different", "will", "result", "in", "approximately", "the", "same", "∠∇", "value", ".", "Based", "on", "this", "analysis", ",", "we", "only", "use", "the", "two", "gradient", "magnitude", "filters", "derived", "from", "the", "four", "gradient", "filters", "[", "reference", "]", "to", "generate", "the", "augmented", "features", ".", "Experiments", "validate", "that", "the", "use", "of", "the", "augmented", "features", "can", "improve", "the", "conventional", "RF", "algorithm", "[", "reference", "]", "to", "achieve", "a", "performance", "gain", "of", "more", "than", "0.1dB", ",", "which", "is", "a", "remarkable", "improvement", ",", "with", "the", "same", "setting", "and", "parameters", ".", "section", ":", "Fine", "-", "grained", "Features", "for", "Regression", "The", "inference", "stage", "of", "the", "RF", "-", "based", "image", "super", "-", "resolution", "process", "is", "similar", "to", "the", "content", "-", "based", "image", "retrieval", "(", "CBIR", ")", "framework", ",", "as", "shown", "in", "Fig", ".", "1", ".", "The", "general", "approximated", "nearest", "neighbor", "(", "ANN", ")", "search", "framework", "[", "reference", "][", "reference", "]", "is", "an", "efficient", "strategy", "for", "large", "-", "scale", "image", "retrieval", ",", "which", "mainly", "consists", "of", "4", "parts", ":", "(", "1", ")", "extracting", "compact", "features", "(", "e.g.", ",", "locality", "-", "sensitive", "Hashing", "(", "LSH", ")", "[", "reference", "]", "feature", ")", "for", "a", "query", "image", ";", "(", "2", ")", "coarse", "-", "level", "search", "using", "Hamming", "distance", "to", "measure", "the", "similarity", "between", "binary", "compact", "Hash", "features", ",", "then", "narrow", "the", "search", "scope", "into", "a", "smaller", "candidate", "group", ";", "(", "3", ")", "fine", "-", "level", "search", "by", "using", "Euclidean", "distance", "to", "measure", "the", "similarity", "between", "their", "corresponding", "feature", "vectors", ";", "and", "(", "4", ")", "finding", "the", "object", "in", "the", "smaller", "candidate", "group", "that", "is", "the", "nearest", "one", "to", "the", "query", "image", ".", "In", "the", "inference", "stage", "of", "conventional", "RF", "-", "based", "SISR", ",", "PCA", "projection", "is", "worked", "as", "a", "Hash", "-", "like", "function", "to", "compress", "the", "feature", "dimension", "for", "decreasing", "the", "search", "range", ",", "which", "can", "speed", "up", "the", "searching", "as", "the", "coarse", "-", "level", "search", "in", "a", "CBIR", "framework", ",", "but", "the", "impact", "of", "using", "PCA", "on", "feature", "dimensionality", "reduction", "has", "been", "overlooked", "in", "previous", "works", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "Inspired", "by", "the", "finelevel", "search", "using", "augmented", "features", "in", "CBIR", "frameworks", ",", "the", "high", "dimensional", "features", "in", "the", "leaf", "nodes", "in", "an", "RF", "can", "further", "improve", "the", "prediction", "accuracy", "in", "the", "regression", "step", ",", "which", "has", "not", "been", "studied", "previously", ".", "Consequently", ",", "we", "use", "the", "original", "features", ",", "rather", "than", "PCA", "or", "the", "LSH", "compressed", "features", ",", "to", "perform", "ridge", "regression", "in", "the", "leaf", "nodes", ".", "Experimental", "results", "show", "that", "the", "new", "RF", "scheme", "can", "greatly", "improve", "the", "quality", "of", "super", "-", "resolved", "images", ",", "by", "using", "this", "augmented", "feature", ".", "Another", "explanation", "for", "this", "is", "that", "the", "regression", "problems", "can", "benefit", "more", "from", "higher", "dimensional", "features", "than", "classification", "problems", ".", "Based", "on", "the", "observation", "that", "the", "original", "edge", "-", "like", "features", "are", "used", "for", "the", "final", "regressors", "in", "the", "leaf", "nodes", "and", "the", "compressed", "features", "(", "either", "produced", "by", "PCA", "or", "LSH", ")", "are", "used", "for", "clustering", "in", "the", "split", "nodes", ",", "a", "new", "clustering", "-", "regression", "-", "based", "SISR", "approach", "can", "be", "designed", "as", "shown", "in", "Fig", ".", "6", ".", "In", "this", "new", "scheme", ",", "the", "original", "-", "compressed", "coupled", "feature", "sets", "are", "worked", "for", "different", "purposes", "at", "different", "stages", ",", "i.e.", ",", "the", "original", "edge", "features", "are", "used", "for", "regression", "in", "the", "leaf", "nodes", ",", "and", "the", "compressed", "features", "derived", "from", "the", "LSH", "-", "like", "functions", "are", "employed", "for", "node", "splitting", "(", "clustering", ")", "in", "the", "training", "stage", ",", "and", "node", "searching", "in", "the", "inference", "stage", "in", "the", "split", "nodes", ".", "section", ":", "Fig", ".", "6", ":", "Augmented", "features", "for", "regressors", "and", "the", "LSH", "compressed", "features", "for", "searching", "in", "a", "random", "forest", "In", "the", "new", "scheme", ",", "we", "unify", "the", "research", "of", "LSH", "-", "based", "SISR", "and", "image", "retrieval", "(", "CBIR", ")", "[", "reference", "][", "reference", "]", ".", "In", "brief", ",", "the", "new", "achievement", "on", "unsupervised", "LSH", "can", "be", "evaluated", "not", "only", "in", "CBIR", "systems", ",", "but", "also", "in", "the", "clustering", "-", "regression", "RF", "-", "based", "SISR", "methods", ".", "Moreover", ",", "as", "evidence", "from", "[", "reference", "]", ",", "proper", "unsupervised", "LSH", "models", ",", "e.g.", ",", "iterative", "quantization", "(", "ITQ", ")", "[", "reference", "]", "used", "for", "feature", "dimension", "reduction", "instead", "of", "PCA", ",", "can", "reduce", "the", "damage", "on", "the", "image", "structure", ".", "This", "can", "further", "improve", "the", "superresolved", "image", "quality", ".", "Different", "from", "[", "reference", "]", "using", "an", "ITQ", "-", "like", "algorithm", "to", "rotate", "the", "original", "features", "into", "a", "new", "feature", "space", ",", "with", "the", "use", "of", "the", "proposed", "original", "-", "compressed", "coupled", "feature", "sets", ",", "any", "unsupervised", "LSH", "generated", "features", "can", "directly", "be", "employed", ".", "section", ":", "Generalized", "Weighted", "Ridge", "Regression", "Model", "In", "this", "sub", "-", "section", ",", "we", "further", "analyze", "the", "ridge", "regression", "employed", "in", "the", "RF", "leaf", "nodes", ".", "The", "anchored", "neighborhood", "regression", "(", "ANR", ")", "[", "reference", "]", "model", "relaxes", "the", "1", "-", "norm", "in", "Eqn", ".", "(", "6", ")", "to", "the", "2", "-", "norm", "constraint", ",", "with", "least", "-", "squares", "minimization", "as", "the", "following", "equation", ",", "Based", "on", "the", "ridge", "regression", "[", "reference", "]", "theory", ",", "this", "2", "-", "norm", "constrained", "least", "square", "regression", "regularized", "problem", "has", "a", "closed", "-", "form", "solution", ",", "according", "to", "the", "Tikhonov", "regularization", "theory", ",", "as", "follows", ":", "With", "the", "assumption", "in", "[", "reference", "]", ",", "where", "HR", "patches", "and", "their", "counterpart", "LR", "patches", "share", "the", "same", "reconstructed", "coefficient", "α", ",", "i.e.", "=", "D", "ℎ", ",", "from", "Eqn", ".", "[", "reference", "]", "we", "have", "If", "we", "define", "as", "a", "pre", "-", "calculated", "projection", "matrix", ",", "as", "follows", ",", "then", "the", "HR", "patches", "can", "be", "reconstructed", "with", "=", ".", "Having", "studied", "the", "model", "in", "Eqn", ".", "(", "18", ")", ",", "the", "authors", "in", "[", "reference", "]", "argued", "that", "different", "weights", "should", "be", "given", "to", "different", "atoms", "when", "reconstructing", "an", "HR", "patch", "so", "as", "to", "emphasize", "the", "similarity", "to", "the", "anchor", "atom", ".", "Based", "on", "this", "idea", ",", "[", "reference", "]", "proposed", "a", "weighted", "collaborative", "representation", "(", "WCR", ")", "model", "by", "generalizing", "the", "normal", "collaborative", "representation", "(", "CR", ")", "model", "in", "the", "ANR", ",", "where", "is", "a", "diagonal", "weight", "matrix", ",", "in", "which", "the", "non", "-", "zero", "entries", "are", "proportional", "to", "the", "similarities", "between", "the", "atoms", "and", "the", "anchor", "atom", ".", "Same", "as", "the", "ANR", "model", ",", "a", "new", "closed", "-", "form", "solution", "can", "be", "computed", "offline", "through", "the", "following", "and", "the", "new", "projection", "matrix", "can", "be", "derived", "as", "The", "WCR", "model", "further", "improves", "the", "ANR", "/", "A", "+", "model", "in", "terms", "of", "image", "quality", ",", "while", "keeping", "the", "same", "level", "of", "computation", ".", "In", "[", "reference", "]", ",", "the", "local", "geometry", "prior", "of", "the", "data", "sub", "-", "space", "is", "used", ".", "However", ",", "all", "the", "weighted", "ridge", "regression", "models", "[", "reference", "][", "reference", "]", "are", "constructed", "based", "on", "an", "existing", "dictionary", ",", "e.g.", ",", "Zeyde", "et", "al", ".", "[", "reference", "]", "used", "K", "-", "SVD", "to", "train", "a", "sparse", "-", "coding", "-", "based", "dictionary", "with", "1024", "items", ".", "This", "limits", "the", "models", "to", "collect", "samples", "in", "a", "smaller", "sub", "-", "space", "when", "constructing", "linear", "regressors", "based", "on", "existing", "anchor", "points", ".", "section", ":", "Fig", ".", "7", ":", "Gaussian", "mixture", "model", "(", "GMM", ")", "is", "used", "to", "generate", "the", "weights", "for", "weighted", "ridge", "regression", ",", "and", "the", "weight", "of", "each", "entry", "lies", "on", "its", "belonging", "cluster", "'s", "weight", "and", "its", "weight", "in", "the", "belonging", "cluster", ".", "When", "training", "the", "regressors", "in", "an", "RF", ",", "there", "is", "no", "existing", "anchor", "point", "in", "the", "clustered", "groups", "of", "the", "leaf", "nodes", ",", "similar", "to", "the", "previous", "models", "[", "reference", "][", "reference", "]", ".", "A", "solution", "to", "mentioned", "problem", "is", "inspired", "from", "the", "work", "on", "image", "classification", "using", "locality", "-", "constrained", "linear", "coding", "(", "LLC", ")", "[", "reference", "]", ",", "where", "Gaussian", "mixture", "model", "(", "GMM", ")", "is", "used", "to", "describe", "the", "locality", "-", "constrained", "affine", "subspace", "coding", "(", "LASC", ")", "[", "reference", "]", ".", "We", "employ", "GMM", "to", "construct", "the", "data", "distribution", "in", "the", "sub", "-", "space", "for", "each", "leaf", "node", ",", "which", "derives", "the", "weights", "of", "all", "the", "entries", "in", "the", "ridge", "regression", "models", ".", "Through", "the", "derived", "weights", ",", "we", "can", "obtain", "a", "generalized", "weighted", "ridge", "regression", "(", "GWRR", ")", "model", "for", "ridge", "regression", ".", "The", "new", "projection", "matrix", "is", "given", "as", "follows", ":", "where", "is", "a", "diagonal", "weight", "matrix", ",", "and", "the", "weight", "of", "each", "diagonal", "entry", "is", "related", "to", "its", "belonging", "cluster", "'s", "weight", "and", "its", "local", "weight", "in", "its", "belongingwhi", "cluster", ",", "as", "illustrated", "in", "the", "right", "part", "of", "Fig", ".", "7", ".", "Obviously", ",", "a", "query", "entry", "falling", "into", "a", "bigger", "cluster", "and", "closer", "to", "the", "center", "of", "the", "belonging", "cluster", "achieves", "a", "larger", "weight", ".", "In", "a", "rough", "form", ",", "the", "diagonal", "weight", "matrix", "is", "given", "as", "follows", ":", "where", "is", "the", "weight", "of", "the", "th", "entry", ",", "is", "number", "of", "samples", "in", "the", "leaf", "nodes", ",", "is", "the", "th", "cluster", "'s", "weight", "for", "the", "th", "entry", ",", "is", "the", "th", "entry", "'s", "local", "weight", "in", "the", "th", "cluster", ",", "which", "is", "approximated", "with", "the", "inverse", "value", "of", "the", "distance", "to", "the", "center", "of", "the", "belonging", "cluster", ",", "and", "is", "the", "number", "of", "clusters", "generated", "by", "the", "GMM", "model", "for", "a", "leaf", "node", ".", "Experimental", "results", "in", "Table", "-", "1", "show", "that", "the", "proposed", "GWRR", "model", "can", "achieve", "the", "same", "level", "of", "performance", "as", "WCR", "[", "reference", "]", ",", "and", "obtain", "0.2dB", "gain", "more", "than", "the", "ANR", "[", "1", "]", "model", ".", "Note", "that", "when", "the", "number", "of", "samples", "in", "a", "leaf", "node", "becomes", "bigger", ",", "the", "performance", "of", "the", "GWRR", "model", "will", "achieve", "less", "advantage", "than", "the", "normal", "regression", "model", ",", "because", "the", "higher", "weights", "will", "be", "averaged", "by", "a", "large", "number", "of", "other", "samples", ".", "Theoretically", ",", "the", "regression", "of", "a", "leaf", "node", "can", "benefit", "from", "the", "GWRR", "model", ",", "particularly", "when", "there", "are", "a", "few", "samples", "falling", "into", "the", "leaf", "node", ".", "section", ":", "Initial", "Estimation", "with", "Iterative", "Back", "Projection", "Generally", "speaking", ",", "SISR", "is", "a", "low", "-", "level", "computer", "vision", "task", ",", "which", "attempts", "to", "restore", "an", "HR", "image", "from", "a", "single", "input", "LR", "image", ".", "A", "mathematical", "model", "for", "image", "degradation", "can", "be", "formulated", "as", "follows", ":", "where", "ℬ", "is", "a", "low", "-", "pass", "(", "blur", ")", "filter", "and", "", "denotes", "the", "down", "-", "sampling", "operator", "with", "factor", ".", "Based", "on", "a", "given", "LR", "image", ",", "how", "to", "achieve", "an", "approximated", "HR", "image", "̂", "is", "a", "classic", "inverse", "problem", ",", "which", "requires", "priors", "based", "on", "the", "Bayesian", "theory", ".", "Irani", "and", "Peleg", "[", "reference", "]", "firstly", "proposed", "an", "iterative", "back", "projection", "(", "IBP", ")", "method", "for", "SR", "reconstruction", ",", "and", "IBP", "is", "the", "most", "effective", "way", "to", "obtain", "an", "HR", "image", "when", "comparing", "it", "with", "other", "SR", "methods", ".", "In", "the", "IBP", "method", ",", "the", "reconstruction", "error", "of", "an", "estimated", "LR", "image", "̂", "is", "the", "difference", "between", "the", "input", "LR", "and", "the", "synthesized", "image", "̂", "generated", "from", "the", "estimated", "HR", "image", "̂", "as", "follows", ":", "IBP", "is", "an", "efficient", "approach", "to", "obtain", "the", "HR", "image", "by", "minimizing", "the", "reconstruction", "error", "defined", "by", "Eqn", ".", "[", "reference", "]", ".", "For", "the", "IBP", "approach", "on", "SISR", ",", "the", "updating", "procedure", "can", "be", "summarized", "as", "the", "following", "two", "steps", ",", "performed", "iteratively", ":", "•", "Compute", "the", "reconstruction", "error", "(", "̂", ")", "with", "the", "following", "equation", ":", "where", "↑", "is", "the", "up", "-", "sampling", "operator", "and", "is", "a", "constant", "back", "-", "projection", "kernel", "to", "approximate", "the", "inverse", "operation", "of", "the", "low", "-", "pass", "filter", "ℬ.", "•", "Update", "the", "estimating", "HR", "image", "̂", "by", "back", "-", "projecting", "errors", "as", "follows", ":", "where", "̂", "is", "the", "estimated", "HR", "image", "at", "the", "-", "th", "iteration", ".", "Most", "learning", "-", "based", "algorithms", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "follow", "the", "milestone", "work", "in", "[", "reference", "]", ",", "which", "uses", "the", "coarse", "estimation", "firstly", "obtained", "via", "bicubic", "interpolation", ".", "As", "we", "know", ",", "the", "classic", "IBP", "algorithm", "is", "an", "efficient", "way", "to", "obtain", "high", "-", "quality", "up", "-", "scaled", "images", ",", "but", "it", "will", "inevitably", "produce", "artifacts", "(", "such", "as", "ringing", ",", "jaggy", "effects", ",", "and", "noise", ")", "at", "the", "output", ",", "because", "the", "kernel", "operator", "in", "Eqn", ".", "(", "29", ")", "is", "hard", "to", "estimate", "accurately", ".", "That", "is", "the", "reason", "why", "algorithms", "with", "IBP", "need", "an", "additional", "denoising", "process", "[", "reference", "][", "reference", "][", "reference", "]", ".", "However", ",", "the", "sparse", "-", "constraint", "-", "based", "approach", "[", "reference", "]", "does", "not", "have", "this", "denoising", "capability", ".", "As", "the", "2", "-", "norm", "constraint", "-", "based", "ridge", "regression", "has", "the", "denoising", "effect", ",", "due", "to", "its", "averaging", "-", "like", "process", ",", "this", "means", "that", "the", "ridge", "regression", "-", "based", "RF", "scheme", "has", "the", "denoise", "capability", "intrinsically", ".", "Based", "on", "this", "observation", ",", "we", "obtain", "the", "coarse", "estimation", "of", "an", "HR", "image", "̂", "by", "applying", "IBP", "to", "the", "corresponding", "input", "LR", "image", ".", "Experimental", "results", "in", "Table", "-", "2", "and", "Table", "-", "3", "validate", "that", "using", "IBP", ",", "instead", "of", "bicubic", ",", "to", "obtain", "the", "initial", "coarse", "estimation", "can", "help", "the", "RF", "-", "based", "SR", "method", "obtain", "a", "remarkable", "improvement", ".", "As", "the", "number", "of", "trees", "is", "an", "important", "parameter", "in", "RF", "-", "based", "approaches", ",", "we", "plot", "the", "performance", "with", "respect", "to", "the", "number", "of", "trees", ".", "As", "shown", "in", "Fig", ".", "8", ",", "the", "performance", "of", "the", "RF", "-", "based", "image", "superresolution", "method", "increases", "as", "expected", ",", "but", "the", "increment", "becomes", "relatively", "smaller", "after", "a", "certain", "number", "of", "trees", "are", "used", ".", "The", "experimental", "results", "in", "Fig", ".", "8", "section", ":", "Fine", "-", "Tuning", "with", "Proper", "Trees", "in", "Random", "Forest", "section", ":", "PSNR", "section", ":", "Set14", "million", "samples", "from", "the", "dataset", "are", "used", "for", "all", "training", "stages", ".", "It", "shows", "that", "using", "45", "trees", "is", "an", "optimal", "number", ",", "as", "a", "trade", "-", "off", "between", "performance", "and", "computational", "cost", ".", "Therefore", ",", "we", "set", "the", "number", "of", "trees", "for", "the", "proposed", "FARF", "method", "at", "45", ",", "and", "our", "method", "with", "this", "number", "is", "denoted", "as", "FARF*.", "The", "performances", "of", "our", "methods", ",", "and", "other", "methods", ",", "are", "tabulated", "in", "Table", "-", "2", "and", "Table", "-", "3", ".", "We", "also", "compare", "our", "methods", "with", "a", "recently", "proposed", "deep", "-", "learning", "-", "based", "algorithm", ",", "SRCNN", "algorithm", "[", "reference", "][", "reference", "]", ",", "and", "our", "methods", "outperform", "it", "in", "some", "cases", ".", "section", ":", "Algorithm", "Workflow", "The", "training", "and", "inference", "stages", "of", "the", "proposed", "FARF", "algorithm", "are", "described", "in", "Algorithm", "1", "and", "Algorithm", "2", ",", "respectively", ".", "To", "help", "the", "readers", "understand", "our", "paper", ",", "the", "source", "code", "of", "our", "algorithm", "will", "be", "available", "at", ":", "https:", "//", "github.com", "/", "HarleyHK", "/", "FARF", ",", "for", "reference", ".", "section", ":", "EXPERIMENTS", "In", "this", "section", ",", "we", "evaluate", "our", "algorithm", "on", "standard", "super", "-", "resolution", "benchmarks", "Set", "5", ",", "Set14", "and", "B100", "[", "reference", "]", ",", "and", "compare", "it", "with", "some", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "They", "are", "bicubic", "interpolation", ",", "adjusted", "anchored", "neighborhood", "regression", "(", "A", "+", ")", "[", "reference", "]", ",", "standard", "RF", "[", "reference", "]", ",", "alternating", "regression", "forests", "(", "ARF", ")", "[", "reference", "]", ",", "and", "the", "convolutional", "neural", "-", "network", "-", "based", "image", "super", "-", "resolution", "(", "SRCNN", ")", "[", "reference", "][", "reference", "]", ",", "as", "listed", "in", "Table", "-", "2", "and", "Table", "-", "Table", "-", "2", ":", "Results", "of", "the", "proposed", "method", "compared", "with", "state", "-", "of", "-", "the", "-", "art", "works", "on", "3", "datasets", "in", "terms", "of", "PSNR", "(", "dB", ")", "using", "three", "different", "magnification", "factors", "(", "#", ")", "(", "×2", ",", "×3", ",", "×4", ")", ".", "Table", "-", "2", "summarizes", "the", "performances", "of", "our", "proposed", "algorithm", "on", "the", "3", "datasets", ",", "in", "terms", "of", "the", "average", "peak", "signal", "to", "noise", "ratio", "(", "PSNR", ")", "scores", ",", "with", "different", "magnification", "factors", "(", "×2", ",", "×3", ",", "×4", ")", ".", "The", "objective", "quality", "metric", ",", "PSNR", ",", "in", "Table", "-", "2", "also", "shows", "that", "the", "fine", "-", "tuned", "FARF", ",", "i.e.", "FARF", "*", ",", "can", "further", "improve", "the", "image", "quality", ",", "which", "is", "comparable", "to", "recently", "proposed", "state", "-", "of", "-", "the", "-", "art", "deeplearning", "-", "based", "algorithms", ",", "such", "as", "SRCNN", "[", "reference", "][", "reference", "]", ".", "Comparing", "our", "proposed", "FARF", "algorithm", "to", "other", "methods", ",", "the", "improved", "visual", "quality", "of", "our", "results", "is", "obvious", ",", "as", "shown", "in", "Fig", ".", "9", ".", "This", "shows", "that", "our", "method", "can", "produce", "more", "details", ",", "particularly", "on", "some", "texture", "-", "rich", "regions", ".", "Fig", ".", "9", ":", "Super", "-", "resolution", "(", "×3", ")", "images", "from", "B100", ",", "bicubic", ",", "A", "+", "(", "ACCV", "-", "2014", ")", "[", "reference", "]", ",", "ARF", "(", "CVPR", "-", "2015", ")", "[", "reference", "]", ",", "SRCNN", "(", "PAMI", "-", "2016", ")", "[", "reference", "]", ",", "our", "proposed", "algorithm", "FARF", ",", "and", "ground", "truth", ".", "The", "results", "show", "that", "our", "FARF", "algorithm", "can", "produce", "more", "details", "and", "its", "performance", "is", "comparable", "to", "a", "recent", "state", "-", "of", "-", "the", "-", "art", "deep", "-", "learning", "method", "[", "reference", "]", ".", "section", ":", "CONCLUSIONS", "This", "paper", "presents", "a", "feature", "-", "augmented", "random", "forest", "(", "FARF", ")", "scheme", "for", "the", "single", "image", "superresolution", "(", "SISR", ")", "task", "by", "augmenting", "features", "and", "redesigning", "the", "inner", "structure", "of", "a", "random", "forest", "(", "RF", ")", ",", "with", "different", "feature", "recipes", "at", "different", "stages", ",", "where", "the", "compressed", "features", "are", "used", "for", "clustering", "in", "the", "split", "nodes", "and", "the", "original", "features", "are", "used", "for", "regression", "in", "the", "leaf", "nodes", ".", "The", "contributions", "of", "this", "paper", "are", "threefold", ":", "(", "1", ")", "the", "more", "discriminative", "gradient", "magnitude", "-", "based", "augmented", "features", "are", "proposed", "for", "clustering", "on", "split", "nodes", "and", "regression", "on", "leaf", "nodes", ";", "(", "2", ")", "By", "extending", "principal", "component", "analysis", "(", "PCA", ")", "to", "a", "generalized", "unsupervised", "locality", "-", "sensitive", "hashing", "(", "LSH", ")", "model", "for", "dimensionality", "reduction", ",", "we", "lay", "out", "an", "original", "compressed", "coupled", "feature", "set", "for", "tackling", "the", "clustering", "-", "regression", "tasks", ",", "which", "unify", "SISR", "and", "content", "-", "based", "image", "retrieval", "(", "CBIR", ")", "for", "LSH", "evaluation", ";", "and", "(", "3", ")", "we", "have", "extended", "WCR", "model", "to", "a", "generalized", "GWRR", "model", "for", "ridge", "regression", ".", "The", "proposed", "FAFR", "scheme", "can", "achieve", "highly", "competitive", "quality", "results", ",", "e.g.", ",", "obtaining", "about", "a", "0.3dB", "gain", "in", "PSNR", ",", "on", "average", ",", "when", "compared", "to", "conventional", "RF", "-", "based", "super", "-", "resolution", "approaches", ".", "Furthermore", ",", "a", "fine", "-", "tuned", "version", "of", "our", "proposed", "FARF", "approach", "is", "provided", ",", "whose", "performance", "is", "comparable", "to", "recent", "state", "-", "of", "-", "the", "-", "art", "deep", "-", "learning", "-", "based", "algorithms", ".", "section", ":"], "entity": [[0, 4, ["Image", "Super", "-", "resolution"], "Task"], [5, 10, ["Feature", "-", "augmented", "Random", "Forest"], "Method"], [14, 18, ["Recent", "random", "-", "forest"], "Method"], [19, 20, ["RF"], "Method"], [31, 37, ["dictionary", "-", "learning", "-", "based", "algorithms"], "Method"], [45, 46, ["RF"], "Method"], [60, 65, ["feature", "-", "augmented", "random", "forest"], "Method"], [66, 67, ["FARF"], "Method"], [69, 73, ["image", "super", "-", "resolution"], "Task"], [77, 81, ["gradient", "-", "based", "features"], "Method"], [97, 98, ["RF"], "Method"], [110, 116, ["dictionary", "-", "learning", "-", "based", "features"], "Method"], [154, 157, ["principal", "component", "analysis"], "Method"], [158, 159, ["PCA"], "Method"], [161, 164, ["feature", "dimensionality", "reduction"], "Task"], [199, 204, ["compressed", "coupled", "feature", "sets", "scheme"], "Method"], [211, 215, ["image", "super", "-", "resolution"], "Task"], [216, 221, ["content", "-", "based", "image", "retrieval"], "Method"], [301, 307, ["deep", "-", "learning", "-", "based", "algorithms"], "Method"], [317, 319, ["random", "forest"], "Method"], [320, 321, ["RF"], "Method"], [329, 333, ["machine", "-", "learning", "tool"], "Method"], [337, 342, ["ensemble", "of", "multiple", "decision", "trees"], "Method"], [348, 352, ["classification", "or", "regression", "problems"], "Task"], [360, 364, ["computer", "-", "vision", "applications"], "Task"], [367, 369, ["object", "recognition"], "Task"], [373, 375, ["face", "alignment"], "Task"], [389, 394, ["single", "image", "super", "-", "resolution"], "Task"], [395, 396, ["SISR"], "Task"], [418, 420, ["binary", "trees"], "Method"], [441, 446, ["ensemble", "of", "multiple", "decision", "trees"], "Method"], [471, 474, ["parallel", "processing", "technology"], "Method"], [478, 482, ["training", "and", "inference", "stages"], "Task"], [488, 493, ["sub", "-", "linear", "search", "complexity"], "Metric"], [508, 510, ["bagging", "strategy"], "Method"], [529, 530, ["regression"], "Task"], [543, 548, ["divide", "and", "conquer", "\"", "strategy"], "Method"], [553, 557, ["classification", "and", "regression", "tasks"], "Task"], [576, 577, ["clustering"], "Method"], [578, 580, ["classificationbased", "method"], "Method"], [595, 596, ["RF"], "Method"], [612, 614, ["Feature", "engineering"], "Method"], [630, 632, ["mapping", "functions"], "Task"], [633, 636, ["low", "-", "resolution"], "Task"], [637, 638, ["LR"], "Task"], [649, 652, ["image", "restoration", "problems"], "Task"], [662, 666, ["high", "-", "pass", "filter"], "Method"], [726, 731, ["approximating", "high", "-", "pass", "filtering"], "Task"], [733, 734, ["RF"], "Method"], [738, 742, ["dictionarylearning", "-", "based", "tool"], "Method"], [750, 756, ["dictionary", "-", "learning", "-", "based", "algorithms"], "Method"], [757, 759, ["feature", "extraction"], "Task"], [763, 765, ["discriminative", "ability"], "Metric"], [767, 771, ["gradient", "-", "based", "features"], "Method"], [772, 774, ["random", "forest"], "Method"], [793, 797, ["gradient", "-", "magnitude", "filters"], "Method"], [802, 804, ["quality", "improvement"], "Metric"], [817, 823, ["dictionary", "-", "learning", "-", "based", "algorithms"], "Method"], [824, 827, ["principal", "component", "analysis"], "Method"], [828, 830, ["PCA", ")"], "Method"], [833, 835, ["dimensionality", "reduction"], "Task"], [836, 840, ["classification", "and", "regression", "processes"], "Task"], [845, 846, ["PCA"], "Method"], [856, 858, ["PCA", "projection"], "Method"], [870, 871, ["clustering"], "Method"], [886, 891, ["content", "-", "based", "image", "retrieval"], "Method"], [892, 893, ["CBIR"], "Method"], [902, 906, ["coarse", "-", "level", "search"], "Method"], [912, 916, ["fine", "-", "level", "search"], "Method"], [938, 939, ["PCA"], "Method"], [960, 961, ["regression"], "Task"], [963, 966, ["image", "quality", "improvement"], "Metric"], [985, 986, ["PCA"], "Method"], [990, 993, ["feature", "dimensionality", "reduction"], "Task"], [1009, 1010, ["clustering"], "Method"], [1020, 1022, ["image", "quality"], "Metric"], [1024, 1026, ["regression", "problems"], "Task"], [1035, 1039, ["generalized", "weighted", "ridge", "regression"], "Method"], [1040, 1041, ["GWRR"], "Method"], [1076, 1078, ["feature", "augmentation"], "Task"], [1084, 1087, ["featureaugmented", "random", "forest"], "Method"], [1088, 1089, ["FARF"], "Method"], [1100, 1102, ["feature", "extraction"], "Task"], [1104, 1106, ["training", "stage"], "Method"], [1108, 1110, ["inference", "stages"], "Method"], [1111, 1112, ["SISR"], "Task"], [1152, 1153, ["PCA"], "Method"], [1160, 1162, ["dimensionality", "reduction"], "Task"], [1170, 1171, ["clustering"], "Method"], [1233, 1234, ["SISR"], "Task"], [1254, 1255, ["FARF"], "Method"], [1263, 1267, ["gradient", "-", "magnitude", "filters"], "Method"], [1269, 1273, ["generalized", "weighted", "ridge", "regression"], "Method"], [1274, 1275, ["GWRR"], "Method"], [1309, 1313, ["IMAGE", "SUPER", "-", "RESOLUTION"], "Task"], [1314, 1316, ["RANDOM", "FOREST"], "Method"], [1318, 1322, ["Image", "Super", "-", "Resolution"], "Task"], [1355, 1358, ["image", "restoration", "area"], "Task"], [1362, 1364, ["interpolation", "algorithms"], "Method"], [1369, 1371, ["bicubic", "interpolation"], "Method"], [1374, 1378, ["edge", "-", "preserving", "algorithms"], "Method"], [1392, 1396, ["super", "-", "resolution", "algorithms"], "Method"], [1399, 1401, ["pixel", "operations"], "Method"], [1424, 1428, ["patch", "-", "based", "algorithms"], "Method"], [1447, 1451, ["non", "-", "local", "means"], "Method"], [1462, 1464, ["manifold", "learning"], "Method"], [1468, 1471, ["block", "-", "matching"], "Method"], [1472, 1474, ["3D", "filtering"], "Method"], [1475, 1477, ["BM3D", ")"], "Method"], [1481, 1483, ["sparse", "representation"], "Method"], [1494, 1497, ["neighbor", "-", "embedding"], "Method"], [1498, 1499, ["NE"], "Method"], [1510, 1516, ["patch", "-", "based", "dictionary", "learning", "methods"], "Method"], [1517, 1518, ["NE"], "Method"], [1520, 1529, ["mapping", "between", "low", "-", "and", "high", "-", "resolution", "patches"], "Task"], [1534, 1536, ["manifold", "learning"], "Method"], [1540, 1547, ["locally", "linear", "embedding", "(", "LLE", ")", "theory"], "Method"], [1556, 1558, ["linear", "combination"], "Method"], [1569, 1571, ["HR", "counterpart"], "Method"], [1576, 1578, ["linear", "combination"], "Method"], [1706, 1709, ["discrete", "cosine", "transform"], "Method"], [1710, 1711, ["DCT"], "Method"], [1735, 1737, ["sparse", "prior"], "Method"], [1738, 1740, ["dictionary", "learning"], "Task"], [1742, 1744, ["sparse", "coding"], "Method"], [1745, 1747, ["image", "representation"], "Task"], [1771, 1773, ["sparse", "coding"], "Method"], [1775, 1778, ["image", "inverse", "problems"], "Task"], [1786, 1792, ["sparse", "-", "coding", "super", "-", "resolution"], "Method"], [1793, 1794, ["ScSR"], "Method"], [1812, 1814, ["coupled", "dictionaries"], "Method"], [1842, 1849, ["low", "and", "high", "-", "resolution", "coupled", "dictionaries"], "Method"], [1895, 1896, ["LR"], "Task"], [1897, 1900, ["HR", "coupled", "dictionaries"], "Method"], [1933, 1937, ["feature", "-", "extraction", "operator"], "Method"], [1974, 1976, ["sparse", "constraint"], "Task"], [1983, 1987, ["NP", "-", "hard", "problem"], "Task"], [2063, 2065, ["training", "dictionaries"], "Task"], [2067, 2068, ["computation"], "Task"], [2106, 2108, ["collaborative", "representation"], "Method"], [2111, 2115, ["image", "classification", "and", "restoration"], "Task"], [2129, 2132, ["anchored", "neighborhood", "regression"], "Method"], [2133, 2134, ["ANR"], "Method"], [2150, 2151, ["ANR"], "Method"], [2153, 2157, ["sparse", "-", "decomposition", "optimization"], "Method"], [2164, 2166, ["ridge", "regression"], "Method"], [2197, 2199, ["offline", "learning"], "Method"], [2205, 2207, ["prediction", "stage"], "Task"], [2244, 2246, ["coupled", "dictionaries"], "Method"], [2274, 2276, ["image", "quality"], "Metric"], [2284, 2285, ["ANR"], "Method"], [2309, 2313, ["dictionary", "-", "learning", "methods"], "Method"], [2339, 2341, ["random", "forest"], "Method"], [2342, 2343, ["RF"], "Method"], [2366, 2367, ["accuracy"], "Metric"], [2368, 2369, ["efficiency"], "Metric"], [2370, 2374, ["classification", "and", "regression", "tasks"], "Task"], [2383, 2385, ["ensemble", "learning"], "Method"], [2386, 2388, ["sublinear", "search"], "Method"], [2401, 2403, ["random", "forest"], "Method"], [2420, 2421, ["SISR"], "Task"], [2450, 2451, ["accuracy"], "Metric"], [2454, 2456, ["computational", "complexity"], "Metric"], [2461, 2463, ["deep", "learning"], "Method"], [2468, 2472, ["image", "super", "-", "resolution"], "Task"], [2492, 2496, ["image", "super", "-", "resolution"], "Task"], [2498, 2500, ["deep", "learning"], "Method"], [2505, 2508, ["convolutional", "neural", "network"], "Method"], [2509, 2510, ["SRCNN"], "Method"], [2516, 2522, ["end", "-", "to", "-", "end", "mapping"], "Task"], [2528, 2532, ["image", "super", "-", "resolution"], "Task"], [2538, 2540, ["deep", "networks"], "Method"], [2541, 2542, ["SISR"], "Task"], [2551, 2553, ["convergence", "rate"], "Metric"], [2555, 2557, ["deep", "network"], "Method"], [2561, 2563, ["residual", "learning"], "Method"], [2566, 2568, ["learning", "rates"], "Metric"], [2581, 2584, ["generative", "adversarial", "network"], "Method"], [2585, 2586, ["GAN"], "Method"], [2588, 2593, ["image", "super", "-", "resolution", "model"], "Task"], [2594, 2596, ["SRGAN", ")"], "Task"], [2612, 2614, ["adversarial", "loss"], "Method"], [2616, 2620, ["deeplearning", "-", "based", "approaches"], "Method"], [2625, 2626, ["SISR"], "Task"], [2629, 2631, ["computational", "requirement"], "Metric"], [2654, 2655, ["GPU"], "Method"], [2694, 2698, ["Image", "Super", "-", "Resolution"], "Task"], [2699, 2701, ["Random", "Forest"], "Method"], [2703, 2705, ["inference", "stage"], "Task"], [2707, 2709, ["decision", "tree"], "Method"], [2733, 2734, ["averaging"], "Method"], [2739, 2741, ["splitting", "function"], "Method"], [2883, 2884, ["RF"], "Method"], [2886, 2888, ["decision", "trees"], "Method"], [2943, 2945, ["stopping", "criterion"], "Metric"], [2987, 2989, ["stopping", "criteria"], "Metric"], [2991, 2993, ["density", "model"], "Method"], [3024, 3026, ["inference", "stage"], "Task"], [3059, 3061, ["Gaussian", "distribution"], "Method"], [3062, 3065, ["kernel", "density", "estimation"], "Method"], [3128, 3130, ["Gini", "index"], "Method"], [3160, 3162, ["regression", "problem"], "Task"], [3164, 3166, ["differential", "entropy"], "Metric"], [3251, 3252, ["classifiers"], "Method"], [3253, 3254, ["SVM"], "Method"], [3255, 3258, ["support", "vector", "machine"], "Method"], [3263, 3264, ["AdaBoost"], "Method"], [3268, 3270, ["Adaptive", "Boosting"], "Method"], [3277, 3278, ["SVM"], "Method"], [3279, 3280, ["AdaBoost"], "Method"], [3294, 3295, ["classifiers"], "Method"], [3299, 3305, ["margin", "-", "based", "global", "loss", "function"], "Metric"], [3306, 3313, ["RF", "-", "based", "image", "super", "-", "resolution"], "Task"], [3314, 3315, ["SR"], "Task"], [3382, 3383, ["regressors"], "Method"], [3409, 3411, ["inference", "stage"], "Task"], [3445, 3448, ["K", "-", "NN"], "Method"], [3454, 3455, ["RF"], "Method"], [3473, 3474, ["regression"], "Method"], [3514, 3515, ["SISR"], "Task"], [3520, 3521, ["accuracy"], "Metric"], [3522, 3523, ["efficiency"], "Metric"], [3526, 3531, ["FEATURE", "-", "AUGMENTED", "RANDOM", "FOREST"], "Method"], [3531, 3532, ["Classification"], "Task"], [3533, 3534, ["regression"], "Task"], [3538, 3540, ["probability", "problems"], "Task"], [3542, 3544, ["statistical", "theory"], "Method"], [3597, 3599, ["Bayesian", "theory"], "Method"], [3678, 3680, ["Bayesian", "rule"], "Method"], [3705, 3707, ["Bayesian", "framework"], "Method"], [3726, 3728, ["inverse", "problems"], "Task"], [3744, 3747, ["image", "restoration", "problems"], "Task"], [3760, 3762, ["feature", "extractors"], "Method"], [3763, 3768, ["deep", "-", "learning", "neural", "networks"], "Method"], [3774, 3777, ["Bayesian", "inference", "framework"], "Method"], [3779, 3780, ["SISR"], "Task"], [3785, 3789, ["ill", "-", "posed", "problem"], "Task"], [3806, 3808, ["computer", "vision"], "Task"], [3809, 3811, ["machine", "learning"], "Task"], [3835, 3837, ["edgedirected", "interpolation"], "Method"], [3838, 3839, ["NEDI"], "Method"], [3844, 3849, ["soft", "-", "decision", "adaptive", "interpolation"], "Method"], [3850, 3851, ["SAI"], "Method"], [3856, 3862, ["directional", "filtering", "and", "data", "-", "fusion"], "Method"], [3863, 3864, ["DFDF"], "Method"], [3870, 3877, ["edge", "-", "directed", "interpolation", "(", "MEDI", ")"], "Method"], [3886, 3888, ["edge", "prior"], "Method"], [3891, 3893, ["image", "processing"], "Task"], [3917, 3923, ["dictionary", "-", "learning", "-", "based", "algorithm"], "Method"], [3945, 3947, ["feature", "engineering"], "Task"], [3976, 3978, ["feature", "filter"], "Method"], [4027, 4030, ["Gradient", "Magnitude", "Filters"], "Method"], [4092, 4094, ["convolutional", "filters"], "Method"], [4098, 4100, ["feature", "filter"], "Method"], [4105, 4109, ["high", "-", "pass", "filter"], "Method"], [4123, 4130, ["first", "and", "second", "-", "order", "gradient", "operators"], "Method"], [4172, 4176, ["1", "-", "D", "filters"], "Method"], [4196, 4202, ["dictionary", "-", "learning", "-", "based", "methods"], "Method"], [4241, 4242, ["RF"], "Method"], [4268, 4269, ["RF"], "Method"], [4273, 4279, ["dictionary", "-", "learning", "-", "based", "methods"], "Method"], [4583, 4586, ["gradient", "magnitude", "filters"], "Method"], [4590, 4592, ["gradient", "filters"], "Method"], [4649, 4650, ["Regression"], "Task"], [4651, 4653, ["inference", "stage"], "Method"], [4667, 4672, ["content", "-", "based", "image", "retrieval"], "Method"], [4673, 4674, ["CBIR"], "Method"], [4686, 4689, ["approximated", "nearest", "neighbor"], "Method"], [4690, 4691, ["ANN"], "Method"], [4704, 4709, ["large", "-", "scale", "image", "retrieval"], "Task"], [4726, 4730, ["locality", "-", "sensitive", "Hashing"], "Method"], [4731, 4732, ["LSH"], "Method"], [4746, 4750, ["coarse", "-", "level", "search"], "Method"], [4756, 4757, ["similarity"], "Metric"], [4777, 4781, ["fine", "-", "level", "search"], "Method"], [4783, 4785, ["Euclidean", "distance"], "Method"], [4819, 4821, ["inference", "stage"], "Task"], [4823, 4827, ["RF", "-", "based", "SISR"], "Task"], [4828, 4830, ["PCA", "projection"], "Method"], [4854, 4855, ["searching"], "Task"], [4857, 4861, ["coarse", "-", "level", "search"], "Task"], [4871, 4872, ["PCA"], "Method"], [4873, 4876, ["feature", "dimensionality", "reduction"], "Task"], [4899, 4901, ["finelevel", "search"], "Method"], [4918, 4919, ["RF"], "Method"], [4923, 4925, ["prediction", "accuracy"], "Metric"], [4927, 4929, ["regression", "step"], "Task"], [4947, 4948, ["PCA"], "Method"], [4956, 4958, ["ridge", "regression"], "Method"], [4975, 4976, ["quality"], "Metric"], [4977, 4981, ["super", "-", "resolved", "images"], "Task"], [4995, 4997, ["regression", "problems"], "Task"], [5005, 5007, ["classification", "problems"], "Task"], [5037, 5038, ["PCA"], "Method"], [5039, 5040, ["LSH"], "Method"], [5044, 5045, ["clustering"], "Method"], [5099, 5100, ["regression"], "Task"], [5119, 5121, ["node", "splitting"], "Task"], [5122, 5123, ["clustering"], "Method"], [5126, 5128, ["training", "stage"], "Task"], [5130, 5132, ["node", "searching"], "Task"], [5134, 5136, ["inference", "stage"], "Task"], [5157, 5158, ["searching"], "Task"], [5160, 5162, ["random", "forest"], "Method"], [5177, 5179, ["image", "retrieval"], "Task"], [5180, 5181, ["CBIR"], "Method"], [5235, 5237, ["iterative", "quantization"], "Method"], [5238, 5239, ["ITQ"], "Method"], [5245, 5248, ["feature", "dimension", "reduction"], "Task"], [5250, 5251, ["PCA"], "Method"], [5266, 5269, ["superresolved", "image", "quality"], "Metric"], [5277, 5281, ["ITQ", "-", "like", "algorithm"], "Method"], [5317, 5322, ["Generalized", "Weighted", "Ridge", "Regression", "Model"], "Method"], [5332, 5334, ["ridge", "regression"], "Method"], [5342, 5345, ["anchored", "neighborhood", "regression"], "Method"], [5346, 5347, ["ANR"], "Method"], [5371, 5375, ["least", "-", "squares", "minimization"], "Method"], [5383, 5385, ["ridge", "regression"], "Method"], [5391, 5400, ["2", "-", "norm", "constrained", "least", "square", "regression", "regularized", "problem"], "Method"], [5410, 5413, ["Tikhonov", "regularization", "theory"], "Method"], [5530, 5533, ["weighted", "collaborative", "representation"], "Method"], [5534, 5535, ["WCR"], "Method"], [5540, 5543, ["normal", "collaborative", "representation"], "Method"], [5544, 5547, ["CR", ")", "model"], "Method"], [5549, 5550, ["ANR"], "Method"], [5586, 5590, ["closed", "-", "form", "solution"], "Method"], [5620, 5622, ["image", "quality"], "Metric"], [5653, 5657, ["weighted", "ridge", "regression", "models"], "Method"], [5680, 5683, ["K", "-", "SVD"], "Method"], [5686, 5692, ["sparse", "-", "coding", "-", "based", "dictionary"], "Method"], [5711, 5713, ["linear", "regressors"], "Method"], [5725, 5728, ["Gaussian", "mixture", "model"], "Method"], [5729, 5730, ["GMM"], "Method"], [5738, 5741, ["weighted", "ridge", "regression"], "Task"], [5769, 5770, ["RF"], "Method"], [5808, 5810, ["image", "classification"], "Task"], [5811, 5816, ["locality", "-", "constrained", "linear", "coding"], "Method"], [5817, 5818, ["LLC"], "Method"], [5824, 5827, ["Gaussian", "mixture", "model"], "Method"], [5828, 5829, ["GMM"], "Method"], [5835, 5841, ["locality", "-", "constrained", "affine", "subspace", "coding"], "Method"], [5842, 5843, ["LASC"], "Method"], [5850, 5851, ["GMM"], "Method"], [5876, 5879, ["ridge", "regression", "models"], "Method"], [5889, 5893, ["generalized", "weighted", "ridge", "regression"], "Method"], [5894, 5895, ["GWRR"], "Method"], [5898, 5900, ["ridge", "regression"], "Task"], [6085, 6086, ["WCR"], "Method"], [6097, 6098, ["ANR"], "Method"], [6129, 6132, ["normal", "regression", "model"], "Method"], [6151, 6156, ["regression", "of", "a", "leaf", "node"], "Task"], [6178, 6180, ["Initial", "Estimation"], "Task"], [6181, 6184, ["Iterative", "Back", "Projection"], "Method"], [6187, 6188, ["SISR"], "Task"], [6190, 6196, ["low", "-", "level", "computer", "vision", "task"], "Task"], [6212, 6214, ["mathematical", "model"], "Method"], [6215, 6217, ["image", "degradation"], "Task"], [6227, 6234, ["low", "-", "pass", "(", "blur", ")", "filter"], "Method"], [6238, 6242, ["down", "-", "sampling", "operator"], "Method"], [6256, 6260, ["approximated", "HR", "image", "̂"], "Task"], [6263, 6265, ["inverse", "problem"], "Task"], [6272, 6274, ["Bayesian", "theory"], "Method"], [6284, 6287, ["iterative", "back", "projection"], "Method"], [6288, 6289, ["IBP"], "Method"], [6296, 6297, ["IBP"], "Method"], [6321, 6323, ["reconstruction", "error"], "Metric"], [6351, 6352, ["IBP"], "Method"], [6359, 6361, ["HR", "image"], "Task"], [6364, 6366, ["reconstruction", "error"], "Metric"], [6379, 6380, ["SISR"], "Task"], [6382, 6384, ["updating", "procedure"], "Method"], [6399, 6401, ["reconstruction", "error"], "Metric"], [6413, 6417, ["up", "-", "sampling", "operator"], "Method"], [6420, 6425, ["constant", "back", "-", "projection", "kernel"], "Method"], [6432, 6436, ["low", "-", "pass", "filter"], "Method"], [6440, 6444, ["estimating", "HR", "image", "̂"], "Task"], [6466, 6470, ["learning", "-", "based", "algorithms"], "Method"], [6491, 6493, ["coarse", "estimation"], "Task"], [6496, 6498, ["bicubic", "interpolation"], "Method"], [6565, 6566, ["IBP"], "Method"], [6569, 6571, ["denoising", "process"], "Method"], [6582, 6588, ["sparse", "-", "constraint", "-", "based", "approach"], "Method"], [6595, 6597, ["denoising", "capability"], "Method"], [6600, 6608, ["2", "-", "norm", "constraint", "-", "based", "ridge", "regression"], "Method"], [6616, 6620, ["averaging", "-", "like", "process"], "Method"], [6645, 6651, ["coarse", "estimation", "of", "an", "HR", "image"], "Task"], [6654, 6655, ["IBP"], "Method"], [6675, 6676, ["IBP"], "Method"], [6684, 6687, ["initial", "coarse", "estimation"], "Task"], [6772, 6775, ["Fine", "-", "Tuning"], "Task"], [6779, 6781, ["Random", "Forest"], "Method"], [6783, 6784, ["PSNR"], "Metric"], [6786, 6787, ["Set14"], "Material"], [6818, 6820, ["computational", "cost"], "Metric"], [6846, 6847, ["FARF*."], "Method"], [6877, 6883, ["deep", "-", "learning", "-", "based", "algorithm"], "Method"], [6884, 6886, ["SRCNN", "algorithm"], "Method"], [6908, 6910, ["inference", "stages"], "Task"], [6913, 6915, ["FARF", "algorithm"], "Method"], [6973, 6975, ["Set", "5"], "Material"], [6976, 6977, ["Set14"], "Material"], [6978, 6979, ["B100"], "Material"], [6999, 7001, ["bicubic", "interpolation"], "Method"], [7015, 7016, ["RF"], "Method"], [7020, 7023, ["alternating", "regression", "forests"], "Method"], [7024, 7025, ["ARF"], "Method"], [7032, 7042, ["convolutional", "neural", "-", "network", "-", "based", "image", "super", "-", "resolution"], "Method"], [7043, 7044, ["SRCNN"], "Method"], [7085, 7086, ["PSNR"], "Metric"], [7124, 7130, ["average", "peak", "signal", "to", "noise", "ratio"], "Metric"], [7131, 7132, ["PSNR"], "Metric"], [7148, 7151, ["objective", "quality", "metric"], "Metric"], [7152, 7153, ["PSNR"], "Metric"], [7165, 7166, ["FARF"], "Method"], [7168, 7169, ["FARF"], "Method"], [7175, 7177, ["image", "quality"], "Metric"], [7191, 7195, ["deeplearning", "-", "based", "algorithms"], "Method"], [7198, 7199, ["SRCNN"], "Method"], [7208, 7210, ["FARF", "algorithm"], "Method"], [7216, 7218, ["visual", "quality"], "Metric"], [7253, 7256, ["Super", "-", "resolution"], "Task"], [7261, 7262, ["B100"], "Material"], [7276, 7277, ["ARF"], "Method"], [7278, 7281, ["CVPR", "-", "2015"], "Method"], [7286, 7287, ["SRCNN"], "Method"], [7288, 7291, ["PAMI", "-", "2016"], "Method"], [7299, 7300, ["FARF"], "Method"], [7302, 7304, ["ground", "truth"], "Metric"], [7310, 7312, ["FARF", "algorithm"], "Method"], [7331, 7335, ["deep", "-", "learning", "method"], "Method"], [7346, 7351, ["feature", "-", "augmented", "random", "forest"], "Method"], [7352, 7353, ["FARF"], "Method"], [7357, 7360, ["single", "image", "superresolution"], "Task"], [7361, 7362, ["SISR"], "Task"], [7374, 7376, ["random", "forest"], "Method"], [7377, 7378, ["RF"], "Method"], [7382, 7384, ["feature", "recipes"], "Method"], [7395, 7396, ["clustering"], "Method"], [7407, 7408, ["regression"], "Task"], [7426, 7433, ["discriminative", "gradient", "magnitude", "-", "based", "augmented", "features"], "Method"], [7441, 7442, ["regression"], "Task"], [7451, 7454, ["principal", "component", "analysis"], "Method"], [7455, 7456, ["PCA"], "Method"], [7470, 7472, ["dimensionality", "reduction"], "Task"], [7492, 7493, ["SISR"], "Task"], [7494, 7499, ["content", "-", "based", "image", "retrieval"], "Method"], [7500, 7501, ["CBIR"], "Method"], [7521, 7523, ["ridge", "regression"], "Method"], [7526, 7528, ["FAFR", "scheme"], "Method"], [7543, 7544, ["PSNR"], "Metric"], [7588, 7594, ["deep", "-", "learning", "-", "based", "algorithms"], "Method"], [148, 149, ["LSH"], "Method"], [207, 208, ["LSH"], "Method"], [222, 223, ["CBIR"], "Method"], [230, 234, ["generalized", "weighted", "ridge", "regression"], "Method"], [235, 236, ["GWRR"], "Method"], [256, 257, ["FARF"], "Method"], [271, 272, ["RF"], "Method"], [282, 283, ["FARF"], "Method"], [384, 385, ["clustering"], "Method"], [408, 409, ["RF"], "Method"], [536, 537, ["clustering"], "Method"], [563, 564, ["RF"], "Method"], [590, 591, ["clustering"], "Method"], [805, 806, ["RF"], "Method"], [807, 808, ["SISR"], "Task"], [979, 980, ["LSH"], "Method"], [1053, 1054, ["GWRR"], "Method"], [1095, 1096, ["FARF"], "Method"], [1122, 1123, ["FARF"], "Method"], [1126, 1127, ["SR"], "Task"], [1158, 1159, ["LSH"], "Method"], [1178, 1179, ["RF"], "Method"], [1201, 1202, ["GWRR"], "Method"], [1237, 1238, ["RF"], "Method"], [1283, 1284, ["FARF"], "Method"], [1293, 1294, ["FARF"], "Method"], [1323, 1324, ["SR"], "Task"], [1338, 1339, ["LR"], "Task"], [1360, 1361, ["SR"], "Task"], [1549, 1550, ["LR"], "Task"], [1585, 1586, ["LR"], "Task"], [1595, 1596, ["NE"], "Method"], [1798, 1799, ["LR"], "Task"], [1826, 1827, ["LR"], "Task"], [1852, 1853, ["LR"], "Task"], [1912, 1913, ["LR"], "Task"], [1917, 1918, ["LR"], "Task"], [1939, 1940, ["LR"], "Task"], [1949, 1950, ["LR"], "Task"], [2226, 2227, ["ANR"], "Method"], [2405, 2406, ["clustering"], "Method"], [2430, 2431, ["RF"], "Method"], [2523, 2524, ["LR"], "Task"], [3239, 3240, ["RF"], "Method"], [3331, 3332, ["SR"], "Task"], [3335, 3336, ["SR"], "Task"], [3339, 3340, ["clustering"], "Method"], [3397, 3398, ["LR"], "Task"], [3413, 3414, ["LR"], "Task"], [3488, 3489, ["clustering"], "Method"], [3493, 3495, ["random", "forest"], "Method"], [3940, 3941, ["clustering"], "Method"], [3991, 3992, ["LR"], "Task"], [4141, 4144, ["low", "-", "resolution"], "Task"], [4614, 4615, ["RF"], "Method"], [4655, 4656, ["RF"], "Method"], [4863, 4864, ["CBIR"], "Method"], [4905, 4906, ["CBIR"], "Method"], [4950, 4951, ["LSH"], "Method"], [4969, 4970, ["RF"], "Method"], [5052, 5053, ["clustering"], "Method"], [5057, 5058, ["SISR"], "Task"], [5112, 5113, ["LSH"], "Method"], [5153, 5154, ["LSH"], "Method"], [5172, 5173, ["LSH"], "Method"], [5175, 5176, ["SISR"], "Task"], [5196, 5197, ["LSH"], "Method"], [5203, 5204, ["CBIR"], "Method"], [5210, 5211, ["clustering"], "Method"], [5213, 5214, ["RF"], "Method"], [5216, 5217, ["SISR"], "Task"], [5230, 5231, ["LSH"], "Method"], [5307, 5308, ["LSH"], "Method"], [5337, 5338, ["RF"], "Method"], [5431, 5432, ["LR"], "Task"], [5581, 5582, ["ANR"], "Method"], [5607, 5608, ["WCR"], "Method"], [5612, 5613, ["ANR"], "Method"], [6058, 6059, ["GMM"], "Method"], [6075, 6076, ["GWRR"], "Method"], [6121, 6122, ["GWRR"], "Method"], [6160, 6161, ["GWRR"], "Method"], [6208, 6209, ["LR"], "Task"], [6249, 6250, ["LR"], "Task"], [6292, 6293, ["SR"], "Task"], [6312, 6313, ["SR"], "Task"], [6317, 6318, ["IBP"], "Method"], [6326, 6327, ["LR"], "Task"], [6335, 6336, ["LR"], "Task"], [6376, 6377, ["IBP"], "Method"], [6505, 6506, ["IBP"], "Method"], [6629, 6630, ["RF"], "Method"], [6659, 6660, ["LR"], "Task"], [6690, 6691, ["RF"], "Method"], [6693, 6694, ["SR"], "Task"], [6710, 6711, ["RF"], "Method"], [6738, 6739, ["RF"], "Method"], [6832, 6833, ["FARF"], "Method"], [6951, 6952, ["FARF"], "Method"], [7003, 7006, ["anchored", "neighborhood", "regression"], "Method"], [7436, 7437, ["clustering"], "Method"], [7466, 7467, ["LSH"], "Method"], [7485, 7486, ["clustering"], "Method"], [7503, 7504, ["LSH"], "Method"], [7513, 7514, ["WCR"], "Method"], [7518, 7519, ["GWRR"], "Method"], [7552, 7553, ["RF"], "Method"], [7570, 7571, ["FARF"], "Method"]]} {"doc_id": "160563abbd75265b19afc8b4169bab9e1eb33d97", "text": ["document", ":", "Massively", "Multilingual", "Sentence", "Embeddings", "for", "Zero", "-", "Shot", "Cross", "-", "Lingual", "Transfer", "and", "Beyond", "We", "introduce", "an", "architecture", "to", "learn", "joint", "multilingual", "sentence", "representations", "for", "93", "languages", ",", "belonging", "to", "more", "than", "30", "different", "language", "families", "and", "written", "in", "28", "different", "scripts", ".", "Our", "system", "uses", "a", "single", "BiLSTM", "encoder", "with", "a", "shared", "BPE", "vocabulary", "for", "all", "languages", ",", "which", "is", "coupled", "with", "an", "auxiliary", "decoder", "and", "trained", "on", "publicly", "available", "parallel", "corpora", ".", "This", "enables", "us", "to", "learn", "a", "classifier", "on", "top", "of", "the", "resulting", "sentence", "embeddings", "using", "English", "annotated", "data", "only", ",", "and", "transfer", "it", "to", "any", "of", "the", "93", "languages", "without", "any", "modification", ".", "Our", "approach", "sets", "a", "new", "state", "-", "of", "-", "the", "-", "art", "on", "zero", "-", "shot", "cross", "-", "lingual", "natural", "language", "inference", "for", "all", "the", "14", "languages", "in", "the", "XNLI", "dataset", "but", "one", ".", "We", "also", "achieve", "very", "competitive", "results", "in", "cross", "-", "lingual", "document", "classification", "(", "MLDoc", "dataset", ")", ".", "Our", "sentence", "embeddings", "are", "also", "strong", "at", "parallel", "corpus", "mining", ",", "establishing", "a", "new", "state", "-", "of", "-", "the", "-", "art", "in", "the", "BUCC", "shared", "task", "for", "3", "of", "its", "4", "language", "pairs", ".", "Finally", ",", "we", "introduce", "a", "new", "test", "set", "of", "aligned", "sentences", "in", "122", "languages", "based", "on", "the", "Tatoeba", "corpus", ",", "and", "show", "that", "our", "sentence", "embeddings", "obtain", "strong", "results", "in", "multilingual", "similarity", "search", "even", "for", "low", "-", "resource", "languages", ".", "Our", "PyTorch", "implementation", ",", "pre", "-", "trained", "encoder", "and", "the", "multilingual", "test", "set", "will", "be", "freely", "available", ".", "section", ":", "Introduction", "While", "the", "recent", "advent", "of", "deep", "learning", "has", "led", "to", "impressive", "progress", "in", "Natural", "Language", "Processing", "(", "NLP", ")", ",", "these", "techniques", "are", "known", "to", "be", "particularly", "data", "hungry", ",", "limiting", "their", "applicability", "in", "many", "practical", "scenarios", ".", "An", "increasingly", "popular", "approach", "to", "alleviate", "this", "issue", "is", "to", "first", "learn", "general", "language", "representations", "on", "unlabeled", "data", ",", "which", "are", "then", "integrated", "in", "task", "-", "specific", "downstream", "systems", ".", "This", "approach", "was", "first", "popularized", "by", "word", "embeddings", "mikolov2013distributed", ",", "pennington2014glove", ",", "but", "has", "recently", "been", "superseded", "by", "sentence", "-", "level", "representations", "Alexis:2017:emnlp", ",", "Peters:2018:naacl_elmo", ",", "Devlin:2018:arxiv_bert", ".", "Nevertheless", ",", "all", "these", "works", "learn", "a", "separate", "model", "for", "each", "language", "and", "are", "thus", "unable", "to", "leverage", "information", "across", "different", "languages", ",", "greatly", "limiting", "their", "potential", "performance", "for", "low", "-", "resource", "languages", ".", "In", "this", "work", ",", "we", "are", "interested", "in", "universal", "language", "agnostic", "sentence", "embeddings", ",", "that", "is", ",", "vector", "representations", "of", "sentences", "that", "are", "general", "with", "respect", "to", "two", "dimensions", ":", "the", "input", "language", "and", "the", "NLP", "task", ".", "The", "motivations", "for", "such", "a", "representation", "are", "multiple", ":", "the", "hope", "that", "languages", "with", "limited", "resources", "benefit", "from", "joint", "training", "over", "many", "languages", ",", "the", "desire", "to", "perform", "zero", "-", "shot", "transfer", "of", "an", "NLP", "model", "from", "one", "language", "(", "e.g.", "English", ")", "to", "another", ",", "and", "the", "possibility", "to", "handle", "code", "-", "switching", ".", "We", "achieve", "this", "by", "using", "a", "single", "encoder", "that", "can", "handle", "multiple", "languages", ",", "so", "that", "semantically", "similar", "sentences", "in", "different", "languages", "are", "close", "in", "the", "resulting", "embedding", "space", ".", "Most", "research", "in", "multilingual", "NLP", "focuses", "on", "high", "-", "resource", "languages", "like", "Chinese", ",", "Arabic", "or", "major", "European", "languages", ",", "and", "is", "usually", "limited", "to", "a", "few", "(", "most", "often", "only", "two", ")", "languages", ".", "In", "contrast", ",", "we", "learn", "joint", "sentence", "representations", "for", "93", "different", "languages", ",", "including", "under", "-", "resourced", "and", "minority", "languages", "(", "see", "Tables", "[", "reference", "]", "and", "[", "reference", "]", ")", ".", "Our", "system", "is", "trained", "on", "freely", "available", "parallel", "texts", "only", ".", "The", "contributions", "of", "this", "paper", "are", "as", "follows", ":", "We", "substantially", "improve", "on", "previous", "work", "to", "learn", "joint", "multilingual", "sentence", "representations", ".", "We", "learn", "one", "shared", "encoder", "that", "can", "handle", "93", "different", "languages", ".", "All", "languages", "are", "jointly", "embedded", "in", "a", "shared", "space", ",", "in", "contrast", "to", "most", "other", "works", "which", "usually", "consider", "separate", "English", "/", "foreign", "alignments", ".", "We", "cover", "34", "language", "families", "and", "28", "different", "scripts", ".", "We", "outperform", "the", "state", "-", "of", "-", "the", "-", "art", "on", "zero", "-", "shot", "cross", "-", "lingual", "natural", "language", "inference", "(", "XNLI", "dataset", ")", "and", "classification", "(", "MLDoc", "dataset", ")", ",", "bitext", "mining", "(", "BUCC", "dataset", ")", "and", "multilingual", "similarity", "search", "(", "Tatoeba", "dataset", ")", ",", "for", "almost", "all", "considered", "languages", ".", "These", "results", "were", "obtained", "with", "a", "single", "pre", "-", "trained", "BiLSTM", "encoder", "for", "all", "93", "languages", "and", "tasks", ",", "without", "any", "fine", "-", "tuning", ".", "We", "define", "a", "new", "test", "set", "based", "on", "the", "freely", "available", "Tatoeba", "corpus", "and", "provide", "baseline", "results", "for", "122", "languages", ".", "We", "report", "accuracy", "for", "multilingual", "similarity", "search", "on", "this", "test", "set", ",", "but", "the", "corpus", "could", "also", "be", "used", "for", "MT", "evaluation", ".", "The", "remaining", "of", "this", "paper", "is", "organized", "as", "follows", ".", "In", "the", "next", "section", ",", "we", "first", "summarize", "related", "work", ".", "Section", "[", "reference", "]", "then", "describe", "our", "approach", "in", "detail", ".", "All", "experimental", "results", "are", "given", "in", "Sections", "[", "reference", "]", "and", "[", "reference", "]", ",", "and", "the", "paper", "concludes", "with", "a", "discussion", "and", "directions", "for", "future", "research", ".", "Dataset", "details", "and", "additional", "result", "analysis", "can", "be", "found", "in", "the", "appendix", ".", "section", ":", "Related", "work", "Following", "the", "success", "of", "word", "embeddings", "mikolov2013distributed", ",", "pennington2014glove", ",", "there", "has", "been", "an", "increasing", "interest", "in", "learning", "continuous", "vector", "representations", "of", "longer", "linguistic", "units", "like", "sentences", ".", "These", "sentence", "embeddings", "are", "commonly", "obtained", "using", "a", "Recurrent", "Neural", "Network", "(", "RNN", ")", "encoder", ",", "which", "is", "typically", "trained", "in", "an", "unsupervised", "way", "over", "large", "collections", "of", "unlabelled", "corpora", ".", "For", "instance", ",", "the", "skip", "-", "thought", "model", "of", "Kiros:2015:nips_skipthought", "couple", "the", "encoder", "with", "an", "auxiliary", "decoder", ",", "and", "train", "the", "entire", "system", "end", "-", "to", "-", "end", "to", "predict", "the", "surrounding", "sentences", "over", "a", "large", "collection", "of", "books", ".", "It", "was", "later", "shown", "that", "more", "competitive", "results", "could", "be", "obtained", "by", "training", "the", "encoder", "over", "labeled", "Natural", "Language", "Inference", "(", "NLI", ")", "data", "Alexis:2017:emnlp", ".", "This", "was", "recently", "extended", "to", "multitask", "learning", ",", "combining", "different", "training", "objectives", "like", "that", "of", "skip", "-", "thought", ",", "NLI", "and", "machine", "translation", "Google:2018:arxiv_srep", ",", "MLIA", "-", "MSR:2018:iclr_srep", ".", "While", "the", "previous", "methods", "consider", "a", "single", "language", "at", "a", "time", ",", "multilingual", "representations", "have", "attracted", "a", "large", "attention", "in", "recent", "times", ".", "Most", "of", "this", "research", "focuses", "on", "cross", "-", "lingual", "word", "embeddings", "ruder2017survey", ",", "which", "are", "commonly", "learned", "jointly", "from", "parallel", "corpora", "gouws2015bilbowa", ",", "luong2015bilingual", ".", "An", "alternative", "approach", "that", "is", "becoming", "increasingly", "popular", "is", "to", "train", "word", "embeddings", "independently", "for", "each", "language", "over", "monolingual", "corpora", ",", "and", "then", "map", "them", "to", "a", "shared", "space", "based", "on", "a", "bilingual", "dictionary", "mikolov2013exploiting", ",", "artetxe2018generalizing", "or", "even", "in", "a", "fully", "unsupervised", "manner", ".", "Cross", "-", "lingual", "word", "embeddings", "are", "often", "used", "to", "build", "bag", "-", "of", "-", "word", "representations", "of", "longer", "linguistic", "units", "by", "taking", "their", "respective", "centroid", "Klementiev:2012:coling_reuters", ".", "While", "this", "approach", "has", "the", "advantage", "of", "requiring", "a", "weak", "(", "or", "even", "no", ")", "cross", "-", "lingual", "signal", ",", "it", "has", "been", "shown", "that", "the", "resulting", "sentence", "embeddings", "works", "rather", "poorly", "in", "practical", "cross", "-", "lingual", "transfer", "settings", "Conneau:2018:emnlp_xnli", ".", "A", "more", "competitive", "approach", "that", "we", "follow", "here", "is", "to", "use", "a", "sequence", "-", "to", "-", "sequence", "encoder", "-", "decoder", "architecture", "Schwenk:2017:repl4nlp", ",", "hassan2018achieving", ".", "The", "full", "system", "is", "trained", "end", "-", "to", "-", "end", "on", "parallel", "corpora", "akin", "to", "neural", "machine", "translation", ":", "the", "encoder", "maps", "the", "source", "sequence", "into", "a", "fixed", "-", "length", "vector", "representation", ",", "which", "is", "used", "by", "the", "decoder", "to", "create", "the", "target", "sequence", ".", "This", "decoder", "is", "then", "discarded", ",", "and", "the", "encoder", "is", "kept", "to", "embed", "sentences", "in", "any", "of", "the", "training", "languages", ".", "While", "some", "proposals", "use", "a", "separate", "encoder", "for", "each", "language", "Schwenk:2017:repl4nlp", ",", "sharing", "a", "single", "encoder", "for", "all", "languages", "also", "gives", "strong", "results", "Schwenk:2018:acl_mine", ".", "Nevertheless", ",", "most", "existing", "work", "is", "either", "limited", "to", "few", ",", "rather", "close", "languages", "or", ",", "more", "commonly", ",", "consider", "pairwise", "joint", "embeddings", "with", "English", "and", "one", "foreign", "language", "only", ".", "To", "the", "best", "of", "our", "knowledge", ",", "all", "existing", "work", "on", "learning", "multilingual", "representations", "for", "a", "large", "number", "of", "languages", "is", "limited", "to", "word", "embeddings", "ammar2016massively", ",", "ours", "being", "the", "first", "paper", "exploring", "massively", "multilingual", "sentence", "representations", ".", "Finally", ",", "while", "all", "the", "previous", "approaches", "learn", "a", "fixed", "-", "length", "representation", "for", "each", "sentence", ",", "a", "recent", "research", "line", "has", "obtained", "very", "strong", "results", "using", "variable", "-", "length", "representations", "instead", ",", "consisting", "of", "contextualized", "embeddings", "of", "the", "words", "in", "the", "sentence", "Peters:2018:naacl_elmo", ",", "howard2018universal", ",", "Devlin:2018:arxiv_bert", ".", "For", "that", "purpose", ",", "these", "methods", "train", "either", "an", "RNN", "or", "self", "-", "attentional", "encoder", "over", "unnanotated", "corpora", "using", "some", "form", "of", "language", "modeling", ".", "A", "classifier", "can", "then", "be", "learned", "on", "top", "of", "the", "resulting", "encoder", ",", "which", "is", "commonly", "further", "fine", "-", "tuned", "during", "this", "supervised", "training", ".", "Despite", "the", "strong", "performance", "of", "these", "approaches", "in", "monolingual", "settings", ",", "we", "argue", "that", "fixed", "-", "length", "approaches", "provide", "a", "more", "generic", ",", "flexible", "and", "compatible", "representation", "form", "for", "our", "multilingual", "scenario", ",", "and", "our", "model", "indeed", "outperforms", "the", "multilingual", "BERT", "model", "Devlin:2018:arxiv_bert", "in", "zero", "-", "shot", "transfer", "(", "see", "Section", "[", "reference", "]", ")", ".", "section", ":", "Proposed", "method", "We", "use", "a", "single", ",", "language", "agnostic", "BiLSTM", "encoder", "to", "build", "our", "sentence", "embeddings", ",", "which", "is", "coupled", "with", "an", "auxiliary", "decoder", "and", "trained", "over", "parallel", "corpora", ".", "From", "Section", "[", "reference", "]", "to", "[", "reference", "]", ",", "we", "describe", "its", "architecture", ",", "our", "training", "strategy", "to", "scale", "to", "up", "to", "93", "languages", ",", "and", "the", "training", "data", "used", "for", "that", "purpose", ".", "subsection", ":", "Architecture", "Figure", "[", "reference", "]", "illustrates", "the", "architecture", "of", "the", "proposed", "system", ",", "which", "is", "based", "on", "Schwenk:2018:acl_mine", ".", "As", "it", "can", "be", "seen", ",", "sentence", "embeddings", "are", "obtained", "by", "applying", "a", "max", "-", "pooling", "operation", "over", "the", "output", "of", "a", "BiLSTM", "encoder", ".", "These", "sentence", "embeddings", "are", "used", "to", "initialize", "the", "decoder", "LSTM", "through", "a", "linear", "transformation", ",", "and", "are", "also", "concatenated", "to", "its", "input", "embeddings", "at", "every", "time", "step", ".", "Note", "that", "there", "is", "no", "other", "connection", "between", "the", "encoder", "and", "the", "decoder", ",", "as", "we", "want", "all", "relevant", "information", "of", "the", "input", "sequence", "to", "be", "captured", "by", "the", "sentence", "embedding", ".", "We", "use", "a", "single", "encoder", "and", "decoder", "in", "our", "system", ",", "which", "are", "shared", "by", "all", "languages", "involved", ".", "For", "that", "purpose", ",", "we", "build", "a", "joint", "byte", "-", "pair", "encoding", "(", "BPE", ")", "vocabulary", "with", "50k", "operations", ",", "which", "is", "learned", "on", "the", "concatenation", "of", "all", "training", "corpora", ".", "This", "way", ",", "the", "encoder", "has", "no", "explicit", "signal", "on", "what", "the", "input", "language", "is", ",", "encouraging", "it", "to", "learn", "language", "independent", "representations", ".", "In", "contrast", ",", "the", "decoder", "takes", "a", "language", "ID", "embedding", "that", "specifies", "the", "language", "to", "generate", ",", "which", "is", "concatenated", "to", "the", "input", "and", "sentence", "embeddings", "at", "every", "time", "step", ".", "Scaling", "up", "to", "almost", "hundred", "languages", ",", "which", "use", "very", "different", "syntax", ",", "writing", "scripts", "and", "linguistic", "concepts", ",", "naturally", "calls", "for", "an", "encoder", "with", "sufficient", "capacity", ".", "In", "this", "paper", ",", "we", "limit", "our", "study", "to", "a", "stacked", "BiLSTM", "with", "1", "to", "5", "layers", ",", "each", "512", "-", "dimensional", ".", "The", "resulting", "sentence", "representations", "(", "after", "concatenating", "both", "directions", ")", "are", "1024", "dimensional", ".", "The", "decoder", "has", "always", "one", "layer", "of", "dimension", "2048", ".", "The", "input", "embedding", "size", "is", "set", "to", "320", ",", "while", "the", "language", "ID", "embedding", "has", "32", "dimensions", ".", "Training", "[", "-", "1pt", "]", "corpus", "[", "-", "2pt", "]", "size", "Tatoeba", "[", "-", "1pt", "]", "test", "set", "[", "-", "2pt", "]", "size", "Training", "[", "-", "1pt", "]", "corpus", "[", "-", "2pt", "]", "size", "Tatoeba", "[", "-", "1pt", "]", "test", "set", "[", "-", "2pt", "]", "size", "subsection", ":", "Training", "strategies", "In", "preceding", "work", ",", "each", "sentence", "at", "the", "input", "was", "jointly", "translated", "into", "all", "other", "languages", ".", "While", "this", "approach", "was", "shown", "to", "learn", "high", "-", "quality", "representations", ",", "it", "poses", "two", "obvious", "drawbacks", "when", "trying", "to", "scale", "to", "a", "large", "number", "of", "languages", ".", "First", ",", "it", "requires", "an", "N", "-", "way", "parallel", "corpus", ",", "which", "is", "difficult", "to", "obtain", "for", "all", "languages", ".", "Second", ",", "it", "has", "a", "quadratic", "cost", "with", "respect", "to", "the", "number", "of", "languages", ",", "making", "training", "prohibitively", "slow", "as", "the", "number", "of", "languages", "is", "increased", ".", "In", "our", "preliminary", "experiments", ",", "we", "observed", "that", "similar", "results", "can", "be", "obtained", "by", "using", "less", "target", "languages", "–", "two", "seem", "to", "be", "enough", ".", "At", "the", "same", "time", ",", "we", "relax", "the", "requirement", "for", "N", "-", "way", "parallel", "corpora", "by", "considering", "independent", "alignments", "with", "the", "two", "target", "languages", ",", "e.g.", "we", "do", "not", "require", "each", "source", "sentence", "to", "be", "translated", "into", "the", "two", "target", "languages", ".", "Training", "minimizes", "the", "cross", "-", "entropy", "loss", "on", "the", "training", "corpus", ",", "alternating", "over", "all", "combinations", "of", "the", "languages", "involved", ".", "For", "that", "purpose", ",", "we", "use", "Adam", "with", "a", "constant", "learning", "rate", "of", "0.001", "and", "dropout", "set", "to", "0.1", ",", "and", "train", "for", "a", "fixed", "number", "of", "epochs", ".", "Our", "implementation", "is", "based", "on", "fairseq", ",", "and", "we", "make", "use", "of", "its", "multi", "-", "GPU", "support", "to", "train", "on", "16", "NVIDIA", "V100", "GPUs", "with", "a", "total", "batch", "size", "of", "128", ",", "000", "tokens", ".", "Unless", "otherwise", "specified", ",", "we", "train", "our", "model", "for", "17", "epochs", ",", "which", "takes", "about", "5", "days", ".", "Stopping", "training", "early", "decreases", "the", "overall", "performance", "only", "slightly", ".", "subsection", ":", "Training", "data", "and", "pre", "-", "processing", "As", "described", "in", "Section", "[", "reference", "]", ",", "training", "requires", "bitexts", "aligned", "with", "two", "target", "languages", ".", "We", "choose", "English", "and", "Spanish", "for", "that", "purpose", ",", "as", "most", "of", "the", "data", "is", "aligned", "with", "these", "languages", ".", "We", "collect", "training", "corpora", "for", "93", "input", "languages", "by", "combining", "the", "Europarl", ",", "United", "Nations", ",", "OpenSubtitles2018", ",", "Global", "Voices", ",", "Tanzil", "and", "Tatoeba", "corpus", ",", "which", "are", "all", "publicly", "available", "on", "the", "OPUS", "website", ".", "Appendix", "[", "reference", "]", "provides", "a", "more", "detailed", "description", "of", "this", "training", "data", ",", "while", "Tables", "[", "reference", "]", "and", "[", "reference", "]", "summarize", "the", "list", "of", "all", "languages", "used", "for", "training", ",", "their", "language", "family", ",", "writing", "script", "and", "the", "size", "of", "the", "bitexts", ".", "Our", "training", "data", "comprises", "a", "total", "of", "223", "million", "parallel", "sentences", ".", "In", "preliminary", "experiments", ",", "we", "observed", "that", "the", "domain", "of", "the", "training", "data", "played", "a", "key", "role", "in", "the", "performance", "of", "our", "sentence", "embeddings", "in", "different", "tasks", ".", "Some", "tasks", "(", "BUCC", ",", "MLDoc", ")", "tend", "to", "perform", "better", "when", "the", "encoder", "is", "trained", "on", "long", "and", "formal", "sentences", ",", "whereas", "other", "tasks", "(", "XNLI", ",", "Tatoeba", ")", "benefit", "from", "training", "on", "shorter", "and", "more", "informal", "sentences", ".", "In", "an", "attempt", "to", "achieve", "a", "general", "purpose", "sentence", "encoder", "that", "performs", "well", "on", "all", "tasks", ",", "we", "aimed", "at", "balancing", "the", "size", "of", "training", "corpora", "with", "long", "and", "short", "sentences", ".", "For", "that", "purpose", ",", "we", "used", "at", "most", "two", "million", "sentences", "from", "OpenSubtitles", ",", "although", "more", "data", "is", "available", "for", "some", "languages", ".", "All", "pre", "-", "processing", "is", "done", "with", "Moses", "tools", ":", "punctuation", "normalization", ",", "removing", "non", "-", "printing", "characters", "and", "tokenization", ".", "As", "the", "only", "exception", ",", "Chinese", "and", "Japanese", "texts", "were", "segmented", "with", "Jieba", "and", "Mecab", ",", "respectively", ".", "All", "the", "languages", "are", "kept", "in", "their", "original", "script", "with", "the", "exception", "of", "Greek", ",", "which", "we", "romanize", "into", "the", "Latin", "alphabet", ".", "section", ":", "Experimental", "evaluation", "In", "contrast", "with", "the", "well", "-", "established", "evaluation", "frameworks", "for", "English", "sentence", "representations", ",", "there", "is", "not", "yet", "a", "commonly", "accepted", "standard", "to", "evaluate", "multilingual", "sentence", "embeddings", ".", "The", "most", "notable", "effort", "in", "this", "regard", "is", "probably", "the", "XNLI", "corpus", ",", "an", "NLI", "test", "set", "similar", "to", "MultiNLI", "for", "which", "the", "premises", "and", "hypotheses", "were", "translated", "into", "14", "languages", "by", "professional", "translators", ".", "We", "train", "an", "NLI", "classifier", "on", "top", "of", "our", "multilingual", "sentence", "embedding", "using", "English", "training", "data", ",", "and", "evaluate", "its", "zero", "-", "shot", "transfer", "performance", "in", "the", "remaining", "languages", "(", "Section", "[", "reference", "]", ")", ".", "So", "as", "to", "obtain", "a", "more", "complete", "picture", "of", "the", "behavior", "of", "our", "multilingual", "sentence", "representations", ",", "we", "also", "evaluate", "them", "in", "cross", "-", "lingual", "document", "classification", "(", "MLDoc", ",", "Section", "[", "reference", "]", ")", ",", "and", "bitext", "mining", "(", "BUCC", ",", "Section", "[", "reference", "]", ")", ".", "However", ",", "all", "these", "datasets", "only", "cover", "a", "subset", "of", "our", "93", "languages", ",", "so", "we", "also", "introduce", "a", "new", "test", "set", "for", "multilingual", "similarity", "search", "in", "122", "languages", ",", "including", "several", "languages", "for", "which", "we", "have", "no", "training", "data", "but", "whose", "language", "family", "is", "covered", "(", "Section", "[", "reference", "]", ")", ".", "We", "remark", "that", "we", "use", "the", "same", "pre", "-", "trained", "BiLSTM", "encoder", "for", "all", "tasks", "and", "languages", "without", "any", "fine", "-", "tuning", ".", "subsection", ":", "XNLI", ":", "cross", "-", "lingual", "NLI", "NLI", "has", "become", "a", "widely", "used", "task", "to", "evaluate", "sentence", "representations", "snli:2015", ",", "multinli:2017", ".", "Given", "two", "sentences", ",", "a", "premise", "and", "a", "hypothesis", ",", "the", "task", "consists", "in", "deciding", "whether", "there", "is", "an", "entailment", ",", "contradiction", "or", "neutral", "relationship", "between", "them", ".", "XNLI", "is", "a", "recent", "effort", "to", "create", "a", "dataset", "similar", "to", "the", "English", "MultiNLI", "for", "several", "languages", ".", "2", ",", "500", "development", "and", "5", ",", "000", "test", "sentences", "have", "been", "translated", "from", "English", "into", "14", "languages", "by", "professional", "translators", ",", "making", "results", "across", "different", "languages", "directly", "comparable", ".", "Note", "that", "no", "human", "translated", "training", "data", "is", "provided", ";", "instead", ",", "different", "systems", "are", "to", "use", "English", "training", "data", "from", "MultiNLI", ",", "and", "their", "transfer", "performance", "is", "evaluated", "on", "the", "rest", "of", "languages", ".", "We", "train", "a", "classifier", "on", "top", "of", "our", "multilingual", "encoder", "using", "the", "usual", "combination", "of", "the", "two", "sentence", "embeddings", ":", ",", "where", "and", "are", "the", "premise", "and", "hypothesis", ".", "For", "that", "purpose", ",", "we", "use", "a", "feed", "-", "forward", "neural", "network", "with", "two", "hidden", "layers", "of", "size", "512", "and", "384", ",", "trained", "with", "Adam", ".", "All", "hyperparameters", "were", "optimized", "on", "the", "English", "XNLI", "development", "corpus", ",", "and", "then", ",", "the", "same", "classifier", "was", "applied", "to", "all", "languages", "of", "the", "XNLI", "test", "set", ".", "As", "such", ",", "we", "did", "not", "use", "any", "training", "or", "development", "data", "in", "any", "of", "the", "foreign", "languages", ".", "Note", ",", "moreover", ",", "that", "the", "multilingual", "sentence", "embeddings", "are", "fixed", "and", "not", "fine", "-", "tuned", "on", "the", "task", "or", "the", "language", ".", "We", "report", "our", "results", "in", "Table", "[", "reference", "]", ",", "along", "with", "several", "baselines", "from", "Conneau:2018:emnlp_xnli", "and", "the", "recently", "released", "multilingual", "BERT", "model", "Devlin:2018:arxiv_bert", ".", "As", "it", "can", "be", "seen", ",", "our", "proposed", "method", "establishes", "a", "new", "state", "-", "of", "-", "the", "-", "art", "in", "zero", "-", "shot", "cross", "-", "lingual", "transfer", "(", "i.e.", "training", "a", "classifier", "on", "English", "data", "and", "applying", "it", "to", "all", "other", "languages", ")", "for", "all", "languages", "but", "Spanish", ".", "Our", "transfer", "results", "are", "strong", "and", "homogeneous", "across", "all", "languages", ":", "for", "11", "of", "them", ",", "the", "zero", "-", "short", "performance", "is", "(", "at", "most", ")", "5", "%", "lower", "than", "the", "one", "on", "English", ",", "including", "distant", "languages", "like", "Arabic", ",", "Chinese", "and", "Vietnamese", ",", "and", "we", "also", "achieve", "remarkable", "good", "results", "on", "low", "-", "resource", "languages", "like", "Swahili", ".", "In", "contrast", ",", "BERT", "achieves", "excellent", "results", "on", "English", ",", "outperforming", "our", "system", "by", "7.5", "points", ",", "but", "its", "zero", "-", "shot", "cross", "-", "lingual", "transfer", "performance", "is", "much", "weaker", ".", "For", "instance", ",", "the", "loss", "in", "accuracy", "for", "both", "Arabic", "and", "Chinese", "is", "2.5", "points", "for", "our", "system", ",", "compared", "to", "19.3", "and", "17.6", "points", "for", "BERT", ".", "Finally", ",", "we", "also", "outperform", "all", "baselines", "of", "Conneau:2018:emnlp_xnli", "by", "a", "substantial", "margin", ",", "with", "the", "additional", "advantage", "that", "we", "use", "a", "single", "pre", "-", "trained", "encoder", ",", "whereas", "X", "-", "BiLSTM", "learns", "a", "separate", "encoder", "for", "each", "language", "by", "aligning", "it", "to", "the", "English", "one", ".", "For", "completeness", ",", "we", "also", "provide", "results", "that", "include", "the", "use", "of", "Machine", "Translation", "(", "MT", ")", ".", "This", "can", "be", "done", "in", "two", "ways", ":", "1", ")", "translate", "the", "test", "data", "into", "English", "and", "apply", "the", "English", "NLI", "classifier", ",", "or", "2", ")", "translate", "the", "English", "training", "data", "and", "train", "a", "language", "specific", "NLI", "classifier", "for", "each", "language", ".", "It", "should", "be", "stressed", "that", "we", "are", "not", "evaluating", "multilingual", "sentence", "embeddings", "anymore", ",", "but", "rather", "the", "quality", "of", "the", "MT", "system", "and", "a", "monolingual", "model", ".", "Moreover", ",", "the", "use", "of", "MT", "incurs", "in", "an", "important", "overhead", "with", "either", "strategy", ":", "translating", "test", "makes", "inference", "substantially", "more", "expensive", ",", "whereas", "translating", "train", "results", "in", "a", "separate", "model", "for", "each", "language", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "our", "approach", "outperforms", "all", "translation", "baselines", "of", "Conneau:2018:emnlp_xnli", "with", "the", "exception", "of", "Urdu", ".", "We", "also", "outperform", "MT", "BERT", "for", "Arabic", "and", "Thai", ",", "and", "are", "very", "close", "for", "Urdu", ".", "Finally", ",", "it", "is", "worth", "mentioning", "that", ",", "thanks", "to", "its", "multilingual", "nature", ",", "our", "system", "can", "also", "handle", "premises", "and", "hypothesis", "in", "different", "languages", ".", "As", "reported", "in", "Appendix", "[", "reference", "]", ",", "the", "proposed", "method", "obtains", "very", "strong", "results", "in", "these", "settings", ",", "even", "for", "distant", "language", "combinations", "like", "French", "-", "Chinese", ".", "subsection", ":", "MLDoc", ":", "cross", "-", "lingual", "classification", "Cross", "-", "lingual", "document", "classification", "is", "a", "typical", "application", "of", "multilingual", "representations", ".", "In", "order", "to", "evaluate", "our", "sentence", "embeddings", "in", "this", "task", ",", "we", "use", "the", "MLDoc", "dataset", "of", "Schwenk:2018:lrec_mldoc", ",", "which", "is", "an", "improved", "version", "of", "the", "Reuters", "benchmark", "Lewis", ":", "Reuters:2004", ",", "Klementiev:2012:coling_reuters", "with", "uniform", "class", "priors", "and", "a", "wider", "language", "coverage", ".", "There", "are", "1", ",", "000", "training", "and", "development", "documents", "and", "4", ",", "000", "test", "documents", "for", "each", "language", ",", "divided", "in", "4", "different", "genders", ".", "Just", "as", "with", "the", "XNLI", "evaluation", ",", "we", "consider", "the", "zero", "-", "shot", "transfer", "scenario", ":", "we", "train", "a", "classifier", "on", "top", "of", "our", "multilingual", "encoder", "using", "the", "English", "training", "data", ",", "optimizing", "hyper", "-", "parameters", "on", "the", "English", "development", "set", ",", "and", "evaluating", "the", "resulting", "system", "in", "the", "remaining", "languages", ".", "We", "use", "a", "feed", "-", "forward", "neural", "network", "with", "one", "hidden", "layer", "of", "10", "units", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "our", "system", "obtains", "the", "best", "published", "results", "for", "5", "of", "the", "7", "transfer", "languages", ".", "We", "believe", "that", "our", "weaker", "performance", "on", "Japanese", "can", "be", "attributed", "to", "the", "domain", "and", "sentence", "length", "mismatch", "between", "MLDoc", "and", "the", "parallel", "corpus", "we", "use", "for", "this", "language", "(", "OpenSubtitles", ")", ".", "subsection", ":", "BUCC", ":", "bitext", "mining", "Bitext", "mining", "is", "another", "natural", "application", "for", "multilingual", "sentence", "embeddings", ".", "Given", "two", "comparable", "corpora", "in", "different", "languages", ",", "the", "task", "consists", "in", "identifying", "sentence", "pairs", "that", "are", "translations", "of", "each", "other", ".", "For", "that", "purpose", ",", "one", "would", "commonly", "score", "sentence", "pairs", "by", "taking", "the", "cosine", "similarity", "of", "their", "respective", "embeddings", ",", "so", "parallel", "sentences", "can", "be", "extracted", "through", "nearest", "neighbor", "retrieval", "and", "filtered", "by", "setting", "a", "fixed", "threshold", "over", "this", "cosine", "score", "Schwenk:2018:acl_mine", ".", "However", ",", "it", "was", "recently", "shown", "that", "this", "approach", "suffers", "from", "scale", "inconsistency", "issues", "guo2018effective", ",", "and", "artetxe2018margin", "proposed", "the", "following", "alternative", "score", "addressing", "it", ":", "where", "and", "are", "the", "source", "and", "target", "sentences", ",", "and", "denotes", "the", "nearest", "neighbors", "of", "in", "the", "other", "language", ".", "The", "paper", "explores", "different", "margin", "functions", ",", "with", "ratio", "(", ")", "yielding", "the", "best", "results", ".", "This", "notion", "of", "margin", "is", "related", "to", "CSLS", "as", "proposed", "in", "Conneau:2018:iclr_muse", ".", "The", "reader", "is", "referred", "to", "artetxe2018margin", "for", "a", "detailed", "discussion", ".", "We", "use", "this", "method", "to", "evaluate", "our", "sentence", "embeddings", "on", "the", "BUCC", "mining", "task", "zweigenbaum2017overview", ",", "zweigenbaum2018overview", ",", "using", "exact", "same", "hyper", "-", "parameters", "as", "artetxe2018margin", ".", "The", "goal", "is", "to", "extract", "parallel", "sentences", "from", "a", "comparable", "corpus", "between", "English", "and", "four", "foreign", "languages", ":", "German", ",", "French", ",", "Russian", "and", "Chinese", ".", "The", "dataset", "consists", "of", "150", "K", "to", "1.2", "M", "sentences", "for", "each", "language", ",", "split", "into", "a", "sample", ",", "training", "and", "test", "set", ",", "with", "about", "2–3", "%", "of", "the", "sentences", "being", "parallel", ".", "As", "shown", "in", "our", "results", "in", "Table", "[", "reference", "]", ",", "our", "sentence", "embeddings", "establish", "a", "new", "state", "-", "of", "-", "the", "-", "art", "for", "all", "language", "pairs", "with", "the", "exception", "of", "English", "-", "Chinese", "test", ".", "Quite", "remarkably", ",", "we", "also", "outperform", "artetxe2018margin", "themselves", ",", "who", "use", "two", "separate", "models", "covering", "4", "languages", "each", "(", "English", "/", "French", "/", "Spanish", "/", "German", "and", "English", "/", "French", "/", "Russian", "/", "Chinese", ")", ".", "The", "average", "performance", "over", "the", "four", "languages", "increased", "from", "93.27", "to", "93.92", ".", "Not", "only", "are", "our", "results", "better", ",", "but", "our", "model", "also", "covers", "many", "more", "languages", ",", "so", "it", "can", "potentially", "be", "used", "to", "mine", "bitext", "for", "any", "combination", "of", "the", "93", "languages", "supported", ".", "subsection", ":", "Tatoeba", ":", "similarity", "search", "While", "XNLI", ",", "MLDoc", "and", "BUCC", "are", "well", "established", "benchmarks", "with", "comparative", "results", "available", ",", "they", "only", "cover", "a", "small", "subset", "of", "our", "93", "languages", ".", "So", "as", "to", "better", "assess", "the", "performance", "of", "our", "model", "in", "all", "these", "different", "languages", ",", "we", "introduce", "a", "new", "test", "set", "of", "similarity", "search", "for", "122", "languages", "based", "on", "the", "Tatoeba", "corpus", ".", "The", "dataset", "consists", "of", "up", "to", "1", ",", "000", "English", "-", "aligned", "sentence", "pairs", "for", "each", "language", ".", "Appendix", "[", "reference", "]", "describes", "how", "the", "dataset", "was", "constructed", "in", "more", "details", ".", "Evaluation", "is", "done", "by", "finding", "the", "nearest", "neighbor", "for", "each", "sentence", "in", "the", "other", "language", "according", "to", "cosine", "similarity", "and", "computing", "the", "error", "rate", ".", "We", "report", "our", "results", "in", "Tables", "[", "reference", "]", "and", "[", "reference", "]", ".", "Contrasting", "these", "results", "with", "those", "of", "XNLI", ",", "one", "would", "assume", "that", "similarity", "error", "rates", "below", "5", "%", "are", "indicative", "of", "strong", "downstream", "performance", ".", "This", "is", "the", "case", "for", "37", "languages", ",", "while", "there", "are", "48", "languages", "with", "an", "error", "rate", "below", "10", "%", "and", "55", "with", "less", "than", "20", "%", ",", "covering", "22", "different", "families", "and", "15", "different", "scripts", ".", "There", "are", "only", "15", "languages", "with", "error", "rates", "above", "50", "%", ".", "We", "believe", "that", "our", "competitive", "results", "for", "many", "low", "-", "resource", "languages", "are", "indicative", "of", "the", "benefits", "of", "joint", "training", ",", "which", "is", "also", "supported", "by", "our", "ablation", "results", "in", "Section", "[", "reference", "]", ".", "In", "relation", "to", "that", ",", "Appendix", "[", "reference", "]", "reports", "similarity", "search", "results", "for", "29", "additional", "languages", "without", "any", "training", "data", ",", "showing", "that", "our", "encoder", "can", "also", "generalize", "to", "unseen", "languages", "to", "some", "extent", "as", "long", "as", "it", "was", "trained", "in", "related", "languages", ".", "section", ":", "Ablation", "experiments", "In", "this", "section", ",", "we", "explore", "different", "variants", "of", "our", "approach", "and", "study", "the", "impact", "on", "the", "performance", "for", "all", "our", "evaluation", "tasks", ".", "We", "report", "average", "results", "across", "all", "languages", ".", "For", "XNLI", ",", "we", "also", "report", "the", "accuracy", "on", "English", ".", "subsection", ":", "Encoder", "depth", "Table", "[", "reference", "]", "reports", "the", "performance", "on", "the", "different", "tasks", "for", "encoders", "with", "one", ",", "three", "or", "five", "layers", ".", "We", "were", "not", "able", "to", "achieve", "good", "convergence", "with", "deeper", "models", ".", "It", "can", "be", "seen", "that", "all", "tasks", "benefit", "from", "deeper", "models", ",", "in", "particular", "XNLI", "and", "Tatoeba", ",", "suggesting", "that", "a", "single", "layer", "BiLSTM", "has", "not", "enough", "capacity", "to", "encode", "so", "many", "languages", ".", "subsection", ":", "Multitask", "learning", "Multitask", "learning", "has", "been", "shown", "to", "be", "helpful", "to", "learn", "English", "sentence", "embeddings", ".", "The", "most", "important", "task", "in", "this", "approach", "is", "arguably", "NLI", ",", "so", "we", "explored", "adding", "an", "additional", "NLI", "objective", "to", "our", "system", "with", "different", "weighting", "schemes", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "the", "NLI", "objective", "leads", "to", "a", "better", "performance", "on", "the", "English", "NLI", "test", "set", ",", "but", "this", "comes", "at", "the", "cost", "of", "a", "worse", "cross", "-", "lingual", "transfer", "performance", "in", "XNLI", "and", "Tatoeba", ".", "The", "effect", "in", "BUCC", "is", "negligible", ".", "subsection", ":", "Number", "of", "training", "languages", "So", "as", "to", "better", "understand", "how", "our", "architecture", "scales", "to", "a", "large", "amount", "of", "languages", ",", "we", "train", "a", "separate", "model", "on", "a", "subset", "of", "18", "evaluation", "languages", ",", "and", "compare", "it", "to", "our", "main", "model", "trained", "on", "93", "languages", ".", "We", "replaced", "the", "Tatoeba", "corpus", "with", "the", "WMT", "2014", "test", "set", "to", "evaluate", "the", "multilingual", "similarity", "error", "rate", ".", "This", "covers", "English", ",", "Czech", ",", "French", ",", "German", "and", "Spanish", ",", "so", "results", "between", "both", "models", "are", "directly", "comparable", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "the", "full", "model", "equals", "or", "outperforms", "the", "one", "covering", "the", "evaluation", "languages", "only", "for", "all", "tasks", "but", "MLDoc", ".", "This", "suggests", "that", "the", "joint", "training", "also", "yields", "to", "overall", "better", "representations", ".", "section", ":", "Conclusions", "In", "this", "paper", ",", "we", "propose", "an", "architecture", "to", "learn", "multilingual", "sentence", "embeddings", "for", "93", "languages", ".", "We", "use", "a", "single", "language", "-", "agnostic", "BiLSTM", "encoder", "for", "all", "languages", ",", "which", "is", "trained", "on", "publicly", "available", "parallel", "corpora", "and", "applied", "to", "different", "downstream", "tasks", "without", "any", "fine", "-", "tuning", ".", "Our", "model", "sets", "a", "new", "state", "-", "of", "-", "the", "-", "art", "for", "most", "languages", "in", "zero", "-", "shot", "cross", "-", "lingual", "natural", "language", "inference", "(", "XNLI", ")", ",", "cross", "-", "lingual", "document", "classification", "(", "MLDoc", ")", ",", "and", "bitext", "mining", "(", "BUCC", ")", ".", "We", "also", "introduce", "a", "new", "test", "set", "of", "cross", "-", "lingual", "similarity", "search", "in", "122", "languages", ",", "and", "show", "that", "our", "approach", "is", "competitive", "even", "for", "low", "-", "resource", "languages", ".", "To", "the", "best", "of", "our", "knowledge", ",", "this", "is", "the", "first", "successful", "exploration", "of", "massively", "multilingual", "sentence", "representations", ".", "In", "the", "future", ",", "we", "would", "like", "to", "explore", "alternative", "architectures", "for", "the", "encoder", ".", "In", "particular", ",", "we", "plan", "to", "replace", "our", "BiLSTM", "with", "the", "Transformer", ",", "which", "has", "been", "shown", "to", "work", "better", "in", "different", "settings", "vaswani2017attention", ",", "Devlin:2018:arxiv_bert", ".", "Moreover", ",", "we", "would", "like", "to", "explore", "possible", "strategies", "to", "exploit", "monolingual", "training", "data", "in", "addition", "to", "parallel", "corpora", ",", "such", "as", "using", "pre", "-", "trained", "word", "embeddings", ",", "backtranslation", "sennrich2016improving", ",", "edunov2018understanding", ",", "or", "other", "ideas", "from", "unsupervised", "machine", "translation", "Artetxe:2018:emnlp_unsupmt", ",", "Lample:2018:emnlp_unsupmt", ".", "Finally", ",", "we", "would", "like", "to", "replace", "our", "language", "-", "specific", "tokenization", "and", "BPE", "segmentation", "with", "a", "language", "agnostic", "approach", "similar", "to", "SentencePiece", ".", "The", "model", "and", "code", "used", "in", "this", "paper", "will", "be", "freely", "available", "in", "the", "framework", "of", "the", "LASER", "toolkit", ".", "bibliography", ":", "References", "appendix", ":", "Training", "data", "Our", "training", "data", "consists", "of", "the", "combination", "of", "the", "following", "publicly", "available", "parallel", "corpora", ":", "Europarl", "provides", "high", "-", "quality", "translations", "for", "21", "European", "languages", ".", "The", "size", "varies", "from", "400k", "to", "2", "M", "sentence", "pairs", ",", "in", "function", "of", "the", "date", "the", "respective", "country", "joined", "the", "European", "Union", ".", "United", "Nations", ":", "More", "than", "11", "million", "sentences", "in", "the", "six", "official", "languages", "of", "the", "United", "Nations", ".", "We", "only", "use", "the", "first", "two", "million", "sentences", "in", "Arabic", ",", "Russian", "and", "Chinese", ".", "OpenSubtitles2018", ":", "A", "collection", "of", "translations", "of", "movie", "subtitles", "in", "57", "languages", ".", "The", "corpus", "size", "varies", "from", "few", "thousand", "sentences", "(", "e.g.", "Armenian", "or", "Kazakh", ")", "to", "more", "than", "50", "million", "(", "e.g.", "Spanish", "or", "Romanian", ")", ".", "We", "keep", "at", "most", "2", "million", "entries", "for", "each", "language", "pair", ".", "Global", "Voices", ":", "A", "parallel", "corpus", "of", "news", "stories", "from", "the", "Global", "Voices", "website", "(", "38", "languages", ")", ".", "This", "is", "a", "rather", "small", "corpus", "with", "less", "than", "100k", "sentence", "in", "most", "of", "the", "languages", ".", "Tanzil", ":", "A", "collection", "of", "Quran", "translations", "in", "42", "languages", ".", "The", "style", "and", "vocabulary", "is", "very", "different", "from", "news", "texts", ".", "The", "average", "size", "is", "135k", "sentences", ".", "Tatoeba", ":", "A", "community", "supported", "collection", "of", "English", "sentences", "and", "translations", "into", "more", "than", "300", "languages", ".", "We", "use", "this", "corpus", "to", "extract", "a", "separate", "test", "set", "of", "up", "to", "1", ",", "000", "sentences", "for", "many", "languages", "(", "see", "Section", "[", "reference", "]", "and", "[", "reference", "]", ")", ".", "For", "languages", "with", "more", "than", "1", ",", "000", "entries", ",", "we", "use", "the", "remaining", "ones", "for", "training", ".", "Using", "all", "these", "corpora", "would", "provide", "parallel", "data", "for", "more", "than", "hundred", "languages", ".", "However", ",", "we", "finally", "only", "kept", "93", "different", "languages", "to", "train", "the", "multilingual", "sentence", "embeddings", ".", "In", "particular", ",", "we", "discarded", "several", "constructed", "languages", "with", "little", "practical", "use", "(", "Klingon", ",", "Kotava", ",", "Lojban", ",", "Toki", "Pona", "and", "Volapük", ")", ".", "appendix", ":", "XNLI", "results", "for", "all", "language", "combinations", "Table", "[", "reference", "]", "reports", "the", "accuracies", "of", "our", "system", "on", "the", "XNLI", "test", "set", "when", "the", "premises", "and", "hypothesis", "are", "in", "a", "different", "language", "(", "e.g.", "premise", "in", "Russian", "and", "hypothesis", "in", "Thai", ")", ".", "The", "numbers", "in", "the", "diagonal", "correspond", "to", "the", "main", "results", "reported", "in", "Table", "[", "reference", "]", ".", "We", "observe", "that", "our", "approach", "seems", "to", "handle", "the", "combination", "of", "different", "languages", "very", "well", ".", "We", "do", "not", "have", "evidence", "that", "very", "distant", "languages", "perform", "considerably", "worse", ".", "It", "rather", "seems", "that", "the", "combined", "performance", "is", "mostly", "bounded", "by", "the", "accuracy", "of", "the", "language", "which", "performs", "worst", "when", "used", "alone", ".", "As", "an", "example", ",", "Greek", "-", "Russian", "achieves", "very", "similar", "results", "than", "Bulgarian", "-", "Russian", ",", "two", "Slavic", "languages", ".", "Combing", "French", "with", "Chinese", ",", "two", "totally", "different", "languages", ",", "is", "only", "1.5", "points", "worse", "than", "French", "/", "Spanish", ",", "two", "very", "close", "languages", ".", "appendix", ":", "Tatoeba", "dataset", "Tatoeba", "is", "an", "open", "collection", "of", "English", "sentences", "and", "high", "quality", "translations", "into", "more", "than", "three", "hundred", "languages", ".", "The", "number", "of", "available", "translations", "is", "updated", "every", "Saturday", ".", "We", "downloaded", "the", "snapshot", "on", "November", "19th", "2018", "and", "performed", "the", "following", "processing", ":", "Removal", "of", "sentences", "that", "contain", "“", "@", "”", "or", "“", "http", "”", ".", "This", "is", "motivated", "by", "the", "fact", "that", "emails", "and", "web", "addresses", "are", "not", "language", "specific", ".", "Removal", "of", "sentences", "with", "less", "than", "three", "words", "(", "before", "tokenization", ")", ".", "These", "are", "usually", "sentences", "with", "limited", "semantic", "information", ".", "Removal", "of", "sentences", "that", "appear", "multiple", "times", ",", "either", "in", "the", "source", "or", "the", "target", ".", "After", "filtering", ",", "we", "created", "test", "sets", "of", "up", "to", "1", ",", "000", "aligned", "sentences", "with", "English", ".", "This", "amount", "of", "texts", "is", "available", "for", "78", "languages", ".", "Limiting", "the", "number", "of", "sentences", "to", "500", ",", "we", "increase", "the", "coverage", "to", "101", "languages", ",", "and", "even", "141", "languages", "with", "100", "parallel", "sentences", ".", "It", "should", "be", "stressed", "that", ",", "in", "general", ",", "the", "English", "sentences", "are", "not", "the", "same", "for", "the", "different", "languages", ".", "This", "implies", "that", "the", "error", "rates", "are", "not", "necessarily", "comparable", "between", "the", "languages", ".", "appendix", ":", "Tatoeba", ":", "result", "analysis", "We", "provide", "here", "some", "analysis", "on", "the", "results", "given", "in", "Tables", "[", "reference", "]", "and", "[", "reference", "]", ".", "We", "have", "48", "languages", "with", "an", "error", "rate", "below", "10", "%", "and", "55", "with", "less", "than", "20", "%", ",", "respectively", "(", "English", "included", ")", ".", "The", "languages", "with", "less", "than", "20", "%", "error", "belong", "to", "20", "different", "families", "and", "use", "12", "different", "scripts", ".", "It", "is", "nice", "to", "find", "six", "languages", "in", "this", "list", "for", "which", "we", "have", "only", "small", "amounts", "of", "bitexts", "(", "less", "than", "400k", ")", ",", "namely", "Esperanto", ",", "Galician", ",", "Hindi", ",", "Interlingua", ",", "Malayam", "and", "Marathi", ".", "The", "two", "constructed", "languages", "probably", "benefit", "from", "their", "inspiration", "by", "other", "European", "languages", ".", "Overall", ",", "we", "observe", "low", "similarity", "error", "rates", "on", "the", "Indo", "-", "Aryan", "languages", ",", "namely", "Hindi", ",", "Bengali", ",", "Marathi", "and", "Urdu", ".", "The", "performance", "on", "Berber", "languages", "(", "“", "ber", "”", "and", "“", "kab", "”", ")", "is", "remarkable", ",", "although", "we", "have", "less", "than", "100", "thousand", "sentences", "to", "train", "them", ".", "This", "is", "a", "typical", "example", "of", "languages", "which", "are", "spoken", "by", "several", "millions", "of", "people", ",", "but", "for", "which", "the", "amount", "of", "written", "resources", "is", "very", "limited", ".", "It", "is", "quite", "unlikely", "that", "we", "would", "be", "able", "to", "train", "a", "good", "sentence", "embedding", "with", "language", "specific", "corpora", "only", ".", "This", "clearly", "shows", "the", "benefit", "of", "joint", "training", "on", "many", "languages", ".", "Fifteen", "languages", "have", "similarity", "error", "rates", "of", "more", "than", "50", "%", ".", "Four", "of", "them", "are", "low", "-", "resource", "languages", "with", "their", "own", "script", "and", "which", "are", "alone", "in", "their", "family", ":", "Amharic", ",", "Armenian", ",", "Khmer", "and", "Georgian", ".", "This", "makes", "it", "difficult", "to", "benefit", "from", "joint", "training", ".", "On", "the", "other", "hand", ",", "one", "can", "also", "argue", "that", "is", "surprising", "that", "a", "language", "like", "Khmer", "performs", "much", "better", "than", "random", "(", "99.9", "%", "error", "rate", ")", "with", "only", "625", "training", "examples", ".", "Khmer", "probably", "benefits", "of", "the", "fact", "that", "he", "have", "trained", "our", "model", "on", "other", "languages", "of", "the", "region", "which", "have", "influenced", "Khmer", ",", "namely", "Thai", "and", "Vietnamese", ".", "There", "are", "also", "several", "Turkic", "languages", "(", "Kazakh", ",", "Tatar", ",", "Uighur", "and", "Uzbek", ")", "and", "Celtic", "languages", "(", "Breton", "and", "Cornish", ")", "with", "high", "error", "rates", ".", "We", "hope", "to", "improve", "their", "performance", "in", "the", "future", ".", "appendix", ":", "Tatoeba", ":", "results", "for", "unseen", "languages", "We", "extend", "our", "Tatoeba", "experiments", "to", "29", "languages", "without", "any", "training", "data", "(", "see", "Table", "[", "reference", "]", ")", ".", "Many", "of", "them", "are", "recognized", "minority", "languages", "spoken", "in", "specific", "regions", ",", "e.g.", "Asturian", ",", "Faroese", ",", "Frisian", ",", "Kashubian", ",", "North", "Moluccan", "Malay", ",", "Piemontese", ",", "Swabian", "or", "Sorbian", ".", "All", "share", "some", "similarities", ",", "at", "various", "degrees", ",", "with", "other", "major", "languages", ",", "but", "also", "differ", "by", "their", "own", "grammar", "or", "specific", "vocabulary", ".", "This", "enables", "our", "encoder", "to", "perform", "reasonably", "well", ".", "We", "can", "probably", "assume", "that", "these", "are", "mainly", "spoken", "languages", "with", "limited", "resources", "in", "written", "form", ".", "The", "six", "languages", "which", "perform", "worst", "are", "Mongolian", ",", "Welsh", ",", "Xhosa", "Pampangan", ",", "Yiddish", "and", "Gaelic", ".", "We", "include", "these", "results", "here", "as", "baseline", "for", "future", "research", ".", "Premise", "Training", "[", "-", "1pt", "]", "corpus", "[", "-", "2pt", "]", "size", "Tatoeba", "[", "-", "1pt", "]", "test", "set", "[", "-", "2pt", "]", "size"], "entity": [[2, 6, ["Massively", "Multilingual", "Sentence", "Embeddings"], "Task"], [7, 14, ["Zero", "-", "Shot", "Cross", "-", "Lingual", "Transfer"], "Task"], [50, 52, ["BiLSTM", "encoder"], "Method"], [66, 68, ["auxiliary", "decoder"], "Method"], [82, 83, ["classifier"], "Method"], [125, 131, ["cross", "-", "lingual", "natural", "language", "inference"], "Task"], [138, 140, ["XNLI", "dataset"], "Material"], [150, 155, ["cross", "-", "lingual", "document", "classification"], "Task"], [156, 158, ["MLDoc", "dataset"], "Material"], [161, 163, ["sentence", "embeddings"], "Method"], [167, 170, ["parallel", "corpus", "mining"], "Task"], [218, 220, ["sentence", "embeddings"], "Method"], [224, 227, ["multilingual", "similarity", "search"], "Task"], [235, 237, ["PyTorch", "implementation"], "Method"], [238, 242, ["pre", "-", "trained", "encoder"], "Method"], [260, 262, ["deep", "learning"], "Method"], [268, 271, ["Natural", "Language", "Processing"], "Task"], [272, 273, ["NLP"], "Task"], [305, 308, ["general", "language", "representations"], "Method"], [317, 322, ["task", "-", "specific", "downstream", "systems"], "Task"], [329, 331, ["word", "embeddings"], "Method"], [341, 345, ["sentence", "-", "level", "representations"], "Method"], [402, 404, ["vector", "representations"], "Method"], [451, 455, ["zero", "-", "shot", "transfer"], "Method"], [474, 477, ["code", "-", "switching"], "Task"], [485, 486, ["encoder"], "Method"], [611, 613, ["shared", "encoder"], "Method"], [669, 675, ["cross", "-", "lingual", "natural", "language", "inference"], "Task"], [676, 678, ["XNLI", "dataset"], "Material"], [680, 681, ["classification"], "Task"], [682, 684, ["MLDoc", "dataset"], "Material"], [686, 688, ["bitext", "mining"], "Task"], [689, 691, ["BUCC", "dataset"], "Material"], [693, 696, ["multilingual", "similarity", "search"], "Task"], [728, 731, ["fine", "-", "tuning"], "Method"], [755, 756, ["accuracy"], "Metric"], [757, 760, ["multilingual", "similarity", "search"], "Task"], [857, 859, ["word", "embeddings"], "Task"], [871, 874, ["continuous", "vector", "representations"], "Method"], [889, 892, ["Recurrent", "Neural", "Network"], "Method"], [893, 894, ["RNN"], "Method"], [903, 905, ["unsupervised", "way"], "Method"], [916, 920, ["skip", "-", "thought", "model"], "Method"], [924, 925, ["encoder"], "Method"], [927, 929, ["auxiliary", "decoder"], "Method"], [966, 967, ["encoder"], "Method"], [969, 972, ["Natural", "Language", "Inference"], "Task"], [973, 974, ["NLI"], "Task"], [983, 985, ["multitask", "learning"], "Task"], [988, 990, ["training", "objectives"], "Metric"], [993, 996, ["skip", "-", "thought"], "Task"], [997, 998, ["NLI"], "Task"], [999, 1001, ["machine", "translation"], "Task"], [1019, 1021, ["multilingual", "representations"], "Method"], [1036, 1041, ["cross", "-", "lingual", "word", "embeddings"], "Task"], [1097, 1099, ["unsupervised", "manner"], "Method"], [1100, 1105, ["Cross", "-", "lingual", "word", "embeddings"], "Method"], [1110, 1116, ["bag", "-", "of", "-", "word", "representations"], "Method"], [1154, 1156, ["sentence", "embeddings"], "Method"], [1161, 1166, ["cross", "-", "lingual", "transfer", "settings"], "Task"], [1180, 1189, ["sequence", "-", "to", "-", "sequence", "encoder", "-", "decoder", "architecture"], "Method"], [1208, 1211, ["neural", "machine", "translation"], "Task"], [1223, 1225, ["vector", "representation"], "Method"], [1231, 1232, ["decoder"], "Method"], [1246, 1247, ["encoder"], "Method"], [1304, 1307, ["pairwise", "joint", "embeddings"], "Method"], [1308, 1309, ["English"], "Material"], [1327, 1329, ["multilingual", "representations"], "Task"], [1338, 1340, ["word", "embeddings"], "Method"], [1362, 1366, ["fixed", "-", "length", "representation"], "Method"], [1380, 1384, ["variable", "-", "length", "representations"], "Method"], [1424, 1426, ["language", "modeling"], "Method"], [1428, 1429, ["classifier"], "Method"], [1438, 1439, ["encoder"], "Method"], [1449, 1451, ["supervised", "training"], "Task"], [1460, 1462, ["monolingual", "settings"], "Task"], [1466, 1470, ["fixed", "-", "length", "approaches"], "Method"], [1482, 1484, ["multilingual", "scenario"], "Task"], [1491, 1494, ["multilingual", "BERT", "model"], "Method"], [1496, 1500, ["zero", "-", "shot", "transfer"], "Task"], [1524, 1526, ["sentence", "embeddings"], "Method"], [1532, 1534, ["auxiliary", "decoder"], "Method"], [1556, 1558, ["training", "strategy"], "Method"], [1609, 1613, ["max", "-", "pooling", "operation"], "Method"], [1618, 1620, ["BiLSTM", "encoder"], "Method"], [1629, 1631, ["decoder", "LSTM"], "Method"], [1633, 1635, ["linear", "transformation"], "Method"], [1658, 1659, ["encoder"], "Method"], [1661, 1662, ["decoder"], "Method"], [1678, 1680, ["sentence", "embedding"], "Method"], [1685, 1688, ["encoder", "and", "decoder"], "Method"], [1708, 1712, ["byte", "-", "pair", "encoding"], "Method"], [1713, 1714, ["BPE"], "Method"], [1751, 1754, ["language", "independent", "representations"], "Method"], [1759, 1760, ["decoder"], "Method"], [1762, 1765, ["language", "ID", "embedding"], "Method"], [1809, 1810, ["encoder"], "Method"], [1824, 1826, ["stacked", "BiLSTM"], "Method"], [1839, 1841, ["sentence", "representations"], "Task"], [1852, 1853, ["decoder"], "Method"], [1863, 1865, ["embedding", "size"], "Metric"], [1872, 1875, ["language", "ID", "embedding"], "Method"], [1999, 2001, ["quadratic", "cost"], "Metric"], [2010, 2011, ["training"], "Task"], [2088, 2089, ["Training"], "Method"], [2091, 2095, ["cross", "-", "entropy", "loss"], "Metric"], [2115, 2116, ["Adam"], "Method"], [2119, 2121, ["learning", "rate"], "Metric"], [2143, 2144, ["fairseq"], "Method"], [2151, 2155, ["multi", "-", "GPU", "support"], "Method"], [2159, 2162, ["NVIDIA", "V100", "GPUs"], "Method"], [2206, 2209, ["pre", "-", "processing"], "Task"], [2217, 2218, ["training"], "Task"], [2228, 2229, ["English"], "Material"], [2362, 2364, ["sentence", "embeddings"], "Method"], [2371, 2372, ["BUCC"], "Material"], [2373, 2374, ["MLDoc"], "Material"], [2381, 2382, ["encoder"], "Method"], [2394, 2395, ["XNLI"], "Material"], [2414, 2418, ["general", "purpose", "sentence", "encoder"], "Method"], [2464, 2467, ["pre", "-", "processing"], "Task"], [2470, 2472, ["Moses", "tools"], "Method"], [2473, 2475, ["punctuation", "normalization"], "Method"], [2536, 2538, ["evaluation", "frameworks"], "Method"], [2567, 2569, ["XNLI", "corpus"], "Material"], [2601, 2604, ["multilingual", "sentence", "embedding"], "Method"], [2605, 2608, ["English", "training", "data"], "Material"], [2612, 2616, ["zero", "-", "shot", "transfer"], "Metric"], [2650, 2655, ["cross", "-", "lingual", "document", "classification"], "Task"], [2656, 2657, ["MLDoc"], "Material"], [2665, 2667, ["bitext", "mining"], "Task"], [2668, 2669, ["BUCC"], "Material"], [2699, 2702, ["multilingual", "similarity", "search"], "Task"], [2754, 2755, ["XNLI"], "Material"], [2760, 2761, ["NLI"], "Task"], [2769, 2771, ["sentence", "representations"], "Task"], [2803, 2804, ["XNLI"], "Material"], [2815, 2817, ["English", "MultiNLI"], "Material"], [2868, 2871, ["English", "training", "data"], "Material"], [2876, 2877, ["transfer"], "Task"], [2889, 2890, ["classifier"], "Method"], [2894, 2896, ["multilingual", "encoder"], "Method"], [2922, 2927, ["feed", "-", "forward", "neural", "network"], "Method"], [2939, 2940, ["Adam"], "Method"], [2942, 2943, ["hyperparameters"], "Method"], [2957, 2958, ["classifier"], "Method"], [3031, 3034, ["multilingual", "BERT", "model"], "Method"], [3056, 3063, ["zero", "-", "shot", "cross", "-", "lingual", "transfer"], "Task"], [3067, 3068, ["classifier"], "Method"], [3069, 3071, ["English", "data"], "Material"], [3083, 3084, ["Spanish"], "Material"], [3102, 3106, ["zero", "-", "short", "performance"], "Metric"], [3118, 3119, ["English"], "Material"], [3148, 3149, ["BERT"], "Method"], [3153, 3154, ["English"], "Material"], [3164, 3172, ["zero", "-", "shot", "cross", "-", "lingual", "transfer", "performance"], "Metric"], [3180, 3181, ["loss"], "Metric"], [3182, 3183, ["accuracy"], "Metric"], [3202, 3203, ["BERT"], "Task"], [3227, 3231, ["pre", "-", "trained", "encoder"], "Method"], [3233, 3236, ["X", "-", "BiLSTM"], "Method"], [3263, 3265, ["Machine", "Translation"], "Task"], [3266, 3267, ["MT"], "Task"], [3284, 3285, ["English"], "Material"], [3297, 3300, ["English", "training", "data"], "Material"], [3335, 3337, ["monolingual", "model"], "Method"], [3343, 3344, ["MT"], "Task"], [3353, 3355, ["translating", "test"], "Task"], [3356, 3357, ["inference"], "Task"], [3362, 3364, ["translating", "train"], "Task"], [3385, 3387, ["translation", "baselines"], "Metric"], [3469, 3470, ["MLDoc"], "Material"], [3471, 3475, ["cross", "-", "lingual", "classification"], "Task"], [3475, 3480, ["Cross", "-", "lingual", "document", "classification"], "Task"], [3485, 3487, ["multilingual", "representations"], "Task"], [3493, 3495, ["sentence", "embeddings"], "Method"], [3502, 3504, ["MLDoc", "dataset"], "Material"], [3566, 3571, ["zero", "-", "shot", "transfer", "scenario"], "Task"], [3575, 3576, ["classifier"], "Method"], [3580, 3582, ["multilingual", "encoder"], "Method"], [3584, 3587, ["English", "training", "data"], "Material"], [3594, 3597, ["English", "development", "set"], "Material"], [3611, 3616, ["feed", "-", "forward", "neural", "network"], "Method"], [3618, 3620, ["hidden", "layer"], "Method"], [3666, 3667, ["MLDoc"], "Material"], [3682, 3683, ["BUCC"], "Material"], [3684, 3686, ["bitext", "mining"], "Task"], [3686, 3688, ["Bitext", "mining"], "Task"], [3746, 3749, ["nearest", "neighbor", "retrieval"], "Method"], [3758, 3760, ["cosine", "score"], "Metric"], [3831, 3832, ["CSLS"], "Method"], [3855, 3857, ["sentence", "embeddings"], "Method"], [3887, 3888, ["English"], "Material"], [3893, 3894, ["German"], "Material"], [3895, 3896, ["French"], "Material"], [3947, 3949, ["sentence", "embeddings"], "Method"], [3967, 3971, ["English", "-", "Chinese", "test"], "Material"], [3999, 4000, ["English"], "Material"], [4059, 4061, ["similarity", "search"], "Task"], [4062, 4063, ["XNLI"], "Material"], [4064, 4065, ["MLDoc"], "Material"], [4066, 4067, ["BUCC"], "Material"], [4110, 4112, ["similarity", "search"], "Task"], [4170, 4172, ["cosine", "similarity"], "Metric"], [4175, 4177, ["error", "rate"], "Metric"], [4198, 4199, ["XNLI"], "Material"], [4204, 4207, ["similarity", "error", "rates"], "Metric"], [4214, 4215, ["downstream"], "Task"], [4232, 4234, ["error", "rate"], "Metric"], [4260, 4262, ["error", "rates"], "Metric"], [4284, 4286, ["joint", "training"], "Method"], [4311, 4313, ["similarity", "search"], "Task"], [4348, 4349, ["Ablation"], "Task"], [4371, 4373, ["evaluation", "tasks"], "Task"], [4383, 4384, ["XNLI"], "Material"], [4389, 4390, ["accuracy"], "Metric"], [4391, 4392, ["English"], "Material"], [4409, 4410, ["encoders"], "Method"], [4427, 4429, ["deeper", "models"], "Method"], [4439, 4441, ["deeper", "models"], "Method"], [4444, 4445, ["XNLI"], "Material"], [4451, 4454, ["single", "layer", "BiLSTM"], "Method"], [4466, 4468, ["Multitask", "learning"], "Task"], [4468, 4470, ["Multitask", "learning"], "Method"], [4478, 4481, ["English", "sentence", "embeddings"], "Method"], [4491, 4492, ["NLI"], "Task"], [4506, 4508, ["weighting", "schemes"], "Method"], [4541, 4546, ["cross", "-", "lingual", "transfer", "performance"], "Metric"], [4547, 4548, ["XNLI"], "Material"], [4554, 4555, ["BUCC"], "Material"], [4619, 4623, ["multilingual", "similarity", "error", "rate"], "Metric"], [4626, 4627, ["English"], "Material"], [4630, 4631, ["French"], "Material"], [4632, 4633, ["German"], "Material"], [4634, 4635, ["Spanish"], "Material"], [4670, 4671, ["MLDoc"], "Material"], [4676, 4678, ["joint", "training"], "Method"], [4730, 4732, ["downstream", "tasks"], "Task"], [4734, 4737, ["fine", "-", "tuning"], "Method"], [4764, 4765, ["XNLI"], "Material"], [4767, 4772, ["cross", "-", "lingual", "document", "classification"], "Task"], [4773, 4774, ["MLDoc"], "Material"], [4777, 4779, ["bitext", "mining"], "Task"], [4780, 4781, ["BUCC"], "Material"], [4791, 4796, ["cross", "-", "lingual", "similarity", "search"], "Task"], [4846, 4847, ["encoder"], "Method"], [4856, 4857, ["BiLSTM"], "Method"], [4859, 4860, ["Transformer"], "Method"], [4901, 4903, ["word", "embeddings"], "Method"], [4913, 4916, ["unsupervised", "machine", "translation"], "Method"], [4928, 4932, ["language", "-", "specific", "tokenization"], "Method"], [4937, 4940, ["language", "agnostic", "approach"], "Method"], [4961, 4963, ["LASER", "toolkit"], "Method"], [5088, 5089, ["Spanish"], "Material"], [5177, 5179, ["English", "sentences"], "Material"], [5306, 5307, ["accuracies"], "Metric"], [5394, 5395, ["accuracy"], "Metric"], [5426, 5427, ["French"], "Material"], [5580, 5581, ["English"], "Material"], [5627, 5629, ["English", "sentences"], "Material"], [5642, 5644, ["error", "rates"], "Metric"], [5683, 5685, ["error", "rate"], "Metric"], [5778, 5781, ["similarity", "error", "rates"], "Metric"], [5867, 5869, ["sentence", "embedding"], "Method"], [5881, 5883, ["joint", "training"], "Method"], [5890, 5893, ["similarity", "error", "rates"], "Metric"], [5934, 5936, ["joint", "training"], "Method"], [5953, 5954, ["Khmer"], "Method"], [5958, 5959, ["random"], "Method"], [5962, 5964, ["error", "rate"], "Metric"], [6024, 6026, ["error", "rates"], "Metric"], [6176, 6178, ["Premise", "Training"], "Task"], [24, 26, ["sentence", "representations"], "Task"], [55, 56, ["BPE"], "Method"], [88, 90, ["sentence", "embeddings"], "Method"], [183, 184, ["BUCC"], "Material"], [396, 398, ["sentence", "embeddings"], "Method"], [420, 421, ["NLP"], "Task"], [457, 458, ["NLP"], "Task"], [512, 513, ["NLP"], "Task"], [549, 551, ["sentence", "representations"], "Task"], [605, 607, ["sentence", "representations"], "Task"], [717, 719, ["BiLSTM", "encoder"], "Method"], [773, 774, ["MT"], "Task"], [882, 884, ["sentence", "embeddings"], "Method"], [1350, 1352, ["sentence", "representations"], "Task"], [1411, 1412, ["RNN"], "Method"], [1519, 1521, ["BiLSTM", "encoder"], "Method"], [1602, 1604, ["sentence", "embeddings"], "Method"], [1622, 1624, ["sentence", "embeddings"], "Method"], [1779, 1781, ["sentence", "embeddings"], "Method"], [2230, 2231, ["Spanish"], "Material"], [2540, 2542, ["sentence", "representations"], "Task"], [2554, 2556, ["sentence", "embeddings"], "Method"], [2571, 2572, ["NLI"], "Task"], [2595, 2596, ["NLI"], "Task"], [2642, 2644, ["sentence", "representations"], "Task"], [2739, 2741, ["BiLSTM", "encoder"], "Method"], [2759, 2760, ["NLI"], "Task"], [2903, 2905, ["sentence", "embeddings"], "Method"], [2948, 2949, ["XNLI"], "Material"], [2965, 2966, ["XNLI"], "Material"], [2995, 2997, ["sentence", "embeddings"], "Method"], [3289, 3290, ["NLI"], "Task"], [3305, 3306, ["NLI"], "Task"], [3321, 3323, ["sentence", "embeddings"], "Method"], [3331, 3332, ["MT"], "Task"], [3398, 3399, ["MT"], "Task"], [3463, 3464, ["French"], "Material"], [3560, 3561, ["XNLI"], "Material"], [3694, 3696, ["sentence", "embeddings"], "Method"], [3859, 3860, ["BUCC"], "Material"], [3993, 3994, ["French"], "Material"], [3995, 3996, ["Spanish"], "Material"], [3997, 3998, ["German"], "Material"], [4001, 4002, ["French"], "Material"], [4499, 4500, ["NLI"], "Task"], [4518, 4519, ["NLI"], "Task"], [4528, 4529, ["NLI"], "Task"], [4699, 4701, ["sentence", "embeddings"], "Method"], [4712, 4714, ["BiLSTM", "encoder"], "Method"], [4757, 4763, ["cross", "-", "lingual", "natural", "language", "inference"], "Task"], [4830, 4832, ["sentence", "representations"], "Task"], [4933, 4934, ["BPE"], "Method"], [5264, 5266, ["sentence", "embeddings"], "Method"], [5294, 5295, ["XNLI"], "Material"], [5312, 5313, ["XNLI"], "Material"], [5441, 5442, ["French"], "Material"], [5443, 5444, ["Spanish"], "Material"]]} {"doc_id": "175f74a09241b6cb5101a2a09978095720db7d5f", "text": ["document", ":", "Image", "Super", "-", "Resolution", "via", "Dual", "-", "State", "Recurrent", "Networks", "Advances", "in", "image", "super", "-", "resolution", "(", "SR", ")", "have", "recently", "benefited", "significantly", "from", "rapid", "developments", "in", "deep", "neural", "networks", ".", "Inspired", "by", "these", "recent", "discoveries", ",", "we", "note", "that", "many", "state", "-", "of", "-", "the", "-", "art", "deep", "SR", "architectures", "can", "be", "reformulated", "as", "a", "single", "-", "state", "recurrent", "neural", "network", "(", "RNN", ")", "with", "finite", "unfoldings", ".", "In", "this", "paper", ",", "we", "explore", "new", "structures", "for", "SR", "based", "on", "this", "compact", "RNN", "view", ",", "leading", "us", "to", "a", "dual", "-", "state", "design", ",", "the", "Dual", "-", "State", "Recurrent", "Network", "(", "DSRN", ")", ".", "Compared", "to", "its", "single", "-", "state", "counterparts", "that", "operate", "at", "a", "fixed", "spatial", "resolution", ",", "DSRN", "exploits", "both", "low", "-", "resolution", "(", "LR", ")", "and", "high", "-", "resolution", "(", "HR", ")", "signals", "jointly", ".", "Recurrent", "signals", "are", "exchanged", "between", "these", "states", "in", "both", "directions", "(", "both", "LR", "to", "HR", "and", "HR", "to", "LR", ")", "via", "delayed", "feedback", ".", "Extensive", "quantitative", "and", "qualitative", "evaluations", "on", "benchmark", "datasets", "and", "on", "a", "recent", "challenge", "demonstrate", "that", "the", "proposed", "DSRN", "performs", "favorably", "against", "state", "-", "of", "-", "the", "-", "art", "algorithms", "in", "terms", "of", "both", "memory", "consumption", "and", "predictive", "accuracy", ".", "section", ":", "Introduction", "In", "the", "problem", "of", "single", "-", "image", "super", "-", "resolution", "(", "SR", ")", ",", "the", "aim", "is", "to", "recover", "a", "high", "-", "resolution", "(", "HR", ")", "image", "from", "a", "single", "low", "-", "resolution", "(", "LR", ")", "image", ".", "In", "recent", "years", ",", "SR", "performance", "has", "been", "significantly", "improved", "due", "to", "rapid", "developments", "in", "deep", "neural", "networks", "(", "DNNs", ")", ".", "Specifically", ",", "convolutional", "neural", "networks", "(", "CNNs", ")", "and", "residual", "learning", "have", "been", "widely", "applied", "in", "much", "recent", "SR", "work", ".", "In", "these", "approaches", ",", "two", "principles", "have", "been", "consistently", "observed", ".", "The", "first", "is", "that", "increasing", "the", "depth", "of", "a", "CNN", "model", "improves", "SR", "performance", ";", "a", "deeper", "model", "with", "more", "parameters", "can", "represent", "a", "more", "complex", "mapping", "from", "LR", "to", "HR", "images", ".", "In", "addition", ",", "increasing", "network", "depth", "enlarges", "the", "size", "of", "receptive", "fields", ",", "providing", "more", "contextual", "information", "that", "can", "be", "exploited", "to", "reconstruct", "missing", "HR", "components", ".", "The", "second", "principle", "is", "that", "adding", "residual", "connections", "(", "globally", ",", "locally", "or", "jointly", ")", "prevents", "the", "problems", "of", "vanishing", "and", "exploding", "gradients", ",", "facilitating", "the", "training", "of", "deep", "models", ".", "While", "these", "recent", "models", "have", "demonstrated", "promising", "results", ",", "there", "are", "also", "drawbacks", ".", "One", "major", "issue", "is", "that", "increasing", "the", "depth", "of", "models", "by", "adding", "new", "layers", "introduces", "more", "parameters", ",", "and", "thus", "raises", "the", "likelihood", "of", "model", "overfitting", ".", "At", "the", "same", "time", ",", "larger", "models", "demand", "more", "storage", "space", ",", "which", "is", "a", "hurdle", "to", "deployment", "in", "resource", "-", "constrained", "environments", "(", "e.g.", "mobile", "systems", ")", ".", "To", "resolve", "this", "issue", ",", "the", "Deep", "Recursive", "Residual", "Network", "(", "DRRN", ")", "inspired", "by", "the", "Deeply", "-", "Recursive", "Convolutional", "Network", "(", "DRCN", ")", "shares", "weights", "across", "different", "residual", "units", "and", "achieves", "state", "-", "of", "-", "the", "-", "art", "performance", "with", "a", "small", "number", "of", "parameters", ".", "Separate", "efforts", "in", "neural", "architectural", "design", "have", "recently", "shown", "that", "commonly", "-", "used", "deep", "structures", "can", "be", "represented", "more", "compactly", "using", "recurrent", "neural", "networks", "(", "RNNs", ")", ".", "Specifically", ",", "Liao", "and", "Poggio", "demonstrated", "that", "a", "weight", "-", "sharing", "Residual", "Neural", "Network", "(", "ResNet", ")", "is", "equivalent", "to", "a", "shallow", "RNN", ".", "Inspired", "by", "their", "findings", ",", "we", "first", "explore", "the", "connections", "between", "the", "neural", "architectures", "of", "existing", "SR", "algorithms", "and", "their", "compact", "RNN", "formulations", ".", "We", "note", "that", "previous", "SR", "models", "with", "recursive", "computation", "and", "weight", "sharing", ",", "including", "DRRN", "and", "DRCN", ",", "work", "at", "a", "single", "spatial", "resolution", "(", "bicubic", "interpolation", "is", "first", "applied", "to", "upscale", "LR", "images", "to", "a", "desired", "spatial", "resolution", ")", ".", "This", "enables", "their", "model", "structures", "to", "be", "represented", "as", "a", "unified", "single", "-", "state", "RNN", ".", "Thus", ",", "both", "DRRN", "and", "DRCN", "can", "be", "viewed", "as", "a", "finite", "unfolding", "in", "time", "of", "the", "same", "RNN", "structure", ",", "but", "with", "different", "transition", "functions", ".", "This", "is", "illustrated", "in", "Figure", "[", "reference", "]", ",", "and", "will", "be", "discussed", "in", "detail", "in", "Section", "[", "reference", "]", ".", "It", "is", "worth", "mentioning", "that", "we", "follow", "the", "terminology", "used", "in", ",", "where", "a", "‘", "‘", "state", "’", "’", "can", "be", "considered", "as", "corresponding", "to", "a", "‘", "‘", "layer", "’", "’", "in", "the", "normal", "RNN", "setting", ".", "Based", "on", "this", "compact", "RNN", "view", "of", "state", "-", "of", "-", "the", "-", "art", "SR", "models", ",", "in", "this", "paper", "we", "explore", "new", "structures", "to", "extend", "the", "frontier", "of", "SR", ".", "The", "first", "approach", "in", "improving", "a", "conventional", "RNN", "model", "is", "generally", "to", "make", "it", "multi", "-", "layer", ".", "We", "apply", "this", "experience", "in", "designing", "the", "SR", "architecture", "in", "our", "compact", "RNN", "view", "by", "adding", "an", "additional", "state", ",", "rendering", "our", "model", "a", "Dual", "-", "State", "Recurrent", "Network", "(", "DSRN", ")", ",", "where", "the", "two", "states", "operate", "at", "different", "spatial", "resolutions", ".", "Specifically", ",", "the", "bottom", "state", "captures", "information", "at", "LR", ",", "while", "the", "top", "state", "operates", "in", "the", "HR", "regime", ".", "As", "with", "a", "conventional", "two", "-", "layer", "stacked", "RNN", ",", "there", "is", "a", "connection", "from", "the", "bottom", "to", "the", "top", "state", "via", "deconvolutional", "operations", ".", "This", "provides", "information", "flow", "from", "LR", "to", "HR", "at", "every", "single", "unrolling", "time", ".", "In", "addition", ",", "to", "allow", "information", "flow", "from", "previously", "predicted", "HR", "features", "to", "LR", "features", ",", "we", "incorporate", "a", "delayed", "feedback", "mechanism", "from", "the", "top", "(", "HR", ")", "state", "to", "the", "bottom", "one", ".", "The", "overall", "structure", "of", "the", "proposed", "DSRN", "is", "shown", "in", "Figure", "[", "reference", "]", ",", "which", "not", "only", "utilizes", "parameters", "efficiently", "but", "also", "allows", "both", "LR", "and", "HR", "signals", "to", "contribute", "jointly", "to", "learning", "the", "mappings", ".", "To", "demonstrate", "the", "effectiveness", "of", "the", "proposed", "method", ",", "we", "compare", "DSRN", "with", "other", "recent", "image", "SR", "approaches", "on", "four", "common", "benchmarks", "as", "well", "as", "on", "the", "DIV2", "K", "dataset", "from", "the", "\"", "New", "Trends", "in", "Image", "Restoration", "and", "Enhancement", "workshop", "and", "challenge", "on", "image", "super", "-", "resolution", "(", "NTIRE", "SR", "2017", ")", "\"", ".", "Extensive", "experimental", "results", "validate", "that", "DSRN", "delivers", "higher", "parameter", "efficiency", ",", "low", "memory", "consumption", "and", "high", "restoration", "accuracy", ".", "section", ":", "Related", "Work", "Single", "image", "SR", "has", "been", "widely", "studied", "in", "the", "past", "few", "decades", "and", "has", "an", "extensive", "literature", ".", "In", "recent", "years", ",", "due", "to", "the", "fast", "development", "of", "deep", "learning", ",", "significant", "progress", "has", "been", "made", "in", "this", "field", ".", "Dong", "et", "al", ".", "first", "exploited", "a", "fully", "convolutional", "neural", "network", ",", "termed", "SRCNN", ",", "to", "predict", "the", "nonlinear", "LR", "-", "HR", "mapping", ".", "It", "demonstrated", "superior", "performance", "to", "many", "other", "example", "-", "based", "learning", "paradigms", ",", "such", "as", "nearest", "neighbor", ",", "sparse", "representation", ",", "neighborhood", "embedding", ",", "random", "forest", ",", "etc", ".", "Although", "all", "layers", "of", "a", "SRCNN", "are", "trained", "jointly", "in", "an", "end", "-", "to", "-", "end", "fashion", ",", "conceptually", "the", "network", "is", "split", "into", "three", "stages", ":", "patch", "representation", ",", "non", "-", "linear", "mapping", ",", "and", "reconstruction", ".", "Much", "of", "the", "later", "work", "follows", "a", "similar", "network", "design", "with", "more", "complicated", "building", "blocks", "or", "advanced", "optimization", "techniques", ".", "Wang", "et", "al", ".", "proposed", "a", "sparse", "coding", "network", "(", "SCN", ")", "that", "encodes", "a", "sparse", "representation", "prior", "for", "image", "SR", "and", "can", "be", "trained", "end", "-", "to", "-", "end", ",", "demonstrating", "the", "benefit", "of", "domain", "expertise", "in", "sparse", "coding", "for", "image", "SR", ".", "Both", "external", "and", "self", "examples", "were", "utilized", "to", "synthesize", "the", "HR", "prediction", "via", "a", "neural", "network", "in", ".", "Inspired", "by", "the", "success", "of", "very", "deep", "models", "on", "ImageNet", "challenges", ",", "Kim", "et", "al", ".", "proposed", "a", "very", "deep", "CNN", ",", "VDSR", ",", "which", "stacks", "20", "convolutional", "layers", "with", "kernels", ".", "Both", "residual", "learning", "and", "adjustable", "gradient", "clipping", "are", "used", "to", "prevent", "vanishing", "and", "exploding", "gradients", ".", "However", ",", "as", "the", "model", "gets", "deeper", ",", "the", "number", "of", "parameters", "increases", ".", "To", "control", "the", "size", "of", "the", "model", ",", "DRCN", "introduces", "16", "recursive", "layers", ",", "each", "with", "the", "same", "structure", "and", "shared", "parameters", ".", "Moreover", ",", "DRCN", "makes", "use", "of", "skip", "connections", "and", "recursive", "supervision", "to", "mitigate", "the", "difficulty", "of", "training", ".", "Tai", "et", "al", ".", "discovered", "that", "many", "residual", "SR", "learning", "algorithms", "are", "based", "on", "either", "global", "residual", "learning", "or", "local", "residual", "learning", ",", "which", "are", "insufficient", "for", "very", "deep", "models", ".", "Instead", ",", "they", "proposed", "the", "DRRN", "that", "applies", "both", "global", "and", "local", "learning", "while", "remaining", "parameter", "efficient", "via", "recursive", "learning", ".", "More", "recently", ",", "Tong", "et", "al", ".", "proposed", "making", "use", "of", "Densely", "Connected", "Networks", "(", "DenseNet", ")", "instead", "of", "ResNet", "as", "the", "building", "block", "for", "image", "SR", ".", "They", "demonstrated", "that", "the", "DenseNet", "structure", "is", "better", "at", "combining", "features", "at", "different", "levels", ",", "which", "boosts", "SR", "performance", ".", "Apart", "from", "deep", "models", "working", "on", "bicubic", "upscaled", "input", "images", ",", "Shi", "et", "al", ".", "used", "a", "compact", "network", "model", "to", "conduct", "convolutions", "on", "LR", "images", "directly", "and", "learned", "upscaling", "filters", "in", "the", "last", "layer", ",", "which", "considerably", "reduces", "the", "computation", "cost", ".", "Similarly", ",", "Dong", "et", "al", ".", "adopted", "deconvolution", "layers", "to", "accelerate", "SRCNN", "in", "combination", "with", "smaller", "filter", "sizes", "and", "more", "convolution", "layers", ".", "However", ",", "these", "networks", "are", "relatively", "small", "and", "have", "difficulty", "capturing", "complicated", "mappings", "owing", "to", "limited", "network", "capacity", ".", "The", "Laplacian", "Pyramid", "Super", "-", "Resolution", "Network", "(", "LapSRN", ")", "works", "on", "LR", "images", "directly", "and", "progressively", "predicts", "sub", "-", "band", "residuals", "on", "various", "scales", ".", "Lim", "et", "al", ".", "proposed", "the", "Enhanced", "Deep", "Super", "-", "Resolution", "(", "EDSR", ")", "network", "and", "a", "multi", "-", "scale", "variant", ",", "which", "learns", "different", "scaled", "mapping", "functions", "in", "parallel", "via", "weight", "sharing", ".", "It", "is", "noteworthy", "that", "most", "SR", "algorithms", "minimize", "the", "mean", "squared", "reconstruction", "error", "(", "i.e.", "via", "loss", ")", ".", "They", "often", "suffer", "from", "regression", "-", "to", "-", "the", "-", "mean", "due", "to", "the", "ill", "-", "posed", "nature", "of", "single", "image", "SR", ",", "resulting", "in", "blurry", "predictions", "and", "poor", "subjective", "scores", ".", "To", "overcome", "this", "drawback", ",", "Generative", "Adversarial", "Networks", "have", "been", "used", "along", "with", "perceptual", "loss", "for", "SR", ".", "Subjective", "evaluation", "by", "mean", "-", "opinion", "-", "score", "showed", "huge", "improvement", "over", "other", "regression", "-", "based", "methods", ".", "Our", "work", "is", "also", "strongly", "related", "to", "and", "built", "upon", "the", "idea", "of", "viewing", "a", "ResNet", "as", "an", "unrolled", "RNN", ".", "It", "was", "first", "proposed", "in", ",", "which", "aids", "understanding", "of", "a", "family", "of", "deep", "structures", "from", "the", "perspective", "of", "RNNs", ".", "Later", ",", "Chen", "et", "al", ".", "unified", "several", "different", "residual", "functions", "to", "provide", "a", "better", "understanding", "of", "the", "design", "of", "DNNs", "with", "high", "learning", "capacity", ".", "Recently", ",", "the", "equivalence", "to", "RNNs", "has", "been", "further", "extended", "to", "DenseNet", ".", "Based", "on", "this", "finding", ",", "Dual", "Path", "Networks", "were", "proposed", "and", "showed", "superior", "performance", "to", "DenseNet", "and", "ResNet", "in", "a", "varity", "of", "applications", ".", "section", ":", "Single", "-", "State", "Recurrent", "Networks", "In", "this", "section", ",", "we", "first", "revisit", "the", "discovery", "that", "a", "ResNet", "with", "shared", "weights", "can", "be", "reformulated", "as", "a", "recurrent", "system", ".", "Then", ",", "based", "on", "this", "view", ",", "we", "unite", "the", "recent", "development", "of", "SR", "models", "with", "such", "RNN", "reformulations", "to", "show", "DRCN", "and", "DRRN", "are", "structurally", "equivalent", "to", "an", "unrolled", "single", "-", "state", "RNN", ".", "To", "establish", "the", "equivalence", ",", "we", "adopt", "the", "commonly", "used", "definition", "of", "a", "RNN", ",", "which", "is", "characterized", "by", "a", "set", "of", "states", "and", "transition", "functions", "among", "the", "states", ".", "A", "RNN", "often", "consists", "of", "the", "input", "state", ",", "output", "state", ",", "and", "the", "recurrent", "states", ".", "Depending", "on", "the", "number", "recurrent", "states", ",", "we", "describe", "RNNs", "as", "‘", "‘", "single", "-", "state", "’", "’", "(", "i.e.", "one", "recurrent", "state", ")", "or", "‘", "‘", "dual", "-", "state", "’", "’", "(", "i.e.", "two", "recurrent", "states", ")", ".", "An", "illustration", "of", "a", "single", "-", "state", "RNN", "is", "shown", "in", "Figure", "[", "reference", "]", "(", "a", ")", ".", "The", "input", ",", "output", ",", "and", "recurrent", "states", "are", "represented", "as", ",", "and", "respectively", ".", "The", "arrow", "link", "indicates", "the", "state", "transition", "function", ".", "The", "square", "on", "the", "directed", "cycle", "indicates", "that", "the", "recurrent", "function", "travels", "one", "time", "step", "forward", "during", "the", "unfolding", ".", "Interested", "readers", "are", "referred", "to", "for", "detailed", "information", "on", "this", "general", "formulation", "of", "a", "RNN", ".", "Based", "on", "Figure", "[", "reference", "]", "(", "a", ")", ",", "we", "unfold", "along", "the", "temporal", "direction", "to", "a", "fixed", "length", ".", "The", "unfolded", "graph", "is", "shown", "in", "figure", "[", "reference", "]", "(", "b", ")", ",", "and", "the", "dynamics", "of", "a", "single", "-", "state", "RNN", "can", "be", "characterized", "by", ":", "where", "the", "upper", "script", "indicates", "the", "-", "th", "unrolling", ".", "The", "parameters", "of", ",", ",", "and", "are", "often", "time", "-", "independent", ",", "which", "means", "these", "parameters", "are", "reused", "at", "every", "unfolding", "step", ".", "This", "allows", "us", "to", "unify", "ResNet", ",", "DRCN", ",", "and", "DRRN", "as", "unrolled", "networks", "with", "the", "same", "recurrent", "structure", "but", "with", "the", "different", "realizations", "of", "and", "different", "rules", "of", "parameter", "sharing", ".", "ResNet", ":", "We", "consider", "a", "ResNet", "in", "its", "simplest", "form", "without", "any", "down", "-", "sampling", "or", "up", "-", "sampling", "operations", ".", "In", "other", "words", ",", "both", "of", "the", "spatial", "dimensions", "and", "feature", "dimensions", "remain", "the", "same", "across", "all", "intermediate", "layers", ".", "To", "render", "Figure", "[", "reference", "]", "(", "b", ")", "equivalent", "to", "a", "ResNet", "with", "residual", "blocks", ",", "one", "possible", "technique", "is", "to", "make", ":", "be", "the", "input", "image", "or", "a", "function", "of", ".", ",", "and", ".", "Thus", ",", "the", "state", "transition", "becomes", ".", "The", "recurrent", "function", "be", "the", "same", "as", "a", "conventional", "residual", "block", ",", "which", "contains", "two", "convolutional", "layers", "with", "skip", "connections", "as", "shown", "in", "Figure", "[", "reference", "]", "(", "c", ")", ".", "Differences", "in", "color", "indicate", "different", "sets", "of", "parameters", ".", "The", "prediction", "state", "be", "calculated", "only", "at", "the", "time", "as", "the", "final", "output", ".", "It", "is", "worth", "mentioning", "that", "the", "only", "difference", "between", "an", "unrolled", "RNN", "following", "the", "above", "definitions", "and", "a", "conventional", "ResNet", "is", "that", "the", "parameters", "in", "need", "to", "be", "reused", "among", "all", "residual", "blocks", ".", "DRCN", ":", "To", "realize", "the", "DRCN", "expressible", "by", "the", "same", "single", "-", "state", "RNN", ",", "we", "define", "and", "in", "the", "same", "way", "as", "for", "the", "ResNet", ".", "Since", "DRCN", "recursively", "applies", "only", "a", "single", "convolutional", "layer", "to", "the", "input", "feature", "map", "16", "times", ",", "with", "the", "parameters", "of", "the", "layer", "reused", "across", "the", "whole", "network", ",", "we", "could", "use", "a", "single", "convolutional", "layer", "to", "express", ".", "The", "graph", "is", "illustrated", "in", "Figure", "[", "reference", "]", "(", "d", ")", ".", "Moreover", ",", "unlike", "the", "ResNet", "where", "the", "output", "is", "predicted", "only", "at", "the", "end", "of", "unfolding", ",", "DRCN", "utilizes", "recursive", "supervision", ",", "which", "generates", "an", "output", "at", "every", "unfolding", ".", "The", "final", "HR", "prediction", "of", "DRCN", "is", "the", "weighted", "sum", "of", "the", "outputs", "at", "every", "unfolding", ".", "DRRN", ":", "The", "recurrent", "structure", "of", "DRRN", "differs", "only", "slightly", "from", "a", "ResNet", ".", "In", "a", "ResNet", ",", "the", "skip", "connection", "comes", "from", "the", "previous", "residual", "block", ",", "whereas", "in", "a", "DRRN", "the", "skip", "connection", "always", "comes", "from", "the", "first", "unrolled", "state", ".", "Figure", "[", "reference", "]", "(", "e", ")", "shows", "the", "equivalent", "recurrent", "function", "for", "a", "DRRN", "with", "one", "recursive", "block", "(", "i.e.", ")", "using", "the", "definition", "in", "the", "original", "paper", ".", "section", ":", "Dual", "-", "State", "Recurrent", "Networks", "Drawing", "on", "the", "connections", "between", "state", "-", "of", "-", "the", "-", "art", "SR", "models", "and", "RNNs", ",", "we", "have", "investigated", "new", "compact", "RNN", "architectures", "for", "image", "SR", ".", "Specifically", ",", "we", "propose", "a", "dual", "-", "state", "design", ",", "which", "adopts", "two", "recurrent", "states", "enable", "use", "of", "features", "from", "both", "LR", "and", "HR", "spaces", ".", "The", "RNN", "view", "of", "our", "DSRN", "is", "shown", "in", "Figure", "[", "reference", "]", "(", "a", ")", "and", "is", "introduced", "as", "follows", ".", "Dual", "-", "state", "design", ":", "Unlike", "single", "-", "state", "models", "working", "at", "the", "same", "spatial", "resolution", ",", "DSRN", "incorporates", "information", "from", "both", "the", "LR", "and", "HR", "spaces", ".", "Specifically", ",", "and", "in", "Figure", "[", "reference", "]", "(", "a", ")", "indicate", "the", "LR", "state", "and", "HR", "state", ",", "respectively", ".", "Four", "colored", "arrows", "indicate", "the", "transition", "functions", "between", "these", "two", "states", ".", "The", "blue", "(", ")", ",", "orange", "(", ")", "and", "yellow", "(", ")", "links", "exist", "in", "a", "conventional", "two", "-", "layer", "RNN", ",", "providing", "information", "flow", "from", "LR", "to", "LR", ",", "HR", "to", "HR", ",", "and", "LR", "to", "HR", ",", "respectively", ".", "To", "further", "enable", "two", "-", "way", "information", "flows", "between", "and", ",", "we", "add", "the", "green", "link", ",", "which", "is", "inspired", "by", "the", "delayed", "feedback", "mechanism", "of", "traditional", "multi", "-", "layer", "RNNs", ".", "Here", ",", "it", "introduces", "a", "delayed", "HR", "to", "LR", "connection", ".", "The", "overall", "dynamics", "of", "our", "DSRN", "is", "given", "as", ":", "Figure", "[", "reference", "]", "(", "b", ")", "demonstrates", "the", "same", "concept", "via", "an", "unfolded", "graph", ",", "where", "the", "top", "row", "represents", "HR", "state", "while", "the", "bottom", "one", "is", "LR", ".", "This", "design", "choice", "encourages", "feature", "specialization", "for", "different", "resolutions", "and", "information", "sharing", "across", "different", "resolutions", ".", "Transition", "functions", ":", "Our", "model", "is", "characterized", "by", "six", "transition", "functions", ".", ",", ",", ",", "and", "as", "illustrated", "in", "Figure", "[", "reference", "]", "(", "b", ")", ".", "Specifically", ",", "we", "use", "the", "standard", "residual", "block", "for", "both", "self", "-", "transitions", ".", "A", "single", "convolutional", "layer", "is", "used", "for", "the", "down", "-", "sampling", "transition", "and", "a", "single", "transposed", "convolutional", "(", "or", "deconvolutional", ")", "layer", "is", "used", "for", "the", "up", "-", "sampling", "transition", ".", "The", "strides", "in", "both", "inter", "-", "state", "layers", "are", "set", "to", "be", "the", "same", "as", "the", "SR", "upscaling", "factor", ".", "Unfolding", "details", ":", "Similarly", "to", "unfolding", "a", "single", "-", "state", "RNN", "to", "obtain", "a", "ResNet", ",", "for", "image", "SR", ",", "we", "let", "have", "no", "contribution", "to", "calculating", "the", "state", "transition", ".", "In", "other", "words", ",", "for", "any", "choice", "of", "(", "e.g.", "choose", ")", ".", "Furthermore", ",", "we", "set", "as", "the", "output", "of", "two", "convolutional", "layers", "with", "skip", "connections", ",", "which", "takes", "the", "LR", "input", "image", "and", "transform", "it", "into", "a", "desired", "feature", "space", ".", "In", "addition", ",", "is", "set", "to", "zero", ".", "Finally", ",", "we", "use", "deep", "supervision", "for", "the", "HR", "prediction", ",", "as", "discussed", "below", ".", "Deep", "supervision", ":", "The", "unrolled", "DSRN", "is", "capable", "of", "making", "a", "prediction", "at", "every", "time", "step", ".", "Denote", "as", "a", "prediction", "at", "the", "unfolding", ",", "where", "is", "characterized", "by", "a", "single", "convolutional", "layer", ".", "Then", ",", "instead", "of", "taking", "the", "prediction", "only", "at", "the", "final", "unfolding", ",", "we", "average", "all", "the", "predictions", "as", "Thus", ",", "every", "unrolled", "layer", "directly", "connects", "to", "the", "loss", "layer", "to", "facilitate", "the", "training", "of", "such", "a", "very", "deep", "network", ".", "Moreover", ",", "the", "model", "predicts", "the", "residual", "image", "and", "minimizes", "the", "following", "mean", "square", "error", "where", "is", "the", "group", "-", "truth", "image", "in", "HR", "and", "is", "the", "residual", "map", "between", "the", "ground", "truth", "and", "bicubic", "upsampled", "LR", "image", ".", "section", ":", "Experiments", "In", "this", "section", ",", "we", "first", "provide", "implementation", "details", ",", "including", "both", "model", "hyper", "-", "parameters", "and", "training", "data", "augmentation", ".", "Then", "we", "analyze", "a", "number", "of", "design", "choices", "and", "their", "contributions", "to", "final", "performance", ".", "Finally", ",", "we", "compare", "DSRN", "to", "other", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "several", "benchmark", "datasets", ".", "subsection", ":", "Datasets", "To", "evaluate", "the", "proposed", "DSRN", "algorithm", ",", "we", "train", "our", "model", "using", "91", "images", "proposed", "in", "and", "test", "on", "the", "following", "datasets", ":", "Set5", ",", "Set14", ",", "B100", "and", "Urban100", ".", "The", "training", "data", "is", "augmented", "in", "a", "similar", "way", "to", "previous", "methods", ",", "which", "includes", "1", ")", "random", "flipping", "along", "the", "vertical", "or", "horizontal", "axis", ";", "2", ")", "random", "rotation", "by", "90", ",", "180", "or", "270", ";", "and", "3", ")", "random", "scaling", "by", "a", "factor", "from", "[", "0.5", ",", "0.6", ",", "0.7", ",", "0.8", ",", "0.9", ",", "1", "]", ".", "Tensorflow", "is", "used", "for", "our", "full", "data", "processing", "pipeline", ";", "the", "LR", "training", "images", "are", "generated", "by", "the", "built", "-", "in", "bicubic", "down", "-", "sampling", "function", ".", "We", "additionally", "test", "our", "algorithm", "on", "the", "DIV2", "K", "dataset", "of", "the", "NTIRE", "SR", "2017", "challenge", ",", "where", "we", "use", "the", "provided", "training", "and", "validation", "sets", "with", "all", "of", "the", "aforementioned", "data", "augmentations", "except", "random", "scaling", ".", "subsection", ":", "Implementation", "Details", "We", "use", "our", "model", "to", "super", "-", "resolve", "only", "the", "luminance", "channel", "of", "images", ",", "and", "use", "bicubic", "interpolation", "to", "upscale", "the", "other", "two", "color", "channels", ",", "following", ".", "We", "train", "independent", "models", "for", "each", "scale", "(", "2", ",", "3", ",", "and", "4", ")", "with", "64", "filters", "on", "the", "first", "input", "convolutional", "layer", "and", "128", "filters", "in", "the", "rest", "of", "the", "network", ".", "All", "layers", "use", "convolution", "filters", ".", "Due", "to", "our", "dual", "-", "state", "design", ",", "the", "feature", "maps", "of", "and", "in", "each", "time", "step", "have", "the", "same", "spatial", "dimensions", "as", "the", "LR", "and", "HR", "images", ",", "respectively", ".", "We", "zero", "-", "pad", "the", "boundaries", "of", "feature", "maps", "to", "ensure", "the", "spatial", "size", "of", "each", "feature", "map", "is", "the", "same", "as", "the", "input", "size", "after", "the", "convolution", "is", "applied", ".", "All", "the", "weights", "in", "the", "network", "are", "initialized", "with", "a", "uniform", "distribution", "using", "the", "method", "proposed", "in", ".", "We", "use", "standard", "stochastic", "gradient", "descent", "(", "SGD", ")", "with", "momentum", "0.95", "as", "our", "optimizer", "to", "minimize", "the", "MSE", "loss", "function", "in", "Equation", "(", "[", "reference", "]", ")", ".", "We", "search", "for", "the", "best", "initial", "learning", "rate", "from", "and", "reduce", "it", "by", "a", "factor", "of", "10", "three", "times", "during", "the", "entire", "training", "process", ".", "This", "learning", "rate", "annealing", "is", "driven", "by", "observing", "that", "the", "loss", "on", "the", "validation", "set", "stops", "decreasing", ".", "Gradient", "clipping", "at", "is", "adopted", "during", "training", "to", "prevent", "the", "gradient", "explosion", ".", "We", "sample", "image", "patches", "with", "a", "size", "of", "and", "use", "a", "mini", "-", "batch", "size", "of", "to", "train", "our", "network", ".", "We", "observe", "that", "the", "recursion", "defined", "in", "Equation", "(", "[", "reference", "]", ")", "may", "lead", "to", "an", "exponential", "increase", "in", "the", "scale", "of", "feature", "values", ",", "especially", "when", "is", "large", ".", "In", ",", "the", "authors", "proposed", "the", "use", "of", "unshared", "batch", "normalization", "at", "every", "unfolding", "time", "to", "resolve", "this", "issue", ".", "Batch", "normalization", "is", "not", "used", "in", "our", "network", ";", "we", "found", "that", "normalizing", "the", "scale", "with", "two", "scalar", "parameters", "was", "sufficient", ".", "Specifically", ",", "we", "use", "one", "unshared", "PReLU", "activation", "for", "each", "recurrent", "state", "after", "every", "unrolling", "step", ".", "All", "other", "layers", "have", "ordinary", "ReLU", "as", "the", "activation", "function", ".", "Ours", "Others", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", ".24", "subsection", ":", "Model", "Analysis", "In", "this", "section", ",", "we", "analyze", "our", "proposed", "model", "in", "the", "following", "respects", ":", "Unrolling", "length", ":", "The", "unrolling", "length", "changes", "the", "maximum", "effective", "depth", "of", "the", "unrolled", "network", ".", "In", "particular", ",", "for", "a", "DSRN", "with", "times", "unrolling", ",", "the", "maximum", "number", "of", "convolution", "layers", "between", "input", "and", "output", "of", "the", "network", "is", ".", "The", "multiplier", "comes", "from", "the", "two", "layers", "in", "a", "residual", "block", ",", "while", "the", "extra", "4", "is", "from", "the", "auxiliary", "input", "and", "output", "layers", ".", "However", ",", "the", "number", "of", "model", "parameters", "remains", "independent", "of", "the", "length", "of", "unrolling", ".", "Essentially", ",", "controls", "the", "trade", "-", "off", "between", "model", "capacity", "and", "computation", "cost", ".", "We", "study", "the", "influence", "of", "by", "training", "the", "model", "with", "different", "unrolling", "lengths", ".", "The", "empirical", "results", "are", "shown", "in", "Figure", "[", "reference", "]", ".", "The", "test", "performance", "increases", "when", "the", "number", "of", "unfolding", "steps", "increases", ",", "but", "the", "benefit", "seems", "to", "diminish", "after", ".", "Unless", "otherwise", "mentioned", ",", "we", "use", "for", "all", "our", "models", ".", "It", "is", "worth", "mentioning", "that", "we", "also", "experimented", "with", "stochastic", "depth", "by", "randomly", "sampling", "during", "training", ",", "but", "we", "observed", "no", "improvement", "in", "validation", "accuracy", ".", "Parameter", "sharing", ":", "We", "empirically", "find", "parameter", "sharing", "to", "be", "crucial", "for", "training", "a", "deep", "recursive", "model", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "the", "same", "model", "with", "untied", "weights", "performs", "much", "more", "poorly", "than", "its", "weight", "-", "sharing", "counterpart", ".", "Specifically", ",", "we", "observe", "around", "0.2dB", "performance", "drop", "across", "all", "three", "upscaling", "scales", "when", "changing", "from", "shared", "weights", "to", "untied", "weights", ".", "We", "speculate", "that", "the", "model", "with", "untied", "weights", "suffers", "a", "larger", "risk", "of", "model", "over", "-", "fitting", "and", "much", "slower", "training", "convergence", ",", "both", "of", "which", "diminish", "the", "model", "’s", "restoration", "accuracy", ".", "Dual", "-", "state", "and", "delayed", "feedback", ":", "We", "compare", "our", "DSRN", "with", "two", "baselines", "under", "the", "same", "unrolling", "time", "steps", "to", "understand", "how", "each", "module", "of", "our", "model", "contributes", "to", "the", "final", "performance", ":", "1", ")", "a", "single", "-", "state", "RNN", "unrolled", "ResNet", ";", "and", "2", ")", "a", "dual", "-", "state", "RNN", "without", "delayed", "feedback", "connections", ".", "The", "quantitative", "comparison", "on", "the", "NTIRE", "SR", "2017", "challenge", "is", "shown", "in", "Table", "[", "reference", "]", ".", "Comparing", "the", "single", "-", "state", "baseline", "and", "the", "DSRN", "without", "feedback", ",", "it", "is", "clear", "that", "considering", "information", "from", "both", "LR", "and", "HR", "spaces", "as", "two", "separated", "states", "provides", "performance", "gains", ".", "In", "addition", ",", "comparing", "our", "models", "with", "and", "without", "feedback", ",", "we", "realize", "that", "incorporating", "such", "an", "information", "flow", "from", "HR", "space", "back", "to", "LR", "space", "consistently", "improves", "performance", "on", "all", "three", "different", "scales", ".", "In", "all", ",", "both", "the", "dual", "-", "state", "and", "delayed", "feedback", "designs", "are", "beneficial", "to", "our", "model", ".", "State", "visualization", "Since", "DSRN", "has", "independent", "scaling", "parameters", "on", "each", "unrolled", "state", ",", "the", "model", "implicitly", "learns", "a", "weighted", "-", "average", "of", "all", "the", "unrolled", "states", "for", "the", "final", "prediction", ".", "Empirically", "we", "observe", "that", "this", "strategy", "performs", "better", "than", "output", "from", "the", "last", "state", "only", ".", "To", "demonstrate", "how", "the", "network", "aggregates", "different", "unrolled", "states", ",", "we", "show", "feature", "response", "maps", "at", "different", "unrolling", "steps", "in", "Figure", "[", "reference", "]", ",", "demonstrating", "that", "the", "network", "distributes", "slightly", "different", "features", "to", "each", "unrolled", "state", ".", "subsection", ":", "Comparison", "with", "the", "State", "-", "of", "-", "the", "-", "Art", "We", "provide", "results", "of", "evaluation", "of", "our", "model", "on", "several", "public", "benchmark", "datasets", "in", "Table", "[", "reference", "]", ",", "with", "three", "commonly", "-", "used", "evaluation", "metrics", ":", "Peak", "Signal", "-", "to", "-", "Noise", "Ratio", "(", "PSNR", ")", ",", "Structural", "SIMilarity", "(", "SSIM", ")", "and", "the", "Information", "Fidelity", "Criterion", "(", "IFC", ")", ".", "Specifically", ",", "we", "perform", "a", "comprehensive", "comparison", "between", "our", "method", "and", "10", "other", "existing", "SR", "algorithms", ",", "including", "both", "deep", "learning", "and", "non", "-", "deep", "-", "learning", "based", "methods", ".", "Note", "that", "many", "recent", "deep", "learning", "based", "competitors", ",", "including", "VDSR", ",", "LapSRN", "and", "DRRN", ",", "use", "291", "training", "samples", "with", "the", "additional", "200", "from", "the", "training", "set", "of", "Berkeley", "Segmentation", "Dataset", ",", "while", "our", "model", "was", "trained", "on", "only", "the", "91", "images", ".", "Still", ",", "our", "DSRN", "method", "achieves", "competitive", "performance", "across", "all", "datasets", "and", "scales", ".", "It", "achieves", "particularly", "strong", "performance", "in", "the", "and", "settings", ".", "In", "addition", ",", "we", "report", "quantitative", "evaluations", "on", "the", "recently", "developed", "DIV2", "K", "dataset", "and", "comparisons", "with", "top", "-", "ranking", "algorithms", "in", "Table", "[", "reference", "]", ".", "Our", "method", "achieves", "competitive", "performance", "with", "the", "best", "algorithm", ",", "EDSR", "+", ",", "and", "outperforms", "all", "the", "other", "algorithms", "by", "a", "large", "margin", ",", "which", "demonstrates", "the", "effectiveness", "of", "our", "proposed", "dual", "-", "state", "recurrent", "structure", ".", "To", "further", "analyze", "the", "proposed", "DSRN", "against", "other", "state", "-", "of", "-", "the", "-", "art", "SR", "approaches", "in", "a", "qualitative", "manner", ",", "in", "Figure", "[", "reference", "]", "we", "present", "several", "visual", "examples", "of", "super", "-", "resolved", "images", "on", "Set14", "with", "upscaling", "among", "different", "SR", "approaches", ".", "For", "these", "competing", "methods", ",", "we", "use", "SR", "results", "publicly", "released", "by", "the", "authors", ".", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "our", "method", "can", "construct", "sharp", "and", "detailed", "structures", "and", "is", "less", "prone", "to", "generating", "spurious", "artifacts", ".", "Furthermore", ",", "the", "proposed", "DSRN", "benefits", "from", "inherent", "parameter", "sharing", "and", "therefore", "obtains", "higher", "parameter", "efficiency", "compared", "to", "other", "methods", ".", "In", "Figure", "[", "reference", "]", ",", "we", "illustrate", "the", "parameters", "-", "to", "-", "PSNR", "relationship", "of", "our", "model", "and", "several", "state", "-", "of", "-", "the", "-", "art", "methods", ",", "including", "SRCNN", ",", "VDSR", ",", "DRCN", ",", "DRRN", "and", "RED30", ".", "Our", "method", "represents", "a", "favorable", "trade", "-", "off", "between", "model", "size", "and", "SR", "performance", ",", "and", "has", "modest", "inference", "time", ".", "The", "DSRN", "takes", "0.4s", "on", "the", "x4", "task", "with", "a", "288x288", "output", "image", "size", ",", "on", "an", "NVIDIA", "Titan", "X", "GPU", ".", "section", ":", "Conclusion", "In", "this", "work", ",", "we", "have", "provided", "a", "unique", "formulation", "that", "expresses", "many", "state", "-", "of", "-", "the", "-", "art", "SR", "models", "as", "a", "finite", "unfolding", "of", "a", "single", "-", "state", "RNN", "with", "various", "recurrent", "functions", ".", "Based", "on", "this", ",", "we", "extend", "existing", "methods", "by", "considering", "a", "dual", "-", "state", "design", ";", "the", "two", "hidden", "states", "of", "our", "proposed", "DSRN", "operate", "at", "different", "spatial", "resolutions", ".", "One", "captures", "the", "LR", "information", "while", "the", "other", "one", "targets", "the", "HR", "domains", ".", "To", "ensure", "two", "-", "way", "communication", "between", "states", ",", "we", "integrate", "a", "delayed", "feedback", "mechanism", ".", "Thus", ",", "the", "predicted", "features", "from", "both", "LR", "and", "HR", "states", "can", "be", "exploited", "jointly", "for", "final", "predictions", ".", "Extensive", "experiments", "on", "benchmark", "datasets", "have", "demonstrated", "that", "the", "proposed", "DSRN", "performs", "favorably", "against", "state", "-", "of", "-", "the", "-", "art", "SR", "models", "in", "terms", "of", "both", "efficiency", "and", "accuracy", ".", "For", "the", "future", "work", ",", "we", "will", "explore", "use", "of", "our", "proposed", "DSRN", "to", "capture", "temporal", "dependencies", "for", "video", "SR", ".", "bibliography", ":", "References"], "entity": [[2, 6, ["Image", "Super", "-", "Resolution"], "Task"], [7, 12, ["Dual", "-", "State", "Recurrent", "Networks"], "Method"], [14, 18, ["image", "super", "-", "resolution"], "Task"], [19, 20, ["SR"], "Task"], [29, 32, ["deep", "neural", "networks"], "Method"], [58, 64, ["single", "-", "state", "recurrent", "neural", "network"], "Method"], [65, 66, ["RNN"], "Method"], [68, 70, ["finite", "unfoldings"], "Method"], [80, 81, ["SR"], "Task"], [84, 87, ["compact", "RNN", "view"], "Method"], [92, 96, ["dual", "-", "state", "design"], "Method"], [98, 103, ["Dual", "-", "State", "Recurrent", "Network"], "Method"], [104, 105, ["DSRN"], "Method"], [122, 123, ["DSRN"], "Method"], [182, 183, ["DSRN"], "Method"], [198, 200, ["memory", "consumption"], "Metric"], [201, 203, ["predictive", "accuracy"], "Metric"], [211, 217, ["single", "-", "image", "super", "-", "resolution"], "Task"], [218, 219, ["SR"], "Task"], [227, 230, ["high", "-", "resolution"], "Task"], [231, 232, ["HR"], "Task"], [249, 250, ["SR"], "Task"], [260, 263, ["deep", "neural", "networks"], "Method"], [264, 265, ["DNNs"], "Method"], [269, 272, ["convolutional", "neural", "networks"], "Method"], [273, 275, ["CNNs", ")"], "Method"], [276, 278, ["residual", "learning"], "Method"], [308, 310, ["CNN", "model"], "Method"], [311, 312, ["SR"], "Task"], [315, 317, ["deeper", "model"], "Method"], [385, 389, ["training", "of", "deep", "models"], "Task"], [428, 430, ["model", "overfitting"], "Task"], [450, 454, ["resource", "-", "constrained", "environments"], "Task"], [456, 458, ["mobile", "systems"], "Task"], [466, 470, ["Deep", "Recursive", "Residual", "Network"], "Method"], [471, 472, ["DRRN"], "Method"], [476, 481, ["Deeply", "-", "Recursive", "Convolutional", "Network"], "Method"], [482, 483, ["DRCN"], "Method"], [510, 513, ["neural", "architectural", "design"], "Task"], [528, 531, ["recurrent", "neural", "networks"], "Method"], [532, 533, ["RNNs"], "Method"], [543, 549, ["weight", "-", "sharing", "Residual", "Neural", "Network"], "Method"], [550, 552, ["ResNet", ")"], "Method"], [556, 558, ["shallow", "RNN"], "Method"], [571, 573, ["neural", "architectures"], "Method"], [579, 582, ["compact", "RNN", "formulations"], "Method"], [590, 592, ["recursive", "computation"], "Method"], [593, 595, ["weight", "sharing"], "Method"], [608, 610, ["bicubic", "interpolation"], "Method"], [635, 639, ["single", "-", "state", "RNN"], "Method"], [722, 724, ["RNN", "setting"], "Method"], [729, 731, ["RNN", "view"], "Method"], [754, 755, ["SR"], "Task"], [763, 765, ["RNN", "model"], "Method"], [770, 773, ["multi", "-", "layer"], "Method"], [785, 788, ["compact", "RNN", "view"], "Method"], [798, 803, ["Dual", "-", "State", "Recurrent", "Network"], "Method"], [841, 846, ["two", "-", "layer", "stacked", "RNN"], "Method"], [859, 861, ["deconvolutional", "operations"], "Method"], [895, 898, ["delayed", "feedback", "mechanism"], "Method"], [916, 917, ["DSRN"], "Method"], [958, 959, ["DSRN"], "Method"], [983, 987, ["Image", "Restoration", "and", "Enhancement"], "Task"], [991, 995, ["image", "super", "-", "resolution"], "Task"], [1007, 1008, ["DSRN"], "Method"], [1010, 1012, ["parameter", "efficiency"], "Metric"], [1013, 1016, ["low", "memory", "consumption"], "Metric"], [1018, 1020, ["restoration", "accuracy"], "Metric"], [1025, 1028, ["Single", "image", "SR"], "Task"], [1053, 1055, ["deep", "learning"], "Method"], [1072, 1076, ["fully", "convolutional", "neural", "network"], "Method"], [1078, 1079, ["SRCNN"], "Method"], [1096, 1101, ["example", "-", "based", "learning", "paradigms"], "Method"], [1104, 1106, ["nearest", "neighbor"], "Method"], [1107, 1109, ["sparse", "representation"], "Method"], [1110, 1112, ["neighborhood", "embedding"], "Method"], [1113, 1115, ["random", "forest"], "Method"], [1123, 1124, ["SRCNN"], "Method"], [1145, 1147, ["patch", "representation"], "Method"], [1148, 1152, ["non", "-", "linear", "mapping"], "Task"], [1154, 1155, ["reconstruction"], "Task"], [1164, 1166, ["network", "design"], "Method"], [1173, 1175, ["optimization", "techniques"], "Method"], [1182, 1188, ["sparse", "coding", "network", "(", "SCN", ")"], "Method"], [1191, 1194, ["sparse", "representation", "prior"], "Method"], [1211, 1213, ["domain", "expertise"], "Method"], [1214, 1216, ["sparse", "coding"], "Method"], [1234, 1236, ["neural", "network"], "Method"], [1244, 1246, ["deep", "models"], "Method"], [1247, 1249, ["ImageNet", "challenges"], "Task"], [1256, 1259, ["very", "deep", "CNN"], "Method"], [1260, 1261, ["VDSR"], "Method"], [1265, 1267, ["convolutional", "layers"], "Method"], [1268, 1269, ["kernels"], "Method"], [1271, 1273, ["residual", "learning"], "Method"], [1274, 1277, ["adjustable", "gradient", "clipping"], "Method"], [1308, 1309, ["DRCN"], "Method"], [1311, 1313, ["recursive", "layers"], "Method"], [1325, 1326, ["DRCN"], "Method"], [1332, 1334, ["recursive", "supervision"], "Method"], [1339, 1340, ["training"], "Task"], [1356, 1359, ["global", "residual", "learning"], "Method"], [1360, 1363, ["local", "residual", "learning"], "Method"], [1369, 1371, ["deep", "models"], "Method"], [1377, 1378, ["DRRN"], "Method"], [1381, 1385, ["global", "and", "local", "learning"], "Method"], [1390, 1392, ["recursive", "learning"], "Method"], [1404, 1407, ["Densely", "Connected", "Networks"], "Method"], [1408, 1409, ["DenseNet"], "Method"], [1412, 1413, ["ResNet"], "Method"], [1415, 1417, ["building", "block"], "Method"], [1438, 1439, ["SR"], "Task"], [1443, 1445, ["deep", "models"], "Method"], [1458, 1461, ["compact", "network", "model"], "Method"], [1463, 1464, ["convolutions"], "Method"], [1470, 1472, ["upscaling", "filters"], "Method"], [1481, 1483, ["computation", "cost"], "Metric"], [1491, 1493, ["deconvolution", "layers"], "Method"], [1495, 1496, ["SRCNN"], "Method"], [1504, 1506, ["convolution", "layers"], "Method"], [1527, 1533, ["Laplacian", "Pyramid", "Super", "-", "Resolution", "Network"], "Method"], [1534, 1535, ["LapSRN"], "Method"], [1558, 1563, ["Enhanced", "Deep", "Super", "-", "Resolution"], "Method"], [1564, 1565, ["EDSR"], "Method"], [1569, 1573, ["multi", "-", "scale", "variant"], "Method"], [1583, 1585, ["weight", "sharing"], "Method"], [1595, 1599, ["mean", "squared", "reconstruction", "error"], "Metric"], [1602, 1603, ["loss"], "Metric"], [1634, 1636, ["subjective", "scores"], "Metric"], [1642, 1645, ["Generative", "Adversarial", "Networks"], "Method"], [1653, 1654, ["SR"], "Task"], [1655, 1657, ["Subjective", "evaluation"], "Metric"], [1658, 1663, ["mean", "-", "opinion", "-", "score"], "Metric"], [1668, 1672, ["regression", "-", "based", "methods"], "Method"], [1688, 1689, ["ResNet"], "Method"], [1691, 1693, ["unrolled", "RNN"], "Method"], [1713, 1714, ["RNNs"], "Method"], [1735, 1736, ["DNNs"], "Method"], [1738, 1740, ["learning", "capacity"], "Metric"], [1746, 1747, ["RNNs"], "Method"], [1759, 1762, ["Dual", "Path", "Networks"], "Method"], [1769, 1770, ["DenseNet"], "Method"], [1771, 1772, ["ResNet"], "Method"], [1780, 1785, ["Single", "-", "State", "Recurrent", "Networks"], "Method"], [1805, 1807, ["recurrent", "system"], "Method"], [1825, 1827, ["RNN", "reformulations"], "Method"], [1829, 1830, ["DRCN"], "Method"], [1831, 1832, ["DRRN"], "Method"], [1837, 1842, ["unrolled", "single", "-", "state", "RNN"], "Method"], [1856, 1857, ["RNN"], "Method"], [1874, 1875, ["RNN"], "Method"], [1899, 1900, ["RNNs"], "Method"], [1933, 1937, ["single", "-", "state", "RNN"], "Method"], [1981, 1983, ["recurrent", "function"], "Method"], [2006, 2007, ["RNN"], "Method"], [2051, 2052, ["RNN"], "Method"], [2095, 2096, ["ResNet"], "Method"], [2097, 2098, ["DRCN"], "Method"], [2100, 2101, ["DRRN"], "Method"], [2102, 2104, ["unrolled", "networks"], "Method"], [2127, 2128, ["ResNet"], "Method"], [2134, 2137, ["down", "-", "sampling"], "Method"], [2138, 2142, ["up", "-", "sampling", "operations"], "Method"], [2175, 2179, ["ResNet", "with", "residual", "blocks"], "Method"], [2221, 2223, ["convolutional", "layers"], "Method"], [2270, 2272, ["unrolled", "RNN"], "Method"], [2279, 2280, ["ResNet"], "Method"], [2294, 2295, ["DRCN"], "Method"], [2299, 2300, ["DRCN"], "Method"], [2304, 2308, ["single", "-", "state", "RNN"], "Method"], [2319, 2320, ["ResNet"], "Method"], [2322, 2324, ["DRCN", "recursively"], "Method"], [2328, 2330, ["convolutional", "layer"], "Method"], [2355, 2357, ["convolutional", "layer"], "Method"], [2377, 2378, ["ResNet"], "Method"], [2390, 2391, ["DRCN"], "Method"], [2392, 2394, ["recursive", "supervision"], "Method"], [2408, 2409, ["DRCN"], "Method"], [2420, 2421, ["DRRN"], "Method"], [2423, 2425, ["recurrent", "structure"], "Method"], [2426, 2427, ["DRRN"], "Method"], [2432, 2433, ["ResNet"], "Method"], [2436, 2437, ["ResNet"], "Method"], [2451, 2452, ["DRRN"], "Method"], [2477, 2478, ["DRRN"], "Method"], [2480, 2482, ["recursive", "block"], "Method"], [2495, 2500, ["Dual", "-", "State", "Recurrent", "Networks"], "Method"], [2515, 2516, ["RNNs"], "Method"], [2521, 2524, ["compact", "RNN", "architectures"], "Method"], [2533, 2537, ["dual", "-", "state", "design"], "Method"], [2555, 2557, ["RNN", "view"], "Method"], [2559, 2560, ["DSRN"], "Method"], [2576, 2580, ["Dual", "-", "state", "design"], "Method"], [2582, 2586, ["single", "-", "state", "models"], "Method"], [2593, 2594, ["DSRN"], "Method"], [2654, 2658, ["two", "-", "layer", "RNN"], "Method"], [2700, 2703, ["delayed", "feedback", "mechanism"], "Method"], [2705, 2709, ["multi", "-", "layer", "RNNs"], "Method"], [2726, 2727, ["DSRN"], "Method"], [2744, 2746, ["unfolded", "graph"], "Method"], [2810, 2812, ["residual", "block"], "Method"], [2814, 2817, ["self", "-", "transitions"], "Task"], [2820, 2822, ["convolutional", "layer"], "Method"], [2826, 2830, ["down", "-", "sampling", "transition"], "Task"], [2833, 2840, ["transposed", "convolutional", "(", "or", "deconvolutional", ")", "layer"], "Method"], [2844, 2848, ["up", "-", "sampling", "transition"], "Task"], [2869, 2871, ["Unfolding", "details"], "Task"], [2876, 2880, ["single", "-", "state", "RNN"], "Method"], [2883, 2884, ["ResNet"], "Method"], [2922, 2924, ["convolutional", "layers"], "Method"], [2966, 2968, ["Deep", "supervision"], "Task"], [2997, 2999, ["convolutional", "layer"], "Method"], [3022, 3024, ["unrolled", "layer"], "Method"], [3028, 3030, ["loss", "layer"], "Method"], [3038, 3040, ["deep", "network"], "Method"], [3053, 3056, ["mean", "square", "error"], "Metric"], [3095, 3099, ["model", "hyper", "-", "parameters"], "Method"], [3100, 3103, ["training", "data", "augmentation"], "Method"], [3123, 3124, ["DSRN"], "Method"], [3165, 3166, ["Set5"], "Material"], [3167, 3168, ["Set14"], "Material"], [3169, 3170, ["B100"], "Material"], [3171, 3172, ["Urban100"], "Material"], [3233, 3234, ["Tensorflow"], "Method"], [3238, 3242, ["full", "data", "processing", "pipeline"], "Task"], [3254, 3259, ["bicubic", "down", "-", "sampling", "function"], "Method"], [3291, 3293, ["data", "augmentations"], "Method"], [3294, 3296, ["random", "scaling"], "Method"], [3318, 3320, ["bicubic", "interpolation"], "Method"], [3332, 3334, ["independent", "models"], "Method"], [3352, 3354, ["convolutional", "layer"], "Method"], [3367, 3369, ["convolution", "filters"], "Method"], [3373, 3377, ["dual", "-", "state", "design"], "Method"], [3428, 3429, ["convolution"], "Method"], [3453, 3456, ["stochastic", "gradient", "descent"], "Method"], [3457, 3458, ["SGD"], "Method"], [3460, 3462, ["momentum", "0.95"], "Method"], [3468, 3471, ["MSE", "loss", "function"], "Metric"], [3485, 3487, ["learning", "rate"], "Metric"], [3505, 3508, ["learning", "rate", "annealing"], "Metric"], [3514, 3515, ["loss"], "Metric"], [3522, 3524, ["Gradient", "clipping"], "Method"], [3528, 3529, ["training"], "Task"], [3595, 3598, ["unshared", "batch", "normalization"], "Method"], [3607, 3609, ["Batch", "normalization"], "Method"], [3651, 3652, ["ReLU"], "Method"], [3726, 3728, ["unrolled", "network"], "Method"], [3734, 3735, ["DSRN"], "Method"], [3736, 3738, ["times", "unrolling"], "Method"], [3802, 3804, ["model", "capacity"], "Metric"], [3805, 3807, ["computation", "cost"], "Metric"], [3876, 3878, ["randomly", "sampling"], "Method"], [3887, 3889, ["validation", "accuracy"], "Metric"], [3890, 3892, ["Parameter", "sharing"], "Method"], [3904, 3907, ["deep", "recursive", "model"], "Method"], [3928, 3932, ["weight", "-", "sharing", "counterpart"], "Method"], [3968, 3972, ["model", "over", "-", "fitting"], "Method"], [3975, 3977, ["training", "convergence"], "Metric"], [3985, 3987, ["restoration", "accuracy"], "Metric"], [3998, 3999, ["DSRN"], "Method"], [4025, 4031, ["single", "-", "state", "RNN", "unrolled", "ResNet"], "Method"], [4036, 4044, ["dual", "-", "state", "RNN", "without", "delayed", "feedback", "connections"], "Method"], [4064, 4068, ["single", "-", "state", "baseline"], "Method"], [4134, 4141, ["dual", "-", "state", "and", "delayed", "feedback", "designs"], "Method"], [4147, 4149, ["State", "visualization"], "Method"], [4150, 4151, ["DSRN"], "Method"], [4165, 4168, ["weighted", "-", "average"], "Method"], [4176, 4177, ["prediction"], "Task"], [4268, 4270, ["evaluation", "metrics"], "Metric"], [4271, 4278, ["Peak", "Signal", "-", "to", "-", "Noise", "Ratio"], "Metric"], [4279, 4280, ["PSNR"], "Metric"], [4282, 4284, ["Structural", "SIMilarity"], "Metric"], [4285, 4286, ["SSIM"], "Metric"], [4289, 4292, ["Information", "Fidelity", "Criterion"], "Metric"], [4293, 4294, ["IFC"], "Method"], [4315, 4325, ["deep", "learning", "and", "non", "-", "deep", "-", "learning", "based", "methods"], "Method"], [4330, 4334, ["deep", "learning", "based", "competitors"], "Method"], [4336, 4337, ["VDSR"], "Method"], [4338, 4339, ["LapSRN"], "Method"], [4340, 4341, ["DRRN"], "Method"], [4411, 4415, ["top", "-", "ranking", "algorithms"], "Method"], [4431, 4433, ["EDSR", "+"], "Method"], [4452, 4457, ["dual", "-", "state", "recurrent", "structure"], "Method"], [4463, 4464, ["DSRN"], "Method"], [4548, 4549, ["DSRN"], "Method"], [4552, 4554, ["parameter", "sharing"], "Method"], [4558, 4560, ["parameter", "efficiency"], "Metric"], [4578, 4580, ["PSNR", "relationship"], "Metric"], [4595, 4596, ["SRCNN"], "Method"], [4597, 4598, ["VDSR"], "Method"], [4599, 4600, ["DRCN"], "Method"], [4601, 4602, ["DRRN"], "Method"], [4603, 4604, ["RED30"], "Method"], [4614, 4616, ["model", "size"], "Metric"], [4623, 4625, ["inference", "time"], "Metric"], [4627, 4628, ["DSRN"], "Method"], [4632, 4634, ["x4", "task"], "Task"], [4675, 4677, ["finite", "unfolding"], "Method"], [4679, 4683, ["single", "-", "state", "RNN"], "Method"], [4699, 4703, ["dual", "-", "state", "design"], "Method"], [4711, 4712, ["DSRN"], "Method"], [4744, 4747, ["delayed", "feedback", "mechanism"], "Method"], [4765, 4766, ["predictions"], "Task"], [4777, 4778, ["DSRN"], "Method"], [4794, 4795, ["efficiency"], "Metric"], [4796, 4797, ["accuracy"], "Metric"], [4810, 4811, ["DSRN"], "Method"], [51, 52, ["SR"], "Task"], [132, 135, ["high", "-", "resolution"], "Task"], [136, 137, ["HR"], "Task"], [155, 156, ["HR"], "Task"], [157, 158, ["HR"], "Task"], [285, 286, ["SR"], "Task"], [329, 330, ["HR"], "Task"], [356, 357, ["HR"], "Task"], [575, 576, ["SR"], "Task"], [587, 588, ["SR"], "Task"], [597, 598, ["DRRN"], "Method"], [643, 644, ["DRRN"], "Method"], [739, 740, ["SR"], "Task"], [781, 782, ["SR"], "Task"], [804, 805, ["DSRN"], "Method"], [834, 835, ["HR"], "Task"], [869, 870, ["HR"], "Task"], [886, 887, ["HR"], "Task"], [902, 903, ["HR"], "Task"], [937, 938, ["HR"], "Task"], [963, 964, ["SR"], "Task"], [997, 998, ["SR"], "Task"], [1086, 1087, ["HR"], "Task"], [1196, 1197, ["SR"], "Task"], [1218, 1219, ["SR"], "Task"], [1230, 1231, ["HR"], "Task"], [1349, 1350, ["SR"], "Task"], [1419, 1420, ["SR"], "Task"], [1591, 1592, ["SR"], "Task"], [1626, 1627, ["SR"], "Task"], [1821, 1822, ["SR"], "Task"], [2405, 2406, ["HR"], "Task"], [2512, 2513, ["SR"], "Task"], [2526, 2527, ["SR"], "Task"], [2551, 2552, ["HR"], "Task"], [2601, 2602, ["HR"], "Task"], [2620, 2621, ["HR"], "Task"], [2667, 2668, ["HR"], "Task"], [2669, 2670, ["HR"], "Task"], [2674, 2675, ["HR"], "Task"], [2716, 2717, ["HR"], "Task"], [2752, 2753, ["HR"], "Task"], [2865, 2866, ["SR"], "Task"], [2887, 2888, ["SR"], "Task"], [2959, 2960, ["HR"], "Task"], [2971, 2972, ["DSRN"], "Method"], [3064, 3065, ["HR"], "Task"], [3146, 3147, ["DSRN"], "Method"], [3273, 3274, ["SR"], "Task"], [3396, 3397, ["HR"], "Task"], [4051, 4052, ["SR"], "Task"], [4070, 4071, ["DSRN"], "Method"], [4084, 4085, ["HR"], "Task"], [4114, 4115, ["HR"], "Task"], [4310, 4311, ["SR"], "Task"], [4373, 4374, ["DSRN"], "Method"], [4473, 4474, ["SR"], "Task"], [4496, 4497, ["Set14"], "Material"], [4501, 4502, ["SR"], "Task"], [4511, 4512, ["SR"], "Task"], [4617, 4618, ["SR"], "Task"], [4671, 4672, ["SR"], "Task"], [4729, 4730, ["HR"], "Task"], [4757, 4758, ["HR"], "Task"], [4788, 4789, ["SR"], "Task"], [4817, 4818, ["SR"], "Task"]]} {"doc_id": "1778e32c18bd611169e64c1805a51abff341ca53", "text": ["document", ":", "Natural", "Language", "Inference", "over", "Interaction", "Space", "Natural", "Language", "Inference", "(", "NLI", ")", "task", "requires", "an", "agent", "to", "determine", "the", "logical", "relationship", "between", "a", "natural", "language", "premise", "and", "a", "natural", "language", "hypothesis", ".", "We", "introduce", "Interactive", "Inference", "Network", "(", "IIN", ")", ",", "a", "novel", "class", "of", "neural", "network", "architectures", "that", "is", "able", "to", "achieve", "high", "-", "level", "understanding", "of", "the", "sentence", "pair", "by", "hierarchically", "extracting", "semantic", "features", "from", "interaction", "space", ".", "We", "show", "that", "an", "interaction", "tensor", "(", "attention", "weight", ")", "contains", "semantic", "information", "to", "solve", "natural", "language", "inference", ",", "and", "a", "denser", "interaction", "tensor", "contains", "richer", "semantic", "information", ".", "One", "instance", "of", "such", "architecture", ",", "Densely", "Interactive", "Inference", "Network", "(", "DIIN", ")", ",", "demonstrates", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "large", "scale", "NLI", "copora", "and", "large", "-", "scale", "NLI", "alike", "corpus", ".", "It", "’s", "noteworthy", "that", "DIIN", "achieve", "a", "greater", "than", "20", "%", "error", "reduction", "on", "the", "challenging", "Multi", "-", "Genre", "NLI", "(", "MultiNLI", ";", ")", "dataset", "with", "respect", "to", "the", "strongest", "published", "system", ".", "section", ":", "Introduction", "Natural", "Language", "Inference", "(", "NLI", "also", "known", "as", "recognizing", "textual", "entiailment", ",", "or", "RTE", ")", "task", "requires", "one", "to", "determine", "whether", "the", "logical", "relationship", "between", "two", "sentences", "is", "among", "entailment", "(", "if", "the", "premise", "is", "true", ",", "then", "the", "hypothesis", "must", "be", "true", ")", ",", "contradiction", "(", "if", "the", "premise", "is", "true", ",", "then", "the", "hypothesis", "must", "be", "false", ")", "and", "neutral", "(", "neither", "entailment", "nor", "contradiction", ")", ".", "NLI", "is", "known", "as", "a", "fundamental", "and", "yet", "challenging", "task", "for", "natural", "language", "understanding", ",", "not", "only", "because", "it", "requires", "one", "to", "identify", "the", "language", "pattern", ",", "but", "also", "to", "understand", "certain", "common", "sense", "knowledge", ".", "In", "Table", "[", "reference", "]", ",", "three", "samples", "from", "MultiNLI", "corpus", "show", "solving", "the", "task", "requires", "one", "to", "handle", "the", "full", "complexity", "of", "lexical", "and", "compositional", "semantics", ".", "The", "previous", "work", "on", "NLI", "(", "or", "RTE", ")", "has", "extensively", "researched", "on", "conventional", "approaches", ".", "Recent", "progress", "on", "NLI", "is", "enabled", "by", "the", "availability", "of", "570k", "human", "annotated", "dataset", "and", "the", "advancement", "of", "representation", "learning", "technique", ".", "Among", "the", "core", "representation", "learning", "techniques", ",", "attention", "mechanism", "is", "broadly", "applied", "in", "many", "NLU", "tasks", "since", "its", "introduction", ":", "machine", "translation", ",", "abstractive", "summarization", ",", "Reading", "Comprehension", ",", "dialog", "system", ",", "etc", ".", "As", "described", "by", ",", "“", "An", "attention", "function", "can", "be", "described", "as", "mapping", "a", "query", "and", "a", "set", "of", "key", "-", "value", "pairs", "to", "an", "output", ",", "where", "the", "query", ",", "keys", ",", "values", ",", "and", "output", "are", "all", "vectors", ".", "The", "output", "is", "computed", "as", "a", "weighted", "sum", "of", "the", "values", ",", "where", "the", "weight", "assigned", "to", "each", "value", "is", "computed", "by", "a", "compatibility", "function", "of", "the", "query", "with", "the", "corresponding", "key", "”", ".", "Attention", "mechanism", "is", "known", "for", "its", "alignment", "between", "representations", ",", "focusing", "one", "part", "of", "representation", "over", "another", ",", "and", "modeling", "the", "dependency", "regardless", "of", "sequence", "length", ".", "Observing", "attention", "’s", "powerful", "capability", ",", "we", "hypothesize", "that", "the", "attention", "weight", "can", "assist", "with", "machine", "to", "understanding", "the", "text", ".", "A", "regular", "attention", "weight", ",", "the", "core", "component", "of", "the", "attention", "mechanism", ",", "encodes", "the", "cross", "-", "sentence", "word", "relationship", "into", "a", "alignment", "matrix", ".", "However", ",", "a", "multi", "-", "head", "attention", "weight", "can", "encode", "such", "interaction", "into", "multiple", "alignment", "matrices", ",", "which", "shows", "a", "more", "powerful", "alignment", ".", "In", "this", "work", ",", "we", "push", "the", "multi", "-", "head", "attention", "to", "a", "extreme", "by", "building", "a", "word", "-", "by", "-", "word", "dimension", "-", "wise", "alignment", "tensor", "which", "we", "call", "interaction", "tensor", ".", "The", "interaction", "tensor", "encodes", "the", "high", "-", "order", "alignment", "relationship", "between", "sentences", "pair", ".", "Our", "experiments", "demonstrate", "that", "by", "capturing", "the", "rich", "semantic", "features", "in", "the", "interaction", "tensor", ",", "we", "are", "able", "to", "solve", "natural", "language", "inference", "task", "well", ",", "especially", "in", "cases", "with", "paraphrase", ",", "antonyms", "and", "overlapping", "words", ".", "We", "dub", "the", "general", "framework", "as", "Interactive", "Inference", "Network", "(", "IIN", ")", ".", "To", "the", "best", "of", "our", "knowledge", ",", "it", "is", "the", "first", "attempt", "to", "solve", "natural", "language", "inference", "task", "in", "the", "interaction", "space", ".", "We", "further", "explore", "one", "instance", "of", "Interactive", "Inference", "Network", ",", "Densely", "Interactive", "Inference", "Network", "(", "DIIN", ")", ",", "which", "achieves", "new", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "both", "SNLI", "and", "MultiNLI", "copora", ".", "To", "test", "the", "generality", "of", "the", "architecture", ",", "we", "interpret", "the", "paraphrase", "identification", "task", "as", "natural", "language", "inference", "task", "where", "matching", "as", "entailment", ",", "not", "-", "matching", "as", "neutral", ".", "We", "test", "the", "model", "on", "Quora", "Question", "Pair", "dataset", ",", "which", "contains", "over", "400k", "real", "world", "question", "pair", ",", "and", "achieves", "new", "state", "-", "of", "-", "the", "-", "art", "performance", ".", "We", "introduce", "the", "related", "work", "in", "Section", "2", ",", "and", "discuss", "the", "general", "framework", "of", "IIN", "along", "with", "a", "specific", "instance", "that", "enjoys", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "multiple", "datasets", "in", "Section", "3", ".", "We", "describe", "experiments", "and", "analysis", "in", "Section", "4", ".", "Finally", ",", "we", "conclude", "and", "discuss", "future", "work", "in", "Section", "5", ".", "section", ":", "Related", "Work", "The", "early", "exploration", "on", "NLI", "mainly", "rely", "on", "conventional", "methods", "and", "small", "scale", "datasets", ".", "The", "availability", "of", "SNLI", "dataset", "with", "570k", "human", "annotated", "sentence", "pairs", "has", "enabled", "a", "good", "deal", "of", "progress", "on", "natural", "language", "understanding", ".", "The", "essential", "representation", "learning", "techniques", "for", "NLU", "such", "as", "attention", ",", "memory", "and", "the", "use", "of", "parse", "structure", "are", "studied", "on", "the", "SNLI", "which", "serves", "as", "an", "important", "benchmark", "for", "sentence", "understanding", ".", "The", "models", "trained", "on", "NLI", "task", "can", "be", "divided", "into", "two", "categories", ":", "(", "i", ")", "sentence", "encoding", "-", "based", "model", "which", "aims", "to", "find", "vector", "representation", "for", "each", "sentence", "and", "classifies", "the", "relation", "by", "using", "the", "concatenation", "of", "two", "vector", "representation", "along", "with", "their", "absolute", "element", "-", "wise", "difference", "and", "element", "-", "wise", "product", ".", "(", "ii", ")", "Joint", "feature", "models", "which", "use", "the", "cross", "sentence", "feature", "or", "attention", "from", "one", "sentence", "to", "another", ".", "After", "neural", "attention", "mechanism", "is", "successfully", "applied", "on", "the", "machine", "translation", "task", ",", "such", "technique", "has", "became", "widely", "used", "in", "both", "natural", "language", "process", "and", "computer", "vision", "domains", ".", "Many", "variants", "of", "attention", "technique", "such", "as", "hard", "-", "attention", ",", "self", "-", "attention", ",", "multi", "-", "hop", "attention", ",", "bidirectional", "attention", "and", "multi", "-", "head", "attention", "are", "also", "introduced", "to", "tackle", "more", "complicated", "tasks", ".", "Before", "this", "work", ",", "neural", "attention", "mechanism", "is", "mainly", "used", "to", "make", "alignment", ",", "focusing", "on", "specific", "part", "of", "the", "representation", ".", "In", "this", "work", ",", "we", "want", "to", "show", "that", "attention", "weight", "contains", "rich", "semantic", "information", "required", "for", "understanding", "the", "logical", "relationship", "between", "sentence", "pair", ".", "Though", "RNN", "or", "LSTM", "are", "very", "good", "for", "variable", "length", "sequence", "modeling", ",", "using", "Convolutional", "neural", "network", "in", "NLU", "tasks", "is", "very", "desirable", "because", "of", "its", "parallelism", "in", "computation", ".", "Convolutional", "structure", "has", "been", "successfully", "applied", "in", "various", "domain", "such", "as", "machine", "translation", ",", "sentence", "classification", ",", "text", "matching", "and", "sentiment", "analysis", ",", "etc", ".", "The", "convolution", "structure", "is", "also", "applied", "on", "different", "level", "of", "granularity", "such", "as", "byte", ",", "character", ",", "word", "and", "sentences", "levels", ".", "section", ":", "Model", "subsection", ":", "Interactive", "Inference", "Network", "The", "Interactive", "Inference", "Network", "(", "IIN", ")", "is", "a", "hierarchical", "multi", "-", "stage", "process", "and", "consists", "of", "five", "components", ".", "Each", "of", "the", "components", "is", "compatible", "with", "different", "type", "of", "implementations", ".", "Potentially", "all", "exiting", "approaches", "in", "machine", "learning", ",", "such", "as", "decision", "tree", ",", "support", "vector", "machine", "and", "neural", "network", "approach", ",", "can", "be", "transfer", "to", "replace", "certain", "component", "in", "this", "architecture", ".", "We", "focus", "on", "neural", "network", "approaches", "below", ".", "Figure", "[", "reference", "]", "provides", "a", "visual", "illustration", "of", "Interactive", "Inference", "Network", ".", "Embedding", "Layer", "converts", "each", "word", "or", "phrase", "to", "a", "vector", "representation", "and", "construct", "the", "representation", "matrix", "for", "sentences", ".", "In", "embedding", "layer", ",", "a", "model", "can", "map", "tokens", "to", "vectors", "with", "the", "pre", "-", "trained", "word", "representation", "such", "as", "GloVe", ",", "word2Vec", "and", "fasttext", ".", "It", "can", "also", "utilize", "the", "pre", "-", "processing", "tool", ",", "e.g.", "named", "entity", "recognizer", ",", "part", "-", "of", "-", "speech", "recognizer", ",", "lexical", "parser", "and", "coreference", "identifier", "etc", ".", ",", "to", "incorporate", "more", "lexical", "and", "syntactical", "information", "into", "the", "feature", "vector", ".", "Encoding", "Layer", "encodes", "the", "representations", "by", "incorporating", "the", "context", "information", "or", "enriching", "the", "representation", "with", "desirable", "features", "for", "future", "use", ".", "For", "instance", ",", "a", "model", "can", "adopt", "bidirectional", "recurrent", "neural", "network", "to", "model", "the", "temporal", "interaction", "on", "both", "direction", ",", "recursive", "neural", "network", "(", "also", "known", "as", "TreeRNN", ")", "to", "model", "the", "compositionality", "and", "the", "recursive", "structure", "of", "language", ",", "or", "self", "-", "attention", "to", "model", "the", "long", "-", "term", "dependency", "on", "sentence", ".", "Different", "components", "of", "encoder", "can", "be", "combined", "to", "obtain", "a", "better", "sentence", "matrix", "representation", ".", "Interaction", "Layer", "creates", "an", "word", "-", "by", "-", "word", "interaction", "tensor", "by", "both", "premise", "and", "hypothesis", "representation", "matrix", ".", "The", "interaction", "can", "be", "modeled", "in", "different", "ways", ".", "A", "common", "approach", "is", "to", "compute", "the", "cosine", "similarity", "or", "dot", "product", "between", "each", "pair", "of", "feature", "vector", ".", "On", "the", "other", "hand", ",", "a", "high", "-", "order", "interaction", "tensor", "can", "be", "constructed", "with", "the", "outer", "product", "between", "two", "matrix", "representations", ".", "Feature", "Extraction", "Layer", "adopts", "feature", "extractor", "to", "extract", "the", "semantic", "feature", "from", "interaction", "tensor", ".", "The", "convolutional", "feature", "extractors", ",", "such", "as", "AlexNet", ",", "VGG", ",", "Inception", ",", "ResNet", "and", "DenseNet", ",", "proven", "work", "well", "on", "image", "recognition", "are", "completely", "compatible", "under", "such", "architecture", ".", "Unlike", "the", "work", "who", "employs", "1", "-", "D", "sliding", "window", ",", "our", "CNN", "architecture", "allows", "2", "-", "D", "kernel", "to", "extract", "semantic", "interaction", "feature", "from", "the", "word", "-", "by", "-", "word", "interaction", "between", "n", "-", "gram", "pair", ".", "Sequential", "or", "tree", "-", "like", "feature", "extractors", "are", "also", "applicable", "in", "the", "feature", "extraction", "layer", ".", "Output", "Layer", "decodes", "the", "acquired", "features", "to", "give", "prediction", ".", "Under", "the", "setting", "of", "NLI", ",", "the", "output", "layer", "predicts", "the", "confidence", "on", "each", "class", ".", "subsection", ":", "Densely", "Interactive", "Inference", "Network", "Here", "we", "introduce", "Densely", "Interactive", "Inference", "Network", "(", "DIIN", ")", ",", "which", "is", "a", "relatively", "simple", "instantiation", "of", "IIN", "but", "produces", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "multiple", "datasets", ".", "paragraph", ":", "Embedding", "Layer", ":", "For", "DIIN", ",", "we", "use", "the", "concatenation", "of", "word", "embedding", ",", "character", "feature", "and", "syntactical", "features", ".", "The", "word", "embedding", "is", "obtained", "by", "mapping", "token", "to", "high", "dimensional", "vector", "space", "by", "pre", "-", "trained", "word", "vector", "(", "840B", "GloVe", ")", ".", "The", "word", "embedding", "is", "updated", "during", "training", ".", "As", "in", ",", "we", "filter", "character", "embedding", "with", "1D", "convolution", "kernel", ".", "The", "character", "convolutional", "feature", "maps", "are", "then", "max", "pooled", "over", "time", "dimension", "for", "each", "token", "to", "obtain", "a", "vector", ".", "The", "character", "features", "supplies", "extra", "information", "for", "some", "out", "-", "of", "-", "vocabulary", "(", "OOV", ")", "words", ".", "Syntactical", "features", "include", "one", "-", "hot", "part", "-", "of", "-", "speech", "(", "POS", ")", "tagging", "feature", "and", "binary", "exact", "match", "(", "EM", ")", "feature", ".", "The", "EM", "value", "is", "activated", "if", "there", "are", "tokens", "with", "same", "stem", "or", "lemma", "in", "the", "other", "sentence", "as", "the", "corresponding", "token", ".", "The", "EM", "feature", "is", "simple", "while", "found", "useful", "as", "in", "reading", "comprehension", "task", ".", "In", "analysis", "section", ",", "we", "study", "how", "EM", "feature", "helps", "text", "understanding", ".", "Now", "we", "have", "premise", "representation", "and", "hypothesis", "representation", ",", "where", "refers", "to", "the", "sequence", "length", "of", "premise", ",", "refers", "to", "the", "sequence", "length", "of", "hypothesis", "and", "means", "the", "dimension", "of", "both", "representation", ".", "The", "1", "-", "D", "convolutional", "neural", "network", "and", "character", "features", "weights", "share", "the", "same", "set", "of", "parameters", "between", "premise", "and", "hypothesis", ".", "paragraph", ":", "Encoding", "Layer", ":", "In", "the", "encoding", "layer", ",", "the", "premise", "representation", "and", "the", "hypothesis", "representation", "are", "passed", "through", "a", "two", "-", "layer", "highway", "network", ",", "thus", "having", "and", "for", "new", "premise", "representation", "and", "new", "hypothesis", "representation", ".", "These", "new", "representation", "are", "then", "passed", "to", "self", "-", "attention", "layer", "to", "take", "into", "account", "the", "word", "order", "and", "context", "information", ".", "Take", "premise", "as", "example", ",", "we", "model", "self", "-", "attention", "by", "where", "is", "a", "weighted", "summation", "of", ".", "We", "choose", ",", "where", "is", "a", "trainable", "weight", ",", "is", "element", "-", "wise", "multiplication", ",", "[", ";", "]", "is", "vector", "concatenation", "across", "row", ",", "and", "the", "implicit", "multiplication", "is", "matrix", "multiplication", ".", "Then", "both", "and", "are", "fed", "into", "a", "semantic", "composite", "fuse", "gate", "(", "fuse", "gate", "in", "short", ")", ",", "which", "acts", "as", "a", "skip", "connection", ".", "The", "fuse", "gate", "is", "implemented", "as", "where", ",", ",", "and", ",", "are", "trainable", "weights", ",", "is", "sigmoid", "nonlinear", "operation", ".", "We", "do", "the", "same", "operation", "on", "hypothesis", "representation", ",", "thus", "having", ".", "The", "weights", "of", "intra", "-", "attention", "and", "fuse", "gate", "for", "premise", "and", "hypothesis", "are", "not", "shared", ",", "but", "the", "difference", "between", "the", "weights", "of", "are", "penalized", ".", "The", "penalization", "aims", "to", "ensure", "the", "parallel", "structure", "learns", "the", "similar", "functionality", "but", "is", "aware", "of", "the", "subtle", "semantic", "difference", "between", "premise", "and", "hypothesis", ".", "paragraph", ":", "Interaction", "Layer", ":", "The", "interaction", "layer", "models", "the", "interaction", "between", "premise", "encoded", "representation", "and", "hypothesis", "encoded", "representation", "as", "follows", ":", "where", "is", "the", "-", "th", "row", "vector", "of", ",", "and", "is", "the", "-", "th", "row", "vector", "of", ".", "Though", "there", "are", "many", "implementations", "of", "interaction", ",", "we", "find", "very", "useful", ".", "paragraph", ":", "Feature", "Extraction", "Layer", ":", "We", "adopt", "DenseNet", "as", "convolutional", "feature", "extractor", "in", "DIIN", ".", "Though", "our", "experiments", "show", "ResNet", "works", "well", "in", "the", "architecture", ",", "we", "choose", "DenseNet", "because", "it", "is", "effective", "in", "saving", "parameters", ".", "One", "interesting", "observation", "with", "ResNet", "is", "that", "if", "we", "remove", "the", "skip", "connection", "in", "residual", "structure", ",", "the", "model", "does", "not", "converge", "at", "all", ".", "We", "found", "batch", "normalization", "delays", "convergence", "without", "contributing", "to", "accuracy", ",", "therefore", "we", "does", "not", "use", "it", "in", "our", "case", ".", "A", "ReLU", "activation", "function", "is", "applied", "after", "all", "convolution", "unless", "otherwise", "noted", ".", "Once", "we", "have", "the", "interaction", "tensor", ",", "we", "use", "a", "convolution", "with", "kernel", "to", "scale", "down", "the", "tensor", "in", "a", "ratio", ",", ",", "without", "following", "ReLU", ".", "If", "the", "input", "channel", "is", "then", "the", "output", "channel", "is", ".", "Then", "the", "generated", "feature", "map", "is", "feed", "into", "three", "sets", "of", "Dense", "block", "and", "transition", "block", "pair", ".", "The", "DenseNet", "block", "contains", "n", "layers", "of", "convolution", "layer", "with", "growth", "rate", "of", "g", ".", "The", "transition", "layer", "has", "a", "convolution", "layer", "with", "kernel", "for", "scaling", "down", "purpose", ",", "followed", "by", "a", "max", "pooling", "layer", "with", "stride", ".", "The", "transition", "scale", "down", "ratio", "in", "transition", "layer", "is", ".", "paragraph", ":", "Output", "Layer", ":", "DIIN", "uses", "a", "linear", "layer", "to", "classify", "final", "flattened", "feature", "representation", "to", "three", "classes", ".", "section", ":", "Experiments", "In", "this", "section", ",", "we", "present", "the", "evaluation", "of", "our", "model", ".", "We", "first", "perform", "quantitative", "evaluation", ",", "comparing", "our", "model", "with", "other", "competitive", "models", ".", "We", "then", "conduct", "some", "qualitative", "analyses", "to", "understand", "how", "DIIN", "achieve", "the", "high", "level", "understanding", "through", "interaction", ".", "subsection", ":", "Data", "Here", "we", "introduce", "three", "datasets", "we", "evaluate", "our", "model", "on", ".", "The", "evaluation", "metric", "for", "all", "dataset", "is", "accuracy", ".", "paragraph", ":", "SNLI", "Stanford", "Natural", "Language", "Inference", "(", "SNLI", ";", ")", "has", "570k", "human", "annotated", "sentence", "pairs", ".", "The", "premise", "data", "is", "draw", "from", "the", "captions", "of", "the", "Flickr30k", "corpus", ",", "and", "the", "hypothesis", "data", "is", "manually", "composed", ".", "The", "labels", "provided", "in", "are", "“", "entailment", "”", ",", "“", "neutral", "’", ",", "“", "contradiction", "”", "and", "“", "-", "”", ".", "“", "-", "”", "shows", "that", "annotators", "can", "not", "reach", "consensus", "with", "each", "other", ",", "thus", "removed", "during", "training", "and", "testing", "as", "in", "other", "works", ".", "We", "use", "the", "same", "data", "split", "as", "in", ".", "paragraph", ":", "MultiNLI", "Multi", "-", "Genre", "NLI", "Corpus", "(", "MultiNLI", ";", ")", "has", "433k", "sentence", "pairs", ",", "whose", "collection", "process", "and", "task", "detail", "are", "modeled", "closely", "to", "SNLI", ".", "The", "premise", "data", "is", "collected", "from", "maximally", "broad", "range", "of", "genre", "of", "American", "English", "such", "as", "written", "non", "-", "fiction", "genres", "(", "SLATE", ",", "OUP", ",", "GOVERNMENT", ",", "VERBATIM", ",", "TRAVEL", ")", ",", "spoken", "genres", "(", "TELEPHONE", ",", "FACE", "-", "TO", "-", "FACE", ")", ",", "less", "formal", "written", "genres", "(", "FICTION", ",", "LETTERS", ")", "and", "a", "specialized", "one", "for", "9", "/", "11", ".", "Half", "of", "these", "selected", "genres", "appear", "in", "training", "set", "while", "the", "rest", "are", "not", ",", "creating", "in", "-", "domain", "(", "matched", ")", "and", "cross", "-", "domain", "(", "mismatched", ")", "development", "/", "test", "sets", ".", "We", "use", "the", "same", "data", "split", "as", "provided", "by", ".", "Since", "test", "set", "labels", "are", "not", "provided", ",", "the", "test", "performance", "is", "obtained", "through", "submission", "on", "Kaggle.com", ".", "Each", "team", "is", "limited", "to", "two", "submissions", "per", "day", ".", "paragraph", ":", "Quora", "question", "pair", "Quora", "question", "pair", "dataset", "contains", "over", "400k", "real", "world", "question", "pair", "selected", "from", "Quora.com", ".", "A", "binary", "annotation", "which", "stands", "for", "match", "(", "duplicate", ")", "or", "not", "match", "(", "not", "duplicate", ")", "is", "provided", "for", "each", "question", "pair", ".", "In", "our", "case", ",", "duplicate", "question", "pair", "can", "be", "interpreted", "as", "entailment", "relation", "and", "not", "duplicate", "as", "neutral", ".", "We", "use", "the", "same", "split", "ratio", "as", "mentioned", "in", ".", "subsection", ":", "Experiments", "setting", "We", "implement", "our", "algorithm", "with", "Tensorflow", "framework", ".", "An", "Adadelta", "optimizer", "with", "as", "0.95", "and", "as", "is", "used", "to", "optimize", "all", "the", "trainable", "weights", ".", "The", "initial", "learning", "rate", "is", "set", "to", "0.5", "and", "batch", "size", "to", "70", ".", "When", "the", "model", "does", "not", "improve", "best", "in", "-", "domain", "performance", "for", "30", ",", "000", "steps", ",", "an", "SGD", "optimizer", "with", "learning", "rate", "of", "is", "used", "to", "help", "model", "to", "find", "a", "better", "local", "optimum", ".", "Dropout", "layers", "are", "applied", "before", "all", "linear", "layers", "and", "after", "word", "-", "embedding", "layer", ".", "We", "use", "an", "exponential", "decayed", "keep", "rate", "during", "training", ",", "where", "the", "initial", "keep", "rate", "is", "1.0", "and", "the", "decay", "rate", "is", "0.977", "for", "every", "10", ",", "000", "step", ".", "We", "initialize", "our", "word", "embeddings", "with", "pre", "-", "trained", "300D", "GloVe", "840B", "vectors", "while", "the", "out", "-", "of", "-", "vocabulary", "word", "are", "randomly", "initialized", "with", "uniform", "distribution", ".", "The", "character", "embeddings", "are", "randomly", "initialized", "with", "100D.", "We", "crop", "or", "pad", "each", "token", "to", "have", "16", "characters", ".", "The", "1D", "convolution", "kernel", "size", "for", "character", "embedding", "is", "5", ".", "All", "weights", "are", "constraint", "by", "L2", "regularization", ",", "and", "the", "L2", "regularization", "at", "step", "is", "calculated", "as", "follows", ":", "where", "determines", "the", "maximum", "L2", "regularization", "ratio", ",", "and", "determines", "at", "which", "step", "the", "maximum", "L2", "regularization", "ratio", "would", "be", "applied", "on", "the", "L2", "regularization", ".", "We", "choose", "as", "and", "as", "100", ",", "000", ".", "The", "ratio", "of", "L2", "penalty", "between", "the", "difference", "of", "two", "encoder", "weights", "is", "set", "to", ".", "For", "a", "dense", "block", "in", "feature", "extraction", "layer", ",", "the", "number", "of", "layer", "is", "set", "to", "and", "growth", "rate", "g", "is", "set", "to", ".", "The", "first", "scale", "down", "ratio", "in", "feature", "extraction", "layer", "is", "set", "to", "and", "transitional", "scale", "down", "ratio", "is", "set", "to", ".", "The", "sequence", "length", "is", "set", "as", "a", "hard", "cutoff", "on", "all", "experiments", ":", "48", "for", "MultiNLI", ",", "32", "for", "SNLI", "and", "24", "for", "Quora", "Question", "Pair", "Dataset", ".", "During", "the", "experiments", "on", "MultiNLI", ",", "we", "use", "15", "%", "of", "data", "from", "SNLI", "as", "in", ".", "We", "select", "the", "parameter", "by", "the", "best", "run", "of", "development", "accuracy", ".", "Our", "ensembling", "approach", "considers", "the", "majority", "vote", "of", "the", "predictions", "given", "by", "multiple", "runs", "of", "the", "same", "model", "under", "different", "random", "parameter", "initialization", ".", "subsection", ":", "Experiment", "on", "MultiNLI", "We", "compare", "our", "result", "with", "all", "other", "published", "systems", "in", "Table", "[", "reference", "]", ".", "Besides", "ESIM", ",", "the", "state", "-", "of", "-", "the", "-", "art", "model", "on", "SNLI", ",", "all", "other", "models", "appear", "at", "RepEval", "2017", "workshop", ".", "RepEval", "2017", "workshop", "requires", "all", "submitted", "model", "to", "be", "sentence", "encoding", "-", "based", "model", "therefore", "alignment", "between", "sentences", "and", "memory", "module", "are", "not", "eligible", "for", "competition", ".", "All", "models", "except", "ours", "share", "one", "common", "feature", "that", "they", "use", "LSTM", "as", "a", "essential", "building", "block", "as", "encoder", ".", "Our", "approach", ",", "without", "using", "any", "recurrent", "structure", ",", "achieves", "the", "new", "state", "-", "of", "-", "the", "-", "art", "performance", "of", "80.0", "%", ",", "exceeding", "current", "state", "-", "of", "-", "the", "-", "art", "performance", "by", "more", "than", "5", "%", ".", "Unlike", "the", "observation", "from", ",", "we", "find", "the", "out", "-", "of", "-", "domain", "test", "performance", "is", "consistently", "lower", "than", "in", "-", "domain", "test", "performance", ".", "Selecting", "parameters", "from", "the", "best", "in", "-", "domain", "development", "accuracy", "partially", "contributes", "to", "this", "result", ".", "subsection", ":", "Experiment", "on", "SNLI", "In", "Table", "[", "reference", "]", ",", "we", "compare", "our", "model", "to", "other", "model", "performance", "on", "SNLI", ".", "Experiments", "(", "2", "-", "7", ")", "are", "sentence", "encoding", "based", "model", ".", "provides", "a", "BiLSTM", "baseline", ".", "adopts", "two", "layer", "GRU", "encoder", "with", "pre", "-", "trained", "”", "skip", "-", "thoughts", "”", "vectors", ".", "To", "capture", "sentence", "-", "level", "semantics", ",", "use", "tree", "-", "based", "CNN", "and", "propose", "a", "stack", "-", "augmented", "parser", "-", "interpreter", "neural", "network", "(", "SPINN", ")", "which", "incorporates", "parsing", "information", "in", "a", "sequential", "manner", ".", "uses", "intra", "-", "attention", "on", "top", "of", "BiLSTM", "to", "generate", "sentence", "representation", ",", "and", "proposes", "an", "memory", "augmented", "neural", "network", "to", "encode", "the", "sentence", ".", "The", "next", "group", "of", "model", ",", "experiments", "(", "8", "-", "18", ")", ",", "uses", "cross", "sentence", "feature", ".", "aligns", "each", "sentence", "word", "-", "by", "-", "word", "with", "attention", "on", "top", "of", "LSTMs", ".", "enforces", "cross", "sentence", "attention", "word", "-", "by", "-", "word", "matching", "with", "the", "proprosed", "mLSTM", "model", ".", "proposes", "long", "short", "-", "term", "memory", "-", "network", "(", "LSTMN", ")", "with", "deep", "attention", "fusion", "that", "links", "the", "current", "word", "to", "previous", "word", "stored", "in", "memory", ".", "decomposes", "the", "task", "into", "sub", "-", "problems", "and", "conquer", "them", "respectively", ".", "proposes", "neural", "tree", "indexer", ",", "a", "full", "n", "-", "ary", "tree", "whose", "subtrees", "can", "be", "overlapped", ".", "Re", "-", "read", "LSTM", "proposed", "by", "considers", "the", "attention", "vector", "of", "one", "sentence", "as", "the", "inner", "-", "state", "of", "LSTM", "for", "another", "sentence", ".", "propose", "a", "sequential", "model", "that", "infers", "locally", ",", "and", "a", "ensemble", "with", "tree", "-", "like", "inference", "module", "that", "further", "improves", "performance", ".", "We", "show", "our", "model", ",", "DIIN", ",", "achieves", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "the", "competitive", "leaderboard", ".", "subsection", ":", "Experiment", "on", "Quora", "Question", "Pair", "dataset", "In", "this", "subsection", ",", "we", "evaluate", "the", "effectiveness", "of", "our", "model", "for", "paraphrase", "identification", "as", "natural", "language", "inference", "task", ".", "Other", "than", "our", "baselines", ",", "we", "compare", "with", "and", ".", "BiMPM", "models", "different", "perspective", "of", "matching", "between", "sentence", "pair", "on", "both", "direction", ",", "then", "aggregates", "matching", "vector", "with", "LSTM", ".", "DecAtt⁢word", "and", "DecAtt⁢char", "uses", "automatically", "collected", "in", "-", "domain", "paraphrase", "data", "to", "noisy", "pretrain", "-", "gram", "word", "embedding", "and", "-", "gram", "subword", "embedding", "correspondingly", "on", "decomposable", "attention", "model", "proposed", "by", ".", "In", "Table", "[", "reference", "]", ",", "our", "experiment", "shows", "DIIN", "has", "better", "performance", "than", "all", "other", "models", "and", "an", "ensemble", "score", "is", "higher", "than", "the", "former", "best", "result", "for", "more", "than", "1", "percent", ".", "subsection", ":", "Analysis", "paragraph", ":", "Ablation", "Study", "We", "conduct", "a", "ablation", "study", "on", "our", "base", "model", "to", "examine", "the", "effectiveness", "of", "each", "component", ".", "We", "study", "our", "model", "on", "MultiNLI", "dataset", "and", "we", "use", "Matched", "validation", "score", "as", "the", "standard", "for", "model", "selection", ".", "The", "result", "is", "shown", "in", "Table", "[", "reference", "]", ".", "We", "studies", "how", "EM", "feature", "contributes", "to", "the", "system", ".", "After", "removing", "the", "exact", "match", "binary", "feature", ",", "we", "find", "the", "performance", "degrade", "to", "78.2", "on", "matched", "score", "on", "development", "set", "and", "78.0", "on", "mismatched", "score", ".", "As", "observed", "in", "reading", "comprehension", "task", ",", "the", "simple", "exact", "match", "feature", "does", "help", "the", "model", "to", "better", "understand", "the", "sentences", ".", "In", "the", "experiment", "3", ",", "we", "remove", "the", "convolutional", "feature", "extractor", "and", "then", "model", "is", "structured", "as", "a", "sentence", "-", "encoding", "based", "model", ".", "The", "sentence", "representation", "matrix", "is", "max", "-", "pooled", "over", "time", "to", "obtain", "a", "feature", "vector", ".", "Once", "we", "have", "the", "feature", "vector", "for", "premise", "and", "for", "hypothesis", ",", "we", "use", "as", "final", "feature", "vector", "to", "classify", "the", "relationship", ".", "We", "obtain", "73.2", "for", "matched", "score", "and", "73.6", "on", "mismatched", "data", ".", "The", "result", "is", "competitive", "among", "other", "sentence", "-", "encoding", "based", "model", ".", "We", "further", "study", "how", "encoding", "layer", "contribute", "in", "enriching", "the", "feature", "space", "in", "interaction", "tensor", ".", "If", "we", "remove", "encoding", "layer", "completely", ",", "then", "we", "’ll", "obtain", "a", "73.5", "for", "matched", "score", "and", "73.2", "for", "mismatched", "score", ".", "The", "result", "demonstrate", "the", "feature", "extraction", "layer", "have", "powerful", "capability", "to", "capture", "the", "semantic", "feature", ".", "In", "experiment", "5", ",", "we", "remove", "both", "self", "-", "attention", "and", "fuse", "gate", ",", "thus", "retaining", "only", "highway", "network", ".", "The", "result", "improves", "to", "77.7", "and", "77.3", "respectively", "on", "matched", "and", "mismatched", "development", "set", ".", "However", ",", "in", "experiment", "6", ",", "when", "we", "only", "remove", "fuse", "gate", ",", "to", "our", "surprise", ",", "the", "performance", "degrade", "to", "73.5", "for", "matched", "score", "and", "73.8", "for", "mismatched", ".", "On", "the", "other", "hand", ",", "if", "we", "use", "the", "addition", "of", "the", "representation", "after", "highway", "network", "and", "the", "representation", "after", "self", "-", "attention", "as", "skip", "connection", "as", "in", "experiment", "7", ",", "the", "performance", "increase", "to", "77.3", "and", "76.3", ".", "The", "comparison", "indicates", "self", "-", "attention", "layer", "makes", "the", "training", "harder", "to", "converge", "while", "a", "skip", "connection", "could", "ease", "the", "gradient", "flow", "for", "both", "highway", "layer", "and", "self", "-", "attention", "layer", ".", "By", "comparing", "the", "base", "model", "and", "the", "model", "the", "in", "experiment", "6", ",", "we", "show", "that", "the", "fuse", "gate", "not", "only", "well", "serves", "as", "a", "skip", "connection", ",", "but", "also", "makes", "good", "decision", "upon", "which", "information", "the", "fuse", "for", "both", "representation", ".", "To", "show", "that", "dense", "interaction", "tensor", "contains", "more", "semantic", "information", ",", "we", "replace", "the", "dense", "interaction", "tensor", "with", "dot", "product", "similarity", "matrix", "between", "the", "encoded", "representation", "of", "premise", "and", "hypothesis", ".", "The", "result", "shows", "that", "the", "dot", "product", "similarity", "matrix", "has", "an", "inferior", "capacity", "of", "semantic", "information", ".", "Another", "dimensionality", "study", "is", "provided", "in", "supplementary", "material", ".", "In", "experiment", "9", ",", "we", "share", "the", "encoding", "layer", "weight", ",", "and", "the", "result", "decrease", "from", "the", "baseline", ".", "The", "result", "shows", "that", "the", "two", "set", "of", "encoding", "weights", "learn", "the", "subtle", "difference", "between", "premise", "and", "hypothesis", ".", "paragraph", ":", "Error", "analysis", "To", "analyze", "the", "model", "prediction", ",", "we", "use", "annotated", "subset", "of", "development", "set", "provided", "by", "that", "consists", "of", "1", ",", "000", "examples", "each", "tagged", "with", "zero", "or", "more", "following", "tags", ":", "CONDITIONAL", ":", "whether", "the", "sentence", "contains", "a", "conditional", ".", "WORD", "OVERLAP", ":", "whether", "both", "sentences", "share", "more", "than", "70", "%", "of", "their", "tokens", ".", "NEGATION", ":", "whether", "a", "negation", "shows", "up", "in", "either", "sentence", ".", "ANTO", ":", "whether", "two", "sentences", "contain", "antonym", "pair", ".", "LONG", "SENTENCE", ":", "whether", "premise", "or", "hypothesis", "is", "longer", "than", "30", "or", "16", "tokens", "respectively", ".", "TENSE", "DIFFERENCE", ":", "whether", "any", "verb", "in", "two", "sentences", "uses", "different", "tense", ".", "ACTIVE", "/", "PASSIVE", ":", "whether", "there", "is", "an", "active", "-", "to", "-", "passive", "(", "or", "vice", "versa", ")", "transformation", "from", "the", "premise", "to", "the", "hypothesis", ".", "PARAPHRASE", ":", "whether", "the", "two", "sentences", "are", "close", "paraphrases", "QUANTITY", "/", "TIME", "REASONING", ":", "whether", "understanding", "the", "pair", "requires", "quantity", "or", "time", "reasoning", ".", "COREF", ":", "Whether", "the", "hypothesis", "contains", "a", "pronoun", "or", "referring", "expression", "that", "needs", "to", "be", "resolved", "using", "the", "premise", ".", "QUANTIFIER", ":", "Whether", "either", "sentence", "contains", "one", "of", "the", "following", "quantifier", ":", "much", ",", "enough", ",", "more", ",", "most", ",", "less", ",", "least", ",", "no", ",", "none", ",", "some", ",", "any", ",", "many", ",", "few", ",", "several", ",", "almost", ",", "nearly", ".", "MODAL", ":", "Whether", "one", "of", "the", "following", "modal", "verbs", "appears", "in", "either", "sentence", ":", "can", ",", "could", ",", "may", ",", "might", ",", "must", ",", "will", ",", "would", ",", "should", ".", "BELIEF", ":", "Whether", "one", "of", "the", "following", "belief", "verbs", "appear", "in", "either", "sentence", ":", "know", ",", "believe", ",", "understand", ",", "doubt", ",", "think", ",", "suppose", ",", "recognize", ",", "forget", ",", "remember", ",", "imagine", ",", "mean", ",", "agree", ",", "disagree", ",", "deny", ",", "promise", ".", "For", "more", "detailed", "descriptions", ",", "please", "resort", "to", ".", "The", "result", "is", "shown", "in", "Table", "[", "reference", "]", ".", "We", "find", "DIIN", "is", "consistently", "better", "on", "sentence", "pair", "with", "WORD", "OVERLAP", ",", "ANTO", ",", "LONG", "SENTENCE", ",", "PARAPHRASE", "and", "BELIEF", "tags", "by", "a", "large", "margin", ".", "During", "investigation", ",", "we", "hypothesize", "exact", "match", "feature", "helps", "the", "model", "to", "better", "understand", "paraphrase", ",", "therefore", "we", "study", "the", "result", "from", "second", "ablation", "ablation", "study", "where", "exact", "match", "feature", "is", "not", "used", ".", "Surprisingly", ",", "the", "model", "without", "exact", "model", "feature", "does", "not", "work", "worse", "on", "PARAPHRASE", ",", "instead", ",", "the", "accuracy", "on", "ANTO", "drops", "about", "10", "%", ".", "DIIN", "is", "also", "work", "well", "on", "LONG", "SENTENCE", ",", "partially", "because", "the", "receptive", "field", "is", "large", "enough", "to", "cover", "all", "tokens", ".", "paragraph", ":", "Visualization", "We", "also", "visualize", "the", "hidden", "representation", "from", "interaction", "tensor", "and", "the", "feature", "map", "from", "first", "dense", "block", "in", "Figure", "[", "reference", "]", ".", "We", "pick", "a", "sentence", "pair", "whose", "premise", "is", "“", "South", "Carolina", "has", "no", "referendum", "right", ",", "so", "the", "Supreme", "Court", "canceled", "the", "vote", "and", "upheld", "the", "ban", ".", "”", "and", "hypothesis", "is", "“", "South", "Carolina", "has", "a", "referendum", "right", ",", "so", "the", "Supreme", "Court", "was", "powerless", "over", "the", "state", ".", "”", ".", "The", "upper", "row", "of", "figures", "are", "sampled", "from", "hidden", "representation", "of", "interaction", "tensor", ".", "We", "observe", "the", "values", "of", "neurons", "are", "highly", "correlated", "row", "-", "wise", "and", "column", "-", "wise", "in", "the", "interaction", "tensor", "and", "different", "channel", "of", "hidden", "representation", "shows", "different", "aspect", "of", "interaction", ".", "Though", "in", "certain", "channel", "same", "words", ",", "“", "referendum", "”", ",", "or", "phrases", ",", "“", "supreme", "court", "”", ",", "cause", "activation", ",", "different", "word", "or", "phrase", "pair", ",", "such", "as", "“", "ban", "”", "and", "“", "powerless", "over", "”", ",", "also", "cause", "activation", "in", "other", "activation", ".", "It", "shows", "the", "model", "’s", "strong", "capacity", "of", "understanding", "text", "in", "different", "perspective", ".", "The", "lower", "row", "of", "Figure", "[", "reference", "]", "shows", "the", "feature", "map", "from", "first", "dense", "block", ".", "After", "being", "convolved", "from", "the", "interaction", "tensor", "and", "previous", "feature", "map", ",", "new", "feature", "maps", "shows", "activation", "in", "different", "position", ",", "demonstrating", "different", "semantic", "features", "are", "found", ".", "The", "first", "figure", "in", "the", "lower", "row", "has", "similar", "pattern", "as", "normal", "attention", "weight", "whereas", "others", "has", "no", "obvious", "pattern", ".", "Different", "channels", "of", "feature", "maps", "indicate", "different", "kinds", "of", "semantic", "feature", ".", "section", ":", "Conclusion", "and", "Future", "Work", "We", "show", "the", "interaction", "tensor", "(", "or", "attention", "weight", ")", "contains", "semantic", "information", "to", "understand", "the", "natural", "language", ".", "We", "introduce", "Interactive", "Inference", "Network", ",", "a", "novel", "class", "of", "architecture", "that", "allows", "the", "model", "to", "solve", "NLI", "or", "NLI", "alike", "tasks", "via", "extracting", "semantic", "feature", "from", "interaction", "tensor", "end", "-", "to", "-", "end", ".", "One", "instance", "of", "such", "architecture", ",", "Densely", "Interactive", "Inference", "Network", "(", "DIIN", ")", ",", "achieves", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "multiple", "datasets", ".", "By", "ablating", "each", "component", "in", "DIIN", "and", "changing", "the", "dimensionality", ",", "we", "show", "the", "effectiveness", "of", "each", "component", "in", "DIIN", ".", "Though", "we", "have", "the", "initial", "exploration", "of", "natural", "language", "inference", "in", "interaction", "space", ",", "the", "full", "potential", "is", "not", "yet", "clear", ".", "We", "will", "keep", "exploring", "the", "potential", "of", "interaction", "space", ".", "Incorporating", "common", "-", "sense", "knowledge", "from", "external", "resources", "such", "as", "knowledge", "base", "to", "leverage", "the", "capacity", "of", "the", "mode", "is", "another", "research", "goal", "of", "ours", ".", "subsubsection", ":", "Acknowledgments", "We", "thank", "Yuchen", "Lu", ",", "Chang", "Huang", "and", "Kai", "Yu", "for", "their", "sincere", "and", "insightful", "advice", ".", "bibliography", ":", "References", "appendix", ":", "Supplementary", "Material", "paragraph", ":", "Dimensionality", "and", "Parameter", "number", "study", "To", "study", "the", "influence", "of", "the", "model", "dimension", "which", "is", "also", "the", "channel", "number", "of", "interaction", "tensor", ",", "we", "design", "experiments", "to", "find", "out", "whether", "dimension", "has", "influence", "on", "performance", ".", "We", "also", "present", "the", "parameter", "count", "of", "these", "models", ".", "The", "dimensionality", "is", "448", "where", "300", "comes", "from", "word", "embedding", ",", "100", "comes", "from", "char", "feature", ",", "47", "comes", "from", "Part", "of", "speech", "tagging", "and", "1", "comes", "from", "the", "binary", "exact", "match", "feature", ".", "Since", "Highway", "network", "sets", "the", "output", "dimensionality", "default", "as", "that", "in", "input", ",", "we", "design", "a", "variant", "to", "highway", "network", "so", "that", "different", "output", "size", "could", "be", "obtained", ".", "The", "variant", "of", "highway", "layer", "is", "designed", "as", "follows", ":", "where", "is", "the", "-", "th", "vector", "of", "input", "matrix", ",", "is", "the", "-", "th", "vector", "of", "output", "matrix", ",", ",", ",", "and", ",", ",", "are", "trainable", "weights", ".", "The", "result", "shows", "that", "higher", "dimension", "number", "have", "better", "performance", "when", "the", "dimension", "number", "is", "lower", "certain", "threshold", ",", "however", ",", "when", "the", "number", "of", "dimensionality", "is", "greater", "than", "the", "threshold", ",", "larger", "number", "of", "parameter", "and", "higher", "dimensionality", "does", "n’t", "contribute", "to", "performance", ".", "In", "the", "case", "of", "SNLI", ",", "due", "to", "its", "simplicity", "in", "language", "pattern", ",", "250D", "would", "be", "suffice", "to", "obtain", "a", "good", "performance", ".", "On", "the", "other", "hand", ",", "it", "requires", "350D", "to", "achieve", "a", "competitive", "performance", "on", "MultiNLI", ".", "We", "fail", "to", "reproduce", "our", "best", "performance", "with", "the", "new", "structure", "on", "MultiNLI", ".", "It", "shows", "that", "the", "additional", "layer", "on", "highway", "network", "does", "n’t", "helps", "convergence", "."], "entity": [[2, 5, ["Natural", "Language", "Inference"], "Task"], [6, 8, ["Interaction", "Space"], "Task"], [8, 11, ["Natural", "Language", "Inference"], "Task"], [12, 13, ["NLI"], "Task"], [36, 39, ["Interactive", "Inference", "Network"], "Method"], [40, 42, ["IIN", ")"], "Method"], [47, 50, ["neural", "network", "architectures"], "Method"], [87, 90, ["natural", "language", "inference"], "Task"], [107, 111, ["Densely", "Interactive", "Inference", "Network"], "Method"], [112, 113, ["DIIN"], "Method"], [142, 143, ["DIIN"], "Method"], [149, 151, ["error", "reduction"], "Metric"], [174, 177, ["Natural", "Language", "Inference"], "Task"], [178, 179, ["NLI"], "Task"], [182, 185, ["recognizing", "textual", "entiailment"], "Task"], [243, 244, ["NLI"], "Task"], [311, 312, ["NLI"], "Task"], [314, 315, ["RTE"], "Task"], [326, 327, ["NLI"], "Task"], [341, 344, ["representation", "learning", "technique"], "Method"], [347, 351, ["core", "representation", "learning", "techniques"], "Method"], [352, 354, ["attention", "mechanism"], "Method"], [365, 367, ["machine", "translation"], "Task"], [368, 370, ["abstractive", "summarization"], "Task"], [371, 373, ["Reading", "Comprehension"], "Task"], [374, 376, ["dialog", "system"], "Task"], [385, 387, ["attention", "function"], "Method"], [454, 456, ["Attention", "mechanism"], "Method"], [512, 514, ["attention", "mechanism"], "Method"], [558, 562, ["multi", "-", "head", "attention"], "Task"], [568, 578, ["word", "-", "by", "-", "word", "dimension", "-", "wise", "alignment", "tensor"], "Method"], [581, 583, ["interaction", "tensor"], "Method"], [642, 644, ["Inference", "Network"], "Method"], [645, 646, ["IIN"], "Method"], [677, 680, ["Interactive", "Inference", "Network"], "Method"], [681, 685, ["Densely", "Interactive", "Inference", "Network"], "Method"], [702, 703, ["SNLI"], "Material"], [718, 721, ["paraphrase", "identification", "task"], "Task"], [727, 728, ["matching"], "Task"], [742, 746, ["Quora", "Question", "Pair", "dataset"], "Material"], [783, 784, ["IIN"], "Method"], [835, 836, ["NLI"], "Task"], [849, 851, ["SNLI", "dataset"], "Material"], [865, 868, ["natural", "language", "understanding"], "Task"], [871, 874, ["representation", "learning", "techniques"], "Method"], [875, 876, ["NLU"], "Task"], [891, 892, ["SNLI"], "Material"], [899, 901, ["sentence", "understanding"], "Task"], [918, 923, ["sentence", "encoding", "-", "based", "model"], "Method"], [927, 929, ["vector", "representation"], "Method"], [942, 944, ["vector", "representation"], "Method"], [961, 964, ["Joint", "feature", "models"], "Method"], [979, 982, ["neural", "attention", "mechanism"], "Method"], [987, 990, ["machine", "translation", "task"], "Task"], [999, 1002, ["natural", "language", "process"], "Task"], [1003, 1006, ["computer", "vision", "domains"], "Task"], [1010, 1012, ["attention", "technique"], "Method"], [1022, 1026, ["multi", "-", "hop", "attention"], "Task"], [1030, 1034, ["multi", "-", "head", "attention"], "Task"], [1047, 1050, ["neural", "attention", "mechanism"], "Method"], [1055, 1056, ["alignment"], "Task"], [1091, 1094, ["RNN", "or", "LSTM"], "Method"], [1098, 1102, ["variable", "length", "sequence", "modeling"], "Task"], [1104, 1107, ["Convolutional", "neural", "network"], "Method"], [1118, 1119, ["computation"], "Task"], [1120, 1122, ["Convolutional", "structure"], "Method"], [1131, 1133, ["machine", "translation"], "Task"], [1134, 1136, ["sentence", "classification"], "Task"], [1137, 1139, ["text", "matching"], "Task"], [1140, 1142, ["sentiment", "analysis"], "Task"], [1146, 1148, ["convolution", "structure"], "Method"], [1172, 1175, ["Interactive", "Inference", "Network"], "Method"], [1176, 1179, ["Interactive", "Inference", "Network"], "Method"], [1180, 1181, ["IIN"], "Method"], [1184, 1189, ["hierarchical", "multi", "-", "stage", "process"], "Method"], [1209, 1211, ["exiting", "approaches"], "Method"], [1212, 1214, ["machine", "learning"], "Task"], [1217, 1219, ["decision", "tree"], "Method"], [1220, 1223, ["support", "vector", "machine"], "Method"], [1224, 1227, ["neural", "network", "approach"], "Method"], [1242, 1245, ["neural", "network", "approaches"], "Method"], [1256, 1259, ["Interactive", "Inference", "Network"], "Method"], [1260, 1262, ["Embedding", "Layer"], "Method"], [1269, 1271, ["vector", "representation"], "Method"], [1274, 1276, ["representation", "matrix"], "Method"], [1280, 1282, ["embedding", "layer"], "Method"], [1295, 1297, ["word", "representation"], "Method"], [1299, 1300, ["GloVe"], "Method"], [1301, 1302, ["word2Vec"], "Method"], [1303, 1304, ["fasttext"], "Method"], [1310, 1314, ["pre", "-", "processing", "tool"], "Method"], [1316, 1319, ["named", "entity", "recognizer"], "Method"], [1320, 1326, ["part", "-", "of", "-", "speech", "recognizer"], "Method"], [1327, 1329, ["lexical", "parser"], "Method"], [1330, 1332, ["coreference", "identifier"], "Method"], [1347, 1349, ["Encoding", "Layer"], "Method"], [1375, 1379, ["bidirectional", "recurrent", "neural", "network"], "Method"], [1388, 1391, ["recursive", "neural", "network"], "Method"], [1395, 1396, ["TreeRNN"], "Method"], [1409, 1412, ["self", "-", "attention"], "Method"], [1425, 1426, ["encoder"], "Method"], [1433, 1436, ["sentence", "matrix", "representation"], "Method"], [1437, 1439, ["Interaction", "Layer"], "Method"], [1452, 1455, ["hypothesis", "representation", "matrix"], "Method"], [1504, 1506, ["matrix", "representations"], "Method"], [1507, 1510, ["Feature", "Extraction", "Layer"], "Method"], [1511, 1513, ["feature", "extractor"], "Method"], [1523, 1526, ["convolutional", "feature", "extractors"], "Method"], [1529, 1530, ["AlexNet"], "Method"], [1531, 1532, ["VGG"], "Method"], [1533, 1534, ["Inception"], "Method"], [1535, 1536, ["ResNet"], "Method"], [1537, 1538, ["DenseNet"], "Method"], [1543, 1545, ["image", "recognition"], "Task"], [1564, 1566, ["CNN", "architecture"], "Method"], [1567, 1571, ["2", "-", "D", "kernel"], "Method"], [1590, 1597, ["Sequential", "or", "tree", "-", "like", "feature", "extractors"], "Method"], [1602, 1605, ["feature", "extraction", "layer"], "Method"], [1606, 1608, ["Output", "Layer"], "Method"], [1614, 1615, ["prediction"], "Task"], [1620, 1621, ["NLI"], "Task"], [1623, 1625, ["output", "layer"], "Method"], [1634, 1638, ["Densely", "Interactive", "Inference", "Network"], "Method"], [1641, 1645, ["Densely", "Interactive", "Inference", "Network"], "Method"], [1656, 1657, ["IIN"], "Method"], [1673, 1675, ["Embedding", "Layer"], "Method"], [1677, 1678, ["DIIN"], "Method"], [1694, 1696, ["word", "embedding"], "Method"], [1718, 1720, ["word", "embedding"], "Method"], [1730, 1732, ["character", "embedding"], "Method"], [1733, 1736, ["1D", "convolution", "kernel"], "Method"], [1738, 1742, ["character", "convolutional", "feature", "maps"], "Method"], [1780, 1791, ["hot", "part", "-", "of", "-", "speech", "(", "POS", ")", "tagging", "feature"], "Method"], [1792, 1795, ["binary", "exact", "match"], "Method"], [1796, 1799, ["EM", ")", "feature"], "Method"], [1824, 1826, ["EM", "feature"], "Method"], [1833, 1836, ["reading", "comprehension", "task"], "Task"], [1844, 1846, ["EM", "feature"], "Method"], [1847, 1849, ["text", "understanding"], "Task"], [1853, 1855, ["premise", "representation"], "Method"], [1856, 1858, ["hypothesis", "representation"], "Method"], [1884, 1890, ["1", "-", "D", "convolutional", "neural", "network"], "Method"], [1907, 1909, ["Encoding", "Layer"], "Method"], [1912, 1914, ["encoding", "layer"], "Method"], [1916, 1918, ["premise", "representation"], "Method"], [1920, 1922, ["hypothesis", "representation"], "Method"], [1926, 1931, ["two", "-", "layer", "highway", "network"], "Method"], [1937, 1939, ["premise", "representation"], "Method"], [1941, 1943, ["hypothesis", "representation"], "Method"], [1951, 1955, ["self", "-", "attention", "layer"], "Method"], [1973, 1976, ["self", "-", "attention"], "Task"], [1980, 1982, ["weighted", "summation"], "Method"], [1994, 1998, ["element", "-", "wise", "multiplication"], "Method"], [2010, 2012, ["implicit", "multiplication"], "Method"], [2013, 2015, ["matrix", "multiplication"], "Method"], [2023, 2027, ["semantic", "composite", "fuse", "gate"], "Method"], [2028, 2030, ["fuse", "gate"], "Method"], [2042, 2044, ["fuse", "gate"], "Method"], [2067, 2069, ["hypothesis", "representation"], "Task"], [2101, 2102, ["penalization"], "Method"], [2131, 2133, ["interaction", "layer"], "Method"], [2137, 2140, ["premise", "encoded", "representation"], "Method"], [2141, 2144, ["hypothesis", "encoded", "representation"], "Method"], [2180, 2183, ["Feature", "Extraction", "Layer"], "Method"], [2186, 2187, ["DenseNet"], "Method"], [2188, 2191, ["convolutional", "feature", "extractor"], "Method"], [2192, 2193, ["DIIN"], "Method"], [2198, 2199, ["ResNet"], "Method"], [2207, 2208, ["DenseNet"], "Method"], [2213, 2215, ["saving", "parameters"], "Task"], [2220, 2221, ["ResNet"], "Method"], [2243, 2245, ["batch", "normalization"], "Method"], [2246, 2247, ["convergence"], "Task"], [2250, 2251, ["accuracy"], "Metric"], [2263, 2266, ["ReLU", "activation", "function"], "Method"], [2270, 2271, ["convolution"], "Method"], [2285, 2288, ["convolution", "with", "kernel"], "Method"], [2300, 2301, ["ReLU"], "Method"], [2338, 2340, ["convolution", "layer"], "Method"], [2347, 2349, ["transition", "layer"], "Method"], [2351, 2353, ["convolution", "layer"], "Method"], [2354, 2355, ["kernel"], "Method"], [2356, 2359, ["scaling", "down", "purpose"], "Task"], [2363, 2366, ["max", "pooling", "layer"], "Method"], [2370, 2374, ["transition", "scale", "down", "ratio"], "Metric"], [2384, 2385, ["DIIN"], "Method"], [2387, 2389, ["linear", "layer"], "Method"], [2392, 2395, ["flattened", "feature", "representation"], "Method"], [2425, 2427, ["competitive", "models"], "Method"], [2437, 2438, ["DIIN"], "Method"], [2461, 2463, ["evaluation", "metric"], "Metric"], [2467, 2468, ["accuracy"], "Metric"], [2471, 2472, ["SNLI"], "Material"], [2472, 2476, ["Stanford", "Natural", "Language", "Inference"], "Material"], [2477, 2478, ["SNLI"], "Material"], [2590, 2591, ["SNLI"], "Material"], [2729, 2732, ["Quora", "question", "pair"], "Material"], [2732, 2736, ["Quora", "question", "pair", "dataset"], "Material"], [2809, 2811, ["Tensorflow", "framework"], "Method"], [2813, 2815, ["Adadelta", "optimizer"], "Method"], [2831, 2833, ["learning", "rate"], "Metric"], [2861, 2863, ["SGD", "optimizer"], "Method"], [2864, 2866, ["learning", "rate"], "Metric"], [2879, 2881, ["Dropout", "layers"], "Method"], [2885, 2887, ["linear", "layers"], "Method"], [2889, 2893, ["word", "-", "embedding", "layer"], "Method"], [2897, 2901, ["exponential", "decayed", "keep", "rate"], "Method"], [2902, 2903, ["training"], "Method"], [2907, 2909, ["keep", "rate"], "Metric"], [2913, 2915, ["decay", "rate"], "Metric"], [2927, 2929, ["word", "embeddings"], "Method"], [2933, 2937, ["300D", "GloVe", "840B", "vectors"], "Method"], [2972, 2976, ["1D", "convolution", "kernel", "size"], "Method"], [2977, 2979, ["character", "embedding"], "Task"], [2987, 2989, ["L2", "regularization"], "Method"], [2992, 2994, ["L2", "regularization"], "Method"], [3024, 3026, ["L2", "regularization"], "Method"], [3057, 3060, ["feature", "extraction", "layer"], "Task"], [3082, 3085, ["feature", "extraction", "layer"], "Method"], [3116, 3117, ["SNLI"], "Material"], [3120, 3124, ["Quora", "Question", "Pair", "Dataset"], "Material"], [3138, 3139, ["SNLI"], "Material"], [3151, 3153, ["development", "accuracy"], "Metric"], [3155, 3157, ["ensembling", "approach"], "Method"], [3174, 3177, ["random", "parameter", "initialization"], "Method"], [3199, 3200, ["ESIM"], "Method"], [3211, 3212, ["SNLI"], "Material"], [3222, 3223, ["RepEval"], "Method"], [3231, 3236, ["sentence", "encoding", "-", "based", "model"], "Method"], [3241, 3243, ["memory", "module"], "Method"], [3260, 3261, ["LSTM"], "Method"], [3267, 3268, ["encoder"], "Method"], [3275, 3277, ["recurrent", "structure"], "Method"], [3317, 3324, ["out", "-", "of", "-", "domain", "test", "performance"], "Metric"], [3339, 3344, ["in", "-", "domain", "development", "accuracy"], "Metric"], [3354, 3355, ["SNLI"], "Material"], [3370, 3371, ["SNLI"], "Material"], [3379, 3383, ["sentence", "encoding", "based", "model"], "Method"], [3386, 3388, ["BiLSTM", "baseline"], "Method"], [3390, 3394, ["two", "layer", "GRU", "encoder"], "Method"], [3413, 3417, ["tree", "-", "based", "CNN"], "Method"], [3420, 3428, ["stack", "-", "augmented", "parser", "-", "interpreter", "neural", "network"], "Method"], [3429, 3430, ["SPINN"], "Method"], [3447, 3448, ["BiLSTM"], "Method"], [3450, 3452, ["sentence", "representation"], "Task"], [3456, 3460, ["memory", "augmented", "neural", "network"], "Method"], [3496, 3497, ["LSTMs"], "Method"], [3499, 3508, ["cross", "sentence", "attention", "word", "-", "by", "-", "word", "matching"], "Task"], [3510, 3513, ["proprosed", "mLSTM", "model"], "Method"], [3515, 3522, ["long", "short", "-", "term", "memory", "-", "network"], "Method"], [3523, 3524, ["LSTMN"], "Method"], [3526, 3529, ["deep", "attention", "fusion"], "Method"], [3554, 3557, ["neural", "tree", "indexer"], "Method"], [3559, 3564, ["full", "n", "-", "ary", "tree"], "Method"], [3570, 3574, ["Re", "-", "read", "LSTM"], "Method"], [3589, 3590, ["LSTM"], "Method"], [3596, 3598, ["sequential", "model"], "Method"], [3604, 3611, ["ensemble", "with", "tree", "-", "like", "inference", "module"], "Method"], [3621, 3622, ["DIIN"], "Method"], [3641, 3645, ["Quora", "Question", "Pair", "dataset"], "Material"], [3657, 3659, ["paraphrase", "identification"], "Task"], [3675, 3676, ["BiMPM"], "Method"], [3693, 3694, ["LSTM"], "Method"], [3708, 3713, ["pretrain", "-", "gram", "word", "embedding"], "Task"], [3714, 3718, ["-", "gram", "subword", "embedding"], "Task"], [3720, 3723, ["decomposable", "attention", "model"], "Method"], [3735, 3736, ["DIIN"], "Method"], [3745, 3747, ["ensemble", "score"], "Metric"], [3765, 3767, ["Ablation", "Study"], "Task"], [3794, 3797, ["Matched", "validation", "score"], "Metric"], [3801, 3803, ["model", "selection"], "Task"], [3840, 3842, ["matched", "score"], "Metric"], [3848, 3850, ["mismatched", "score"], "Metric"], [3854, 3857, ["reading", "comprehension", "task"], "Task"], [3881, 3884, ["convolutional", "feature", "extractor"], "Method"], [3891, 3896, ["sentence", "-", "encoding", "based", "model"], "Method"], [3898, 3901, ["sentence", "representation", "matrix"], "Method"], [3940, 3942, ["matched", "score"], "Metric"], [3954, 3959, ["sentence", "-", "encoding", "based", "model"], "Method"], [3964, 3966, ["encoding", "layer"], "Method"], [3990, 3992, ["matched", "score"], "Metric"], [4002, 4005, ["feature", "extraction", "layer"], "Method"], [4031, 4033, ["highway", "network"], "Method"], [4072, 4074, ["matched", "score"], "Metric"], [4093, 4095, ["highway", "network"], "Method"], [4121, 4125, ["self", "-", "attention", "layer"], "Method"], [4142, 4144, ["highway", "layer"], "Method"], [4145, 4149, ["self", "-", "attention", "layer"], "Method"], [4167, 4169, ["fuse", "gate"], "Method"], [4216, 4218, ["encoded", "representation"], "Method"], [4228, 4232, ["dot", "product", "similarity", "matrix"], "Method"], [4289, 4291, ["Error", "analysis"], "Method"], [4357, 4358, ["ANTO"], "Method"], [4421, 4422, ["PARAPHRASE"], "Method"], [4429, 4434, ["paraphrases", "QUANTITY", "/", "TIME", "REASONING"], "Task"], [4445, 4446, ["COREF"], "Method"], [4465, 4466, ["QUANTIFIER"], "Method"], [4507, 4508, ["MODAL"], "Method"], [4602, 4603, ["DIIN"], "Method"], [4650, 4653, ["ablation", "ablation", "study"], "Task"], [4666, 4669, ["exact", "model", "feature"], "Method"], [4674, 4675, ["PARAPHRASE"], "Task"], [4679, 4680, ["accuracy"], "Metric"], [4681, 4682, ["ANTO"], "Method"], [4687, 4688, ["DIIN"], "Method"], [4711, 4712, ["Visualization"], "Task"], [4795, 4800, ["hidden", "representation", "of", "interaction", "tensor"], "Method"], [4825, 4827, ["hidden", "representation"], "Method"], [4919, 4921, ["feature", "map"], "Method"], [4923, 4925, ["feature", "maps"], "Method"], [4998, 5001, ["Interactive", "Inference", "Network"], "Method"], [5013, 5014, ["NLI"], "Task"], [5015, 5016, ["NLI"], "Task"], [5037, 5041, ["Densely", "Interactive", "Inference", "Network"], "Method"], [5063, 5064, ["DIIN"], "Method"], [5077, 5078, ["DIIN"], "Method"], [5232, 5236, ["Part", "of", "speech", "tagging"], "Method"], [5247, 5249, ["Highway", "network"], "Method"], [5264, 5266, ["highway", "network"], "Method"], [5278, 5280, ["highway", "layer"], "Method"], [5362, 5363, ["SNLI"], "Material"], [5389, 5390, ["350D"], "Method"], [5419, 5421, ["highway", "network"], "Method"], [128, 129, ["NLI"], "Task"], [134, 135, ["NLI"], "Task"], [157, 158, ["NLI"], "Task"], [187, 188, ["RTE"], "Task"], [254, 257, ["natural", "language", "understanding"], "Task"], [359, 360, ["NLU"], "Task"], [618, 621, ["natural", "language", "inference"], "Task"], [662, 665, ["natural", "language", "inference"], "Task"], [686, 687, ["DIIN"], "Method"], [722, 725, ["natural", "language", "inference"], "Task"], [906, 907, ["NLI"], "Task"], [1108, 1109, ["NLU"], "Task"], [1646, 1647, ["DIIN"], "Method"], [2332, 2333, ["DenseNet"], "Method"], [2569, 2570, ["NLI"], "Task"], [3660, 3663, ["natural", "language", "inference"], "Task"], [4295, 4296, ["prediction"], "Task"], [4613, 4614, ["ANTO"], "Method"], [5042, 5043, ["DIIN"], "Method"], [5086, 5089, ["natural", "language", "inference"], "Task"]]} {"doc_id": "178275dbdcfa267e41a9d5efe386ee5874c6d23f", "text": ["document", ":", "Fraternal", "Dropout", "Recurrent", "neural", "networks", "(", "RNNs", ")", "form", "an", "important", "class", "of", "architectures", "among", "neural", "networks", "useful", "for", "language", "modeling", "and", "sequential", "prediction", ".", "However", ",", "optimizing", "RNNs", "is", "known", "to", "be", "harder", "compared", "to", "feed", "-", "forward", "neural", "networks", ".", "A", "number", "of", "techniques", "have", "been", "proposed", "in", "literature", "to", "address", "this", "problem", ".", "In", "this", "paper", "we", "propose", "a", "simple", "technique", "called", "fraternal", "dropout", "that", "takes", "advantage", "of", "dropout", "to", "achieve", "this", "goal", ".", "Specifically", ",", "we", "propose", "to", "train", "two", "identical", "copies", "of", "an", "RNN", "(", "that", "share", "parameters", ")", "with", "different", "dropout", "masks", "while", "minimizing", "the", "difference", "between", "their", "(", "pre", "-", "softmax", ")", "predictions", ".", "In", "this", "way", "our", "regularization", "encourages", "the", "representations", "of", "RNNs", "to", "be", "invariant", "to", "dropout", "mask", ",", "thus", "being", "robust", ".", "We", "show", "that", "our", "regularization", "term", "is", "upper", "bounded", "by", "the", "expectation", "-", "linear", "dropout", "objective", "which", "has", "been", "shown", "to", "address", "the", "gap", "due", "to", "the", "difference", "between", "the", "train", "and", "inference", "phases", "of", "dropout", ".", "We", "evaluate", "our", "model", "and", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", "in", "sequence", "modeling", "tasks", "on", "two", "benchmark", "datasets", "–", "Penn", "Treebank", "and", "Wikitext", "-", "2", ".", "We", "also", "show", "that", "our", "approach", "leads", "to", "performance", "improvement", "by", "a", "significant", "margin", "in", "image", "captioning", "(", "Microsoft", "COCO", ")", "and", "semi", "-", "supervised", "(", "CIFAR", "-", "10", ")", "tasks", ".", "section", ":", "Introduction", "Recurrent", "neural", "networks", "(", "RNNs", ")", "like", "long", "short", "-", "term", "memory", "(", "LSTM", ";", ")", "networks", "and", "gated", "recurrent", "unit", "(", "GRU", ";", ")", "are", "popular", "architectures", "for", "sequence", "modeling", "tasks", "like", "language", "generation", ",", "translation", ",", "speech", "synthesis", ",", "and", "machine", "comprehension", ".", "However", ",", "they", "are", "harder", "to", "optimize", "compared", "to", "feed", "-", "forward", "networks", "due", "to", "challenges", "like", "variable", "length", "input", "sequences", ",", "repeated", "application", "of", "the", "same", "transition", "operator", "at", "each", "time", "step", ",", "and", "largely", "-", "dense", "embedding", "matrix", "that", "depends", "on", "the", "vocabulary", "size", ".", "Due", "to", "these", "optimization", "challenges", "in", "RNNs", ",", "the", "application", "of", "batch", "normalization", "and", "its", "variants", "(", "layer", "normalization", ",", "recurrent", "batch", "normalization", ",", "recurrent", "normalization", "propagation", ")", "have", "not", "been", "as", "successful", "as", "their", "counterparts", "in", "feed", "-", "forward", "networks", ",", "although", "they", "do", "considerably", "provide", "performance", "gains", ".", "Similarly", ",", "naive", "application", "of", "dropout", "has", "been", "shown", "to", "be", "ineffective", "in", "RNNs", ".", "Therefore", ",", "regularization", "techniques", "for", "RNNs", "is", "an", "active", "area", "of", "research", ".", "To", "address", "these", "challenges", ",", "proposed", "to", "apply", "dropout", "only", "to", "the", "non", "-", "recurrent", "connections", "in", "multi", "-", "layer", "RNNs", ".", "Variational", "dropout", "(", ")", "uses", "the", "same", "dropout", "mask", "throughout", "a", "sequence", "during", "training", ".", "DropConnect", "applies", "the", "dropout", "operation", "on", "the", "weight", "matrices", ".", "Zoneout", "(", ")", ",", "in", "a", "similar", "spirit", "with", "dropout", ",", "randomly", "chooses", "to", "use", "the", "previous", "time", "step", "hidden", "state", "instead", "of", "using", "the", "current", "one", ".", "Similarly", "as", "a", "substitute", "for", "batch", "normalization", ",", "layer", "normalization", "normalizes", "the", "hidden", "units", "within", "each", "sample", "to", "have", "zero", "mean", "and", "unit", "standard", "deviation", ".", "Recurrent", "batch", "normalization", "applies", "batch", "normalization", "but", "with", "unshared", "mini", "-", "batch", "statistics", "for", "each", "time", "step", ".", "and", "on", "the", "other", "hand", "show", "that", "activity", "regularization", "(", "AR", ")", "and", "temporal", "activation", "regularization", "(", "TAR", ")", "are", "also", "effective", "methods", "for", "regularizing", "LSTMs", ".", "Another", "more", "recent", "way", "of", "regularizing", "RNNs", ",", "that", "is", "similar", "in", "spirit", "to", "the", "approach", "we", "take", ",", "involves", "minimizing", "the", "difference", "between", "the", "hidden", "states", "of", "the", "original", "and", "the", "auxiliary", "network", ".", "In", "this", "paper", "we", "propose", "a", "simple", "regularization", "based", "on", "dropout", "that", "we", "call", "fraternal", "dropout", ",", "where", "we", "minimize", "an", "equally", "weighted", "sum", "of", "prediction", "losses", "from", "two", "identical", "copies", "of", "the", "same", "LSTM", "with", "different", "dropout", "masks", ",", "and", "add", "as", "a", "regularization", "the", "difference", "between", "the", "predictions", "(", "pre", "-", "softmax", ")", "of", "the", "two", "networks", ".", "We", "analytically", "show", "that", "our", "regularization", "objective", "is", "equivalent", "to", "minimizing", "the", "variance", "in", "predictions", "from", "different", "i.i.d", ".", "dropout", "masks", ";", "thus", "encouraging", "the", "predictions", "to", "be", "invariant", "to", "dropout", "masks", ".", "We", "also", "discuss", "how", "our", "regularization", "is", "related", "to", "expectation", "linear", "dropout", ",", "-", "model", "and", "activity", "regularization", ",", "and", "empirically", "show", "that", "our", "method", "provides", "non", "-", "trivial", "gains", "over", "these", "related", "methods", "which", "we", "explain", "furthermore", "in", "our", "ablation", "study", "(", "Section", "[", "reference", "]", ")", ".", "section", ":", "Fraternal", "dropout", "Dropout", "is", "a", "powerful", "regularization", "for", "neural", "networks", ".", "It", "is", "usually", "more", "effective", "on", "densely", "connected", "layers", "because", "they", "suffer", "more", "from", "overfitting", "compared", "with", "convolution", "layers", "where", "the", "parameters", "are", "shared", ".", "For", "this", "reason", "dropout", "is", "an", "important", "regularization", "for", "RNNs", ".", "However", ",", "dropout", "has", "a", "gap", "between", "its", "training", "and", "inference", "phase", "since", "the", "latter", "phase", "assumes", "linear", "activations", "to", "correct", "for", "the", "factor", "by", "which", "the", "expected", "value", "of", "each", "activation", "would", "be", "different", ".", "In", "addition", ",", "the", "prediction", "of", "models", "with", "dropout", "generally", "vary", "with", "different", "dropout", "mask", ".", "However", ",", "the", "desirable", "property", "in", "such", "cases", "would", "be", "to", "have", "final", "predictions", "be", "invariant", "to", "dropout", "masks", ".", "As", "such", ",", "the", "idea", "behind", "fraternal", "dropout", "is", "to", "train", "a", "neural", "network", "model", "in", "a", "way", "that", "encourages", "the", "variance", "in", "predictions", "under", "different", "dropout", "masks", "to", "be", "as", "small", "as", "possible", ".", "Specifically", ",", "consider", "we", "have", "an", "RNN", "model", "denoted", "by", "that", "takes", "as", "input", ",", "where", "denotes", "the", "model", "parameters", ".", "Let", "be", "the", "prediction", "of", "the", "model", "for", "input", "sample", "at", "time", ",", "for", "dropout", "mask", "and", "current", "input", ",", "where", "is", "a", "function", "of", "and", "the", "hidden", "states", "corresponding", "to", "the", "previous", "time", "steps", ".", "Similarly", ",", "let", "be", "the", "corresponding", "time", "step", "loss", "value", "for", "the", "overall", "input", "-", "target", "sample", "pair", ".", "Then", "in", "fraternal", "dropout", ",", "we", "simultaneously", "feed", "-", "forward", "the", "input", "sample", "through", "two", "identical", "copies", "of", "the", "RNN", "that", "share", "the", "same", "parameters", "but", "with", "different", "dropout", "masks", "and", "at", "each", "time", "step", ".", "This", "yields", "two", "loss", "values", "at", "each", "time", "step", "given", "by", ",", "and", ".", "Then", "the", "overall", "loss", "function", "of", "fraternal", "dropout", "is", "given", "by", ",", "where", "is", "the", "regularization", "coefficient", ",", "is", "the", "dimensions", "of", "and", "is", "the", "fraternal", "dropout", "regularization", "given", "by", ",", "We", "use", "Monte", "Carlo", "sampling", "to", "approximate", "where", "and", "are", "the", "same", "as", "the", "one", "used", "to", "calculate", "values", ".", "Hence", ",", "the", "additional", "computation", "is", "negligible", ".", "We", "note", "that", "the", "regularization", "term", "of", "our", "objective", "is", "equivalent", "to", "minimizing", "the", "variance", "in", "the", "prediction", "function", "with", "different", "dropout", "masks", "as", "shown", "below", "(", "proof", "in", "the", "appendix", ")", ".", "theorem", ":", ".", "Let", "sit", "and", "sjt", "be", "i.i.d", ".", "dropout", "masks", "and", "∈⁢pt", "(", "zt", ",", "sit;θ", ")", "Rm", "be", "the", "prediction", "function", "as", "described", "above", ".", "Then", ",", "Note", "that", "a", "generalization", "of", "our", "approach", "would", "be", "to", "minimize", "the", "difference", "between", "the", "predictions", "of", "the", "two", "networks", "with", "different", "data", "/", "model", "augmentations", ".", "However", ",", "in", "this", "paper", "we", "focus", "on", "using", "different", "dropout", "masks", "and", "experiment", "mainly", "with", "RNNs", ".", "section", ":", "Related", "Work", "subsection", ":", "Relation", "to", "Expectation", "Linear", "Dropout", "(", "ELD", ")", "analytically", "showed", "that", "the", "expected", "error", "(", "over", "samples", ")", "between", "a", "model", "’s", "expected", "prediction", "over", "all", "dropout", "masks", ",", "and", "the", "prediction", "using", "the", "average", "mask", ",", "is", "upper", "bounded", ".", "Based", "on", "this", "result", ",", "they", "propose", "to", "explicitly", "minimize", "the", "difference", "(", "we", "have", "adapted", "their", "regularization", "to", "our", "notations", ")", ",", "where", "is", "the", "dropout", "mask", ".", "However", ",", "due", "to", "feasibility", "consideration", ",", "they", "instead", "propose", "to", "use", "the", "following", "regularization", "in", "practice", ",", "Specifically", ",", "this", "is", "achieved", "by", "feed", "-", "forwarding", "the", "input", "twice", "through", "the", "network", ",", "with", "and", "without", "dropout", "mask", ",", "and", "minimizing", "the", "main", "network", "loss", "(", "with", "dropout", ")", "along", "with", "the", "regularization", "term", "specified", "above", "(", "but", "without", "back", "-", "propagating", "the", "gradients", "through", "the", "network", "without", "dropout", ")", ".", "The", "goal", "of", "is", "to", "minimize", "the", "network", "loss", "along", "with", "the", "expected", "difference", "between", "the", "prediction", "from", "individual", "dropout", "mask", "and", "the", "prediction", "from", "the", "expected", "dropout", "mask", ".", "We", "note", "that", "our", "regularization", "objective", "is", "upper", "bounded", "by", "the", "expectation", "-", "linear", "dropout", "regularization", "as", "shown", "below", "(", "proof", "in", "the", "appendix", ")", ".", "theorem", ":", ".", ".", "This", "result", "shows", "that", "minimizing", "the", "ELD", "objective", "indirectly", "minimizes", "our", "regularization", "term", ".", "Finally", "as", "indicated", "above", ",", "they", "apply", "the", "target", "loss", "only", "on", "the", "network", "with", "dropout", ".", "In", "fact", ",", "in", "our", "own", "ablation", "studies", "(", "see", "Section", "[", "reference", "]", ")", "we", "find", "that", "back", "-", "propagating", "target", "loss", "through", "the", "network", "(", "without", "dropout", ")", "makes", "optimizing", "the", "model", "harder", ".", "However", ",", "in", "our", "setting", ",", "simultaneously", "back", "-", "propagating", "target", "loss", "through", "both", "networks", "yields", "both", "performance", "gain", "as", "well", "as", "convergence", "gain", ".", "We", "believe", "convergence", "is", "faster", "for", "our", "regularization", "because", "network", "weights", "are", "more", "likely", "to", "get", "target", "based", "updates", "from", "back", "-", "propagation", "in", "our", "case", ".", "This", "is", "especially", "true", "for", "weight", "dropout", "since", "in", "this", "case", "dropped", "weights", "do", "not", "get", "updated", "in", "the", "training", "iteration", ".", "subsection", ":", "Relation", "to", "-", "model", "propose", "-", "model", "with", "the", "goal", "of", "improving", "performance", "on", "classification", "tasks", "in", "the", "semi", "-", "supervised", "setting", ".", "They", "propose", "a", "model", "similar", "to", "ours", "(", "considering", "the", "equivalent", "deep", "feed", "-", "forward", "version", "of", "our", "model", ")", "except", "they", "apply", "target", "loss", "only", "on", "one", "of", "the", "networks", "and", "use", "time", "-", "dependent", "weighting", "function", "(", "while", "we", "use", "constant", ")", ".", "The", "intuition", "in", "their", "case", "is", "to", "leverage", "unlabeled", "data", "by", "using", "them", "to", "minimize", "the", "difference", "in", "prediction", "between", "the", "two", "copies", "of", "the", "network", "with", "different", "dropout", "masks", ".", "Further", ",", "they", "also", "test", "their", "model", "in", "the", "supervised", "setting", "but", "fail", "to", "explain", "the", "improvements", "they", "obtain", "by", "using", "this", "regularization", ".", "We", "note", "that", "in", "our", "case", "we", "analytically", "show", "that", "minimizing", "our", "regularizer", "(", "also", "used", "in", "-", "model", ")", "is", "equivalent", "to", "minimizing", "the", "variance", "in", "the", "model", "predictions", "(", "Remark", "[", "reference", "]", ")", ".", "Furthermore", ",", "we", "also", "show", "the", "relation", "of", "our", "regularizer", "to", "expectation", "linear", "dropout", "(", "Proposition", "[", "reference", "]", ")", ".", "In", "Section", "[", "reference", "]", ",", "we", "study", "the", "effects", "of", "target", "based", "loss", "on", "both", "networks", ",", "which", "is", "not", "used", "in", "the", "-", "model", ".", "We", "find", "that", "applying", "target", "loss", "on", "both", "the", "networks", "leads", "to", "significantly", "faster", "convergence", ".", "Finally", ",", "we", "bring", "to", "attention", "that", "temporal", "embedding", "(", "another", "model", "proposed", "by", ",", "claimed", "to", "be", "a", "better", "version", "of", "-", "model", "for", "semi", "-", "supervised", ",", "learning", ")", "is", "intractable", "in", "natural", "language", "processing", "applications", "because", "storing", "averaged", "predictions", "over", "all", "of", "the", "time", "steps", "would", "be", "memory", "exhaustive", "(", "since", "predictions", "are", "usually", "huge", "-", "tens", "of", "thousands", "values", ")", ".", "On", "a", "final", "note", ",", "we", "argue", "that", "in", "the", "supervised", "case", ",", "using", "a", "time", "-", "dependent", "weighting", "function", "instead", "of", "a", "constant", "value", "is", "not", "needed", ".", "Since", "the", "ground", "truth", "labels", "are", "known", ",", "we", "have", "not", "observed", "the", "problem", "mentioned", "by", ",", "that", "the", "network", "gets", "stuck", "in", "a", "degenerate", "solution", "when", "is", "too", "large", "in", "earlier", "epochs", "of", "training", ".", "We", "note", "that", "it", "is", "much", "easier", "to", "search", "for", "an", "optimal", "constant", "value", ",", "which", "is", "true", "in", "our", "case", ",", "as", "opposed", "to", "tuning", "the", "time", "-", "dependent", "function", ".", "Similarity", "to", "-", "model", "makes", "our", "method", "related", "to", "other", "semi", "-", "supervised", "works", ",", "mainly", "and", ".", "Since", "semi", "-", "supervised", "learning", "is", "not", "a", "primary", "focus", "of", "this", "paper", ",", "we", "refer", "to", "for", "more", "details", ".", "We", "note", "that", "the", "idea", "of", "adding", "a", "penalty", "encouraging", "the", "representation", "to", "be", "similar", "for", "two", "different", "masks", "was", "previously", "implemented", "by", "the", "authors", "of", "a", "Multi", "-", "Prediction", "Deep", "Boltzmann", "Machines", ".", "Nevertheless", ",", "the", "idea", "is", "not", "discussed", "in", "their", "paper", ".", "Another", "way", "to", "address", "the", "gap", "between", "the", "train", "and", "evaluation", "mode", "of", "dropout", "is", "to", "perform", "Monte", "Carlo", "sampling", "of", "masks", "and", "average", "the", "predictions", "during", "evaluation", ",", "and", "this", "has", "been", "used", "for", "feed", "-", "forward", "networks", ".", "We", "find", "that", "this", "technique", "does", "not", "work", "well", "for", "RNNs", ".", "The", "details", "of", "these", "experiments", "can", "be", "found", "in", "the", "appendix", ".", "section", ":", "Experiments", "subsection", ":", "Language", "Models", "In", "the", "case", "of", "language", "modeling", "we", "test", "our", "model", "on", "two", "benchmark", "datasets", "–", "Penn", "Tree", "-", "bank", "(", "PTB", ")", "dataset", "and", "WikiText", "-", "2", "(", "WT2", ")", "dataset", ".", "We", "use", "preprocessing", "as", "specified", "by", "(", "for", "PTB", "corpus", ")", "and", "Moses", "tokenizer", "(", "for", "the", "WT2", "dataset", ")", ".", "For", "both", "datasets", "we", "use", "the", "AWD", "-", "LSTM", "3", "-", "layer", "architecture", "described", "in", "which", "we", "call", "the", "baseline", "model", ".", "The", "number", "of", "parameters", "in", "the", "model", "used", "for", "PTB", "is", "24", "million", "as", "compared", "to", "34", "million", "in", "the", "case", "of", "WT2", "because", "WT2", "has", "a", "larger", "vocabulary", "size", "for", "which", "we", "use", "a", "larger", "embedding", "matrix", ".", "Apart", "from", "those", "differences", ",", "the", "architectures", "are", "identical", ".", "When", "we", "use", "fraternal", "dropout", ",", "we", "simply", "add", "our", "regularization", "on", "top", "of", "this", "baseline", "model", ".", "Word", "level", "Penn", "Treebank", "(", "PTB", ")", ".", "Influenced", "by", ",", "our", "goal", "here", "is", "to", "make", "sure", "that", "fraternal", "dropout", "outperforms", "existing", "methods", "not", "simply", "because", "of", "extensive", "hyper", "-", "parameter", "grid", "search", "but", "rather", "due", "to", "its", "regularization", "effects", ".", "Hence", ",", "in", "our", "experiments", "we", "leave", "a", "vast", "majority", "of", "hyper", "-", "parameters", "used", "in", "the", "baseline", "model", "unchanged", "i.e.", "embedding", "and", "hidden", "states", "sizes", ",", "gradient", "clipping", "value", ",", "weight", "decay", "and", "the", "values", "used", "for", "all", "dropout", "layers", "(", "dropout", "on", "the", "word", "vectors", ",", "the", "output", "between", "LSTM", "layers", ",", "the", "output", "of", "the", "final", "LSTM", ",", "and", "embedding", "dropout", ")", ".", "However", ",", "a", "few", "changes", "are", "necessary", ":", "the", "coefficients", "for", "AR", "and", "TAR", "needed", "to", "be", "altered", "because", "fraternal", "dropout", "also", "affects", "RNNs", "activation", "(", "as", "explained", "in", "Subsection", "[", "reference", "]", ")", "–", "we", "did", "not", "run", "grid", "search", "to", "obtain", "the", "best", "values", "but", "simply", "deactivated", "AR", "and", "TAR", "regularizers", ";", "since", "fraternal", "dropout", "needs", "twice", "as", "much", "memory", ",", "batch", "size", "is", "halved", "so", "the", "model", "needs", "approximately", "the", "same", "amount", "of", "memory", "and", "hence", "fits", "on", "the", "same", "GPU", ".", "The", "final", "change", "in", "hyper", "-", "parameters", "is", "to", "alter", "the", "non", "-", "monotone", "interval", "used", "in", "non", "-", "monotonically", "triggered", "averaged", "SGD", "(", "NT", "-", "ASGD", ")", "optimizer", ".", "We", "run", "a", "grid", "search", "on", "and", "obtain", "very", "similar", "results", "for", "the", "largest", "values", "(", "40", ",", "50", "and", "60", ")", "in", "the", "candidate", "set", ".", "Hence", ",", "our", "model", "is", "trained", "longer", "using", "ordinary", "SGD", "optimizer", "as", "compared", "to", "the", "baseline", "model", ".", "We", "evaluate", "our", "model", "using", "the", "perplexity", "metric", "and", "compare", "the", "results", "that", "we", "obtain", "against", "the", "existing", "state", "-", "of", "-", "the", "-", "art", "results", ".", "The", "results", "are", "reported", "in", "Table", "[", "reference", "]", ".", "Our", "approach", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "compared", "with", "existing", "benchmarks", ".", "To", "confirm", "that", "the", "gains", "are", "robust", "to", "initialization", ",", "we", "run", "ten", "experiments", "for", "the", "baseline", "model", "with", "different", "seeds", "(", "without", "fine", "-", "tuning", ")", "for", "PTB", "dataset", "to", "compute", "confidence", "intervals", ".", "The", "average", "best", "validation", "perplexity", "is", "with", "the", "minimum", "value", "equals", ".", "The", "same", "for", "test", "perplexity", "is", "and", ",", "respectively", ".", "Our", "score", "(", "validation", "and", "test", "perplexity", ")", "beats", "ordinal", "dropout", "minimum", "values", ".", "We", "also", "perform", "experiments", "using", "fraternal", "dropout", "with", "a", "grid", "search", "on", "all", "the", "hyper", "-", "parameters", "and", "find", "that", "it", "leads", "to", "further", "improvements", "in", "performance", ".", "The", "details", "of", "this", "experiment", "can", "be", "found", "in", "section", "[", "reference", "]", ".", "Word", "level", "WikiText", "-", "2", "(", "WT2", ")", ".", "In", "the", "case", "of", "WikiText", "-", "2", "language", "modeling", "task", ",", "we", "outperform", "the", "current", "state", "-", "of", "-", "the", "-", "art", "using", "the", "perplexity", "metric", "by", "a", "significant", "margin", ".", "Due", "to", "the", "lack", "of", "computational", "power", ",", "we", "run", "a", "single", "training", "procedure", "for", "fraternal", "dropout", "on", "WT2", "dataset", "because", "it", "is", "larger", "than", "PTB", ".", "In", "this", "experiment", ",", "we", "use", "the", "best", "hyper", "-", "parameters", "found", "for", "PTB", "dataset", "(", ",", "non", "-", "monotone", "interval", "and", "halved", "batch", "size", ";", "the", "rest", "of", "the", "hyper", "-", "parameters", "are", "the", "same", "as", "described", "in", "for", "WT2", ")", ".", "The", "final", "results", "are", "presented", "in", "Table", "[", "reference", "]", ".", "subsection", ":", "Image", "captioning", "We", "also", "apply", "fraternal", "dropout", "on", "an", "image", "captioning", "task", ".", "We", "use", "the", "well", "-", "known", "show", "and", "tell", "model", "as", "a", "baseline", ".", "We", "emphasize", "that", "in", "the", "image", "captioning", "task", ",", "the", "image", "encoder", "and", "sentence", "decoder", "architectures", "are", "usually", "learned", "together", ".", "Since", "we", "want", "to", "focus", "on", "the", "benefits", "of", "using", "fraternal", "dropout", "in", "RNNs", "we", "use", "frozen", "pretrained", "ResNet", "-", "101", "model", "as", "our", "image", "encoder", ".", "It", "means", "that", "our", "results", "are", "not", "directly", "comparable", "with", "other", "state", "-", "of", "-", "the", "-", "art", "methods", ",", "however", "we", "report", "results", "for", "the", "original", "methods", "so", "readers", "can", "see", "that", "our", "baseline", "performs", "well", ".", "The", "final", "results", "are", "presented", "in", "Table", "[", "reference", "]", ".", "We", "argue", "that", "in", "this", "task", "smaller", "values", "are", "optimal", "because", "the", "image", "captioning", "encoder", "is", "given", "all", "information", "in", "the", "beginning", "and", "hence", "the", "variance", "of", "consecutive", "predictions", "is", "smaller", "that", "in", "unconditioned", "natural", "language", "processing", "tasks", ".", "Fraternal", "dropout", "may", "benefits", "here", "mainly", "due", "to", "averaging", "gradients", "for", "different", "mask", "and", "hence", "updating", "weights", "more", "frequently", ".", "section", ":", "Ablation", "Studies", "In", "this", "section", ",", "the", "goal", "is", "to", "study", "existing", "methods", "closely", "related", "to", "ours", "–", "expectation", "linear", "dropout", ",", "-", "model", "and", "activity", "regularization", ".", "All", "of", "our", "experiments", "for", "ablation", "studies", ",", "which", "apply", "a", "single", "layer", "LSTM", ",", "use", "the", "same", "hyper", "-", "parameters", "and", "model", "architecture", "as", ".", "subsection", ":", "Expectation", "-", "linear", "dropout", "(", "ELD", ")", "The", "relation", "with", "expectation", "-", "linear", "dropout", "has", "been", "discussed", "in", "Section", "[", "reference", "]", ".", "Here", "we", "perform", "experiments", "to", "study", "the", "difference", "in", "performance", "when", "using", "the", "ELD", "regularization", "versus", "our", "regularization", "(", "FD", ")", ".", "In", "addition", "to", "ELD", ",", "we", "also", "study", "a", "modification", "(", "ELDM", ")", "of", "ELD", "which", "applies", "target", "loss", "to", "both", "copies", "of", "LSTMs", "in", "ELD", "similar", "to", "FD", "(", "notice", "in", "their", "case", "they", "only", "have", "dropout", "on", "one", "LSTM", ")", ".", "Finally", "we", "also", "evaluate", "a", "baseline", "model", "without", "any", "of", "these", "regularizations", ".", "The", "learning", "dynamics", "curves", "are", "shown", "in", "Figure", "[", "reference", "]", ".", "Our", "regularization", "performs", "better", "in", "terms", "of", "convergence", "compared", "with", "other", "methods", ".", "In", "terms", "of", "generalization", ",", "we", "find", "that", "FD", "is", "similar", "to", "ELD", ",", "but", "baseline", "and", "ELDM", "are", "much", "worse", ".", "Interestingly", ",", "looking", "at", "the", "train", "and", "validation", "curves", "together", ",", "ELDM", "seems", "to", "be", "suffering", "from", "optimization", "problems", ".", "subsection", ":", "-", "model", "Since", "-", "model", "is", "similar", "to", "our", "algorithm", "(", "even", "though", "it", "is", "designed", "for", "semi", "-", "supervised", "learning", "in", "feed", "-", "forward", "networks", ")", ",", "we", "study", "the", "difference", "in", "performance", "with", "-", "model", "both", "qualitatively", "and", "quantitatively", "to", "establish", "the", "advantage", "of", "our", "approach", ".", "First", ",", "we", "run", "both", "single", "layer", "LSTM", "and", "3", "-", "layer", "AWD", "-", "LSTM", "on", "PTB", "task", "to", "check", "how", "their", "model", "compares", "with", "ours", "in", "the", "case", "of", "language", "modeling", ".", "The", "results", "are", "shown", "in", "Figure", "[", "reference", "]", "and", "[", "reference", "]", ".", "We", "find", "that", "our", "model", "converges", "significantly", "faster", "than", "-", "model", ".", "We", "believe", "this", "happens", "because", "we", "back", "-", "propagate", "the", "target", "loss", "through", "both", "networks", "(", "in", "contrast", "to", "-", "model", ")", "that", "leads", "to", "weights", "getting", "updated", "using", "target", "-", "based", "gradients", "more", "often", ".", "Even", "though", "we", "designed", "our", "algorithm", "specifically", "to", "address", "problems", "in", "RNNs", ",", "to", "have", "a", "fair", "comparison", ",", "we", "compare", "with", "-", "model", "on", "a", "semi", "-", "supervised", "task", "which", "is", "their", "goal", ".", "Specifically", ",", "we", "use", "the", "CIFAR", "-", "10", "dataset", "that", "consists", "of", "images", "from", "10", "classes", ".", "Following", "the", "usual", "splits", "used", "in", "semi", "-", "supervised", "learning", "literature", ",", "we", "use", "4", "thousand", "labeled", "and", "41", "thousand", "unlabeled", "samples", "for", "training", ",", "5", "thousand", "labeled", "samples", "for", "validation", "and", "10", "thousand", "labeled", "samples", "for", "test", "set", ".", "We", "use", "the", "original", "ResNet", "-", "56", "architecture", ".", "We", "run", "grid", "search", "on", ",", "dropout", "rates", "in", "and", "leave", "the", "rest", "of", "the", "hyper", "-", "parameters", "unchanged", ".", "We", "additionally", "check", "importance", "of", "using", "unlabeled", "data", ".", "The", "results", "are", "reported", "in", "Table", "[", "reference", "]", ".", "We", "find", "that", "our", "algorithm", "performs", "at", "par", "with", "-", "model", ".", "When", "unlabeled", "data", "is", "not", "used", ",", "fraternal", "dropout", "provides", "slightly", "better", "results", "as", "compared", "to", "traditional", "dropout", ".", "subsection", ":", "Activity", "regularization", "and", "temporal", "activity", "regularization", "analysis", "The", "authors", "of", "study", "the", "importance", "of", "activity", "regularization", "(", "AR", ")", "and", "temporal", "activity", "regularization", "(", "TAR", ")", "in", "LSTMs", "given", "as", ",", "where", "is", "the", "LSTM", "’s", "output", "activation", "at", "time", "step", "(", "hence", "depends", "on", "both", "current", "input", "and", "the", "model", "parameters", ")", ".", "Notice", "that", "AR", "and", "TAR", "regularizations", "are", "applied", "on", "the", "output", "of", "the", "LSTM", ",", "while", "our", "regularization", "is", "applied", "on", "the", "pre", "-", "softmax", "output", "of", "the", "LSTM", ".", "However", ",", "since", "our", "regularization", "can", "be", "decomposed", "as", "and", "encapsulates", "an", "term", "along", "with", "the", "dot", "product", "term", ",", "we", "perform", "experiments", "to", "confirm", "that", "the", "gains", "in", "our", "approach", "is", "not", "due", "to", "the", "regularization", "alone", ".", "A", "similar", "argument", "goes", "for", "the", "TAR", "objective", ".", "We", "run", "a", "grid", "search", "on", ",", ",", "which", "include", "the", "hyper", "-", "parameters", "mentioned", "in", ".", "For", "our", "regularization", ",", "we", "use", ".", "Furthermore", ",", "we", "also", "compare", "with", "a", "regularization", "(", "PR", ")", "that", "regularizes", "to", "further", "rule", "-", "out", "any", "gains", "only", "from", "regularization", ".", "Based", "on", "this", "grid", "search", ",", "we", "pick", "the", "best", "model", "on", "the", "validation", "set", "for", "all", "the", "regularizations", ",", "and", "additionally", "report", "a", "baseline", "model", "without", "any", "of", "these", "four", "mentioned", "regularizations", ".", "The", "learning", "dynamics", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "Our", "regularization", "performs", "better", "both", "in", "terms", "of", "convergence", "and", "generalization", "compared", "with", "other", "methods", ".", "Average", "hidden", "state", "activation", "is", "reduced", "when", "any", "of", "the", "regularizer", "described", "is", "applied", "(", "see", "Figure", "[", "reference", "]", ")", ".", "subsection", ":", "Improvements", "using", "fine", "-", "tuning", "We", "confirm", "that", "models", "trained", "with", "fraternal", "dropout", "benefit", "from", "the", "NT", "-", "ASGD", "fine", "-", "tuning", "step", "(", "as", "also", "used", "in", ")", ".", "However", ",", "this", "is", "a", "very", "time", "-", "consuming", "practice", "and", "since", "different", "hyper", "-", "parameters", "may", "be", "used", "in", "this", "additional", "part", "of", "the", "learning", "procedure", ",", "the", "probability", "of", "obtaining", "better", "results", "due", "to", "the", "extensive", "grid", "search", "is", "higher", ".", "Hence", ",", "in", "our", "experiments", "we", "use", "the", "same", "fine", "-", "tuning", "procedure", "as", "implemented", "in", "the", "official", "repository", "(", "even", "fraternal", "dropout", "was", "not", "used", ")", ".", "We", "present", "the", "importance", "of", "fine", "-", "tuning", "in", "Table", "[", "reference", "]", ".", "subsection", ":", "Fraternal", "dropout", "and", "expectation", "linear", "dropout", "comparison", "We", "perform", "extensive", "grid", "search", "for", "the", "baseline", "model", "from", "Subsection", "[", "reference", "]", "(", "an", "AWD", "-", "LSTM", "3", "-", "layer", "architecture", ")", "trained", "with", "either", "fraternal", "dropout", "or", "expectation", "linear", "dropout", "regularizations", ",", "to", "further", "contrast", "the", "performance", "of", "these", "two", "methods", ".", "The", "experiments", "are", "run", "without", "fine", "-", "tuning", "on", "the", "PTB", "dataset", ".", "In", "each", "run", ",", "all", "five", "dropout", "rates", "are", "randomly", "altered", "(", "they", "are", "set", "to", "their", "original", "value", ",", "as", "in", ",", "multiplied", "by", "a", "value", "drawn", "from", "the", "uniform", "distribution", "on", "the", "interval", ")", "and", "the", "rest", "of", "the", "hyper", "-", "parameters", "are", "drawn", "as", "shown", "in", "Table", "[", "reference", "]", ".", "As", "in", "Subsection", "[", "reference", "]", ",", "AR", "and", "TAR", "regularizers", "are", "deactivated", ".", "Together", "we", "run", "more", "than", "400", "experiments", ".", "The", "results", "are", "presented", "in", "Table", "[", "reference", "]", ".", "Both", "FD", "and", "ELD", "perform", "better", "than", "the", "baseline", "model", "that", "instead", "uses", "AR", "and", "TAR", "regularizers", ".", "Hence", ",", "we", "confirm", "our", "previous", "finding", "(", "see", "Subsection", "[", "reference", "]", ")", "that", "both", "FD", "and", "ELD", "are", "better", ".", "However", ",", "as", "found", "previously", "for", "smaller", "model", "in", "Subsection", "[", "reference", "]", ",", "the", "convergence", "of", "FD", "is", "faster", "than", "that", "of", "ELD", ".", "Additionally", ",", "fraternal", "dropout", "is", "more", "robust", "to", "different", "hyper", "-", "parameters", "choice", "(", "more", "runs", "performing", "better", "than", "the", "baseline", "and", "better", "average", "for", "top", "performing", "runs", ")", ".", "section", ":", "Conclusion", "In", "this", "paper", "we", "propose", "a", "simple", "regularization", "method", "for", "RNNs", "called", "fraternal", "dropout", "that", "acts", "as", "a", "regularization", "by", "reducing", "the", "variance", "in", "model", "predictions", "across", "different", "dropout", "masks", ".", "We", "show", "that", "our", "model", "achieves", "state", "-", "of", "-", "the", "-", "art", "results", "on", "benchmark", "language", "modeling", "tasks", "along", "with", "faster", "convergence", ".", "We", "also", "analytically", "study", "the", "relationship", "between", "our", "regularization", "and", "expectation", "linear", "dropout", ".", "We", "perform", "a", "number", "of", "ablation", "studies", "to", "evaluate", "our", "model", "from", "different", "aspects", "and", "carefully", "compare", "it", "with", "related", "methods", "both", "qualitatively", "and", "quantitatively", ".", "section", ":", "Acknowledgements", "The", "authors", "would", "like", "to", "acknowledge", "the", "support", "of", "the", "following", "agencies", "for", "research", "funding", "and", "computing", "support", ":", "NSERC", ",", "CIFAR", ",", "and", "IVADO", ".", "We", "would", "like", "to", "thank", "Rosemary", "Nan", "Ke", "and", "Philippe", "Lacaille", "for", "their", "thoughts", "and", "comments", "throughout", "the", "project", ".", "We", "would", "also", "like", "to", "thank", "Stanisław", "Jastrzębski", "and", "Evan", "Racah", "for", "useful", "discussions", ".", "bibliography", ":", "References", "appendix", ":", "Appendix", "subsection", ":", "Monte", "Carlo", "evaluation", "A", "well", "known", "way", "to", "address", "the", "gap", "between", "the", "train", "and", "evaluation", "mode", "of", "dropout", "is", "to", "perform", "Monte", "Carlo", "sampling", "of", "masks", "and", "average", "the", "predictions", "during", "evaluation", "(", "MC", "-", "eval", ")", ",", "and", "this", "has", "been", "used", "for", "feed", "-", "forward", "networks", ".", "Since", "fraternal", "dropout", "addresses", "the", "same", "problem", ",", "we", "would", "like", "to", "clarify", "that", "it", "is", "not", "straight", "-", "forward", "and", "feasible", "to", "apply", "MC", "-", "eval", "for", "RNNs", ".", "In", "feed", "-", "forward", "networks", ",", "we", "average", "the", "output", "prediction", "scores", "from", "different", "masks", ".", "However", ",", "in", "the", "case", "RNNs", "(", "for", "next", "step", "predictions", ")", ",", "there", "is", "more", "than", "one", "way", "to", "perform", "such", "evaluation", ",", "but", "each", "one", "is", "problematic", ".", "They", "are", "as", "follows", ":", "1", ".", "Online", "averaging", "Consider", "that", "we", "first", "make", "the", "prediction", "at", "time", "step", "1", "using", "different", "masks", "by", "averaging", "the", "prediction", "score", ".", "Then", "we", "use", "this", "output", "to", "feed", "as", "input", "to", "the", "time", "step", "2", ",", "then", "use", "different", "masks", "at", "time", "step", "2", "to", "generate", "the", "output", "at", "time", "step", "2", ",", "and", "so", "on", ".", "But", "in", "order", "to", "do", "so", ",", "because", "of", "the", "way", "RNNs", "work", ",", "we", "also", "need", "to", "feed", "the", "previous", "time", "hidden", "state", "to", "time", "step", "2", ".", "One", "way", "would", "be", "to", "average", "the", "hidden", "states", "over", "different", "masks", "at", "time", "step", "1", ".", "But", "the", "hidden", "space", "can", "in", "general", "be", "highly", "nonlinear", ",", "and", "it", "is", "not", "clear", "if", "averaging", "in", "this", "space", "is", "a", "good", "strategy", ".", "This", "approach", "is", "not", "justified", ".", "Besides", ",", "this", "strategy", "as", "a", "whole", "is", "extremely", "time", "consuming", "because", "we", "would", "need", "to", "sequentially", "make", "predictions", "with", "multiple", "masks", "at", "each", "time", "step", ".", "2", ".", "Sequence", "averaging", "Let", "’s", "consider", "that", "we", "use", "a", "different", "mask", "each", "time", "we", "want", "to", "generate", "a", "sequence", ",", "and", "then", "we", "average", "the", "prediction", "scores", ",", "and", "compute", "the", "argmax", "(", "at", "each", "time", "step", ")", "to", "get", "the", "actual", "generated", "sequence", ".", "In", "this", "case", ",", "notice", "it", "is", "not", "guaranteed", "that", "the", "predicted", "word", "at", "time", "step", "due", "to", "averaging", "the", "predictions", "would", "lead", "to", "the", "next", "word", "(", "generated", "by", "the", "same", "process", ")", "if", "we", "were", "to", "feed", "the", "time", "step", "output", "as", "input", "to", "the", "time", "step", ".", "For", "example", ",", "with", "different", "dropout", "masks", ",", "if", "the", "probability", "of", "time", "step", "outputs", "are", ":", "I", "40", "%", ")", ",", "he", "(", "30", "%", ")", ",", "she", "(", "30", "%", ")", ",", "and", "the", "probability", "of", "the", "2nd", "time", "step", "outputs", "are", ":", "am", "(", "30", "%", ")", ",", "is", "(", "60", "%", ")", ",", "was", "(", "10", "%", ")", ".", "Then", "the", "averaged", "prediction", "score", "followed", "by", "argmax", "will", "result", "in", "the", "prediction", "‘", "‘", "I", "is", "’", "’", ",", "but", "this", "would", "be", "incorrect", ".", "A", "similar", "concern", "applies", "for", "output", "predictions", "varying", "in", "temporal", "length", ".", "Hence", ",", "this", "approach", "can", "not", "be", "used", "to", "generate", "a", "sequence", "(", "it", "has", "to", "be", "done", "by", "by", "sampling", "a", "mask", "and", "generating", "a", "single", "sequence", ")", ".", "However", ",", "this", "approach", "may", "be", "used", "to", "estimate", "the", "probability", "assigned", "by", "the", "model", "to", "a", "given", "sequence", ".", "Nonetheless", ",", "we", "run", "experiments", "on", "the", "PTB", "dataset", "using", "MC", "-", "eval", "(", "the", "results", "are", "summarized", "in", "Table", "[", "reference", "]", ")", ".", "We", "start", "with", "a", "simple", "comparison", "that", "compares", "fraternal", "dropout", "with", "the", "averaged", "mask", "and", "the", "AWD", "-", "LSTM", "3", "-", "layer", "baseline", "with", "a", "single", "fixed", "mask", "that", "we", "call", "MC1", ".", "The", "MC1", "model", "performs", "much", "worse", "than", "fraternal", "dropout", ".", "Hence", ",", "it", "would", "be", "hard", "to", "use", "MC1", "model", "in", "practice", "because", "a", "single", "sample", "is", "inaccurate", ".", "We", "also", "check", "MC", "-", "eval", "for", "a", "larger", "number", "of", "models", "(", "MC50", ")", "(", "50", "models", "were", "used", "since", "we", "were", "not", "able", "to", "fit", "more", "models", "simultaneously", "on", "a", "single", "GPU", ")", ".", "The", "final", "results", "for", "MC50", "are", "worse", "than", "the", "baseline", "which", "uses", "the", "averaged", "mask", ".", "For", "comparison", ",", "we", "also", "evaluate", "MC10", ".", "Note", "that", "no", "fine", "-", "tuning", "is", "used", "for", "the", "above", "experiments", ".", "subsection", ":", "Reasons", "for", "focusing", "on", "RNNs", "The", "fraternal", "dropout", "method", "is", "general", "and", "may", "be", "applied", "in", "feed", "-", "forward", "architectures", "(", "as", "shown", "in", "Subsection", "[", "reference", "]", "for", "CIFAR", "-", "10", "semisupervised", "example", ")", ".", "However", ",", "we", "believe", "that", "it", "is", "more", "powerful", "in", "the", "case", "of", "RNNs", "because", ":", "Variance", "in", "prediction", "accumulates", "among", "time", "steps", "in", "RNNs", "and", "since", "we", "share", "parameters", "for", "all", "time", "steps", ",", "one", "may", "use", "the", "same", "value", "at", "each", "step", ".", "In", "feed", "-", "forward", "networks", "the", "layers", "usually", "do", "not", "share", "parameters", "and", "hence", "one", "may", "want", "to", "use", "different", "values", "for", "different", "layers", "(", "which", "may", "be", "hard", "to", "tune", ")", ".", "The", "simple", "way", "to", "alleviate", "this", "problem", "is", "to", "apply", "the", "regularization", "term", "on", "the", "pre", "-", "softmax", "predictions", "only", "(", "as", "shown", "in", "the", "paper", ")", "or", "use", "the", "same", "value", "for", "all", "layers", ".", "However", ",", "we", "believe", "that", "it", "may", "limit", "possible", "gains", ".", "The", "best", "performing", "RNN", "architectures", "(", "state", "-", "of", "-", "the", "-", "art", ")", "usually", "use", "some", "kind", "of", "dropout", "(", "embedding", "dropout", ",", "word", "dropout", ",", "weight", "dropout", "etc", ".", ")", ",", "very", "often", "with", "high", "dropout", "rates", "(", "even", "larger", "than", "50", "%", "for", "input", "word", "embedding", "in", "NLP", "tasks", ")", ".", "However", ",", "this", "is", "not", "true", "for", "feed", "-", "forward", "networks", ".", "For", "instance", ",", "ResNet", "architectures", "very", "often", "do", "not", "use", "dropout", "at", "all", "(", "probably", "because", "batch", "normalization", "is", "often", "better", "to", "use", ")", ".", "It", "can", "be", "seen", "in", "the", "paper", "(", "Subsection", "[", "reference", "]", ",", "semisupervised", "CIFAR", "-", "10", "task", ")", "that", "when", "unlabeled", "data", "is", "not", "used", "the", "regular", "dropout", "hurts", "performance", "and", "using", "fraternal", "dropout", "seems", "to", "improve", "just", "a", "little", ".", "On", "a", "final", "note", ",", "the", "Monte", "Carlo", "sampling", "(", "a", "well", "known", "method", "that", "adresses", "the", "gap", "betweem", "the", "train", "and", "evaluation", "mode", "of", "dropout", ")", "can", "not", "be", "easily", "applied", "for", "RNNs", "and", "fraternal", "dropout", "may", "be", "seen", "as", "an", "alternative", ".", "To", "conclude", ",", "we", "believe", "that", "when", "the", "use", "of", "dropout", "benefits", "in", "a", "given", "architecture", ",", "applying", "fraternal", "dropout", "should", "improve", "performance", "even", "more", ".", "As", "mentioned", "before", ",", "in", "image", "recognition", "tasks", ",", "one", "may", "experiment", "with", "something", "what", "we", "would", "temporarily", "dub", "fraternal", "augmentation", "(", "even", "though", "dropout", "is", "not", "used", ",", "one", "can", "use", "random", "data", "augmentation", "such", "as", "random", "crop", "or", "random", "flip", ")", ".", "Hence", ",", "one", "may", "force", "a", "given", "neural", "network", "to", "have", "the", "same", "predictions", "for", "different", "augmentations", ".", "subsection", ":", "Proofs", "theorem", ":", ".", "Let", "sit", "and", "sjt", "be", "i.i.d", ".", "dropout", "masks", "and", "∈⁢pt", "(", "zt", ",", "sit;θ", ")", "Rm", "be", "the", "prediction", "function", "as", "described", "above", ".", "Then", ",", "proof", ":", "Proof", ".", "For", "simplicity", "of", "notation", ",", "we", "omit", "the", "time", "index", ".", "∎", "theorem", ":", ".", ".", "proof", ":", "Proof", ".", "Let", ",", "then", "Then", "using", "Jensen", "’s", "inequality", ",", "∎"], "entity": [[4, 7, ["Recurrent", "neural", "networks"], "Method"], [8, 9, ["RNNs"], "Method"], [17, 19, ["neural", "networks"], "Method"], [21, 23, ["language", "modeling"], "Task"], [24, 26, ["sequential", "prediction"], "Task"], [29, 31, ["optimizing", "RNNs"], "Task"], [38, 43, ["feed", "-", "forward", "neural", "networks"], "Method"], [67, 69, ["fraternal", "dropout"], "Method"], [73, 74, ["dropout"], "Method"], [90, 91, ["RNN"], "Method"], [117, 118, ["regularization"], "Method"], [122, 123, ["RNNs"], "Method"], [138, 140, ["regularization", "term"], "Method"], [145, 150, ["expectation", "-", "linear", "dropout", "objective"], "Method"], [164, 170, ["train", "and", "inference", "phases", "of", "dropout"], "Method"], [186, 189, ["sequence", "modeling", "tasks"], "Task"], [194, 196, ["Penn", "Treebank"], "Material"], [197, 200, ["Wikitext", "-", "2"], "Material"], [216, 218, ["image", "captioning"], "Task"], [223, 232, ["semi", "-", "supervised", "(", "CIFAR", "-", "10", ")", "tasks"], "Task"], [236, 239, ["Recurrent", "neural", "networks"], "Method"], [240, 241, ["RNNs"], "Method"], [243, 253, ["long", "short", "-", "term", "memory", "(", "LSTM", ";", ")", "networks"], "Method"], [254, 257, ["gated", "recurrent", "unit"], "Method"], [258, 259, ["GRU"], "Method"], [265, 268, ["sequence", "modeling", "tasks"], "Task"], [269, 271, ["language", "generation"], "Task"], [272, 273, ["translation"], "Task"], [274, 276, ["speech", "synthesis"], "Task"], [278, 280, ["machine", "comprehension"], "Task"], [290, 294, ["feed", "-", "forward", "networks"], "Method"], [331, 333, ["optimization", "challenges"], "Task"], [334, 335, ["RNNs"], "Task"], [339, 341, ["batch", "normalization"], "Method"], [345, 347, ["layer", "normalization"], "Method"], [348, 351, ["recurrent", "batch", "normalization"], "Method"], [352, 355, ["recurrent", "normalization", "propagation"], "Method"], [365, 369, ["feed", "-", "forward", "networks"], "Method"], [380, 384, ["naive", "application", "of", "dropout"], "Method"], [391, 392, ["RNNs"], "Task"], [395, 397, ["regularization", "techniques"], "Method"], [398, 399, ["RNNs"], "Method"], [414, 416, ["dropout", "only"], "Method"], [423, 427, ["multi", "-", "layer", "RNNs"], "Method"], [428, 430, ["Variational", "dropout"], "Method"], [443, 444, ["DropConnect"], "Method"], [446, 448, ["dropout", "operation"], "Method"], [453, 455, ["Zoneout", "("], "Method"], [462, 463, ["dropout"], "Method"], [486, 488, ["batch", "normalization"], "Method"], [489, 491, ["layer", "normalization"], "Method"], [507, 510, ["Recurrent", "batch", "normalization"], "Method"], [511, 513, ["batch", "normalization"], "Method"], [532, 534, ["activity", "regularization"], "Method"], [535, 536, ["AR"], "Method"], [538, 541, ["temporal", "activation", "regularization"], "Method"], [542, 543, ["TAR"], "Method"], [549, 551, ["regularizing", "LSTMs"], "Method"], [557, 559, ["regularizing", "RNNs"], "Method"], [584, 586, ["auxiliary", "network"], "Method"], [594, 595, ["regularization"], "Method"], [597, 598, ["dropout"], "Method"], [601, 603, ["fraternal", "dropout"], "Method"], [621, 622, ["LSTM"], "Method"], [652, 654, ["regularization", "objective"], "Metric"], [685, 686, ["regularization"], "Method"], [689, 692, ["expectation", "linear", "dropout"], "Method"], [693, 698, ["-", "model", "and", "activity", "regularization"], "Method"], [733, 734, ["Dropout"], "Method"], [737, 738, ["regularization"], "Method"], [739, 741, ["neural", "networks"], "Method"], [770, 771, ["dropout"], "Method"], [774, 775, ["regularization"], "Method"], [776, 777, ["RNNs"], "Task"], [780, 781, ["dropout"], "Method"], [788, 790, ["inference", "phase"], "Task"], [856, 858, ["fraternal", "dropout"], "Method"], [862, 865, ["neural", "network", "model"], "Method"], [891, 893, ["RNN", "model"], "Method"], [963, 965, ["fraternal", "dropout"], "Method"], [980, 981, ["RNN"], "Method"], [1013, 1016, ["overall", "loss", "function"], "Metric"], [1017, 1019, ["fraternal", "dropout"], "Method"], [1036, 1039, ["fraternal", "dropout", "regularization"], "Method"], [1044, 1047, ["Monte", "Carlo", "sampling"], "Method"], [1125, 1127, ["prediction", "function"], "Method"], [1155, 1159, ["data", "/", "model", "augmentations"], "Method"], [1176, 1177, ["RNNs"], "Method"], [1186, 1191, ["Expectation", "Linear", "Dropout", "(", "ELD"], "Method"], [1196, 1198, ["expected", "error"], "Metric"], [1268, 1269, ["regularization"], "Method"], [1298, 1300, ["network", "loss"], "Metric"], [1302, 1303, ["dropout"], "Method"], [1323, 1324, ["dropout"], "Method"], [1360, 1362, ["regularization", "objective"], "Metric"], [1367, 1372, ["expectation", "-", "linear", "dropout", "regularization"], "Method"], [1392, 1394, ["ELD", "objective"], "Metric"], [1415, 1416, ["dropout"], "Method"], [1445, 1446, ["dropout"], "Method"], [1475, 1477, ["convergence", "gain"], "Metric"], [1480, 1481, ["convergence"], "Metric"], [1485, 1486, ["regularization"], "Method"], [1498, 1501, ["back", "-", "propagation"], "Method"], [1510, 1512, ["weight", "dropout"], "Task"], [1543, 1545, ["classification", "tasks"], "Task"], [1547, 1551, ["semi", "-", "supervised", "setting"], "Task"], [1563, 1568, ["deep", "feed", "-", "forward", "version"], "Method"], [1575, 1577, ["target", "loss"], "Method"], [1605, 1607, ["unlabeled", "data"], "Material"], [1637, 1639, ["supervised", "setting"], "Task"], [1650, 1651, ["regularization"], "Method"], [1664, 1665, ["regularizer"], "Method"], [1698, 1699, ["regularizer"], "Method"], [1700, 1703, ["expectation", "linear", "dropout"], "Method"], [1721, 1724, ["target", "based", "loss"], "Metric"], [1751, 1752, ["convergence"], "Metric"], [1760, 1762, ["temporal", "embedding"], "Method"], [1778, 1783, ["semi", "-", "supervised", ",", "learning"], "Task"], [1787, 1791, ["natural", "language", "processing", "applications"], "Task"], [1828, 1830, ["supervised", "case"], "Task"], [1925, 1929, ["semi", "-", "supervised", "works"], "Task"], [1934, 1938, ["semi", "-", "supervised", "learning"], "Task"], [1981, 1987, ["Multi", "-", "Prediction", "Deep", "Boltzmann", "Machines"], "Method"], [2007, 2011, ["train", "and", "evaluation", "mode"], "Metric"], [2012, 2013, ["dropout"], "Method"], [2016, 2021, ["Monte", "Carlo", "sampling", "of", "masks"], "Method"], [2026, 2027, ["evaluation"], "Task"], [2034, 2038, ["feed", "-", "forward", "networks"], "Method"], [2049, 2050, ["RNNs"], "Method"], [2068, 2070, ["Language", "Models"], "Method"], [2074, 2076, ["language", "modeling"], "Task"], [2085, 2093, ["Penn", "Tree", "-", "bank", "(", "PTB", ")", "dataset"], "Material"], [2094, 2101, ["WikiText", "-", "2", "(", "WT2", ")", "dataset"], "Material"], [2104, 2105, ["preprocessing"], "Method"], [2110, 2112, ["PTB", "corpus"], "Material"], [2114, 2116, ["Moses", "tokenizer"], "Material"], [2119, 2121, ["WT2", "dataset"], "Material"], [2129, 2136, ["AWD", "-", "LSTM", "3", "-", "layer", "architecture"], "Method"], [2142, 2144, ["baseline", "model"], "Method"], [2154, 2155, ["PTB"], "Method"], [2167, 2168, ["WT2"], "Method"], [2169, 2170, ["WT2"], "Method"], [2197, 2199, ["fraternal", "dropout"], "Method"], [2204, 2205, ["regularization"], "Method"], [2217, 2218, ["PTB"], "Material"], [2231, 2233, ["fraternal", "dropout"], "Method"], [2241, 2246, ["hyper", "-", "parameter", "grid", "search"], "Method"], [2293, 2295, ["dropout", "layers"], "Method"], [2313, 2314, ["LSTM"], "Method"], [2316, 2318, ["embedding", "dropout"], "Method"], [2359, 2361, ["grid", "search"], "Method"], [2369, 2373, ["AR", "and", "TAR", "regularizers"], "Method"], [2375, 2377, ["fraternal", "dropout"], "Method"], [2422, 2434, ["non", "-", "monotonically", "triggered", "averaged", "SGD", "(", "NT", "-", "ASGD", ")", "optimizer"], "Method"], [2438, 2440, ["grid", "search"], "Method"], [2471, 2473, ["SGD", "optimizer"], "Method"], [2486, 2488, ["perplexity", "metric"], "Metric"], [2550, 2552, ["baseline", "model"], "Method"], [2557, 2560, ["fine", "-", "tuning"], "Method"], [2562, 2564, ["PTB", "dataset"], "Material"], [2570, 2574, ["average", "best", "validation", "perplexity"], "Metric"], [2584, 2586, ["test", "perplexity"], "Task"], [2594, 2598, ["validation", "and", "test", "perplexity"], "Metric"], [2610, 2612, ["fraternal", "dropout"], "Method"], [2614, 2616, ["grid", "search"], "Method"], [2647, 2652, ["Word", "level", "WikiText", "-", "2"], "Method"], [2653, 2654, ["WT2"], "Method"], [2660, 2666, ["WikiText", "-", "2", "language", "modeling", "task"], "Task"], [2680, 2682, ["perplexity", "metric"], "Metric"], [2699, 2701, ["training", "procedure"], "Method"], [2702, 2704, ["fraternal", "dropout"], "Task"], [2705, 2707, ["WT2", "dataset"], "Material"], [2727, 2729, ["PTB", "dataset"], "Material"], [2754, 2755, ["WT2"], "Method"], [2770, 2772, ["Image", "captioning"], "Task"], [2775, 2777, ["fraternal", "dropout"], "Method"], [2779, 2782, ["image", "captioning", "task"], "Task"], [2789, 2793, ["show", "and", "tell", "model"], "Method"], [2802, 2805, ["image", "captioning", "task"], "Task"], [2807, 2813, ["image", "encoder", "and", "sentence", "decoder", "architectures"], "Method"], [2828, 2830, ["fraternal", "dropout"], "Method"], [2831, 2832, ["RNNs"], "Method"], [2834, 2840, ["frozen", "pretrained", "ResNet", "-", "101", "model"], "Method"], [2842, 2844, ["image", "encoder"], "Method"], [2906, 2909, ["image", "captioning", "encoder"], "Method"], [2927, 2932, ["unconditioned", "natural", "language", "processing", "tasks"], "Task"], [2933, 2935, ["Fraternal", "dropout"], "Method"], [2955, 2957, ["Ablation", "Studies"], "Task"], [2973, 2976, ["expectation", "linear", "dropout"], "Method"], [2977, 2982, ["-", "model", "and", "activity", "regularization"], "Method"], [2988, 2990, ["ablation", "studies"], "Task"], [2994, 2997, ["single", "layer", "LSTM"], "Method"], [3005, 3007, ["model", "architecture"], "Method"], [3011, 3017, ["Expectation", "-", "linear", "dropout", "(", "ELD"], "Method"], [3021, 3025, ["expectation", "-", "linear", "dropout"], "Method"], [3047, 3049, ["ELD", "regularization"], "Method"], [3051, 3055, ["regularization", "(", "FD", ")"], "Method"], [3059, 3060, ["ELD"], "Method"], [3067, 3068, ["ELDM"], "Method"], [3070, 3071, ["ELD"], "Method"], [3079, 3080, ["LSTMs"], "Method"], [3084, 3085, ["FD"], "Method"], [3096, 3097, ["LSTM"], "Method"], [3104, 3106, ["baseline", "model"], "Method"], [3125, 3126, ["regularization"], "Method"], [3131, 3132, ["convergence"], "Metric"], [3140, 3141, ["generalization"], "Task"], [3145, 3146, ["FD"], "Method"], [3149, 3150, ["ELD"], "Method"], [3152, 3153, ["baseline"], "Method"], [3154, 3155, ["ELDM"], "Method"], [3164, 3168, ["train", "and", "validation", "curves"], "Metric"], [3170, 3171, ["ELDM"], "Method"], [3176, 3178, ["optimization", "problems"], "Task"], [3198, 3207, ["semi", "-", "supervised", "learning", "in", "feed", "-", "forward", "networks"], "Task"], [3235, 3245, ["single", "layer", "LSTM", "and", "3", "-", "layer", "AWD", "-", "LSTM"], "Method"], [3246, 3248, ["PTB", "task"], "Task"], [3260, 3262, ["language", "modeling"], "Task"], [3336, 3337, ["RNNs"], "Task"], [3351, 3355, ["semi", "-", "supervised", "task"], "Task"], [3365, 3369, ["CIFAR", "-", "10", "dataset"], "Material"], [3383, 3388, ["semi", "-", "supervised", "learning", "literature"], "Task"], [3421, 3425, ["ResNet", "-", "56", "architecture"], "Method"], [3428, 3430, ["grid", "search"], "Method"], [3452, 3454, ["unlabeled", "data"], "Material"], [3478, 3480, ["unlabeled", "data"], "Material"], [3484, 3486, ["fraternal", "dropout"], "Method"], [3494, 3495, ["dropout"], "Method"], [3498, 3500, ["Activity", "regularization"], "Method"], [3501, 3505, ["temporal", "activity", "regularization", "analysis"], "Method"], [3512, 3514, ["activity", "regularization"], "Method"], [3515, 3516, ["AR"], "Method"], [3518, 3521, ["temporal", "activity", "regularization"], "Method"], [3522, 3523, ["TAR"], "Method"], [3525, 3526, ["LSTMs"], "Method"], [3554, 3558, ["AR", "and", "TAR", "regularizations"], "Method"], [3565, 3566, ["LSTM"], "Method"], [3580, 3581, ["LSTM"], "Method"], [3586, 3587, ["regularization"], "Method"], [3618, 3619, ["regularization"], "Method"], [3627, 3629, ["TAR", "objective"], "Metric"], [3633, 3635, ["grid", "search"], "Method"], [3649, 3650, ["regularization"], "Task"], [3661, 3662, ["regularization"], "Method"], [3663, 3664, ["PR"], "Method"], [3676, 3677, ["regularization"], "Method"], [3681, 3683, ["grid", "search"], "Method"], [3696, 3697, ["regularizations"], "Method"], [3713, 3715, ["learning", "dynamics"], "Method"], [3724, 3725, ["regularization"], "Method"], [3731, 3732, ["convergence"], "Metric"], [3733, 3734, ["generalization"], "Task"], [3749, 3751, ["regularizer", "described"], "Method"], [3765, 3768, ["fine", "-", "tuning"], "Method"], [3774, 3776, ["fraternal", "dropout"], "Method"], [3779, 3786, ["NT", "-", "ASGD", "fine", "-", "tuning", "step"], "Method"], [3818, 3820, ["learning", "procedure"], "Method"], [3831, 3833, ["grid", "search"], "Method"], [3845, 3849, ["fine", "-", "tuning", "procedure"], "Method"], [3853, 3855, ["official", "repository"], "Material"], [3880, 3882, ["Fraternal", "dropout"], "Method"], [3883, 3887, ["expectation", "linear", "dropout", "comparison"], "Method"], [3890, 3892, ["grid", "search"], "Task"], [3903, 3910, ["AWD", "-", "LSTM", "3", "-", "layer", "architecture"], "Method"], [3914, 3916, ["fraternal", "dropout"], "Method"], [3917, 3921, ["expectation", "linear", "dropout", "regularizations"], "Method"], [3937, 3940, ["fine", "-", "tuning"], "Method"], [3942, 3944, ["PTB", "dataset"], "Material"], [3951, 3953, ["dropout", "rates"], "Metric"], [4008, 4010, ["TAR", "regularizers"], "Method"], [4032, 4033, ["FD"], "Method"], [4034, 4035, ["ELD"], "Method"], [4044, 4048, ["AR", "and", "TAR", "regularizers"], "Method"], [4065, 4066, ["FD"], "Method"], [4067, 4068, ["ELD"], "Method"], [4088, 4089, ["FD"], "Method"], [4094, 4095, ["ELD"], "Method"], [4098, 4100, ["fraternal", "dropout"], "Method"], [4136, 4138, ["regularization", "method"], "Method"], [4139, 4140, ["RNNs"], "Method"], [4141, 4143, ["fraternal", "dropout"], "Method"], [4147, 4148, ["regularization"], "Method"], [4176, 4179, ["language", "modeling", "tasks"], "Task"], [4182, 4183, ["convergence"], "Metric"], [4192, 4193, ["regularization"], "Method"], [4194, 4197, ["expectation", "linear", "dropout"], "Method"], [4246, 4247, ["NSERC"], "Material"], [4248, 4249, ["CIFAR"], "Material"], [4251, 4252, ["IVADO"], "Material"], [4296, 4299, ["Monte", "Carlo", "evaluation"], "Task"], [4309, 4315, ["train", "and", "evaluation", "mode", "of", "dropout"], "Metric"], [4318, 4323, ["Monte", "Carlo", "sampling", "of", "masks"], "Method"], [4328, 4329, ["evaluation"], "Task"], [4330, 4333, ["MC", "-", "eval"], "Method"], [4341, 4345, ["feed", "-", "forward", "networks"], "Method"], [4370, 4373, ["MC", "-", "eval"], "Method"], [4374, 4375, ["RNNs"], "Method"], [4377, 4381, ["feed", "-", "forward", "networks"], "Method"], [4397, 4398, ["RNNs"], "Method"], [4400, 4403, ["next", "step", "predictions"], "Task"], [4429, 4431, ["Online", "averaging"], "Task"], [4448, 4450, ["prediction", "score"], "Metric"], [4498, 4499, ["RNNs"], "Method"], [4594, 4596, ["Sequence", "averaging"], "Method"], [4625, 4626, ["argmax"], "Method"], [4754, 4757, ["averaged", "prediction", "score"], "Metric"], [4759, 4760, ["argmax"], "Method"], [4783, 4785, ["output", "predictions"], "Task"], [4847, 4849, ["PTB", "dataset"], "Material"], [4850, 4853, ["MC", "-", "eval"], "Method"], [4873, 4875, ["fraternal", "dropout"], "Method"], [4881, 4888, ["AWD", "-", "LSTM", "3", "-", "layer", "baseline"], "Method"], [4896, 4897, ["MC1"], "Method"], [4899, 4901, ["MC1", "model"], "Method"], [4905, 4907, ["fraternal", "dropout"], "Method"], [4916, 4918, ["MC1", "model"], "Method"], [4967, 4968, ["MC50"], "Method"], [4985, 4986, ["MC10"], "Method"], [4990, 4993, ["fine", "-", "tuning"], "Method"], [5006, 5007, ["RNNs"], "Task"], [5008, 5011, ["fraternal", "dropout", "method"], "Method"], [5018, 5022, ["feed", "-", "forward", "architectures"], "Method"], [5031, 5036, ["CIFAR", "-", "10", "semisupervised", "example"], "Material"], [5051, 5052, ["RNNs"], "Method"], [5056, 5057, ["prediction"], "Task"], [5062, 5063, ["RNNs"], "Method"], [5084, 5088, ["feed", "-", "forward", "networks"], "Method"], [5166, 5168, ["RNN", "architectures"], "Method"], [5182, 5183, ["dropout"], "Method"], [5184, 5186, ["embedding", "dropout"], "Method"], [5200, 5202, ["dropout", "rates"], "Metric"], [5209, 5212, ["input", "word", "embedding"], "Task"], [5213, 5215, ["NLP", "tasks"], "Task"], [5224, 5228, ["feed", "-", "forward", "networks"], "Method"], [5232, 5234, ["ResNet", "architectures"], "Method"], [5239, 5240, ["dropout"], "Method"], [5245, 5247, ["batch", "normalization"], "Method"], [5267, 5272, ["semisupervised", "CIFAR", "-", "10", "task"], "Task"], [5275, 5277, ["unlabeled", "data"], "Material"], [5281, 5283, ["regular", "dropout"], "Method"], [5287, 5289, ["fraternal", "dropout"], "Method"], [5302, 5305, ["Monte", "Carlo", "sampling"], "Method"], [5316, 5322, ["train", "and", "evaluation", "mode", "of", "dropout"], "Method"], [5329, 5330, ["RNNs"], "Method"], [5331, 5333, ["fraternal", "dropout"], "Method"], [5358, 5360, ["fraternal", "dropout"], "Method"], [5371, 5374, ["image", "recognition", "tasks"], "Task"], [5385, 5387, ["fraternal", "augmentation"], "Task"], [5390, 5391, ["dropout"], "Method"], [5398, 5401, ["random", "data", "augmentation"], "Method"], [5417, 5419, ["neural", "network"], "Method"], [5453, 5455, ["prediction", "function"], "Method"], [2214, 2216, ["Penn", "Treebank"], "Material"]]} {"doc_id": "178631e0f0e624b1607c7a7a2507ed30d4e83a42", "text": ["document", ":", "Speech", "Recognition", "with", "Deep", "Recurrent", "Neural", "Networks", "Recurrent", "neural", "networks", "(", "RNNs", ")", "are", "a", "powerful", "model", "for", "sequential", "data", ".", "End", "-", "to", "-", "end", "training", "methods", "such", "as", "Connectionist", "Temporal", "Classification", "make", "it", "possible", "to", "train", "RNNs", "for", "sequence", "labelling", "problems", "where", "the", "input", "-", "output", "alignment", "is", "unknown", ".", "The", "combination", "of", "these", "methods", "with", "the", "Long", "Short", "-", "term", "Memory", "RNN", "architecture", "has", "proved", "particularly", "fruitful", ",", "delivering", "state", "-", "of", "-", "the", "-", "art", "results", "in", "cursive", "handwriting", "recognition", ".", "However", "RNN", "performance", "in", "speech", "recognition", "has", "so", "far", "been", "disappointing", ",", "with", "better", "results", "returned", "by", "deep", "feedforward", "networks", ".", "This", "paper", "investigates", "deep", "recurrent", "neural", "networks", ",", "which", "combine", "the", "multiple", "levels", "of", "representation", "that", "have", "proved", "so", "effective", "in", "deep", "networks", "with", "the", "flexible", "use", "of", "long", "range", "context", "that", "empowers", "RNNs", ".", "When", "trained", "end", "-", "to", "-", "end", "with", "suitable", "regularisation", ",", "we", "find", "that", "deep", "Long", "Short", "-", "term", "Memory", "RNNs", "achieve", "a", "test", "set", "error", "of", "17.7", "%", "on", "the", "TIMIT", "phoneme", "recognition", "benchmark", ",", "which", "to", "our", "knowledge", "is", "the", "best", "recorded", "score", ".", "AlexGraves", ",", "Abdel", "-", "rahmanMohamedandGeoffreyHinton", "DepartmentofComputerScience", ",", "UniversityofToronto", "recurrent", "neural", "networks", ",", "deep", "neural", "networks", ",", "speech", "recognition", "section", ":", "Introduction", "Neural", "networks", "have", "a", "long", "history", "in", "speech", "recognition", ",", "usually", "in", "combination", "with", "hidden", "Markov", "models", ".", "They", "have", "gained", "attention", "in", "recent", "years", "with", "the", "dramatic", "improvements", "in", "acoustic", "modelling", "yielded", "by", "deep", "feedforward", "networks", ".", "Given", "that", "speech", "is", "an", "inherently", "dynamic", "process", ",", "it", "seems", "natural", "to", "consider", "recurrent", "neural", "networks", "(", "RNNs", ")", "as", "an", "alternative", "model", ".", "HMM", "-", "RNN", "systems", "have", "also", "seen", "a", "recent", "revival", ",", "but", "do", "not", "currently", "perform", "as", "well", "as", "deep", "networks", ".", "Instead", "of", "combining", "RNNs", "with", "HMMs", ",", "it", "is", "possible", "to", "train", "RNNs", "‘", "end", "-", "to", "-", "end", "’", "for", "speech", "recognition", ".", "This", "approach", "exploits", "the", "larger", "state", "-", "space", "and", "richer", "dynamics", "of", "RNNs", "compared", "to", "HMMs", ",", "and", "avoids", "the", "problem", "of", "using", "potentially", "incorrect", "alignments", "as", "training", "targets", ".", "The", "combination", "of", "Long", "Short", "-", "term", "Memory", ",", "an", "RNN", "architecture", "with", "an", "improved", "memory", ",", "with", "end", "-", "to", "-", "end", "training", "has", "proved", "especially", "effective", "for", "cursive", "handwriting", "recognition", ".", "However", "it", "has", "so", "far", "made", "little", "impact", "on", "speech", "recognition", ".", "RNNs", "are", "inherently", "deep", "in", "time", ",", "since", "their", "hidden", "state", "is", "a", "function", "of", "all", "previous", "hidden", "states", ".", "The", "question", "that", "inspired", "this", "paper", "was", "whether", "RNNs", "could", "also", "benefit", "from", "depth", "in", "space", ";", "that", "is", "from", "stacking", "multiple", "recurrent", "hidden", "layers", "on", "top", "of", "each", "other", ",", "just", "as", "feedforward", "layers", "are", "stacked", "in", "conventional", "deep", "networks", ".", "To", "answer", "this", "question", "we", "introduce", "deep", "Long", "Short", "-", "term", "Memory", "RNNs", "and", "assess", "their", "potential", "for", "speech", "recognition", ".", "We", "also", "present", "an", "enhancement", "to", "a", "recently", "introduced", "end", "-", "to", "-", "end", "learning", "method", "that", "jointly", "trains", "two", "separate", "RNNs", "as", "acoustic", "and", "linguistic", "models", ".", "Sections", "[", "reference", "]", "and", "[", "reference", "]", "describe", "the", "network", "architectures", "and", "training", "methods", ",", "sec", ":", "experiments", "provides", "experimental", "results", "and", "concluding", "remarks", "are", "given", "in", "sec", ":", "conclusion", ".", "section", ":", "Recurrent", "Neural", "Networks", "Given", "an", "input", "sequence", ",", "a", "standard", "recurrent", "neural", "network", "(", "RNN", ")", "computes", "the", "hidden", "vector", "sequence", "and", "output", "vector", "sequence", "by", "iterating", "the", "following", "equations", "from", "to", ":", "where", "the", "terms", "denote", "weight", "matrices", "(", "e.g.", "is", "the", "input", "-", "hidden", "weight", "matrix", ")", ",", "the", "terms", "denote", "bias", "vectors", "(", "e.g.", "is", "hidden", "bias", "vector", ")", "and", "is", "the", "hidden", "layer", "function", ".", "is", "usually", "an", "elementwise", "application", "of", "a", "sigmoid", "function", ".", "However", "we", "have", "found", "that", "the", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "architecture", ",", "which", "uses", "purpose", "-", "built", "memory", "cells", "to", "store", "information", ",", "is", "better", "at", "finding", "and", "exploiting", "long", "range", "context", ".", "fig", ":", "lstm", "illustrates", "a", "single", "LSTM", "memory", "cell", ".", "For", "the", "version", "of", "LSTM", "used", "in", "this", "paper", "is", "implemented", "by", "the", "following", "composite", "function", ":", "where", "is", "the", "logistic", "sigmoid", "function", ",", "and", ",", ",", "and", "are", "respectively", "the", "input", "gate", ",", "forget", "gate", ",", "output", "gate", "and", "cell", "activation", "vectors", ",", "all", "of", "which", "are", "the", "same", "size", "as", "the", "hidden", "vector", ".", "The", "weight", "matrices", "from", "the", "cell", "to", "gate", "vectors", "(", "e.g.", ")", "are", "diagonal", ",", "so", "element", "in", "each", "gate", "vector", "only", "receives", "input", "from", "element", "of", "the", "cell", "vector", ".", "One", "shortcoming", "of", "conventional", "RNNs", "is", "that", "they", "are", "only", "able", "to", "make", "use", "of", "previous", "context", ".", "In", "speech", "recognition", ",", "where", "whole", "utterances", "are", "transcribed", "at", "once", ",", "there", "is", "no", "reason", "not", "to", "exploit", "future", "context", "as", "well", ".", "Bidirectional", "RNNs", "(", "BRNNs", ")", "do", "this", "by", "processing", "the", "data", "in", "both", "directions", "with", "two", "separate", "hidden", "layers", ",", "which", "are", "then", "fed", "forwards", "to", "the", "same", "output", "layer", ".", "As", "illustrated", "in", "fig", ":", "brnn", ",", "a", "BRNN", "computes", "the", "forward", "hidden", "sequence", ",", "the", "backward", "hidden", "sequence", "and", "the", "output", "sequence", "by", "iterating", "the", "backward", "layer", "from", "to", ",", "the", "forward", "layer", "from", "to", "and", "then", "updating", "the", "output", "layer", ":", "Combing", "BRNNs", "with", "LSTM", "gives", "bidirectional", "LSTM", ",", "which", "can", "access", "long", "-", "range", "context", "in", "both", "input", "directions", ".", "A", "crucial", "element", "of", "the", "recent", "success", "of", "hybrid", "HMM", "-", "neural", "network", "systems", "is", "the", "use", "of", "deep", "architectures", ",", "which", "are", "able", "to", "build", "up", "progressively", "higher", "level", "representations", "of", "acoustic", "data", ".", "Deep", "RNNs", "can", "be", "created", "by", "stacking", "multiple", "RNN", "hidden", "layers", "on", "top", "of", "each", "other", ",", "with", "the", "output", "sequence", "of", "one", "layer", "forming", "the", "input", "sequence", "for", "the", "next", ".", "Assuming", "the", "same", "hidden", "layer", "function", "is", "used", "for", "all", "layers", "in", "the", "stack", ",", "the", "hidden", "vector", "sequences", "are", "iteratively", "computed", "from", "to", "and", "to", ":", "where", "we", "define", ".", "The", "network", "outputs", "are", "Deep", "bidirectional", "RNNs", "can", "be", "implemented", "by", "replacing", "each", "hidden", "sequence", "with", "the", "forward", "and", "backward", "sequences", "and", ",", "and", "ensuring", "that", "every", "hidden", "layer", "receives", "input", "from", "both", "the", "forward", "and", "backward", "layers", "at", "the", "level", "below", ".", "If", "LSTM", "is", "used", "for", "the", "hidden", "layers", "we", "get", "deep", "bidirectional", "LSTM", ",", "the", "main", "architecture", "used", "in", "this", "paper", ".", "As", "far", "as", "we", "are", "aware", "this", "is", "the", "first", "time", "deep", "LSTM", "has", "been", "applied", "to", "speech", "recognition", ",", "and", "we", "find", "that", "it", "yields", "a", "dramatic", "improvement", "over", "single", "-", "layer", "LSTM", ".", "section", ":", "Network", "Training", "We", "focus", "on", "end", "-", "to", "-", "end", "training", ",", "where", "RNNs", "learn", "to", "map", "directly", "from", "acoustic", "to", "phonetic", "sequences", ".", "One", "advantage", "of", "this", "approach", "is", "that", "it", "removes", "the", "need", "for", "a", "predefined", "(", "and", "error", "-", "prone", ")", "alignment", "to", "create", "the", "training", "targets", ".", "The", "first", "step", "is", "to", "to", "use", "the", "network", "outputs", "to", "parameterise", "a", "differentiable", "distribution", "over", "all", "possible", "phonetic", "output", "sequences", "given", "an", "acoustic", "input", "sequence", ".", "The", "log", "-", "probability", "of", "the", "target", "output", "sequence", "can", "then", "be", "differentiated", "with", "respect", "to", "the", "network", "weights", "using", "backpropagation", "through", "time", ",", "and", "the", "whole", "system", "can", "be", "optimised", "with", "gradient", "descent", ".", "We", "now", "describe", "two", "ways", "to", "define", "the", "output", "distribution", "and", "hence", "train", "the", "network", ".", "We", "refer", "throughout", "to", "the", "length", "of", "as", ",", "the", "length", "of", "as", ",", "and", "the", "number", "of", "possible", "phonemes", "as", ".", "subsection", ":", "Connectionist", "Temporal", "Classification", "The", "first", "method", ",", "known", "as", "Connectionist", "Temporal", "Classification", "(", "CTC", ")", ",", "uses", "a", "softmax", "layer", "to", "define", "a", "separate", "output", "distribution", "at", "every", "step", "along", "the", "input", "sequence", ".", "This", "distribution", "covers", "the", "phonemes", "plus", "an", "extra", "blank", "symbol", "which", "represents", "a", "non", "-", "output", "(", "the", "softmax", "layer", "is", "therefore", "size", ")", ".", "Intuitively", "the", "network", "decides", "whether", "to", "emit", "any", "label", ",", "or", "no", "label", ",", "at", "every", "timestep", ".", "Taken", "together", "these", "decisions", "define", "a", "distribution", "over", "alignments", "between", "the", "input", "and", "target", "sequences", ".", "CTC", "then", "uses", "a", "forward", "-", "backward", "algorithm", "to", "sum", "over", "all", "possible", "alignments", "and", "determine", "the", "normalised", "probability", "of", "the", "target", "sequence", "given", "the", "input", "sequence", ".", "Similar", "procedures", "have", "been", "used", "elsewhere", "in", "speech", "and", "handwriting", "recognition", "to", "integrate", "out", "over", "possible", "segmentations", ";", "however", "CTC", "differs", "in", "that", "it", "ignores", "segmentation", "altogether", "and", "sums", "over", "single", "-", "timestep", "label", "decisions", "instead", ".", "RNNs", "trained", "with", "CTC", "are", "generally", "bidirectional", ",", "to", "ensure", "that", "every", "depends", "on", "the", "entire", "input", "sequence", ",", "and", "not", "just", "the", "inputs", "up", "to", ".", "In", "this", "work", "we", "focus", "on", "deep", "bidirectional", "networks", ",", "with", "defined", "as", "follows", ":", "where", "is", "the", "element", "of", "the", "length", "unnormalised", "output", "vector", ",", "and", "is", "the", "number", "of", "bidirectional", "levels", ".", "subsection", ":", "RNN", "Transducer", "CTC", "defines", "a", "distribution", "over", "phoneme", "sequences", "that", "depends", "only", "on", "the", "acoustic", "input", "sequence", ".", "It", "is", "therefore", "an", "acoustic", "-", "only", "model", ".", "A", "recent", "augmentation", ",", "known", "as", "an", "RNN", "transducer", "combines", "a", "CTC", "-", "like", "network", "with", "a", "separate", "RNN", "that", "predicts", "each", "phoneme", "given", "the", "previous", "ones", ",", "thereby", "yielding", "a", "jointly", "trained", "acoustic", "and", "language", "model", ".", "Joint", "LM", "-", "acoustic", "training", "has", "proved", "beneficial", "in", "the", "past", "for", "speech", "recognition", ".", "Whereas", "CTC", "determines", "an", "output", "distribution", "at", "every", "input", "timestep", ",", "an", "RNN", "transducer", "determines", "a", "separate", "distribution", "for", "every", "combination", "of", "input", "timestep", "and", "output", "timestep", ".", "As", "with", "CTC", ",", "each", "distribution", "covers", "the", "phonemes", "plus", ".", "Intuitively", "the", "network", "‘", "decides", "’", "what", "to", "output", "depending", "both", "on", "where", "it", "is", "in", "the", "input", "sequence", "and", "the", "outputs", "it", "has", "already", "emitted", ".", "For", "a", "length", "target", "sequence", ",", "the", "complete", "set", "of", "decisions", "jointly", "determines", "a", "distribution", "over", "all", "possible", "alignments", "between", "and", ",", "which", "can", "then", "be", "integrated", "out", "with", "a", "forward", "-", "backward", "algorithm", "to", "determine", ".", "In", "the", "original", "formulation", "was", "defined", "by", "taking", "an", "‘", "acoustic", "’", "distribution", "from", "the", "CTC", "network", ",", "a", "‘", "linguistic", "’", "distribution", "from", "the", "prediction", "network", ",", "then", "multiplying", "the", "two", "together", "and", "renormalising", ".", "An", "improvement", "introduced", "in", "this", "paper", "is", "to", "instead", "feed", "the", "hidden", "activations", "of", "both", "networks", "into", "a", "separate", "feedforward", "output", "network", ",", "whose", "outputs", "are", "then", "normalised", "with", "a", "softmax", "function", "to", "yield", ".", "This", "allows", "a", "richer", "set", "of", "possibilities", "for", "combining", "linguistic", "and", "acoustic", "information", ",", "and", "appears", "to", "lead", "to", "better", "generalisation", ".", "In", "particular", "we", "have", "found", "that", "the", "number", "of", "deletion", "errors", "encountered", "during", "decoding", "is", "reduced", ".", "Denote", "by", "and", "the", "uppermost", "forward", "and", "backward", "hidden", "sequences", "of", "the", "CTC", "network", ",", "and", "by", "the", "hidden", "sequence", "of", "the", "prediction", "network", ".", "At", "each", "the", "output", "network", "is", "implemented", "by", "feeding", "and", "to", "a", "linear", "layer", "to", "generate", "the", "vector", ",", "then", "feeding", "and", "to", "a", "hidden", "layer", "to", "yield", ",", "and", "finally", "feeding", "to", "a", "size", "softmax", "layer", "to", "determine", ":", "where", "is", "the", "element", "of", "the", "length", "unnormalised", "output", "vector", ".", "For", "simplicity", "we", "constrained", "all", "non", "-", "output", "layers", "to", "be", "the", "same", "size", "(", ";", "however", "they", "could", "be", "varied", "independently", ".", "RNN", "transducers", "can", "be", "trained", "from", "random", "initial", "weights", ".", "However", "they", "appear", "to", "work", "better", "when", "initialised", "with", "the", "weights", "of", "a", "pretrained", "CTC", "network", "and", "a", "pretrained", "next", "-", "step", "prediction", "network", "(", "so", "that", "only", "the", "output", "network", "starts", "from", "random", "weights", ")", ".", "The", "output", "layers", "(", "and", "all", "associated", "weights", ")", "used", "by", "the", "networks", "during", "pretraining", "are", "removed", "during", "retraining", ".", "In", "this", "work", "we", "pretrain", "the", "prediction", "network", "on", "the", "phonetic", "transcriptions", "of", "the", "audio", "training", "data", ";", "however", "for", "large", "-", "scale", "applications", "it", "would", "make", "more", "sense", "to", "pretrain", "on", "a", "separate", "text", "corpus", ".", "subsection", ":", "Decoding", "RNN", "transducers", "can", "be", "decoded", "with", "beam", "search", "to", "yield", "an", "n", "-", "best", "list", "of", "candidate", "transcriptions", ".", "In", "the", "past", "CTC", "networks", "have", "been", "decoded", "using", "either", "a", "form", "of", "best", "-", "first", "decoding", "known", "as", "prefix", "search", ",", "or", "by", "simply", "taking", "the", "most", "active", "output", "at", "every", "timestep", ".", "In", "this", "work", "however", "we", "exploit", "the", "same", "beam", "search", "as", "the", "transducer", ",", "with", "the", "modification", "that", "the", "output", "label", "probabilities", "do", "not", "depend", "on", "the", "previous", "outputs", "(", "so", ")", ".", "We", "find", "beam", "search", "both", "faster", "and", "more", "effective", "than", "prefix", "search", "for", "CTC", ".", "Note", "the", "n", "-", "best", "list", "from", "the", "transducer", "was", "originally", "sorted", "by", "the", "length", "normalised", "log", "-", "probabilty", ";", "in", "the", "current", "work", "we", "dispense", "with", "the", "normalisation", "(", "which", "only", "helps", "when", "there", "are", "many", "more", "deletions", "than", "insertions", ")", "and", "sort", "by", ".", "subsection", ":", "Regularisation", "Regularisation", "is", "vital", "for", "good", "performance", "with", "RNNs", ",", "as", "their", "flexibility", "makes", "them", "prone", "to", "overfitting", ".", "Two", "regularisers", "were", "used", "in", "this", "paper", ":", "early", "stopping", "and", "weight", "noise", "(", "the", "addition", "of", "Gaussian", "noise", "to", "the", "network", "weights", "during", "training", ")", ".", "Weight", "noise", "was", "added", "once", "per", "training", "sequence", ",", "rather", "than", "at", "every", "timestep", ".", "Weight", "noise", "tends", "to", "‘", "simplify", "’", "neural", "networks", ",", "in", "the", "sense", "of", "reducing", "the", "amount", "of", "information", "required", "to", "transmit", "the", "parameters", ",", "which", "improves", "generalisation", ".", "section", ":", "Experiments", "Phoneme", "recognition", "experiments", "were", "performed", "on", "the", "TIMIT", "corpus", ".", "The", "standard", "462", "speaker", "set", "with", "all", "SA", "records", "removed", "was", "used", "for", "training", ",", "and", "a", "separate", "development", "set", "of", "50", "speakers", "was", "used", "for", "early", "stopping", ".", "Results", "are", "reported", "for", "the", "24", "-", "speaker", "core", "test", "set", ".", "The", "audio", "data", "was", "encoded", "using", "a", "Fourier", "-", "transform", "-", "based", "filter", "-", "bank", "with", "40", "coefficients", "(", "plus", "energy", ")", "distributed", "on", "a", "mel", "-", "scale", ",", "together", "with", "their", "first", "and", "second", "temporal", "derivatives", ".", "Each", "input", "vector", "was", "therefore", "size", "123", ".", "The", "data", "were", "normalised", "so", "that", "every", "element", "of", "the", "input", "vectors", "had", "zero", "mean", "and", "unit", "variance", "over", "the", "training", "set", ".", "All", "61", "phoneme", "labels", "were", "used", "during", "training", "and", "decoding", "(", "so", ")", ",", "then", "mapped", "to", "39", "classes", "for", "scoring", ".", "Note", "that", "all", "experiments", "were", "run", "only", "once", ",", "so", "the", "variance", "due", "to", "random", "weight", "initialisation", "and", "weight", "noise", "is", "unknown", ".", "As", "shown", "in", "tab", ":", "timit", ",", "nine", "RNNs", "were", "evaluated", ",", "varying", "along", "three", "main", "dimensions", ":", "the", "training", "method", "used", "(", "CTC", ",", "Transducer", "or", "pretrained", "Transducer", ")", ",", "the", "number", "of", "hidden", "levels", "(", "1–5", ")", ",", "and", "the", "number", "of", "LSTM", "cells", "in", "each", "hidden", "layer", ".", "Bidirectional", "LSTM", "was", "used", "for", "all", "networks", "except", "CTC", "-", "3l", "-", "500h", "-", "tanh", ",", "which", "had", "units", "instead", "of", "LSTM", "cells", ",", "and", "CTC", "-", "3l", "-", "421h", "-", "uni", "where", "the", "LSTM", "layers", "were", "unidirectional", ".", "All", "networks", "were", "trained", "using", "stochastic", "gradient", "descent", ",", "with", "learning", "rate", ",", "momentum", "and", "random", "initial", "weights", "drawn", "uniformly", "from", ".", "All", "networks", "except", "CTC", "-", "3l", "-", "500h", "-", "tanh", "and", "PreTrans", "-", "3l", "-", "250h", "were", "first", "trained", "with", "no", "noise", "and", "then", ",", "starting", "from", "the", "point", "of", "highest", "log", "-", "probability", "on", "the", "development", "set", ",", "retrained", "with", "Gaussian", "weight", "noise", "(", ")", "until", "the", "point", "of", "lowest", "phoneme", "error", "rate", "on", "the", "development", "set", ".", "PreTrans", "-", "3l", "-", "250h", "was", "initialised", "with", "the", "weights", "of", "CTC", "-", "3l", "-", "250h", ",", "along", "with", "the", "weights", "of", "a", "phoneme", "prediction", "network", "(", "which", "also", "had", "a", "hidden", "layer", "of", "250", "LSTM", "cells", ")", ",", "both", "of", "which", "were", "trained", "without", "noise", ",", "retrained", "with", "noise", ",", "and", "stopped", "at", "the", "point", "of", "highest", "log", "-", "probability", ".", "PreTrans", "-", "3l", "-", "250h", "was", "trained", "from", "this", "point", "with", "noise", "added", ".", "CTC", "-", "3l", "-", "500h", "-", "tanh", "was", "entirely", "trained", "without", "weight", "noise", "because", "it", "failed", "to", "learn", "with", "noise", "added", ".", "Beam", "search", "decoding", "was", "used", "for", "all", "networks", ",", "with", "a", "beam", "width", "of", "100", ".", "The", "advantage", "of", "deep", "networks", "is", "immediately", "obvious", ",", "with", "the", "error", "rate", "for", "CTC", "dropping", "from", "23.9", "%", "to", "18.4", "%", "as", "the", "number", "of", "hidden", "levels", "increases", "from", "one", "to", "five", ".", "The", "four", "networks", "CTC", "-", "3l", "-", "500h", "-", "tanh", ",", "CTC", "-", "1l", "-", "622h", ",", "CTC", "-", "3l", "-", "421h", "-", "uni", "and", "CTC", "-", "3l", "-", "250h", "all", "had", "approximately", "the", "same", "number", "of", "weights", ",", "but", "give", "radically", "different", "results", ".", "The", "three", "main", "conclusions", "we", "can", "draw", "from", "this", "are", "(", "a", ")", "LSTM", "works", "much", "better", "than", "for", "this", "task", ",", "(", "b", ")", "bidirectional", "LSTM", "has", "a", "slight", "advantage", "over", "unidirectional", "LSTMand", "(", "c", ")", "depth", "is", "more", "important", "than", "layer", "size", "(", "which", "supports", "previous", "findings", "for", "deep", "networks", ")", ".", "Although", "the", "advantage", "of", "the", "transducer", "is", "slight", "when", "the", "weights", "are", "randomly", "initialised", ",", "it", "becomes", "more", "substantial", "when", "pretraining", "is", "used", ".", "section", ":", "Conclusions", "and", "future", "work", "We", "have", "shown", "that", "the", "combination", "of", "deep", ",", "bidirectional", "Long", "Short", "-", "term", "Memory", "RNNs", "with", "end", "-", "to", "-", "end", "training", "and", "weight", "noise", "gives", "state", "-", "of", "-", "the", "-", "art", "results", "in", "phoneme", "recognition", "on", "the", "TIMIT", "database", ".", "An", "obvious", "next", "step", "is", "to", "extend", "the", "system", "to", "large", "vocabulary", "speech", "recognition", ".", "Another", "interesting", "direction", "would", "be", "to", "combine", "frequency", "-", "domain", "convolutional", "neural", "networks", "with", "deep", "LSTM", ".", "bibliography", ":", "References"], "entity": [[2, 4, ["Speech", "Recognition"], "Task"], [5, 9, ["Deep", "Recurrent", "Neural", "Networks"], "Method"], [9, 12, ["Recurrent", "neural", "networks"], "Method"], [13, 14, ["RNNs"], "Method"], [27, 30, ["end", "training", "methods"], "Method"], [32, 35, ["Connectionist", "Temporal", "Classification"], "Method"], [40, 41, ["RNNs"], "Method"], [42, 45, ["sequence", "labelling", "problems"], "Task"], [61, 68, ["Long", "Short", "-", "term", "Memory", "RNN", "architecture"], "Method"], [88, 89, ["RNN"], "Method"], [104, 107, ["deep", "feedforward", "networks"], "Method"], [111, 115, ["deep", "recurrent", "neural", "networks"], "Method"], [129, 131, ["deep", "networks"], "Method"], [141, 142, ["RNNs"], "Method"], [152, 153, ["regularisation"], "Method"], [157, 164, ["deep", "Long", "Short", "-", "term", "Memory", "RNNs"], "Method"], [166, 169, ["test", "set", "error"], "Metric"], [174, 178, ["TIMIT", "phoneme", "recognition", "benchmark"], "Task"], [197, 200, ["recurrent", "neural", "networks"], "Method"], [201, 204, ["deep", "neural", "networks"], "Method"], [210, 212, ["Neural", "networks"], "Method"], [224, 227, ["hidden", "Markov", "models"], "Method"], [240, 242, ["acoustic", "modelling"], "Task"], [244, 247, ["deep", "feedforward", "networks"], "Method"], [254, 256, ["dynamic", "process"], "Task"], [262, 265, ["recurrent", "neural", "networks"], "Method"], [266, 267, ["RNNs"], "Method"], [273, 277, ["HMM", "-", "RNN", "systems"], "Method"], [292, 294, ["deep", "networks"], "Method"], [298, 299, ["RNNs"], "Method"], [300, 301, ["HMMs"], "Method"], [307, 308, ["RNNs"], "Method"], [331, 332, ["RNNs"], "Method"], [334, 335, ["HMMs"], "Method"], [352, 357, ["Long", "Short", "-", "term", "Memory"], "Method"], [359, 361, ["RNN", "architecture"], "Method"], [364, 365, ["memory"], "Method"], [367, 373, ["end", "-", "to", "-", "end", "training"], "Task"], [394, 395, ["RNNs"], "Method"], [422, 423, ["RNNs"], "Method"], [447, 449, ["feedforward", "layers"], "Method"], [453, 455, ["deep", "networks"], "Method"], [462, 469, ["deep", "Long", "Short", "-", "term", "Memory", "RNNs"], "Method"], [486, 493, ["end", "-", "to", "-", "end", "learning", "method"], "Method"], [498, 499, ["RNNs"], "Method"], [500, 504, ["acoustic", "and", "linguistic", "models"], "Method"], [515, 517, ["network", "architectures"], "Method"], [518, 520, ["training", "methods"], "Method"], [539, 542, ["Recurrent", "Neural", "Networks"], "Method"], [549, 552, ["recurrent", "neural", "network"], "Method"], [553, 554, ["RNN"], "Method"], [611, 613, ["elementwise", "application"], "Method"], [615, 617, ["sigmoid", "function"], "Method"], [624, 629, ["Long", "Short", "-", "Term", "Memory"], "Method"], [630, 631, ["LSTM"], "Method"], [657, 658, ["lstm"], "Method"], [679, 681, ["composite", "function"], "Method"], [685, 688, ["logistic", "sigmoid", "function"], "Method"], [756, 757, ["RNNs"], "Method"], [794, 796, ["Bidirectional", "RNNs"], "Method"], [797, 798, ["BRNNs"], "Method"], [830, 831, ["brnn"], "Method"], [833, 834, ["BRNN"], "Method"], [851, 853, ["backward", "layer"], "Method"], [857, 859, ["forward", "layer"], "Method"], [869, 870, ["BRNNs"], "Method"], [871, 872, ["LSTM"], "Method"], [873, 875, ["bidirectional", "LSTM"], "Method"], [896, 902, ["hybrid", "HMM", "-", "neural", "network", "systems"], "Method"], [906, 908, ["deep", "architectures"], "Method"], [916, 919, ["higher", "level", "representations"], "Method"], [923, 925, ["Deep", "RNNs"], "Method"], [931, 934, ["RNN", "hidden", "layers"], "Method"], [990, 993, ["Deep", "bidirectional", "RNNs"], "Method"], [1013, 1015, ["hidden", "layer"], "Method"], [1030, 1031, ["LSTM"], "Method"], [1062, 1064, ["deep", "LSTM"], "Method"], [1088, 1090, ["Network", "Training"], "Method"], [1093, 1099, ["end", "-", "to", "-", "end", "training"], "Task"], [1101, 1102, ["RNNs"], "Method"], [1186, 1187, ["backpropagation"], "Method"], [1198, 1200, ["gradient", "descent"], "Method"], [1241, 1244, ["Connectionist", "Temporal", "Classification"], "Method"], [1250, 1253, ["Connectionist", "Temporal", "Classification"], "Method"], [1254, 1255, ["CTC"], "Method"], [1259, 1261, ["softmax", "layer"], "Method"], [1293, 1295, ["softmax", "layer"], "Method"], [1334, 1335, ["CTC"], "Method"], [1338, 1342, ["forward", "-", "backward", "algorithm"], "Method"], [1369, 1370, ["speech"], "Task"], [1371, 1373, ["handwriting", "recognition"], "Task"], [1387, 1388, ["segmentation"], "Task"], [1399, 1400, ["RNNs"], "Method"], [1402, 1403, ["CTC"], "Method"], [1432, 1435, ["deep", "bidirectional", "networks"], "Method"], [1462, 1464, ["RNN", "Transducer"], "Method"], [1464, 1465, ["CTC"], "Method"], [1467, 1471, ["distribution", "over", "phoneme", "sequences"], "Method"], [1484, 1488, ["acoustic", "-", "only", "model"], "Method"], [1496, 1498, ["RNN", "transducer"], "Method"], [1507, 1508, ["RNN"], "Method"], [1522, 1526, ["acoustic", "and", "language", "model"], "Method"], [1527, 1532, ["Joint", "LM", "-", "acoustic", "training"], "Method"], [1543, 1544, ["CTC"], "Method"], [1554, 1556, ["RNN", "transducer"], "Method"], [1572, 1573, ["CTC"], "Method"], [1638, 1642, ["forward", "-", "backward", "algorithm"], "Method"], [1655, 1658, ["acoustic", "’", "distribution"], "Method"], [1665, 1668, ["linguistic", "’", "distribution"], "Method"], [1670, 1672, ["prediction", "network"], "Method"], [1700, 1703, ["feedforward", "output", "network"], "Method"], [1747, 1749, ["deletion", "errors"], "Metric"], [1751, 1752, ["decoding"], "Task"], [1777, 1779, ["prediction", "network"], "Method"], [1783, 1785, ["output", "network"], "Method"], [1792, 1794, ["linear", "layer"], "Method"], [1804, 1806, ["hidden", "layer"], "Method"], [1815, 1817, ["softmax", "layer"], "Method"], [1854, 1856, ["RNN", "transducers"], "Method"], [1882, 1888, ["pretrained", "next", "-", "step", "prediction", "network"], "Method"], [1893, 1895, ["output", "network"], "Method"], [1915, 1916, ["pretraining"], "Task"], [1919, 1920, ["retraining"], "Task"], [1927, 1929, ["prediction", "network"], "Method"], [1941, 1945, ["large", "-", "scale", "applications"], "Task"], [1960, 1961, ["Decoding"], "Task"], [1961, 1963, ["RNN", "transducers"], "Method"], [1967, 1969, ["beam", "search"], "Method"], [1993, 1997, ["best", "-", "first", "decoding"], "Method"], [1999, 2001, ["prefix", "search"], "Method"], [2022, 2024, ["beam", "search"], "Method"], [2026, 2027, ["transducer"], "Method"], [2049, 2051, ["beam", "search"], "Method"], [2057, 2059, ["prefix", "search"], "Method"], [2060, 2061, ["CTC"], "Method"], [2070, 2071, ["transducer"], "Method"], [2090, 2091, ["normalisation"], "Method"], [2110, 2111, ["Regularisation"], "Method"], [2111, 2112, ["Regularisation"], "Method"], [2118, 2119, ["RNNs"], "Method"], [2130, 2131, ["regularisers"], "Method"], [2137, 2139, ["early", "stopping"], "Task"], [2178, 2180, ["neural", "networks"], "Method"], [2198, 2199, ["generalisation"], "Task"], [2203, 2205, ["Phoneme", "recognition"], "Task"], [2210, 2212, ["TIMIT", "corpus"], "Material"], [2239, 2241, ["early", "stopping"], "Task"], [2261, 2269, ["Fourier", "-", "transform", "-", "based", "filter", "-", "bank"], "Method"], [2332, 2333, ["decoding"], "Task"], [2343, 2344, ["scoring"], "Task"], [2373, 2374, ["timit"], "Material"], [2376, 2377, ["RNNs"], "Method"], [2387, 2389, ["training", "method"], "Method"], [2391, 2392, ["CTC"], "Method"], [2393, 2394, ["Transducer"], "Method"], [2395, 2397, ["pretrained", "Transducer"], "Method"], [2463, 2466, ["stochastic", "gradient", "descent"], "Method"], [2468, 2470, ["learning", "rate"], "Metric"], [2491, 2496, ["PreTrans", "-", "3l", "-", "250h"], "Method"], [2521, 2524, ["Gaussian", "weight", "noise"], "Method"], [2539, 2544, ["PreTrans", "-", "3l", "-", "250h"], "Method"], [2562, 2565, ["phoneme", "prediction", "network"], "Method"], [2601, 2606, ["PreTrans", "-", "3l", "-", "250h"], "Method"], [2637, 2640, ["Beam", "search", "decoding"], "Method"], [2656, 2658, ["deep", "networks"], "Method"], [2664, 2666, ["error", "rate"], "Metric"], [2667, 2668, ["CTC"], "Method"], [2745, 2746, ["LSTM"], "Method"], [2757, 2759, ["bidirectional", "LSTM"], "Method"], [2764, 2766, ["unidirectional", "LSTMand"], "Method"], [2782, 2784, ["deep", "networks"], "Method"], [2806, 2807, ["pretraining"], "Task"], [2823, 2832, ["deep", ",", "bidirectional", "Long", "Short", "-", "term", "Memory", "RNNs"], "Method"], [2833, 2839, ["end", "-", "to", "-", "end", "training"], "Method"], [2852, 2854, ["phoneme", "recognition"], "Task"], [2856, 2858, ["TIMIT", "database"], "Material"], [2881, 2887, ["frequency", "-", "domain", "convolutional", "neural", "networks"], "Method"], [2888, 2890, ["deep", "LSTM"], "Method"], [84, 86, ["handwriting", "recognition"], "Task"], [91, 92, ["speech"], "Task"], [205, 206, ["speech"], "Task"], [217, 218, ["speech"], "Task"], [250, 251, ["speech"], "Task"], [316, 317, ["speech"], "Task"], [379, 381, ["handwriting", "recognition"], "Task"], [391, 392, ["speech"], "Task"], [474, 475, ["speech"], "Task"], [661, 662, ["LSTM"], "Method"], [669, 670, ["LSTM"], "Method"], [771, 772, ["speech"], "Task"], [1041, 1042, ["LSTM"], "Method"], [1068, 1069, ["speech"], "Task"], [1084, 1085, ["LSTM"], "Method"], [1381, 1382, ["CTC"], "Method"], [1500, 1501, ["CTC"], "Method"], [1539, 1540, ["speech"], "Task"], [1660, 1661, ["CTC"], "Method"], [1767, 1768, ["CTC"], "Method"], [1878, 1879, ["CTC"], "Method"], [1983, 1984, ["CTC"], "Method"], [2412, 2413, ["LSTM"], "Method"], [2420, 2421, ["LSTM"], "Method"], [2427, 2428, ["CTC"], "Method"], [2440, 2441, ["LSTM"], "Method"], [2444, 2445, ["CTC"], "Method"], [2453, 2454, ["LSTM"], "Method"], [2483, 2484, ["CTC"], "Method"], [2532, 2534, ["error", "rate"], "Metric"], [2550, 2551, ["CTC"], "Method"], [2574, 2575, ["LSTM"], "Method"], [2615, 2616, ["CTC"], "Method"], [2690, 2691, ["CTC"], "Method"], [2698, 2699, ["CTC"], "Method"], [2704, 2705, ["CTC"], "Method"], [2712, 2713, ["CTC"], "Method"], [2871, 2872, ["speech"], "Task"]]} {"doc_id": "18168aea48a22f6fe2fe407c0ff70083cba225a7", "text": ["document", ":", "Image", "Restoration", "Using", "Very", "Deep", "Convolutional", "Encoder", "-", "Decoder", "Networks", "with", "Symmetric", "Skip", "Connections", "In", "this", "paper", ",", "we", "propose", "a", "very", "deep", "fully", "convolutional", "encoding", "-", "decoding", "framework", "for", "image", "restoration", "such", "as", "denoising", "and", "super", "-", "resolution", ".", "The", "network", "is", "composed", "of", "multiple", "layers", "of", "convolution", "and", "de", "-", "convolution", "operators", ",", "learning", "end", "-", "to", "-", "end", "mappings", "from", "corrupted", "images", "to", "the", "original", "ones", ".", "The", "convolutional", "layers", "act", "as", "the", "feature", "extractor", ",", "which", "capture", "the", "abstraction", "of", "image", "contents", "while", "eliminating", "noises", "/", "corruptions", ".", "De", "-", "convolutional", "layers", "are", "then", "used", "to", "recover", "the", "image", "details", ".", "We", "propose", "to", "symmetrically", "link", "convolutional", "and", "de", "-", "convolutional", "layers", "with", "skip", "-", "layer", "connections", ",", "with", "which", "the", "training", "converges", "much", "faster", "and", "attains", "a", "higher", "-", "quality", "local", "optimum", ".", "First", ",", "The", "skip", "connections", "allow", "the", "signal", "to", "be", "back", "-", "propagated", "to", "bottom", "layers", "directly", ",", "and", "thus", "tackles", "the", "problem", "of", "gradient", "vanishing", ",", "making", "training", "deep", "networks", "easier", "and", "achieving", "restoration", "performance", "gains", "consequently", ".", "Second", ",", "these", "skip", "connections", "pass", "image", "details", "from", "convolutional", "layers", "to", "de", "-", "convolutional", "layers", ",", "which", "is", "beneficial", "in", "recovering", "the", "original", "image", ".", "Significantly", ",", "with", "the", "large", "capacity", ",", "we", "can", "handle", "different", "levels", "of", "noises", "using", "a", "single", "model", ".", "Experimental", "results", "show", "that", "our", "network", "achieves", "better", "performance", "than", "all", "previously", "reported", "state", "-", "of", "-", "the", "-", "art", "methods", ".", ".", "/", "figs", "/", "section", ":", "Introduction", "The", "task", "of", "image", "restoration", "is", "to", "recover", "an", "clean", "image", "from", "its", "corrupted", "observation", ",", "which", "is", "known", "to", "be", "an", "ill", "-", "posed", "inverse", "problem", ".", "By", "accommodating", "different", "types", "of", "corruption", "distributions", ",", "the", "same", "mathematical", "model", "applies", "to", "problems", "such", "as", "image", "denoising", "and", "super", "-", "resolution", ".", "Recently", ",", "deep", "neural", "networks", "(", "DNNs", ")", "have", "shown", "their", "superior", "performance", "in", "image", "processing", "and", "computer", "vision", "tasks", ",", "ranging", "from", "high", "-", "level", "recognition", ",", "semantic", "segmentation", "to", "low", "-", "level", "denoising", ",", "super", "-", "resolution", ",", "deblur", ",", "inpainting", "and", "recovering", "raw", "images", "from", "compressed", "images", ".", "Despite", "the", "progress", "that", "DNNs", "achieve", ",", "there", "still", "are", "some", "problems", ".", "For", "example", ",", "can", "a", "deeper", "network", "in", "general", "achieve", "better", "performance", ";", "can", "we", "design", "a", "single", "model", "to", "handle", "different", "levels", "of", "corruption", ".", "Observing", "recent", "superior", "performance", "of", "DNNs", "on", "image", "processing", "tasks", ",", "we", "propose", "a", "convolutional", "neural", "network", "(", "CNN", ")-", "based", "framework", "for", "image", "restoration", ".", "We", "observe", "that", "in", "order", "to", "obtain", "good", "restoration", "performance", ",", "it", "is", "beneficial", "to", "train", "a", "very", "deep", "model", ".", "Meanwhile", ",", "we", "show", "that", "it", "is", "possible", "to", "achieve", "good", "performance", "with", "a", "single", "network", "when", "processing", "multiple", "different", "levels", "of", "corruptions", "due", "to", "the", "benefits", "of", "large", "-", "capacity", "networks", ".", "Specifically", ",", "the", "proposed", "framework", "learns", "end", "-", "to", "-", "end", "fully", "convolutional", "mappings", "from", "corrupted", "images", "to", "the", "clean", "ones", ".", "The", "network", "is", "composed", "of", "multiple", "layers", "of", "convolution", "and", "de", "-", "convolution", "operators", ".", "As", "deeper", "networks", "tend", "to", "be", "more", "difficult", "to", "train", ",", "we", "propose", "to", "symmetrically", "link", "convolutional", "and", "de", "-", "convolutional", "layers", "with", "skip", "-", "layer", "connections", ",", "with", "which", "the", "training", "converges", "much", "faster", "and", "attains", "a", "higher", "-", "quality", "local", "optimum", ".", "Our", "main", "contributions", "are", "briefly", "outlined", "as", "follows", ":", "1", ")", "A", "very", "deep", "network", "architecture", ",", "which", "consists", "of", "a", "chain", "of", "symmetric", "convolutional", "and", "deconvolutional", "layers", ",", "for", "image", "restoration", "is", "proposed", "in", "this", "paper", ".", "The", "convolutional", "layers", "act", "as", "the", "feature", "extractor", "which", "encode", "the", "primary", "components", "of", "image", "contents", "while", "eliminating", "the", "corruption", ".", "The", "deconvolutional", "layers", "then", "decode", "the", "image", "abstraction", "to", "recover", "the", "image", "content", "details", ".", "2", ")", "We", "propose", "to", "add", "skip", "connections", "between", "corresponding", "convolutional", "and", "de", "-", "convolutional", "layers", ".", "These", "skip", "connections", "help", "to", "back", "-", "propagate", "the", "gradients", "to", "bottom", "layers", "and", "pass", "image", "details", "to", "the", "top", "layers", ",", "making", "training", "of", "the", "end", "-", "to", "-", "end", "mapping", "more", "easier", "and", "effective", ",", "and", "thus", "achieve", "performance", "improvement", "while", "the", "network", "going", "deeper", ".", "3", ")", "Relying", "on", "the", "large", "capacity", "and", "fitting", "ability", "of", "our", "very", "deep", "network", ",", "we", "propose", "to", "handle", "different", "level", "of", "noises", "/", "corruption", "using", "a", "single", "model", ".", "To", "our", "knowledge", ",", "this", "is", "the", "first", "approach", "that", "achieves", "good", "accuracy", "for", "processing", "different", "levels", "of", "noises", "with", "a", "single", "model", ".", "4", ")", "Experimental", "results", "demonstrate", "the", "advantages", "of", "our", "network", "over", "other", "recent", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "image", "denoising", "and", "super", "-", "resolution", ",", "setting", "new", "records", "on", "these", "topics", ".", "Related", "work", "Extensive", "work", "has", "been", "done", "on", "image", "restoration", "in", "the", "literature", ".", "See", "detail", "reviews", "in", "a", "survey", ".", "Traditional", "methods", "such", "as", "Total", "variation", ",", "BM3D", "algorithm", "and", "dictionary", "learning", "based", "methods", "have", "shown", "very", "good", "performance", "on", "image", "restoration", "topics", "such", "as", "image", "denoising", "and", "super", "-", "resolution", ".", "Since", "image", "restoration", "is", "in", "general", "an", "ill", "-", "posed", "problem", ",", "the", "use", "of", "regularization", "has", "been", "proved", "to", "be", "essential", ".", "An", "active", "(", "and", "probably", "more", "promising", ")", "category", "for", "image", "restoration", "is", "the", "DNN", "based", "methods", ".", "Stacked", "denoising", "auto", "-", "encoder", "is", "one", "of", "the", "most", "well", "-", "known", "DNN", "models", "which", "can", "be", "used", "for", "image", "restoration", ".", "Xie", "et", "al", ".", "combined", "sparse", "coding", "and", "DNN", "pre", "-", "trained", "with", "denoising", "auto", "-", "encoder", "for", "low", "-", "level", "vision", "tasks", "such", "as", "image", "denoising", "and", "inpainting", ".", "Other", "neural", "networks", "based", "methods", "such", "as", "multi", "-", "layer", "perceptron", "and", "CNN", "for", "image", "denoising", ",", "as", "well", "as", "DNN", "for", "image", "or", "video", "super", "-", "resolution", "and", "compression", "artifacts", "reduction", "have", "been", "actively", "studied", "in", "these", "years", ".", "Burger", "et", "al", ".", "presented", "a", "patch", "-", "based", "algorithm", "learned", "with", "a", "plain", "multi", "-", "layer", "perceptron", ".", "They", "also", "concluded", "that", "with", "large", "networks", ",", "large", "training", "data", ",", "neural", "networks", "can", "achieve", "state", "-", "of", "-", "the", "-", "art", "image", "denoising", "performance", ".", "Jain", "and", "Seung", "proposed", "fully", "convolutional", "CNN", "for", "denoising", ".", "They", "found", "that", "CNN", "provide", "comparable", "or", "even", "superior", "performance", "to", "wavelet", "and", "Markov", "Random", "Field", "(", "MRF", ")", "methods", ".", "Cui", "et", "al", ".", "employed", "non", "-", "local", "self", "-", "similarity", "(", "NLSS", ")", "search", "on", "the", "input", "image", "in", "multi", "-", "scale", ",", "and", "then", "used", "collaborative", "local", "auto", "-", "encoder", "for", "super", "-", "resolution", "in", "a", "layer", "by", "layer", "fashion", ".", "Dong", "et", "al", ".", "proposed", "to", "directly", "learn", "an", "end", "-", "to", "-", "end", "mapping", "between", "the", "low", "/", "high", "-", "resolution", "images", ".", "Wang", "et", "al", ".", "argued", "that", "domain", "expertise", "represented", "by", "the", "conventional", "sparse", "coding", "can", "be", "combined", "to", "achieve", "further", "improved", "results", ".", "In", "general", ",", "DNN", "-", "based", "methods", "learn", "restoration", "parameters", "directly", "from", "data", ",", "which", "tends", "to", "been", "more", "effective", "in", "real", "-", "world", "image", "restoration", "applications", ".", "An", "advantage", "of", "DNN", "methods", "is", "that", "these", "methods", "are", "purely", "data", "driven", "and", "no", "assumption", "about", "the", "noise", "distributions", "are", "made", ".", "section", ":", "Very", "deep", "RED", "-", "Net", "for", "Image", "Restoration", "The", "proposed", "framework", "mainly", "contains", "a", "chain", "of", "convolutional", "layers", "and", "symmetric", "deconvolutional", "layers", ",", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "We", "term", "our", "method", "‘", "‘", "RED", "-", "Net’’—very", "deep", "Residual", "Encoder", "-", "Decoder", "Networks", ".", "subsection", ":", "Architecture", "The", "framework", "is", "fully", "convolutional", "and", "deconvolutional", ".", "Rectification", "layers", "are", "added", "after", "each", "convolution", "and", "deconvolution", ".", "The", "convolutional", "layers", "act", "as", "feature", "extractor", ",", "which", "preserve", "the", "primary", "components", "of", "objects", "in", "the", "image", "and", "meanwhile", "eliminating", "the", "corruptions", ".", "The", "deconvolutional", "layers", "are", "then", "combined", "to", "recover", "the", "details", "of", "image", "contents", ".", "The", "output", "of", "the", "deconvolutional", "layers", "is", "the", "‘", "‘", "clean", "’", "’", "version", "of", "the", "input", "image", ".", "Moreover", ",", "skip", "connections", "are", "also", "added", "from", "a", "convolutional", "layer", "to", "its", "corresponding", "mirrored", "deconvolutional", "layer", ".", "The", "passed", "convolutional", "feature", "maps", "are", "summed", "to", "the", "deconvolutional", "feature", "maps", "element", "-", "wise", ",", "and", "passed", "to", "the", "next", "layer", "after", "rectification", ".", "For", "low", "-", "level", "image", "restoration", "problems", ",", "we", "use", "neither", "pooling", "nor", "unpooling", "in", "the", "network", "as", "usually", "pooling", "discards", "useful", "image", "details", "that", "are", "essential", "for", "these", "tasks", ".", "Motivated", "by", "the", "VGG", "model", ",", "the", "kernel", "size", "for", "convolution", "and", "deconvolution", "is", "set", "to", ",", "which", "has", "shown", "excellent", "image", "recognition", "performance", ".", "It", "is", "worth", "mentioning", "that", "the", "size", "of", "input", "image", "can", "be", "arbitrary", "since", "our", "network", "is", "essentially", "a", "pixel", "-", "wise", "prediction", ".", "The", "input", "and", "output", "of", "the", "network", "are", "images", "of", "the", "same", "size", ",", "where", ",", "and", "are", "width", ",", "height", "and", "number", "of", "channels", ".", "In", "this", "paper", ",", "we", "use", "although", "it", "is", "straightforward", "to", "apply", "to", "images", "with", ".", "We", "found", "that", "using", "64", "feature", "maps", "for", "convolutional", "and", "deconvolutional", "layers", "achieves", "satisfactory", "results", ",", "although", "more", "feature", "maps", "leads", "to", "slightly", "better", "performance", ".", "Deriving", "from", "the", "above", "architecture", ",", "we", "propose", "two", "networks", ",", "which", "are", "20", "-", "layer", "and", "30", "-", "layer", "respectively", ".", "subsubsection", ":", "Deconvolution", "decoder", "Architectures", "combining", "layers", "of", "convolution", "and", "deconvolution", "have", "been", "proposed", "for", "semantic", "segmentation", "lately", ".", "In", "contrast", "to", "convolutional", "layers", ",", "in", "which", "multiple", "input", "activations", "within", "a", "filter", "window", "are", "fused", "to", "output", "a", "single", "activation", ",", "deconvolutional", "layers", "associate", "a", "single", "input", "activation", "with", "multiple", "outputs", ".", "One", "can", "simply", "replace", "deconvolution", "with", "convolution", ",", "which", "results", "in", "a", "architecture", "that", "is", "very", "similar", "to", "recently", "proposed", "very", "deep", "fully", "convolutional", "neural", "networks", ".", "However", ",", "there", "exist", "essential", "differences", "between", "a", "fully", "convolution", "model", "and", "our", "model", ".", "In", "the", "fully", "convolution", "case", ",", "the", "noise", "is", "eliminated", "step", "by", "step", ",", "i.e.", ",", "the", "noise", "level", "is", "reduced", "after", "each", "layer", ".", "During", "this", "process", ",", "the", "details", "of", "the", "image", "content", "may", "be", "lost", ".", "Nevertheless", ",", "in", "our", "network", ",", "convolution", "preserves", "the", "primary", "image", "content", ".", "Then", "deconvolution", "is", "used", "to", "compensate", "the", "details", ".", "We", "compare", "the", "5", "-", "layer", "and", "10", "-", "layer", "fully", "convolutional", "network", "with", "our", "network", "(", "combining", "convolution", "and", "deconvolution", ",", "but", "without", "skip", "connection", ")", ".", "For", "fully", "convolutional", "networks", ",", "we", "use", "padding", "and", "up", "-", "sample", "the", "input", "to", "make", "the", "input", "and", "output", "the", "same", "size", ".", "For", "our", "network", ",", "the", "first", "5", "layers", "are", "convolutional", "and", "the", "second", "5", "layers", "are", "deconvolutional", ".", "All", "the", "other", "parameters", "for", "training", "are", "the", "same", ",", "i.e.", ",", "trained", "with", "SGD", "and", "learning", "rate", "of", ",", "noise", "level", ".", "In", "terms", "of", "PSNR", ",", "using", "deconvolution", "works", "better", "than", "the", "fully", "convolutional", "counterpart", ".", "We", "see", "that", ",", "the", "fully", "convolutional", "network", "reduces", "noise", "layer", "by", "layer", ",", "and", "our", "network", "preserve", "primary", "image", "contents", "by", "convolution", "and", "recover", "some", "details", "by", "using", "deconvolution", ".", "Detailed", "results", "are", "in", "the", "supplementary", "materials", ".", "subsubsection", ":", "Skip", "connections", "An", "intuitive", "question", "is", "that", ",", "is", "deconvolution", "able", "to", "recover", "image", "details", "from", "the", "image", "abstraction", "only", "?", "We", "find", "that", "in", "shallow", "networks", "with", "only", "a", "few", "layers", "of", "convolution", ",", "deconvolution", "is", "able", "to", "recover", "the", "details", ".", "However", ",", "when", "the", "network", "goes", "deeper", "or", "using", "operations", "such", "as", "max", "pooling", ",", "deconvolution", "does", "not", "work", "so", "well", ",", "possibly", "because", "too", "much", "details", "are", "already", "lost", "in", "the", "convolution", ".", "The", "second", "question", "is", "that", ",", "when", "our", "network", "goes", "deeper", ",", "does", "it", "achieve", "performance", "gain", "?", "We", "observe", "that", "deeper", "networks", "often", "suffer", "from", "gradients", "vanishing", "and", "become", "hard", "to", "train", "—", "a", "problem", "that", "is", "well", "addressed", "in", "the", "literature", ".", "To", "address", "the", "above", "two", "problems", ",", "inspired", "by", "highway", "networks", "and", "deep", "residual", "networks", ",", "we", "add", "skip", "connections", "between", "two", "corresponding", "convolutional", "and", "deconvolutional", "layers", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "A", "building", "block", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "There", "are", "two", "reasons", "for", "using", "such", "connections", ".", "First", ",", "when", "the", "network", "goes", "deeper", ",", "as", "mentioned", "above", ",", "image", "details", "can", "be", "lost", ",", "making", "deconvolution", "weaker", "in", "recovering", "them", ".", "However", ",", "the", "feature", "maps", "passed", "by", "skip", "connections", "carry", "much", "image", "detail", ",", "which", "helps", "deconvolution", "to", "recover", "a", "better", "clean", "image", ".", "Second", ",", "the", "skip", "connections", "also", "achieve", "benefits", "on", "back", "-", "propagating", "the", "gradient", "to", "bottom", "layer", ",", "which", "makes", "training", "deeper", "network", "much", "easier", "as", "observed", "in", "and", ".", "Note", "that", "our", "skip", "layer", "connections", "are", "very", "different", "from", "the", "ones", "proposed", "in", "and", ",", "where", "the", "only", "concern", "is", "on", "the", "optimization", "side", ".", "In", "our", "case", ",", "we", "want", "to", "pass", "information", "of", "the", "convolutional", "feature", "maps", "to", "the", "corresponding", "deconvolutional", "layers", ".", "Instead", "of", "directly", "learning", "the", "mappings", "from", "input", "to", "the", "output", ",", "we", "would", "like", "the", "network", "to", "fit", "the", "residual", "of", "the", "problem", ",", "which", "is", "denoted", "as", ".", "Such", "a", "learning", "strategy", "is", "applied", "to", "inner", "blocks", "of", "the", "encoding", "-", "decoding", "network", "to", "make", "training", "more", "effective", ".", "Skip", "connections", "are", "passed", "every", "two", "convolutional", "layers", "to", "their", "mirrored", "deconvolutional", "layers", ".", "Other", "configurations", "are", "possible", "and", "our", "experiments", "show", "that", "this", "configuration", "already", "works", "very", "well", ".", "Using", "such", "shortcuts", "makes", "the", "network", "easier", "to", "be", "trained", "and", "gains", "restoration", "performance", "via", "increasing", "network", "depth", ".", "The", "very", "deep", "highway", "networks", "are", "essentially", "feed", "-", "forward", "long", "short", "-", "term", "memory", "(", "LSTMs", ")", "with", "forget", "gates", ";", "and", "the", "CNN", "layers", "of", "deep", "residual", "network", "are", "feed", "-", "forward", "LSTMs", "without", "gates", ".", "Note", "that", "our", "deep", "residual", "networks", "are", "in", "general", "not", "in", "the", "format", "of", "standard", "feed", "-", "forward", "LSTMs", ".", "subsection", ":", "Discussions", "Training", "with", "symmetric", "skip", "connections", "As", "mentioned", "above", ",", "using", "skip", "connections", "mainly", "has", "two", "benefits", ":", "(", "1", ")", "passing", "image", "detail", "forwardly", ",", "which", "helps", "recovering", "clean", "images", "and", "(", "2", ")", "passing", "gradient", "backwardly", ",", "which", "helps", "finding", "better", "local", "minimum", ".", "We", "design", "experiments", "to", "show", "these", "observations", ".", "We", "first", "compare", "two", "networks", "trained", "for", "denoising", "noises", "of", ".", "In", "the", "first", "network", ",", "we", "use", "5", "layers", "of", "convolution", "with", "stride", "3", ".", "The", "input", "size", "of", "training", "data", "is", ",", "which", "results", "in", "a", "vector", "after", "5", "layers", "of", "convolution", ".", "Then", "deconvolution", "is", "used", "to", "recover", "the", "input", ".", "The", "second", "network", "uses", "the", "same", "settings", "as", "the", "first", "one", ",", "except", "for", "adding", "skip", "connections", ".", "The", "results", "are", "show", "in", "Figure", "[", "reference", "]", "(", "a", ")", ".", "We", "can", "observe", "that", "it", "is", "hard", "for", "deconvolution", "to", "recover", "details", "from", "only", "a", "vector", "encoding", "the", "abstraction", "of", "the", "input", ",", "which", "shows", "that", "the", "ability", "on", "recovering", "image", "details", "for", "deconvolution", "is", "limited", ".", "However", ",", "if", "we", "use", "skip", "connections", ",", "the", "network", "can", "still", "recover", "the", "input", ",", "because", "details", "are", "passed", "to", "topper", "layers", "in", "the", "network", ".", "We", "also", "train", "five", "networks", "to", "show", "that", "using", "skip", "connections", "help", "to", "back", "-", "propagate", "gradient", "in", "training", "to", "better", "fit", "the", "end", "-", "to", "-", "end", "mapping", ",", "as", "shown", "in", "Figure", "[", "reference", "]", "(", "b", ")", ".", "The", "five", "networks", "are", ":", "10", ",", "20", "and", "30", "layer", "networks", "without", "skip", "connections", ",", "and", "20", ",", "30", "layer", "networks", "with", "skip", "connections", ".", "As", "we", "can", "see", ",", "the", "training", "loss", "increases", "when", "the", "network", "going", "deeper", "without", "shortcuts", "(", "similar", "phenomenon", "is", "also", "observed", "in", ")", ",", "but", "we", "obtain", "smaller", "loss", "when", "using", "skip", "connections", ".", "Comparison", "with", "deep", "residual", "networks", "[", "]", "One", "may", "use", "different", "types", "of", "skip", "connections", "in", "our", "network", ",", "a", "straightforward", "alternate", "is", "that", "in", ".", "In", ",", "the", "skip", "connections", "are", "added", "to", "divide", "the", "network", "into", "sequential", "blocks", ".", "A", "benefit", "of", "our", "model", "is", "that", "our", "skip", "connections", "have", "element", "-", "wise", "correspondence", ",", "which", "can", "be", "very", "important", "in", "pixel", "-", "wise", "prediction", "problems", ".", "We", "carry", "out", "experiments", "to", "compare", "the", "two", "types", "of", "skip", "connections", ".", "Here", "the", "block", "size", "indicates", "the", "span", "of", "the", "connections", ".", "The", "results", "are", "shown", "in", "Figure", "[", "reference", "]", "(", "c", ")", ".", "We", "can", "observe", "that", "our", "connections", "often", "converge", "to", "a", "better", "optimum", ",", "demonstrating", "that", "element", "-", "wise", "correspondence", "can", "be", "important", ".", "Dealing", "with", "different", "levels", "of", "noises", "/", "corruption", "An", "important", "question", "is", ",", "can", "we", "handle", "different", "levels", "of", "corruption", "with", "a", "single", "model", ".", "Almost", "all", "existing", "methods", "need", "to", "train", "different", "models", "for", "different", "levels", "of", "corruption", "and", "estimate", "the", "corruption", "level", "at", "first", ".", "We", "use", "a", "trained", "model", "in", ",", "to", "denoise", "different", "levels", "of", "noises", "with", "being", "10", ",", "30", ",", "50", "and", "70", ".", "The", "obtained", "average", "PSNR", "on", "the", "14", "images", "are", "29.95dB", ",", "27.81dB", ",", "18.62dB", "and", "14.84dB", ",", "respectively", ".", "The", "results", "show", "that", "the", "parameters", "trained", "on", "a", "single", "noise", "level", "can", "not", "handle", "different", "levels", "of", "noises", "well", ".", "Therefore", ",", "in", "this", "paper", ",", "we", "aim", "to", "train", "a", "single", "model", "for", "recovering", "different", "levels", "of", "corruption", ",", "which", "are", "different", "noise", "levels", "in", "the", "task", "of", "image", "denoising", "and", "different", "scaling", "parameters", "in", "image", "super", "-", "resolution", ".", "The", "large", "capacity", "of", "the", "network", "is", "the", "key", "to", "this", "success", ".", "subsection", ":", "Training", "Learning", "the", "end", "-", "to", "-", "end", "mapping", "from", "corrupted", "images", "to", "clean", "ones", "needs", "to", "estimate", "the", "weights", "represented", "by", "the", "convolutional", "and", "deconvolutional", "kernels", ".", "This", "is", "achieved", "by", "minimizing", "the", "Euclidean", "loss", "between", "the", "outputs", "of", "the", "network", "and", "the", "clean", "image", ".", "In", "specific", ",", "given", "a", "collection", "of", "training", "sample", "pairs", ",", "where", "is", "a", "corrupted", "image", "and", "is", "the", "clean", "version", "as", "the", "groundtruth", ".", "We", "minimize", "the", "following", "Mean", "Squared", "Error", "(", "MSE", ")", ":", "We", "implement", "and", "train", "our", "network", "using", "Caffe", ".", "In", "practice", ",", "we", "find", "that", "using", "Adam", "with", "learning", "rate", "for", "training", "converges", "faster", "than", "traditional", "stochastic", "gradient", "descent", "(", "SGD", ")", ".", "The", "base", "learning", "rate", "for", "all", "layers", "are", "the", "same", ",", "different", "from", ",", "in", "which", "a", "smaller", "learning", "rate", "is", "set", "for", "the", "last", "layer", ".", "This", "trick", "is", "not", "necessary", "in", "our", "network", ".", "As", "general", "settings", "in", "the", "literature", ",", "we", "use", "gray", "-", "scale", "image", "for", "denoising", "and", "the", "luminance", "channel", "for", "super", "-", "resolution", "in", "this", "paper", ".", "300", "images", "from", "the", "Berkeley", "Segmentation", "Dataset", "(", "BSD", ")", "are", "used", "to", "generate", "the", "training", "set", ".", "For", "each", "image", ",", "patches", "of", "size", "are", "sampled", "as", "ground", "truth", ".", "For", "denoising", ",", "we", "add", "additive", "Gaussian", "noise", "to", "the", "patches", "multiple", "times", "to", "generate", "a", "large", "training", "set", "(", "about", "0.5", "M", ")", ".", "For", "super", "-", "resolution", ",", "we", "first", "down", "-", "sample", "a", "patch", "and", "then", "up", "-", "sample", "it", "to", "its", "original", "size", ",", "obtaining", "a", "low", "-", "resolution", "version", "as", "the", "input", "of", "the", "network", ".", "subsection", ":", "Testing", "Although", "trained", "on", "local", "patches", ",", "our", "network", "can", "perform", "denoising", "and", "super", "-", "resolution", "on", "images", "of", "arbitrary", "size", ".", "Given", "a", "testing", "image", ",", "one", "can", "simply", "go", "forward", "through", "the", "network", ",", "which", "is", "able", "to", "obtain", "a", "better", "performance", "than", "existing", "methods", ".", "To", "achieve", "more", "smooth", "results", ",", "we", "propose", "to", "process", "a", "corrupted", "image", "on", "multiple", "orientations", ".", "Different", "from", "segmentation", ",", "the", "filter", "kernels", "in", "our", "network", "only", "eliminate", "the", "corruptions", ",", "which", "is", "not", "sensitive", "to", "the", "orientation", "of", "image", "contents", ".", "Therefore", ",", "we", "can", "rotate", "and", "mirror", "flip", "the", "kernels", "and", "perform", "forward", "multiple", "times", ",", "and", "then", "average", "the", "output", "to", "get", "a", "more", "smooth", "image", ".", "We", "see", "that", "this", "can", "lead", "to", "slightly", "better", "denoising", "and", "super", "-", "resolution", "performance", ".", "section", ":", "Experiments", "In", "this", "section", ",", "we", "provide", "evaluation", "of", "denoising", "and", "super", "-", "resolution", "performance", "of", "our", "models", "against", "a", "few", "existing", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "Denoising", "experiments", "are", "performed", "on", "two", "datasets", ":", "14", "common", "benchmark", "images", "and", "the", "BSD200", "dataset", ".", "We", "test", "additive", "Gaussian", "noises", "with", "zero", "mean", "and", "standard", "deviation", "10", ",", "30", ",", "50", "and", "70", "respectively", ".", "BM3D", ",", "NCSR", ",", "EPLL", ",", "PCLR", ",", "PDPD", "and", "WMMN", "are", "compared", "with", "our", "method", ".", "For", "super", "-", "resolution", ",", "we", "compare", "our", "network", "with", "SRCNN", ",", "NBSRF", ",", "CSCN", ",", "CSC", ",", "TSE", "and", "ARFL", "+", "on", "three", "dataset", ":", "Set5", ",", "Set14", "and", "BSD100", ".", "The", "scaling", "parameter", "are", "tested", "with", "2", ",", "3", "and", "4", ".", "Peak", "Signal", "-", "to", "-", "Noise", "Ratio", "(", "PSNR", ")", "and", "Structural", "SIMilarity", "(", "SSIM", ")", "index", "are", "calculated", "for", "evaluation", ".", "For", "our", "method", ",", "which", "is", "denoted", "as", "RED", "-", "Net", ",", "we", "implement", "three", "versions", ":", "RED10", "contains", "5", "convolutional", "and", "deconvolutional", "layers", "without", "shortcuts", ",", "RED20", "contains", "10", "convolutional", "and", "deconvolutional", "layers", "with", "shortcuts", ",", "and", "RED30", "contains", "15", "convolutional", "and", "deconvolutional", "layers", "with", "shortcuts", ".", "subsection", ":", "Image", "Denoising", "Evaluation", "on", "the", "14", "images", "Table", "[", "reference", "]", "presents", "the", "PSNR", "and", "SSIM", "results", "of", "10", ",", "30", ",", "50", ",", "and", "70", ".", "We", "can", "make", "some", "observations", "from", "the", "results", ".", "First", "of", "all", ",", "the", "10", "layer", "convolutional", "and", "deconvolutional", "network", "has", "already", "achieved", "better", "results", "than", "the", "state", "-", "of", "-", "the", "-", "art", "methods", ",", "which", "demonstrates", "that", "combining", "convolution", "and", "deconvolution", "for", "denoising", "works", "well", ",", "even", "without", "any", "skip", "connections", ".", "Moreover", ",", "when", "the", "network", "goes", "deeper", ",", "the", "skip", "connections", "proposed", "in", "this", "paper", "help", "to", "achieve", "even", "better", "denoising", "performance", ",", "which", "exceeds", "the", "existing", "best", "method", "WNNM", "by", "0.32dB", ",", "0.43dB", ",", "0.49dB", "and", "0.51dB", "on", "noise", "levels", "of", "being", "10", ",", "30", ",", "50", "and", "70", "respectively", ".", "While", "WNNM", "is", "only", "slightly", "better", "than", "the", "second", "best", "existing", "method", "PCLR", "by", "0.01dB", ",", "0.06dB", ",", "0.03dB", "and", "0.01dB", "respectively", ",", "which", "shows", "the", "large", "improvement", "of", "our", "model", ".", "Last", ",", "we", "can", "observe", "that", "the", "more", "complex", "the", "noise", "is", ",", "the", "more", "improvement", "our", "model", "achieves", "than", "other", "methods", ".", "Similar", "observations", "can", "be", "made", "on", "the", "evaluation", "of", "SSIM", ".", "Evaluation", "on", "BSD200", "For", "testing", "efficiency", ",", "we", "convert", "the", "images", "to", "gray", "-", "scale", "and", "resize", "them", "to", "smaller", "ones", "on", "BSD", "-", "200", ".", "Then", "all", "the", "methods", "are", "run", "on", "these", "images", "to", "get", "average", "PSNR", "and", "SSIM", "results", "of", "10", ",", "30", ",", "50", ",", "and", "70", ",", "as", "shown", "in", "Table", "[", "reference", "]", ".", "For", "existing", "methods", ",", "their", "denoising", "performance", "does", "not", "differ", "much", ",", "while", "our", "model", "achieves", "0.38dB", ",", "0.47dB", ",", "0.49dB", "and", "0.42dB", "higher", "of", "PSNR", "over", "WNNM", ".", "subsection", ":", "Image", "super", "-", "resolution", "The", "evaluation", "on", "Set5", "is", "shown", "in", "Table", "[", "reference", "]", ".", "Our", "10", "-", "layer", "network", "outperforms", "the", "compared", "methods", "already", ",", "and", "we", "achieve", "better", "performance", "with", "deeper", "networks", ".", "The", "30", "-", "layer", "network", "exceeds", "the", "second", "best", "method", "CSCN", "for", "0.52dB", ",", "0.56dB", "and", "0.47dB", "on", "scale", "2", ",", "3", "and", "4", "respectively", ".", "The", "evaluation", "on", "Set14", "is", "shown", "in", "Table", "[", "reference", "]", ".", "The", "improvement", "on", "Set14", "in", "not", "as", "significant", "as", "that", "on", "Set5", ",", "but", "we", "can", "still", "observe", "that", "the", "30", "layer", "network", "achieves", "higher", "PSNR", "than", "the", "second", "best", "CSCN", "for", "0.23dB", ",", "0.06dB", "and", "0.1dB.", "The", "results", "on", "BSD100", ",", "as", "shown", "in", "Table", "[", "reference", "]", ",", "is", "similar", "than", "that", "on", "Set5", ".", "The", "second", "best", "method", "is", "still", "CSCN", ",", "the", "performance", "of", "which", "is", "not", "as", "good", "as", "our", "10", "layer", "network", ".", "Our", "deeper", "network", "obtains", "much", "more", "performance", "gain", "than", "the", "others", ".", "subsection", ":", "Evaluation", "with", "a", "single", "model", "To", "construct", "the", "training", "set", ",", "we", "extract", "image", "patches", "with", "different", "noise", "levels", "and", "scaling", "parameters", "for", "denoising", "and", "super", "-", "resolution", ".", "Then", "a", "30", "-", "layer", "network", "is", "trained", "for", "the", "two", "tasks", "respectively", ".", "The", "evaluation", "results", "are", "shown", "in", "Table", "[", "reference", "]", "and", "Table", "[", "reference", "]", ".", "Although", "training", "with", "different", "levels", "of", "corruption", ",", "we", "can", "observe", "that", "the", "performance", "of", "our", "network", "only", "slightly", "degrades", "comparing", "to", "the", "case", "in", "which", "using", "separate", "models", "for", "denoising", "and", "super", "-", "resolution", ".", "This", "may", "be", "due", "the", "fact", "that", "the", "network", "has", "to", "fit", "much", "more", "complex", "mappings", ".", "Except", "that", "CSCN", "works", "slightly", "better", "on", "super", "-", "resolution", "with", "scales", "3", "and", "4", ",", "our", "network", "still", "beats", "the", "existing", "methods", ",", "showing", "that", "our", "network", "works", "much", "better", "in", "image", "denoising", "and", "super", "-", "resolution", "even", "using", "only", "one", "single", "model", "to", "deal", "with", "complex", "corruption", ".", "section", ":", "Conclusions", "In", "this", "paper", "we", "have", "proposed", "a", "deep", "encoding", "and", "decoding", "framework", "for", "image", "restoration", ".", "Convolution", "and", "deconvolution", "are", "combined", ",", "modeling", "the", "restoration", "problem", "by", "extracting", "primary", "image", "content", "and", "recovering", "details", ".", "More", "importantly", ",", "we", "propose", "to", "use", "skip", "connections", ",", "which", "helps", "on", "recovering", "clean", "images", "and", "tackles", "the", "optimization", "difficulty", "caused", "by", "gradient", "vanishing", ",", "and", "thus", "obtains", "performance", "gains", "when", "the", "network", "goes", "deeper", ".", "Experimental", "results", "and", "our", "analysis", "show", "that", "our", "network", "achieves", "better", "performance", "than", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "image", "denoising", "and", "super", "-", "resolution", ".", "X.", "-", "J.", "Mao", "’s", "contribution", "was", "made", "when", "visiting", "The", "University", "of", "Adelaide", ".", "This", "work", "was", "in", "part", "supported", "by", "ARC", "Future", "Fellowship", "(", "FT120100969", ")", ".", "Correspondence", "should", "be", "addressed", "to", "C.", "Shen", ".", "bibliography", ":", "References"], "entity": [[2, 4, ["Image", "Restoration"], "Task"], [6, 12, ["Deep", "Convolutional", "Encoder", "-", "Decoder", "Networks"], "Method"], [13, 16, ["Symmetric", "Skip", "Connections"], "Method"], [24, 31, ["deep", "fully", "convolutional", "encoding", "-", "decoding", "framework"], "Method"], [32, 34, ["image", "restoration"], "Task"], [36, 37, ["denoising"], "Task"], [38, 41, ["super", "-", "resolution"], "Task"], [48, 56, ["layers", "of", "convolution", "and", "de", "-", "convolution", "operators"], "Method"], [73, 75, ["convolutional", "layers"], "Method"], [78, 80, ["feature", "extractor"], "Method"], [94, 98, ["De", "-", "convolutional", "layers"], "Method"], [112, 118, ["convolutional", "and", "de", "-", "convolutional", "layers"], "Method"], [119, 123, ["skip", "-", "layer", "connections"], "Method"], [137, 139, ["local", "optimum"], "Metric"], [164, 166, ["gradient", "vanishing"], "Task"], [169, 171, ["deep", "networks"], "Method"], [174, 175, ["restoration"], "Task"], [188, 190, ["convolutional", "layers"], "Method"], [191, 195, ["de", "-", "convolutional", "layers"], "Method"], [256, 258, ["image", "restoration"], "Task"], [275, 280, ["ill", "-", "posed", "inverse", "problem"], "Task"], [291, 293, ["mathematical", "model"], "Method"], [301, 304, ["super", "-", "resolution"], "Task"], [307, 310, ["deep", "neural", "networks"], "Method"], [311, 312, ["DNNs"], "Method"], [319, 321, ["image", "processing"], "Task"], [322, 325, ["computer", "vision", "tasks"], "Task"], [328, 332, ["high", "-", "level", "recognition"], "Task"], [333, 335, ["semantic", "segmentation"], "Task"], [341, 344, ["super", "-", "resolution"], "Task"], [345, 346, ["deblur"], "Task"], [347, 348, ["inpainting"], "Task"], [349, 352, ["recovering", "raw", "images"], "Task"], [360, 361, ["DNNs"], "Method"], [374, 376, ["deeper", "network"], "Method"], [400, 401, ["DNNs"], "Method"], [402, 405, ["image", "processing", "tasks"], "Task"], [409, 412, ["convolutional", "neural", "network"], "Method"], [413, 414, ["CNN"], "Method"], [418, 420, ["image", "restoration"], "Task"], [429, 430, ["restoration"], "Task"], [439, 441, ["deep", "model"], "Method"], [470, 474, ["large", "-", "capacity", "networks"], "Method"], [486, 489, ["fully", "convolutional", "mappings"], "Method"], [502, 511, ["multiple", "layers", "of", "convolution", "and", "de", "-", "convolution", "operators"], "Method"], [513, 515, ["deeper", "networks"], "Method"], [528, 534, ["convolutional", "and", "de", "-", "convolutional", "layers"], "Method"], [569, 572, ["deep", "network", "architecture"], "Method"], [577, 584, ["chain", "of", "symmetric", "convolutional", "and", "deconvolutional", "layers"], "Method"], [586, 588, ["image", "restoration"], "Task"], [595, 597, ["convolutional", "layers"], "Method"], [600, 602, ["feature", "extractor"], "Method"], [616, 618, ["deconvolutional", "layers"], "Method"], [621, 623, ["image", "abstraction"], "Task"], [640, 646, ["convolutional", "and", "de", "-", "convolutional", "layers"], "Method"], [673, 679, ["end", "-", "to", "-", "end", "mapping"], "Task"], [708, 710, ["deep", "network"], "Method"], [738, 739, ["accuracy"], "Metric"], [775, 778, ["super", "-", "resolution"], "Task"], [794, 796, ["image", "restoration"], "Task"], [811, 813, ["Total", "variation"], "Method"], [814, 816, ["BM3D", "algorithm"], "Method"], [817, 821, ["dictionary", "learning", "based", "methods"], "Method"], [827, 830, ["image", "restoration", "topics"], "Task"], [835, 838, ["super", "-", "resolution"], "Task"], [840, 842, ["image", "restoration"], "Task"], [846, 850, ["ill", "-", "posed", "problem"], "Task"], [854, 855, ["regularization"], "Method"], [872, 874, ["image", "restoration"], "Task"], [876, 879, ["DNN", "based", "methods"], "Method"], [893, 895, ["DNN", "models"], "Method"], [900, 902, ["image", "restoration"], "Task"], [908, 910, ["sparse", "coding"], "Method"], [911, 913, ["DNN", "pre"], "Method"], [921, 926, ["low", "-", "level", "vision", "tasks"], "Task"], [931, 932, ["inpainting"], "Task"], [934, 938, ["neural", "networks", "based", "methods"], "Method"], [953, 954, ["DNN"], "Method"], [962, 965, ["compression", "artifacts", "reduction"], "Task"], [979, 983, ["patch", "-", "based", "algorithm"], "Method"], [986, 991, ["plain", "multi", "-", "layer", "perceptron"], "Method"], [997, 999, ["large", "networks"], "Method"], [1004, 1006, ["neural", "networks"], "Method"], [1027, 1028, ["denoising"], "Task"], [1032, 1033, ["CNN"], "Method"], [1040, 1049, ["wavelet", "and", "Markov", "Random", "Field", "(", "MRF", ")", "methods"], "Method"], [1055, 1065, ["non", "-", "local", "self", "-", "similarity", "(", "NLSS", ")", "search"], "Method"], [1077, 1082, ["collaborative", "local", "auto", "-", "encoder"], "Method"], [1083, 1086, ["super", "-", "resolution"], "Task"], [1102, 1108, ["end", "-", "to", "-", "end", "mapping"], "Task"], [1129, 1131, ["sparse", "coding"], "Method"], [1143, 1147, ["DNN", "-", "based", "methods"], "Method"], [1163, 1167, ["world", "image", "restoration", "applications"], "Task"], [1171, 1173, ["DNN", "methods"], "Method"], [1195, 1198, ["RED", "-", "Net"], "Method"], [1199, 1201, ["Image", "Restoration"], "Task"], [1207, 1211, ["chain", "of", "convolutional", "layers"], "Method"], [1212, 1215, ["symmetric", "deconvolutional", "layers"], "Method"], [1230, 1239, ["RED", "-", "Net’’—very", "deep", "Residual", "Encoder", "-", "Decoder", "Networks"], "Method"], [1247, 1250, ["convolutional", "and", "deconvolutional"], "Method"], [1251, 1253, ["Rectification", "layers"], "Method"], [1257, 1260, ["convolution", "and", "deconvolution"], "Method"], [1262, 1264, ["convolutional", "layers"], "Method"], [1266, 1268, ["feature", "extractor"], "Method"], [1286, 1288, ["deconvolutional", "layers"], "Method"], [1303, 1305, ["deconvolutional", "layers"], "Method"], [1327, 1329, ["convolutional", "layer"], "Method"], [1332, 1335, ["mirrored", "deconvolutional", "layer"], "Method"], [1359, 1360, ["rectification"], "Task"], [1362, 1368, ["low", "-", "level", "image", "restoration", "problems"], "Task"], [1372, 1373, ["pooling"], "Method"], [1374, 1375, ["unpooling"], "Method"], [1380, 1381, ["pooling"], "Method"], [1395, 1397, ["VGG", "model"], "Method"], [1399, 1401, ["kernel", "size"], "Method"], [1402, 1405, ["convolution", "and", "deconvolution"], "Method"], [1413, 1415, ["image", "recognition"], "Task"], [1436, 1440, ["pixel", "-", "wise", "prediction"], "Method"], [1491, 1495, ["convolutional", "and", "deconvolutional", "layers"], "Method"], [1531, 1532, ["subsubsection"], "Method"], [1533, 1535, ["Deconvolution", "decoder"], "Method"], [1535, 1536, ["Architectures"], "Method"], [1537, 1542, ["layers", "of", "convolution", "and", "deconvolution"], "Method"], [1546, 1548, ["semantic", "segmentation"], "Task"], [1553, 1555, ["convolutional", "layers"], "Method"], [1573, 1575, ["deconvolutional", "layers"], "Method"], [1588, 1589, ["deconvolution"], "Method"], [1590, 1591, ["convolution"], "Method"], [1605, 1610, ["deep", "fully", "convolutional", "neural", "networks"], "Method"], [1619, 1622, ["fully", "convolution", "model"], "Method"], [1628, 1631, ["fully", "convolution", "case"], "Task"], [1671, 1672, ["convolution"], "Method"], [1679, 1680, ["deconvolution"], "Method"], [1690, 1700, ["5", "-", "layer", "and", "10", "-", "layer", "fully", "convolutional", "network"], "Method"], [1705, 1708, ["convolution", "and", "deconvolution"], "Method"], [1716, 1719, ["fully", "convolutional", "networks"], "Method"], [1748, 1749, ["convolutional"], "Method"], [1755, 1756, ["deconvolutional"], "Method"], [1771, 1772, ["SGD"], "Method"], [1773, 1775, ["learning", "rate"], "Metric"], [1777, 1779, ["noise", "level"], "Metric"], [1783, 1784, ["PSNR"], "Metric"], [1786, 1787, ["deconvolution"], "Method"], [1791, 1794, ["fully", "convolutional", "counterpart"], "Method"], [1800, 1803, ["fully", "convolutional", "network"], "Method"], [1804, 1808, ["noise", "layer", "by", "layer"], "Method"], [1817, 1818, ["convolution"], "Method"], [1824, 1825, ["deconvolution"], "Method"], [1834, 1835, ["subsubsection"], "Method"], [1836, 1838, ["Skip", "connections"], "Method"], [1845, 1846, ["deconvolution"], "Method"], [1861, 1863, ["shallow", "networks"], "Method"], [1869, 1870, ["convolution"], "Method"], [1871, 1872, ["deconvolution"], "Method"], [1891, 1893, ["max", "pooling"], "Method"], [1894, 1895, ["deconvolution"], "Method"], [1911, 1912, ["convolution"], "Method"], [1934, 1936, ["deeper", "networks"], "Method"], [1966, 1968, ["highway", "networks"], "Method"], [1969, 1972, ["deep", "residual", "networks"], "Method"], [1980, 1984, ["convolutional", "and", "deconvolutional", "layers"], "Method"], [2031, 2032, ["deconvolution"], "Method"], [2044, 2046, ["skip", "connections"], "Method"], [2053, 2054, ["deconvolution"], "Task"], [2082, 2084, ["deeper", "network"], "Method"], [2094, 2097, ["skip", "layer", "connections"], "Method"], [2134, 2136, ["deconvolutional", "layers"], "Method"], [2169, 2171, ["learning", "strategy"], "Method"], [2178, 2182, ["encoding", "-", "decoding", "network"], "Method"], [2184, 2185, ["training"], "Task"], [2194, 2196, ["convolutional", "layers"], "Method"], [2198, 2201, ["mirrored", "deconvolutional", "layers"], "Method"], [2230, 2231, ["restoration"], "Task"], [2239, 2242, ["deep", "highway", "networks"], "Method"], [2244, 2255, ["feed", "-", "forward", "long", "short", "-", "term", "memory", "(", "LSTMs", ")"], "Method"], [2256, 2258, ["forget", "gates"], "Method"], [2264, 2267, ["deep", "residual", "network"], "Method"], [2268, 2274, ["feed", "-", "forward", "LSTMs", "without", "gates"], "Method"], [2278, 2281, ["deep", "residual", "networks"], "Method"], [2290, 2294, ["feed", "-", "forward", "LSTMs"], "Method"], [2372, 2373, ["convolution"], "Method"], [2394, 2395, ["convolution"], "Method"], [2397, 2398, ["deconvolution"], "Method"], [2444, 2445, ["deconvolution"], "Task"], [2465, 2468, ["recovering", "image", "details"], "Task"], [2469, 2470, ["deconvolution"], "Task"], [2527, 2529, ["end", "mapping"], "Task"], [2551, 2553, ["layer", "networks"], "Method"], [2560, 2563, ["30", "layer", "networks"], "Method"], [2573, 2575, ["training", "loss"], "Metric"], [2596, 2597, ["loss"], "Metric"], [2604, 2607, ["deep", "residual", "networks"], "Method"], [2665, 2670, ["pixel", "-", "wise", "prediction", "problems"], "Task"], [2804, 2805, ["PSNR"], "Metric"], [2900, 2906, ["end", "-", "to", "-", "end", "mapping"], "Task"], [2920, 2924, ["convolutional", "and", "deconvolutional", "kernels"], "Method"], [2973, 2979, ["Mean", "Squared", "Error", "(", "MSE", ")"], "Metric"], [2987, 2988, ["Caffe"], "Method"], [2996, 2997, ["Adam"], "Method"], [2998, 3000, ["learning", "rate"], "Method"], [3001, 3002, ["training"], "Task"], [3006, 3009, ["stochastic", "gradient", "descent"], "Method"], [3010, 3011, ["SGD"], "Method"], [3015, 3017, ["learning", "rate"], "Metric"], [3031, 3033, ["learning", "rate"], "Metric"], [3063, 3064, ["denoising"], "Task"], [3069, 3072, ["super", "-", "resolution"], "Task"], [3080, 3083, ["Berkeley", "Segmentation", "Dataset"], "Material"], [3084, 3085, ["BSD"], "Material"], [3108, 3109, ["denoising"], "Task"], [3133, 3136, ["super", "-", "resolution"], "Task"], [3237, 3238, ["segmentation"], "Task"], [3240, 3242, ["filter", "kernels"], "Method"], [3338, 3339, ["Denoising"], "Task"], [3375, 3376, ["BM3D"], "Method"], [3377, 3378, ["NCSR"], "Method"], [3379, 3380, ["EPLL"], "Method"], [3381, 3382, ["PCLR"], "Method"], [3383, 3384, ["PDPD"], "Method"], [3385, 3386, ["WMMN"], "Method"], [3393, 3396, ["super", "-", "resolution"], "Task"], [3402, 3403, ["SRCNN"], "Method"], [3404, 3405, ["NBSRF"], "Method"], [3406, 3407, ["CSCN"], "Method"], [3408, 3409, ["CSC"], "Method"], [3410, 3411, ["TSE"], "Method"], [3412, 3414, ["ARFL", "+"], "Method"], [3418, 3419, ["Set5"], "Material"], [3422, 3423, ["BSD100"], "Material"], [3436, 3443, ["Peak", "Signal", "-", "to", "-", "Noise", "Ratio"], "Metric"], [3444, 3445, ["PSNR"], "Metric"], [3447, 3449, ["Structural", "SIMilarity"], "Metric"], [3450, 3451, ["SSIM"], "Metric"], [3466, 3469, ["RED", "-", "Net"], "Method"], [3475, 3476, ["RED10"], "Method"], [3478, 3482, ["convolutional", "and", "deconvolutional", "layers"], "Method"], [3483, 3484, ["shortcuts"], "Method"], [3485, 3486, ["RED20"], "Method"], [3488, 3492, ["convolutional", "and", "deconvolutional", "layers"], "Method"], [3493, 3494, ["shortcuts"], "Method"], [3496, 3497, ["RED30"], "Method"], [3499, 3503, ["convolutional", "and", "deconvolutional", "layers"], "Method"], [3504, 3505, ["shortcuts"], "Method"], [3508, 3510, ["Image", "Denoising"], "Task"], [3521, 3522, ["PSNR"], "Metric"], [3523, 3524, ["SSIM"], "Metric"], [3549, 3555, ["10", "layer", "convolutional", "and", "deconvolutional", "network"], "Method"], [3575, 3578, ["convolution", "and", "deconvolution"], "Method"], [3579, 3580, ["denoising"], "Task"], [3609, 3610, ["denoising"], "Task"], [3618, 3619, ["WNNM"], "Method"], [3628, 3630, ["noise", "levels"], "Metric"], [3642, 3643, ["WNNM"], "Method"], [3653, 3654, ["PCLR"], "Method"], [3705, 3706, ["SSIM"], "Metric"], [3709, 3710, ["BSD200"], "Method"], [3729, 3732, ["BSD", "-", "200"], "Method"], [3745, 3746, ["PSNR"], "Metric"], [3747, 3748, ["SSIM"], "Metric"], [3772, 3773, ["denoising"], "Task"], [3792, 3793, ["PSNR"], "Metric"], [3794, 3795, ["WNNM"], "Method"], [3798, 3802, ["Image", "super", "-", "resolution"], "Task"], [3805, 3806, ["Set5"], "Material"], [3815, 3819, ["10", "-", "layer", "network"], "Method"], [3831, 3833, ["deeper", "networks"], "Method"], [3835, 3839, ["30", "-", "layer", "network"], "Method"], [3844, 3845, ["CSCN"], "Method"], [3883, 3884, ["Set5"], "Material"], [3892, 3895, ["30", "layer", "network"], "Method"], [3897, 3898, ["PSNR"], "Metric"], [3902, 3903, ["CSCN"], "Method"], [3912, 3913, ["BSD100"], "Material"], [3927, 3928, ["Set5"], "Material"], [3935, 3936, ["CSCN"], "Method"], [3947, 3950, ["10", "layer", "network"], "Method"], [3952, 3954, ["deeper", "network"], "Method"], [3988, 3989, ["denoising"], "Task"], [3990, 3993, ["super", "-", "resolution"], "Task"], [3996, 4000, ["30", "-", "layer", "network"], "Method"], [4054, 4055, ["denoising"], "Task"], [4056, 4059, ["super", "-", "resolution"], "Task"], [4079, 4080, ["CSCN"], "Method"], [4084, 4087, ["super", "-", "resolution"], "Task"], [4112, 4115, ["super", "-", "resolution"], "Task"], [4137, 4142, ["deep", "encoding", "and", "decoding", "framework"], "Method"], [4143, 4145, ["image", "restoration"], "Task"], [4146, 4147, ["Convolution"], "Method"], [4148, 4149, ["deconvolution"], "Method"], [4154, 4156, ["restoration", "problem"], "Task"], [4184, 4186, ["optimization", "difficulty"], "Task"], [4227, 4230, ["super", "-", "resolution"], "Task"], [299, 300, ["denoising"], "Task"], [339, 340, ["denoising"], "Task"], [773, 774, ["denoising"], "Task"], [833, 834, ["denoising"], "Task"], [881, 882, ["denoising"], "Task"], [916, 917, ["denoising"], "Task"], [929, 930, ["denoising"], "Task"], [945, 946, ["CNN"], "Method"], [948, 949, ["denoising"], "Task"], [958, 961, ["super", "-", "resolution"], "Task"], [1016, 1017, ["denoising"], "Task"], [1025, 1026, ["CNN"], "Method"], [2261, 2262, ["CNN"], "Method"], [2358, 2359, ["denoising"], "Task"], [2871, 2872, ["denoising"], "Task"], [2878, 2881, ["super", "-", "resolution"], "Task"], [3181, 3182, ["denoising"], "Task"], [3183, 3186, ["super", "-", "resolution"], "Task"], [3298, 3299, ["denoising"], "Task"], [3300, 3303, ["super", "-", "resolution"], "Task"], [3316, 3317, ["denoising"], "Task"], [3318, 3321, ["super", "-", "resolution"], "Task"], [4110, 4111, ["denoising"], "Task"], [4225, 4226, ["denoising"], "Task"]]} {"doc_id": "193089d56758ab88391d846edd08d359b1f9a863", "text": ["document", ":", "A", "Discriminatively", "Learned", "CNN", "Embedding", "for", "Person", "Re", "-", "identification", "In", "this", "paper", ",", "we", "revisit", "two", "popular", "convolutional", "neural", "networks", "(", "CNN", ")", "in", "person", "re", "-", "identification", "(", "re", "-", "ID", ")", ",", "i.e.", ",", "verification", "and", "identification", "models", ".", "The", "two", "models", "have", "their", "respective", "advantages", "and", "limitations", "due", "to", "different", "loss", "functions", ".", "In", "this", "paper", ",", "we", "shed", "light", "on", "how", "to", "combine", "the", "two", "models", "to", "learn", "more", "discriminative", "pedestrian", "descriptors", ".", "Specifically", ",", "we", "propose", "a", "siamese", "network", "that", "simultaneously", "computes", "the", "identification", "loss", "and", "verification", "loss", ".", "Given", "a", "pair", "of", "training", "images", ",", "the", "network", "predicts", "the", "identities", "of", "the", "two", "input", "images", "and", "whether", "they", "belong", "to", "the", "same", "identity", ".", "Our", "network", "learns", "a", "discriminative", "embedding", "and", "a", "similarity", "measurement", "at", "the", "same", "time", ",", "thus", "making", "full", "usage", "of", "the", "re", "-", "ID", "annotations", ".", "Our", "method", "can", "be", "easily", "applied", "on", "different", "pre", "-", "trained", "networks", ".", "Albeit", "simple", ",", "the", "learned", "embedding", "improves", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "two", "public", "person", "re", "-", "ID", "benchmarks", ".", "Further", ",", "we", "show", "our", "architecture", "can", "also", "be", "applied", "in", "image", "retrieval", ".", "Large", "-", "scale", "Person", "Re", "-", "identification", ",", "Convolutional", "Neural", "Networks", ".", "section", ":", "Introduction", "Person", "re", "-", "identification", "(", "re", "-", "ID", ")", "is", "usually", "viewed", "as", "an", "image", "retrieval", "problem", ",", "which", "matches", "pedestrians", "from", "different", "cameras", ".", "Given", "a", "person", "-", "of", "-", "interest", "(", "query", ")", ",", "person", "re", "-", "ID", "determines", "whether", "the", "person", "has", "been", "observed", "by", "another", "camera", ".", "Recent", "progress", "in", "this", "area", "has", "been", "due", "to", "two", "factors", ":", "1", ")", "the", "availability", "of", "the", "large", "-", "scale", "pedestrian", "datasets", ".", "The", "datasets", "contain", "the", "general", "visual", "variance", "of", "pedestrian", "and", "provide", "a", "comprehensive", "evaluation", ".", "2", ")", "the", "learned", "embedding", "of", "pedestrian", "using", "a", "convolutional", "neural", "network", "(", "CNN", ")", ".", "Recently", ",", "the", "convolutional", "neural", "network", "(", "CNN", ")", "has", "shown", "potential", "for", "learning", "state", "-", "of", "-", "the", "-", "art", "feature", "embeddings", "or", "deep", "metrics", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "there", "are", "two", "major", "types", "of", "CNN", "structures", ",", "i.e.", ",", "verification", "models", "and", "identification", "models", ".", "The", "two", "models", "are", "different", "in", "terms", "of", "input", ",", "feature", "extraction", "and", "loss", "function", "for", "training", ".", "Our", "motivation", "is", "to", "combine", "the", "strengths", "of", "the", "two", "models", "and", "learn", "a", "more", "discriminative", "pedestrian", "embedding", ".", "Verification", "models", "take", "a", "pair", "of", "images", "as", "input", "and", "determine", "whether", "they", "belong", "to", "the", "same", "person", "or", "not", ".", "A", "number", "of", "previous", "works", "treat", "person", "re", "-", "ID", "as", "a", "binary", "-", "class", "classification", "task", "or", "a", "similarity", "regression", "task", ".", "Given", "a", "label", ",", "the", "verification", "network", "forces", "two", "images", "of", "the", "same", "person", "to", "be", "mapped", "to", "nearby", "points", "in", "the", "feature", "space", ".", "If", "the", "images", "are", "of", "different", "people", ",", "the", "points", "are", "far", "apart", ".", "However", ",", "the", "major", "problem", "in", "the", "verification", "models", "is", "that", "they", "only", "use", "weak", "re", "-", "ID", "labels", ",", "and", "do", "not", "take", "all", "the", "annotated", "information", "into", "consideration", ".", "Therefore", ",", "the", "verification", "network", "lacks", "the", "consideration", "of", "the", "relationship", "between", "the", "image", "pairs", "and", "other", "images", "in", "the", "dataset", ".", "In", "the", "attempt", "to", "take", "full", "advantages", "of", "the", "re", "-", "ID", "labels", ",", "identification", "models", "which", "treat", "person", "re", "-", "identification", "as", "a", "multi", "-", "class", "recognition", "task", ",", "are", "employed", "for", "feature", "learning", ".", "They", "directly", "learn", "the", "non", "-", "linear", "functions", "from", "an", "input", "image", "to", "the", "person", "ID", "and", "the", "cross", "-", "entropy", "loss", "is", "used", "following", "the", "final", "layer", ".", "During", "testing", ",", "the", "feature", "is", "extracted", "from", "a", "fully", "connected", "layer", "and", "then", "normalized", ".", "The", "similarity", "of", "two", "images", "is", "thus", "computed", "by", "the", "Euclidean", "distance", "between", "their", "normalized", "CNN", "embeddings", ".", "The", "major", "drawback", "of", "the", "identification", "model", "is", "that", "the", "training", "objective", "is", "different", "from", "the", "testing", "procedure", ",", "i.e.", ",", "it", "does", "not", "account", "for", "the", "similarity", "measurement", "between", "image", "pairs", ",", "which", "can", "be", "problematic", "during", "the", "pedestrian", "retrieval", "process", ".", "The", "above", "-", "mentioned", "observations", "demonstrate", "that", "the", "two", "types", "of", "models", "have", "complementary", "advantages", "and", "limitations", "as", "shown", "in", "Table", "[", "reference", "]", ".", "Motivated", "by", "these", "properties", ",", "this", "work", "proposes", "to", "combine", "the", "strengths", "of", "the", "two", "networks", "and", "leverage", "their", "complementary", "nature", "to", "improve", "the", "discriminative", "ability", "of", "the", "learned", "embeddings", ".", "The", "proposed", "model", "is", "a", "siamese", "network", "that", "predicts", "person", "identities", "and", "similarity", "scores", "at", "the", "same", "time", ".", "Compared", "to", "previous", "networks", ",", "we", "take", "full", "advantages", "of", "the", "annotated", "data", "in", "terms", "of", "pair", "-", "wise", "similarity", "and", "image", "identities", ".", "During", "testing", ",", "the", "final", "convolutional", "activations", "are", "extracted", "for", "Euclidepdfan", "distance", "based", "pedestrian", "retrieval", ".", "To", "summarize", ",", "our", "contributions", "are", ":", "We", "propose", "a", "siamese", "network", "that", "has", "two", "losses", ":", "identification", "loss", "and", "verification", "loss", ".", "This", "network", "simultaneously", "learns", "a", "discriminative", "CNN", "embedding", "and", "a", "similarity", "metric", ",", "thus", "improving", "pedestrian", "retrieval", "accuracy", ".", "We", "report", "competitive", "accuracy", "compared", "to", "the", "state", "-", "of", "-", "art", "methods", "on", "two", "large", "-", "scale", "person", "re", "-", "ID", "datasets", "(", "Market1501", "and", "CUHK03", ")", "and", "one", "instance", "retrieval", "dataset", "(", "Oxford5k", ")", ".", "The", "paper", "is", "organized", "as", "follows", ".", "We", "first", "review", "some", "related", "works", "in", "Section", "[", "reference", "]", ".", "In", "Section", "[", "reference", "]", ",", "we", "describe", "how", "we", "combine", "the", "two", "losses", "and", "define", "the", "CNN", "structure", ".", "The", "implementation", "details", "are", "provided", ".", "In", "Section", "[", "reference", "]", ",", "we", "present", "the", "experimental", "results", "on", "two", "large", "-", "scale", "person", "re", "-", "identification", "datasets", "and", "one", "instance", "retrieval", "dataset", ".", "We", "conclude", "this", "paper", "in", "Section", "[", "reference", "]", ".", "section", ":", "Related", "Work", "In", "this", "section", "we", "describe", "previous", "works", "relevant", "to", "the", "approach", "discussed", "in", "this", "paper", ".", "They", "are", "mainly", "based", "on", "verification", "models", "or", "identification", "models", ".", "subsection", ":", "Verification", "Models", "In", "1993", ",", "Bromley", "et", "al", ".", "first", "used", "verification", "models", "to", "deep", "metric", "learning", "in", "signature", "verification", ".", "Verification", "models", "usually", "take", "a", "pair", "of", "images", "as", "input", "and", "output", "a", "similarity", "score", "by", "calculating", "the", "cosine", "distance", "between", "low", "-", "dimensional", "features", ",", "which", "can", "be", "penalized", "by", "the", "contrastive", "loss", ".", "Recently", "researchers", "have", "begun", "to", "apply", "verification", "models", "to", "person", "re", "-", "identification", "with", "a", "focus", "on", "data", "augmentation", "and", "image", "matching", ".", "Yi", "et", "al", ".", "split", "a", "pedestrian", "image", "into", "three", "horizontal", "parts", "and", "train", "three", "part", "-", "CNNs", "to", "extract", "features", ".", "The", "similarity", "of", "two", "images", "is", "computed", "by", "the", "cosine", "distance", "of", "their", "features", ".", "Similarly", ",", "Cheng", "et", "al", ".", "split", "the", "convolutional", "map", "into", "four", "parts", "and", "fuse", "the", "part", "features", "with", "the", "global", "features", ".", "Li", "et", "al", ".", "add", "a", "patch", "-", "matching", "layer", "that", "multiplies", "the", "activation", "of", "two", "images", "in", "different", "horizontal", "stripes", ".", "They", "use", "it", "to", "find", "similar", "locations", "and", "treat", "similarity", "regression", "as", "binary", "-", "class", "penalized", "by", "softmax", "loss", ".", "Later", ",", "Ahmed", "et", "al", ".", "improve", "the", "verification", "model", "by", "adding", "a", "different", "matching", "layer", "that", "compares", "the", "activation", "of", "two", "images", "in", "neighboring", "pixels", ".", "Besides", ",", "Wu", "et", "al", ".", "use", "smaller", "filters", "and", "a", "deeper", "network", "to", "extract", "features", ".", "Varior", "et", "al", ".", "combine", "CNN", "with", "some", "gate", "functions", ",", "similar", "to", "long", "-", "short", "-", "term", "memory", "(", "LSTM", ")", "in", "spirit", ",", "which", "aims", "to", "adaptively", "focus", "on", "the", "similar", "parts", "of", "input", "image", "pairs", ".", "But", "it", "is", "limited", "by", "the", "computational", "inefficiency", "because", "the", "query", "image", "has", "to", "pair", "with", "every", "gallery", "image", "to", "pass", "through", "the", "network", ".", "Moreover", ",", "Ding", "et", "al", ".", "use", "triplet", "samples", "for", "training", "the", "network", "which", "considers", "the", "images", "from", "the", "same", "people", "and", "the", "different", "people", "at", "the", "same", "time", ".", "subsection", ":", "Identification", "Models", "Recent", "datasets", "such", "as", "CUHK03", "and", "Market1501", "provide", "large", "-", "scale", "training", "sets", ",", "which", "make", "it", "possible", "to", "train", "a", "deeper", "classification", "model", "without", "over", "-", "fitting", ".", "Every", "identity", "has", "9.6", "training", "images", "on", "average", "in", "CUHK03", "and", "has", "17.2", "images", "in", "Market1501", ".", "CNN", "can", "learn", "discriminative", "embeddings", "by", "itself", "without", "part", "-", "matching", ".", "Zheng", "et", "al", ".", "directly", "use", "a", "conventional", "fine", "-", "tuning", "approach", "on", "Market1501", ",", "PRW", "and", "MARS", "and", "outperform", "many", "recent", "results", ".", "Wu", "et", "al", ".", "combine", "CNN", "embeddings", "with", "the", "hand", "-", "crafted", "features", "in", "the", "FC", "layer", ".", "Besides", ",", "Xiao", "et", "al", ".", "jointly", "train", "a", "classification", "model", "using", "multiple", "datasets", "and", "propose", "a", "new", "dropout", "function", "to", "deal", "with", "the", "hundreds", "of", "classes", ".", "In", ",", "Xiao", "et", "al", ".", "train", "a", "classification", "model", "similar", "to", "the", "faster", "-", "RCNN", "method", "and", "automatically", "predict", "the", "location", "of", "the", "candidate", "pedestrian", "from", "the", "whole", "image", ",", "which", "alleviates", "the", "pedestrian", "detection", "errors", ".", "subsection", ":", "Verification", "-", "identification", "Models", "In", "face", "recognition", ",", "the", "“", "DeepID", "networks", "”", "train", "the", "network", "with", "the", "verification", "and", "identification", "losses", ",", "which", "is", "similar", "to", "our", "network", ".", "In", ",", "Sun", "et", "al", ".", "jointly", "train", "face", "identification", "and", "verification", ".", "Then", "more", "verification", "supervision", "is", "added", "into", "the", "model", "and", "a", "deeper", "network", "is", "used", ".", "Our", "method", "is", "different", "from", "their", "models", "in", "the", "following", "aspects", ".", "First", ",", "in", "face", "recognition", ",", "the", "training", "dataset", "contains", "202", ",", "599", "face", "images", "of", "10", ",", "177", "identities", "while", "the", "current", "largest", "person", "re", "-", "i", "d", "training", "dataset", "contains", "12", ",", "936", "images", "of", "751", "identities", ".", "DeepID", "networks", "apply", "contrastive", "loss", "to", "the", "verification", "problem", ",", "wile", "our", "model", "uses", "the", "cross", "-", "entropy", "loss", ".", "We", "find", "that", "the", "contrastive", "loss", "leads", "to", "over", "-", "fitting", "when", "the", "number", "of", "images", "is", "limited", ".", "In", "the", "experiment", ",", "we", "show", "the", "proposed", "method", "learns", "more", "robust", "person", "representative", "and", "outperforms", "using", "contrastive", "loss", ".", "Second", ",", "dropout", "can", "not", "be", "applied", "on", "the", "embedding", "before", "the", "contrastive", "loss", ",", "which", "introduces", "zero", "values", "at", "random", "locations", ".", "On", "the", "contrary", ",", "we", "can", "add", "dropout", "regularization", "on", "the", "embedding", "in", "the", "proposed", "model", ".", "Third", ",", "the", "DeepID", "networks", "are", "trained", "from", "scratch", ",", "while", "our", "model", "benefits", "from", "the", "networks", "pretrained", "on", "ImageNet", ".", "Finally", ",", "we", "evaluate", "our", "method", "on", "the", "tasks", "of", "person", "re", "-", "ID", "and", "instance", "retrieval", ",", "providing", "more", "insights", "in", "the", "verification", "-", "classification", "models", ".", "section", ":", "Proposed", "Method", "subsection", ":", "Preview", "Fig", ".", "[", "reference", "]", "(", "a", ")", "and", "Fig", ".", "[", "reference", "]", "(", "b", ")", "illustrate", "the", "relational", "graph", "built", "by", "verification", "and", "identification", "models", ".", "In", "a", "sample", "batch", "of", "size", ",", "red", "edges", "represent", "the", "positive", "pairs", "(", "the", "same", "person", ")", "and", "blue", "edges", "represent", "the", "negative", "pairs", "(", "different", "persons", ")", ".", "The", "dotted", "edges", "denote", "implicit", "relationships", "built", "by", "the", "identification", "loss", "and", "the", "solid", "edges", "denote", "explicit", "relationships", "built", "by", "the", "verification", "loss", ".", "In", "verification", "models", ",", "there", "are", "several", "operations", "between", "the", "two", "inputs", ".", "The", "explicit", "relationship", "between", "data", "is", "built", "by", "the", "pair", "-", "wise", "comparison", ",", "such", "as", "part", "matching", "or", "contrastive", "loss", ".", "For", "example", ",", "contrastive", "loss", "directly", "calculates", "the", "Euclidean", "distance", "between", "two", "embeddings", ".", "In", "identification", "models", ",", "the", "input", "is", "independent", "to", "each", "other", ".", "But", "there", "is", "implicit", "relationship", "between", "the", "learned", "embeddings", "built", "by", "the", "cross", "-", "entropy", "loss", ".", "The", "cross", "-", "entropy", "loss", "can", "be", "formulated", "as", ".", "is", "the", "weight", "of", "the", "linear", "function", ".", "are", "the", "embeddings", "of", "the", "two", "images", "from", "the", "same", "class", ".", "To", "maximize", ",", ",", "the", "network", "converges", "when", "and", "have", "similar", "vector", "direction", "with", ".", "In", ",", "similar", "observation", "and", "visualization", "are", "shown", ".", "So", "the", "learned", "embeddings", "are", "eventually", "close", "for", "images", "within", "the", "same", "class", "and", "far", "away", "for", "images", "in", "the", "different", "classes", ".", "The", "relationship", "is", "implicitly", "built", "between", "and", "bridged", "by", "the", "weight", ".", "Due", "to", "the", "usage", "of", "the", "weak", "labels", ",", "verification", "models", "take", "limited", "relationships", "into", "consideration", ".", "On", "the", "other", "hand", ",", "classification", "models", "do", "not", "explicitly", "consider", "similarity", "measurements", ".", "Fig", ".", "[", "reference", "]", "(", "c", ")", "illustrates", "how", "our", "model", "works", "in", "a", "batch", ".", "We", "benefit", "from", "simultaneously", "considering", "the", "verification", "and", "identification", "losses", ".", "The", "proposed", "model", "thus", "combines", "the", "strength", "of", "the", "two", "models", "(", "see", "Table", "[", "reference", "]", ")", ".", "subsection", ":", "Overall", "Network", "Our", "network", "is", "basically", "a", "convolutional", "siamese", "network", "that", "combines", "the", "verification", "and", "identification", "losses", ".", "Fig", ".", "[", "reference", "]", "briefly", "illustrates", "the", "architecture", "of", "the", "proposed", "network", ".", "Given", "an", "input", "pair", "of", "images", "resized", "to", ",", "the", "proposed", "network", "simultaneously", "predicts", "the", "IDs", "of", "the", "two", "images", "and", "the", "similarity", "score", ".", "The", "network", "consists", "of", "two", "ImageNet", "pre", "-", "trained", "CNN", "models", ",", "three", "additional", "Convolutional", "Layers", ",", "one", "Square", "Layer", "and", "three", "losses", ".", "It", "is", "supervised", "by", "the", "identification", "label", "and", "the", "verification", "label", ".", "The", "pre", "-", "trained", "CNN", "model", "can", "be", "CaffeNet", ",", "VGG16", "or", "ResNet", "-", "50", ",", "from", "which", "we", "have", "removed", "the", "final", "fully", "-", "connected", "(", "FC", ")", "layer", ".", "The", "re", "-", "ID", "performance", "of", "the", "three", "models", "is", "comprehensively", "evaluated", "in", "Section", "[", "reference", "]", ".", "Here", ",", "we", "do", "not", "provide", "detailed", "descriptions", "of", "the", "architecture", "of", "the", "CNN", "models", "and", "only", "take", "CaffeNet", "as", "an", "example", "in", "the", "following", "subsections", ".", "The", "three", "optimization", "objectives", "include", "two", "identification", "losses", "and", "one", "verification", "loss", ".", "We", "use", "the", "final", "convolutional", "activations", "as", "the", "discriminative", "descriptor", "for", "person", "re", "-", "ID", ",", "which", "is", "directly", "supervised", "by", "three", "objectives", ".", "subsection", ":", "Identification", "Loss", "There", "are", "two", "CaffeNets", "in", "our", "architecture", ".", "They", "share", "weights", "and", "predict", "the", "two", "identity", "labels", "of", "the", "input", "image", "pair", "simultaneously", ".", "In", "order", "to", "fine", "-", "tune", "the", "network", "on", "a", "new", "dataset", ",", "we", "replace", "the", "final", "fully", "-", "connected", "layer", "(", "1", ",", "000", "-", "dim", ")", "of", "the", "pre", "-", "trained", "CNN", "model", "with", "a", "convolutional", "layer", ".", "The", "number", "of", "the", "training", "identities", "in", "Market", "-", "1501", "is", "751", ".", "So", "this", "convolutional", "layer", "has", "kernels", "of", "size", "connected", "to", "the", "output", "of", "CaffeNet", "and", "then", "we", "add", "a", "softmax", "unit", "to", "normalize", "the", "output", ".", "The", "size", "of", "the", "result", "tensor", "is", ".", "The", "Rectified", "Linear", "Unit", "(", "ReLU", ")", "is", "not", "added", "after", "this", "convolution", ".", "Similar", "to", "conventional", "multi", "-", "class", "recognition", "approaches", ",", "we", "use", "the", "cross", "-", "entropy", "loss", "for", "identity", "prediction", ",", "which", "is", "Here", "denotes", "the", "convolutional", "operation", ".", "is", "a", "tensor", ",", "is", "the", "target", "class", "and", "denotes", "the", "parameters", "of", "the", "added", "convolutional", "layer", ".", "is", "the", "predicted", "probability", ",", "is", "the", "target", "probability", ".", "for", "all", "except", ".", "subsection", ":", "Verification", "Loss", "While", "some", "previous", "works", "contain", "a", "matching", "function", "in", "the", "intermediate", "layers", ",", "our", "work", "directly", "compares", "the", "high", "-", "level", "features", "for", "similarity", "estimation", ".", "The", "high", "-", "level", "feature", "from", "the", "fine", "-", "tuned", "CNN", "has", "shown", "a", "discriminative", "ability", "and", "it", "is", "more", "compact", "than", "the", "activations", "in", "the", "intermediate", "layers", ".", "So", "in", "our", "model", ",", "the", "pedestrian", "descriptor", "in", "the", "identification", "model", "are", "directly", "supervised", "by", "the", "verification", "loss", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "we", "introduce", "a", "non", "-", "parametric", "layer", "called", "Square", "Layer", "to", "compare", "the", "high", "-", "level", "features", ".", "It", "takes", "two", "tensors", "as", "inputs", "and", "outputs", "one", "tensor", "after", "subtracting", "and", "squaring", "element", "-", "wisely", ".", "The", "Square", "Layer", "is", "denoted", "as", ",", "where", "are", "the", "4", ",", "096", "-", "dim", "embeddings", "and", "is", "the", "output", "tensor", "of", "the", "Square", "Layer", ".", "We", "then", "add", "a", "convolutional", "layer", "and", "the", "softmax", "output", "function", "to", "embed", "the", "resulting", "tensor", "to", "a", "2", "-", "dim", "vector", "(", ",", ")", "which", "represents", "the", "predicted", "probability", "of", "the", "two", "input", "images", "belonging", "to", "the", "same", "identity", ".", "The", "convolutional", "layer", "takes", "as", "input", "and", "filters", "it", "with", "kernels", "of", "size", ".", "The", "ReLU", "is", "not", "added", "after", "this", "convolution", ".", "We", "treat", "pedestrian", "verification", "as", "a", "binary", "classification", "problem", "and", "use", "the", "cross", "-", "entropy", "loss", "that", "is", "similar", "to", "the", "one", "in", "the", "identification", "loss", ",", "which", "is", "Here", "are", "the", "two", "tensors", "of", "size", ".", "is", "the", "target", "class", "(", "same", "/", "different", ")", ",", "denotes", "the", "parameters", "of", "the", "added", "convolutional", "layer", "and", "is", "the", "predicted", "probability", ".", "If", "the", "image", "pair", "depicts", "the", "same", "person", ",", ";", "otherwise", ",", ".", "Departing", "from", ",", "we", "do", "not", "use", "the", "contrastive", "loss", ".", "On", "the", "one", "hand", ",", "the", "contrastive", "loss", ",", "as", "a", "regression", "loss", ",", "forces", "the", "same", "-", "class", "embeddings", "to", "be", "as", "close", "as", "possible", ".", "It", "may", "make", "the", "model", "over", "-", "fitting", "because", "the", "number", "of", "training", "of", "each", "identity", "is", "limited", "in", "person", "re", "-", "ID", ".", "On", "the", "other", "hand", ",", "dropout", ",", "which", "introduces", "zero", "values", "at", "random", "locations", ",", "can", "not", "be", "applied", "on", "the", "embedding", "before", "the", "contrastive", "loss", ".", "But", "the", "cross", "-", "entropy", "loss", "in", "our", "model", "can", "work", "with", "dropout", "to", "regularize", "the", "model", ".", "In", "Section", "[", "reference", "]", ",", "we", "show", "that", "the", "result", "using", "contrastive", "loss", "is", "4.39", "%", "and", "6.55", "%", "lower", "than", "the", "one", "using", "the", "cross", "-", "entropy", "loss", "on", "rank", "-", "1", "accuracy", "and", "mAP", "respectively", ".", "subsection", ":", "Identification", "vs.", "Verification", "The", "proposed", "network", "is", "trained", "to", "minimize", "the", "three", "cross", "-", "entropy", "losses", "jointly", ".", "To", "figure", "out", "which", "objective", "contributes", "more", ",", "we", "train", "the", "identification", "model", "and", "verification", "model", "separately", ".", "Following", "the", "learning", "rate", "setting", "in", "Section", "[", "reference", "]", ",", "we", "train", "the", "models", "until", "convergence", ".", "We", "also", "train", "the", "network", "with", "the", "two", "losses", "jointly", "until", "two", "objectives", "both", "converge", ".", "As", "the", "quantitative", "results", "shown", "in", "Table", "[", "reference", "]", ",", "the", "fine", "-", "tuned", "CNN", "model", "with", "two", "kinds", "of", "losses", "outperforms", "the", "one", "trained", "individually", ".", "This", "result", "has", "been", "confirmed", "on", "the", "three", "different", "network", "structures", ".", "Further", ",", "we", "visualize", "the", "intermediate", "feature", "maps", "that", "are", "trained", "using", "ResNet", "-", "50", "as", "the", "pretrained", "model", "and", "try", "to", "find", "the", "differences", "between", "identification", "loss", "and", "verification", "loss", ".", "We", "select", "three", "test", "images", "in", "the", "Market1501", ".", "One", "image", "is", "considered", "to", "be", "well", "detected", "and", "the", "other", "two", "images", "are", "not", "well", "aligned", ".", "Given", "one", "image", "as", "input", ",", "we", "get", "its", "activation", "in", "the", "intermediate", "layer", "“", "res4fx", "”", ",", "the", "size", "of", "which", "is", ".", "We", "visualize", "the", "sum", "of", "several", "activation", "maps", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "the", "identification", "and", "the", "verification", "networks", "exhibit", "different", "activation", "patterns", "to", "the", "pedestrian", ".", "We", "find", "that", "if", "we", "use", "only", "one", "kind", "of", "loss", ",", "the", "network", "tends", "to", "find", "one", "discriminative", "part", ".", "The", "proposed", "model", "takes", "advantages", "of", "both", "networks", ",", "so", "the", "new", "activation", "map", "is", "mostly", "a", "union", "of", "the", "two", "individual", "maps", ".", "This", "also", "illustrates", "the", "complementary", "nature", "of", "the", "two", "baseline", "networks", ".", "The", "proposed", "model", "makes", "more", "neurons", "activated", ".", "Moreover", ",", "as", "shown", "in", "Fig", ".", "[", "reference", "]", "we", "visualize", "the", "embedding", "by", "plot", "them", "to", "the", "2", "-", "dimension", "map", ".", "In", "regard", "to", "Fig", ".", "[", "reference", "]", ",", "we", "find", "the", "network", "usually", "has", "strong", "attention", "on", "the", "center", "part", "of", "the", "human", "(", "usually", "clothes", ")", "and", "it", "also", "illustrates", "the", "color", "of", "the", "clothes", "is", "the", "major", "clue", "for", "the", "person", "re", "-", "identification", ".", "subsection", ":", "Training", "and", "Optimization", "Input", "preparation", ".", "We", "resize", "all", "the", "training", "images", "to", ".", "The", "mean", "image", "computed", "from", "all", "the", "training", "images", "is", "subtracted", "from", "all", "the", "images", ".", "During", "training", ",", "all", "the", "images", "are", "randomly", "cropped", "to", "for", "CaffeNet", "and", "mirrored", "horizontally", ".", "For", "ResNet", "-", "50", "and", "VGG16", ",", "we", "randomly", "crop", "images", "to", ".", "We", "shuffle", "the", "dataset", "and", "use", "a", "random", "order", "of", "the", "images", ".", "Then", "we", "sample", "another", "image", "from", "the", "same", "/", "different", "class", "to", "compose", "a", "positive", "/", "negative", "pair", ".", "The", "initial", "ratio", "between", "negative", "pairs", "and", "positive", "pairs", "is", "to", "alleviate", "the", "prediction", "bias", "and", "we", "multiple", "it", "by", "a", "factor", "of", "every", "epoch", "until", "it", "reaches", ",", "since", "the", "number", "of", "positive", "pairs", "is", "so", "limited", "that", "the", "network", "risks", "over", "-", "fitting", ".", "Training", ".", "We", "use", "the", "Matconvnet", "package", "for", "training", "and", "testing", "the", "embedding", "with", "CaffeNet", ",", "VGG16", "and", "ResNet", "-", "50", ",", "respectively", ".", "The", "maximum", "number", "of", "training", "epochs", "is", "set", "to", "75", "for", "ResNet", "-", "50", ",", "65", "for", "VGG16net", "and", "155", "for", "CaffeNet", ".", "The", "batch", "size", "(", "in", "image", "pairs", ")", "is", "set", "to", "128", "for", "CaffeNet", ",", "48", "for", "VGG16", "and", "ResNet", "-", "50", ".", "The", "learning", "rate", "is", "initialized", "as", "0.001", "and", "then", "set", "to", "0.0001", "for", "the", "final", "5", "epochs", ".", "We", "adopt", "the", "mini", "-", "batch", "stochastic", "gradient", "descent", "(", "SGD", ")", "to", "update", "the", "parameters", "of", "the", "network", ".", "There", "are", "three", "objectives", "in", "our", "network", ".", "Therefore", ",", "we", "first", "compute", "all", "the", "gradients", "produced", "by", "every", "objectives", "respectively", "and", "add", "the", "weighted", "gradients", "together", "to", "update", "the", "network", ".", "We", "assign", "a", "weight", "of", "1", "to", "the", "gradient", "produced", "by", "the", "verification", "loss", "and", "0.5", "for", "the", "two", "gradients", "produced", "by", "two", "identification", "losses", ".", "Moreover", ",", "we", "insert", "the", "dropout", "function", "before", "the", "final", "convolutional", "layer", ".", "Testing", ".", "We", "adopt", "an", "efficient", "method", "to", "extract", "features", "as", "well", "as", "the", "activation", "in", "the", "intermediate", "layer", ".", "Because", "two", "CaffeNet", "share", "weights", ",", "our", "model", "has", "nearly", "the", "same", "memory", "consumption", "with", "the", "pretrained", "model", ".", "So", "we", "extract", "features", "by", "only", "activating", "one", "fine", "-", "tuned", "model", ".", "Given", "a", "image", ",", "we", "feed", "forward", "the", "image", "to", "one", "CaffeNet", "in", "our", "network", "and", "obtain", "a", "4", ",", "096", "-", "dim", "pedestrian", "descriptor", ".", "Once", "the", "descriptors", "for", "the", "gallery", "sets", "are", "obtained", ",", "they", "are", "stored", "offline", ".", "Given", "a", "query", "image", ",", "its", "descriptor", "is", "extracted", "online", ".", "We", "sort", "the", "cosine", "distance", "between", "the", "query", "and", "all", "the", "gallery", "features", "to", "obtain", "the", "final", "ranking", "result", ".", "Note", "that", "the", "cosine", "distance", "is", "equivalent", "to", "Euclidean", "distance", "when", "the", "feature", "is", "L2", "-", "normalized", ".", "section", ":", "Experiments", "We", "mainly", "verify", "the", "proposed", "model", "on", "two", "large", "-", "scale", "datasets", "Market1501", "and", "CUHK03", ".", "We", "report", "the", "results", "trained", "by", "three", "network", "structures", ".", "Besides", ",", "we", "also", "report", "the", "result", "on", "Market1501", "+", "500k", "dataset", ".", "Meanwhile", ",", "the", "proposed", "architecture", "is", "also", "applied", "on", "the", "image", "retrieval", "task", ".", "We", "modify", "our", "model", "and", "test", "it", "on", "a", "popular", "image", "retrieval", "dataset", ",", "i.e.", ",", "Oxford", "Buildings", ".", "The", "performance", "is", "comparable", "to", "the", "state", "of", "the", "art", ".", "subsection", ":", "Dataset", "Market1501", "contains", "32", ",", "668", "annotated", "bounding", "boxes", "of", "1", ",", "501", "identities", ".", "Images", "of", "each", "identity", "are", "captured", "by", "at", "most", "six", "cameras", ".", "According", "to", "the", "dataset", "setting", ",", "the", "training", "set", "contains", "12", ",", "936", "cropped", "images", "of", "751", "identities", "and", "testing", "set", "contains", "19", ",", "732", "cropped", "images", "of", "750", "identities", "and", "distractors", ".", "They", "are", "directly", "detected", "by", "the", "Deformable", "Part", "Model", "(", "DPM", ")", "instead", "of", "using", "hand", "-", "drawn", "bboxes", ",", "which", "is", "closer", "to", "the", "realistic", "setting", ".", "For", "each", "query", ",", "we", "aim", "to", "retrieve", "the", "ground", "truth", "images", "from", "the", "19", ",", "732", "candidate", "images", ".", "The", "searching", "pool", "(", "gallery", ")", "is", "important", "to", "person", "re", "-", "identification", ".", "In", "the", "realistic", "setting", ",", "the", "scale", "of", "the", "gallery", "is", "usually", "large", ".", "The", "distractor", "dataset", "of", "Market1501", "provides", "extra", "500", ",", "000", "bboxes", ",", "consisting", "of", "false", "alarms", "on", "the", "background", "as", "well", "as", "the", "persons", "not", "belonging", "to", "any", "of", "the", "original", "1", ",", "501", "identities", ".", "When", "testing", ",", "we", "add", "the", "500k", "images", "to", "the", "original", "gallery", ",", "which", "makes", "the", "retrieval", "more", "difficult", ".", "CUHK03", "dataset", "contains", "14", ",", "097", "cropped", "images", "of", "1", ",", "467", "identities", "collected", "in", "the", "CUHK", "campus", ".", "Each", "identity", "is", "observed", "by", "two", "camera", "views", "and", "has", "4.8", "images", "in", "average", "for", "each", "view", ".", "The", "Author", "provides", "two", "kinds", "of", "bounding", "boxes", ".", "We", "evaluate", "our", "model", "on", "the", "bounding", "boxes", "detected", "by", "DPM", ",", "which", "is", "closer", "to", "the", "realistic", "setting", ".", "Following", "the", "setting", "of", "the", "dataset", ",", "the", "dataset", "is", "partitioned", "into", "a", "training", "set", "of", "1", ",", "367", "persons", "and", "a", "testing", "set", "of", "100", "persons", ".", "The", "experiment", "is", "repeated", "with", "20", "random", "splits", ".", "Both", "the", "single", "-", "shot", "and", "multiple", "-", "shot", "results", "will", "be", "reported", ".", "Oxford5k", "buildings", "consists", "of", "5062", "images", "collected", "from", "the", "internet", "and", "corresponding", "to", "particular", "Oxford", "landmarks", ".", "Some", "images", "have", "complex", "structures", "and", "may", "contain", "other", "buildings", ".", "The", "images", "corresponding", "to", "11", "Oxford", "landmarks", "are", "manually", "annotated", "and", "a", "set", "of", "55", "queries", "for", "11", "different", "landmarks", "are", "provided", ".", "This", "benchmark", "contains", "many", "high", "-", "resolution", "images", "and", "the", "mean", "image", "size", "of", "this", "dataset", "is", ".", "We", "use", "the", "rank", "-", "1", "accuracy", "and", "mean", "average", "precision", "(", "mAP", ")", "for", "performance", "evaluation", "on", "Market1501", "(", "+", "100k", ")", "and", "CUHK03", ",", "while", "on", "Oxford", ",", "we", "use", "mAP", ".", "subsection", ":", "Person", "Re", "-", "i", "d", "Evaluation", "Comparison", "with", "the", "CNN", "baseline", ".", "We", "train", "the", "baseline", "networks", "according", "the", "conventional", "fine", "-", "tuning", "method", ".", "The", "baseline", "networks", "are", "pretrained", "on", "ImageNet", "and", "fine", "-", "tuned", "to", "predict", "the", "person", "identities", ".", "As", "shown", "in", "Tab", ".", "[", "reference", "]", ",", "we", "obtain", "50.89", "%", ",", "65.02", "%", "and", "73.69", "%", "rank", "-", "1", "accuracy", "by", "CaffeNet", ",", "VGG16", "and", "ResNet", "-", "50", ",", "respectively", "on", "Market1501", ".", "Note", "that", "using", "the", "baseline", "alone", "exceeds", "many", "previous", "works", ".", "Our", "model", "further", "improves", "these", "baselines", "on", "Market1501", ".", "The", "improvement", "can", "be", "observed", "on", "three", "network", "architectures", ".", "To", "be", "specific", ",", "we", "obtain", "11.25", "%", ",", "5.14", "%", "and", "5.82", "%", "improvement", ",", "respectively", ",", "using", "CaffeNet", ",", "VGG16", "and", "ResNet", "-", "50", "on", "Market1501", ".", "Similarly", ",", "we", "observe", "35.8", "%", ",", "49.1", "%", "and", "71.5", "%", "baseline", "rank", "-", "1", "accuracy", "on", "CUHK03", "in", "single", "-", "shot", "setting", ".", "As", "show", "in", "Tab", ".", "[", "reference", "]", ",", "these", "baseline", "results", "exceed", "some", "previous", "works", "as", "well", ".", "We", "further", "get", "14.0", "%", ",", "22.7", "%", "and", "11.9", "%", "improvement", "on", "the", "baseline", "by", "our", "method", ".", "These", "results", "show", "that", "our", "method", "can", "work", "with", "different", "networks", "and", "improve", "their", "results", ".", "It", "indicates", "that", "the", "proposed", "model", "helps", "the", "network", "to", "learn", "more", "discriminative", "features", ".", "Cross", "-", "entropy", "vs.", "Contrastive", "loss", ".", "We", "replace", "the", "cross", "-", "entropy", "loss", "with", "the", "contrastive", "loss", "as", "used", "in", "“", "DeepID", "network", "”", ".", "However", ",", "we", "find", "a", "4.39", "%", "and", "6.55", "%", "drop", "in", "rank", "-", "1", "and", "mAP", ".", "The", "ResNet", "-", "50", "model", "using", "the", "contrastive", "loss", "has", "75.12", "%", "rank", "-", "1", "accuracy", "and", "53.32", "%", "mAP", ".", "We", "speculate", "that", "the", "contrastive", "loss", "tends", "to", "over", "-", "fit", "on", "the", "re", "-", "ID", "dataset", "because", "no", "regularization", "is", "added", "to", "the", "verification", ".", "Cross", "-", "entropy", "loss", "designed", "in", "our", "model", "can", "work", "with", "the", "dropout", "function", "and", "avoid", "the", "over", "-", "fitting", ".", "Comparison", "with", "the", "state", "of", "the", "art", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "we", "compare", "our", "method", "with", "other", "state", "-", "of", "-", "the", "-", "art", "algorithms", "in", "terms", "of", "mean", "average", "precision", "(", "mAP", ")", "and", "rank", "-", "1", "accuracy", "on", "Market1501", ".", "We", "report", "the", "single", "-", "query", "as", "well", "as", "multiple", "-", "query", "evaluation", "results", ".", "Our", "model", "(", "CaffeNet", ")", "achieves", "62.14", "%", "rank", "-", "1", "accuracy", "and", "39.61", "%", "mAP", ",", "which", "is", "comparable", "to", "the", "state", "of", "the", "art", "65.88", "%", "rank", "-", "1", "accuracy", "and", "39.55", "%", "mAP", ".", "Our", "model", "using", "ResNet", "-", "50", "produces", "the", "best", "performance", "79.51", "%", "in", "rank", "-", "1", "accuracy", "and", "59.87", "%", "in", "mAP", ",", "which", "outperforms", "other", "state", "-", "of", "-", "the", "-", "art", "algorithms", ".", "For", "CUHK03", ",", "we", "evaluate", "our", "method", "in", "the", "single", "-", "shot", "setting", "as", "shown", "in", "Tab", ".", "[", "reference", "]", ".", "There", "is", "only", "one", "right", "image", "in", "the", "searching", "pool", ".", "In", "the", "evaluation", ",", "we", "randomly", "select", "100", "images", "from", "100", "identities", "under", "the", "other", "camera", "as", "gallery", ".", "The", "proposed", "model", "yields", "83.4", "%", "rank", "-", "1", "and", "86.4", "%", "mAP", "and", "outperforms", "the", "state", "-", "of", "-", "the", "-", "art", "performance", ".", "As", "shown", "in", "Tab", ".", "[", "reference", "]", ",", "we", "also", "report", "the", "results", "in", "the", "multi", "-", "shot", "setting", ",", "which", "uses", "all", "the", "images", "from", "the", "other", "camera", "as", "gallery", "and", "the", "number", "of", "the", "gallery", "images", "is", "about", "500", ".", "We", "think", "this", "setting", "is", "much", "closer", "to", "image", "retrieval", "and", "alleviate", "the", "unstable", "effect", "caused", "by", "the", "random", "searching", "pool", "under", "single", "-", "shot", "settings", ".", "Fig", ".", "[", "reference", "]", "presents", "some", "re", "-", "ID", "samples", "on", "CUHK03", "dataset", ".", "The", "images", "in", "the", "first", "column", "are", "the", "query", "images", ".", "The", "retrieval", "images", "are", "sorted", "according", "to", "the", "similarity", "scores", "from", "left", "to", "right", ".", "Most", "ground", "-", "truth", "candidate", "images", "are", "correctly", "retrieved", ".", "Although", "the", "model", "retrieves", "some", "incorrect", "candidates", "on", "the", "third", "row", ",", "we", "find", "it", "is", "a", "reasonable", "prediction", "since", "the", "man", "with", "red", "hat", "and", "blue", "coat", "is", "similar", "to", "the", "query", ".", "The", "proposed", "model", "yields", "88.3", "%", "rank", "-", "1", "and", "85.0", "%", "mAP", "and", "also", "outperforms", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "in", "the", "multi", "-", "shot", "setting", ".", "Results", "between", "camera", "pairs", ".", "CUHK03", "only", "contains", "two", "camera", "views", ".", "So", "this", "experiment", "is", "evaluated", "on", "Market1501", "since", "it", "contains", "six", "different", "cameras", ".", "We", "provide", "the", "re", "-", "identification", "results", "between", "all", "camera", "pairs", "in", "Fig", ".", "[", "reference", "]", ".", "Although", "camera", "-", "6", "is", "a", "low", "-", "resolution", "camera", "and", "captures", "distinct", "background", "with", "the", "other", "HD", "cameras", ",", "the", "re", "-", "ID", "accuracy", "between", "camera", "6", "and", "the", "others", "is", "relatively", "high", ".", "We", "also", "compute", "the", "cross", "-", "camera", "average", "mAP", "and", "average", "rank", "-", "1", "accuracy", ":", "48.42", "%", "and", "54.42", "%", "respectively", ".", "Comparing", "to", "the", "previous", "reported", "results", ",", "i.e.", ",", "10.51", "%", "and", "13.72", "%", "in", ",", "our", "method", "largely", "improves", "the", "performance", "and", "observes", "a", "smaller", "standard", "deviation", "between", "cameras", ".", "It", "suggests", "that", "the", "discriminatively", "learned", "embedding", "works", "under", "different", "viewpoints", ".", "Further", ",", "Fig", ".", "[", "reference", "]", "shows", "the", "Barnes", "-", "Hut", "t", "-", "SNE", "visualization", "on", "the", "learned", "embeddings", "of", "our", "model", ".", "By", "the", "clustering", "algorithm", ",", "the", "persons", "wearing", "the", "similar", "-", "color", "clothes", "are", "quit", "clustered", "together", "and", "are", "apart", "from", "other", "persons", ".", "The", "learned", "pedestrian", "descriptor", "pay", "more", "attention", "to", "the", "color", "and", "it", "is", "robust", "to", "some", "illusion", "and", "viewpoint", "variations", ".", "In", "realistic", "setting", ",", "we", "think", "color", "provides", "the", "most", "important", "information", "to", "figure", "out", "the", "person", ".", "Large", "-", "scale", "experiments", ".", "The", "Market1501", "dataset", "also", "provides", "an", "additional", "distractor", "set", "with", "500k", "images", "to", "enlarge", "the", "gallery", ".", "In", "general", ",", "more", "candidate", "images", "may", "confuse", "the", "image", "retrieval", ".", "The", "re", "-", "ID", "performance", "of", "our", "model", "(", "ResNet", ")", "on", "the", "large", "-", "scale", "dataset", "is", "presented", "in", "Tab", ".", "[", "reference", "]", ".", "As", "the", "searching", "pool", "gets", "larger", ",", "the", "accuracy", "drops", ".", "With", "the", "gallery", "size", "of", ",", "we", "still", "achieve", "68.26", "%", "rank1", "accuracy", "and", "45.24", "%", "mAP", ".", "A", "relative", "drop", "24.4", "%", "from", "59.87", "%", "to", "45.24", "%", "on", "mAP", "is", "observed", ",", "compared", "to", "a", "relative", "drop", "37.88", "%", "from", "13.94", "%", "to", "8.66", "%", "in", "our", "previous", "work", ".", "Besides", ",", "we", "also", "compare", "our", "result", "with", "the", "performance", "of", "the", "ResNet", "Baseline", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "it", "is", "interesting", "that", "the", "re", "-", "ID", "precision", "of", "our", "model", "decreases", "more", "quickly", "comparing", "to", "the", "baseline", "model", ".", "We", "speculate", "that", "the", "Market1501", "training", "set", "is", "relatively", "small", "in", "covering", "the", "pedestrian", "variations", "encountered", "in", "a", "much", "larger", "test", "set", ".", "In", "fact", ",", "the", "500k", "dataset", "was", "collected", "in", "a", "different", "time", "(", "the", "same", "location", ")", "with", "the", "Market1501", "dataset", ",", "so", "the", "transfer", "effect", "is", "large", "enough", "that", "the", "learned", "embedding", "is", "inferior", "to", "the", "baseline", "on", "the", "scale", "of", "500", "k", "images", ".", "In", "the", "future", ",", "we", "will", "look", "into", "this", "interesting", "problem", "and", "design", "more", "robust", "descriptors", "for", "the", "transfer", "dataset", ".", "subsection", ":", "Instance", "Retrieval", "We", "apply", "the", "identification", "-", "verification", "model", "to", "the", "generic", "image", "retrieval", "task", ".", "Oxford5k", "is", "a", "testing", "dataset", "containing", "buildings", "in", "the", "Oxford", "University", ".", "We", "train", "the", "network", "on", "another", "scene", "dataset", "proposed", "in", ",", "which", "comprises", "of", "a", "number", "of", "buildings", "without", "overlapping", "with", "the", "Oxford5k", ".", "Similarly", ",", "the", "model", "is", "trained", "to", "not", "only", "tell", "which", "building", "the", "image", "depicts", "but", "also", "determine", "whether", "the", "two", "input", "images", "are", "from", "the", "same", "architecture", ".", "The", "training", "data", "is", "high", "-", "resolution", ".", "In", "order", "to", "obtain", "more", "information", "from", "the", "high", "-", "resolution", "building", "images", ",", "we", "modify", "the", "final", "pooling", "layer", "of", "our", "model", "to", "a", "MAC", "layer", ",", "which", "outputs", "the", "maximum", "value", "over", "the", "whole", "activation", "map", ".", "This", "layer", "helps", "us", "to", "handle", "large", "images", "without", "resizing", "them", "to", "a", "fixed", "size", "and", "output", "a", "fixed", "-", "dimension", "feature", "to", "retrieve", "the", "images", ".", "During", "training", ",", "the", "input", "image", "is", "randomly", "cropped", "to", "from", "and", "mirrored", "horizontally", ".", "During", "testing", ",", "we", "keep", "the", "original", "size", "of", "the", "images", "that", "are", "not", "cropped", "or", "resized", "and", "extract", "the", "feature", ".", "In", "Table", "[", "reference", "]", ",", "many", "previous", "works", "are", "based", "on", "CaffeNet", "or", "VGG16", ".", "For", "fair", "comparison", ",", "we", "report", "the", "baseline", "results", "and", "the", "results", "of", "our", "model", "based", "on", "these", "two", "network", "structures", ",", "respectively", ".", "Our", "model", "which", "uses", "CaffeNet", "as", "pretrained", "model", "outperforms", "the", "state", "of", "the", "art", ".", "Meanwhile", ",", "the", "model", "using", "VGG16", "is", "comparable", "to", "the", "state", "-", "of", "-", "the", "-", "arts", "methods", ".", "The", "proposed", "method", "show", "a", "6.0", "%", "and", "6.6", "%", "improvement", "over", "the", "baseline", "networks", "CaffeNet", "and", "VGG16", ",", "respectively", ".", "We", "visualize", "some", "retrieval", "results", "in", "Fig", ".", "[", "reference", "]", ".", "The", "images", "in", "the", "first", "column", "are", "the", "query", "images", ".", "The", "retrieval", "images", "are", "sorted", "according", "to", "the", "similarity", "scores", "from", "left", "to", "right", ".", "The", "main", "difficulty", "in", "the", "image", "retrieval", "is", "various", "object", "sizes", "in", "the", "image", ".", "In", "the", "first", "row", ",", "we", "use", "the", "roof", "(", "part", "of", "the", "building", ")", "to", "retrieve", "the", "images", "and", "the", "top", "five", "images", "are", "correct", "candidate", "images", ".", "The", "other", "retrieval", "samples", "also", "show", "our", "model", "is", "robust", "to", "the", "scale", "variations", ".", "section", ":", "Conclusion", "In", "this", "work", ",", "we", "propose", "a", "siamese", "network", "that", "simultaneously", "considers", "the", "identification", "loss", "and", "the", "verification", "loss", ".", "The", "proposed", "model", "learns", "a", "discriminative", "embedding", "and", "a", "similarity", "measurement", "at", "the", "same", "time", ".", "It", "outperforms", "the", "state", "of", "the", "art", "on", "two", "popular", "person", "re", "-", "ID", "benchmarks", "and", "shows", "potential", "ability", "to", "apply", "on", "the", "generic", "instance", "retrieval", "task", ".", "Future", "work", "includes", "exploring", "more", "novel", "applications", "of", "the", "proposed", "method", ",", "such", "as", "car", "recognition", "and", "fine", "-", "grained", "classification", ".", "Besides", ",", "we", "will", "investigate", "how", "to", "learn", "a", "robust", "descriptor", "to", "further", "improve", "the", "performance", "of", "the", "person", "re", "-", "identification", "on", "large", "-", "scale", "testing", "set", ".", "bibliography", ":", "References"], "entity": [[3, 7, ["Discriminatively", "Learned", "CNN", "Embedding"], "Method"], [8, 12, ["Person", "Re", "-", "identification"], "Task"], [20, 23, ["convolutional", "neural", "networks"], "Method"], [24, 25, ["CNN"], "Method"], [27, 31, ["person", "re", "-", "identification"], "Task"], [32, 35, ["re", "-", "ID"], "Task"], [39, 40, ["verification"], "Task"], [41, 43, ["identification", "models"], "Task"], [76, 79, ["discriminative", "pedestrian", "descriptors"], "Task"], [85, 87, ["siamese", "network"], "Method"], [91, 93, ["identification", "loss"], "Metric"], [131, 133, ["similarity", "measurement"], "Method"], [157, 161, ["pre", "-", "trained", "networks"], "Method"], [167, 168, ["embedding"], "Method"], [198, 200, ["image", "retrieval"], "Task"], [209, 212, ["Convolutional", "Neural", "Networks"], "Method"], [216, 220, ["Person", "re", "-", "identification"], "Task"], [230, 233, ["image", "retrieval", "problem"], "Task"], [315, 318, ["convolutional", "neural", "network"], "Method"], [319, 320, ["CNN"], "Method"], [325, 328, ["convolutional", "neural", "network"], "Method"], [329, 330, ["CNN"], "Method"], [343, 345, ["feature", "embeddings"], "Metric"], [346, 348, ["deep", "metrics"], "Metric"], [372, 374, ["identification", "models"], "Task"], [385, 387, ["feature", "extraction"], "Task"], [412, 414, ["Verification", "models"], "Method"], [445, 450, ["binary", "-", "class", "classification", "task"], "Task"], [452, 455, ["similarity", "regression", "task"], "Task"], [562, 564, ["identification", "models"], "Task"], [566, 570, ["person", "re", "-", "identification"], "Task"], [572, 577, ["multi", "-", "class", "recognition", "task"], "Task"], [581, 583, ["feature", "learning"], "Method"], [602, 606, ["cross", "-", "entropy", "loss"], "Metric"], [622, 625, ["fully", "connected", "layer"], "Method"], [652, 654, ["identification", "model"], "Method"], [686, 689, ["pedestrian", "retrieval", "process"], "Task"], [739, 741, ["discriminative", "ability"], "Metric"], [751, 753, ["siamese", "network"], "Method"], [781, 785, ["pair", "-", "wise", "similarity"], "Metric"], [786, 788, ["image", "identities"], "Metric"], [799, 804, ["Euclidepdfan", "distance", "based", "pedestrian", "retrieval"], "Task"], [815, 817, ["siamese", "network"], "Method"], [820, 821, ["losses"], "Metric"], [822, 824, ["identification", "loss"], "Metric"], [838, 840, ["similarity", "metric"], "Metric"], [843, 846, ["pedestrian", "retrieval", "accuracy"], "Metric"], [849, 851, ["competitive", "accuracy"], "Metric"], [871, 872, ["Market1501"], "Material"], [994, 996, ["identification", "models"], "Task"], [999, 1001, ["Verification", "Models"], "Method"], [1013, 1016, ["deep", "metric", "learning"], "Method"], [1020, 1022, ["Verification", "models"], "Method"], [1033, 1035, ["similarity", "score"], "Metric"], [1064, 1068, ["person", "re", "-", "identification"], "Task"], [1072, 1074, ["data", "augmentation"], "Task"], [1075, 1077, ["image", "matching"], "Task"], [1093, 1096, ["part", "-", "CNNs"], "Method"], [1144, 1148, ["patch", "-", "matching", "layer"], "Method"], [1169, 1171, ["similarity", "regression"], "Method"], [1177, 1179, ["softmax", "loss"], "Method"], [1194, 1196, ["matching", "layer"], "Method"], [1215, 1216, ["filters"], "Method"], [1218, 1220, ["deeper", "network"], "Method"], [1229, 1230, ["CNN"], "Method"], [1232, 1234, ["gate", "functions"], "Method"], [1237, 1243, ["long", "-", "short", "-", "term", "memory"], "Method"], [1244, 1245, ["LSTM"], "Method"], [1269, 1271, ["computational", "inefficiency"], "Metric"], [1320, 1322, ["Identification", "Models"], "Method"], [1328, 1329, ["Market1501"], "Material"], [1344, 1346, ["classification", "model"], "Method"], [1368, 1369, ["CNN"], "Method"], [1376, 1379, ["part", "-", "matching"], "Method"], [1388, 1392, ["fine", "-", "tuning", "approach"], "Method"], [1393, 1394, ["Market1501"], "Material"], [1431, 1433, ["classification", "model"], "Method"], [1440, 1442, ["dropout", "function"], "Method"], [1458, 1460, ["classification", "model"], "Method"], [1463, 1467, ["faster", "-", "RCNN", "method"], "Method"], [1484, 1487, ["pedestrian", "detection", "errors"], "Task"], [1490, 1494, ["Verification", "-", "identification", "Models"], "Method"], [1495, 1497, ["face", "recognition"], "Task"], [1500, 1502, ["DeepID", "networks"], "Method"], [1544, 1546, ["deeper", "network"], "Method"], [1564, 1566, ["face", "recognition"], "Task"], [1601, 1603, ["DeepID", "networks"], "Method"], [1604, 1606, ["contrastive", "loss"], "Method"], [1616, 1620, ["cross", "-", "entropy", "loss"], "Metric"], [1629, 1632, ["over", "-", "fitting"], "Task"], [1657, 1659, ["contrastive", "loss"], "Method"], [1662, 1663, ["dropout"], "Method"], [1669, 1670, ["embedding"], "Method"], [1690, 1692, ["dropout", "regularization"], "Method"], [1703, 1705, ["DeepID", "networks"], "Method"], [1779, 1780, ["verification"], "Task"], [1781, 1783, ["identification", "models"], "Task"], [1860, 1864, ["pair", "-", "wise", "comparison"], "Method"], [1867, 1869, ["part", "matching"], "Method"], [1870, 1872, ["contrastive", "loss"], "Method"], [1888, 1890, ["identification", "models"], "Task"], [1911, 1915, ["cross", "-", "entropy", "loss"], "Method"], [1917, 1921, ["cross", "-", "entropy", "loss"], "Metric"], [1931, 1933, ["linear", "function"], "Method"], [2027, 2029, ["classification", "models"], "Method"], [2092, 2095, ["convolutional", "siamese", "network"], "Method"], [2139, 2141, ["similarity", "score"], "Metric"], [2156, 2158, ["Convolutional", "Layers"], "Method"], [2160, 2162, ["Square", "Layer"], "Method"], [2186, 2187, ["CaffeNet"], "Method"], [2188, 2189, ["VGG16"], "Method"], [2190, 2193, ["ResNet", "-", "50"], "Method"], [2201, 2204, ["fully", "-", "connected"], "Method"], [2205, 2206, ["FC"], "Method"], [2210, 2213, ["re", "-", "ID"], "Task"], [2256, 2258, ["optimization", "objectives"], "Metric"], [2275, 2277, ["discriminative", "descriptor"], "Method"], [2293, 2295, ["Identification", "Loss"], "Task"], [2298, 2299, ["CaffeNets"], "Method"], [2356, 2358, ["convolutional", "layer"], "Method"], [2374, 2376, ["convolutional", "layer"], "Method"], [2385, 2386, ["CaffeNet"], "Method"], [2391, 2393, ["softmax", "unit"], "Method"], [2407, 2410, ["Rectified", "Linear", "Unit"], "Method"], [2411, 2413, ["ReLU", ")"], "Method"], [2418, 2419, ["convolution"], "Method"], [2423, 2428, ["multi", "-", "class", "recognition", "approaches"], "Method"], [2432, 2436, ["cross", "-", "entropy", "loss"], "Metric"], [2437, 2439, ["identity", "prediction"], "Task"], [2445, 2447, ["convolutional", "operation"], "Method"], [2463, 2465, ["convolutional", "layer"], "Method"], [2482, 2484, ["Verification", "Loss"], "Metric"], [2490, 2492, ["matching", "function"], "Method"], [2507, 2509, ["similarity", "estimation"], "Task"], [2545, 2547, ["pedestrian", "descriptor"], "Method"], [2549, 2551, ["identification", "model"], "Method"], [2571, 2575, ["non", "-", "parametric", "layer"], "Method"], [2576, 2578, ["Square", "Layer"], "Method"], [2605, 2607, ["Square", "Layer"], "Method"], [2627, 2629, ["Square", "Layer"], "Method"], [2634, 2636, ["convolutional", "layer"], "Method"], [2672, 2674, ["convolutional", "layer"], "Method"], [2686, 2687, ["ReLU"], "Method"], [2692, 2693, ["convolution"], "Method"], [2700, 2703, ["binary", "classification", "problem"], "Task"], [2706, 2710, ["cross", "-", "entropy", "loss"], "Metric"], [2718, 2720, ["identification", "loss"], "Task"], [2747, 2749, ["convolutional", "layer"], "Method"], [2785, 2787, ["contrastive", "loss"], "Method"], [2790, 2792, ["regression", "loss"], "Method"], [2835, 2836, ["dropout"], "Method"], [2851, 2852, ["embedding"], "Method"], [2859, 2863, ["cross", "-", "entropy", "loss"], "Metric"], [2869, 2870, ["dropout"], "Method"], [2887, 2889, ["contrastive", "loss"], "Method"], [2901, 2905, ["cross", "-", "entropy", "loss"], "Metric"], [2906, 2910, ["rank", "-", "1", "accuracy"], "Metric"], [2911, 2912, ["mAP"], "Metric"], [2916, 2917, ["Identification"], "Task"], [2918, 2919, ["Verification"], "Task"], [2928, 2932, ["cross", "-", "entropy", "losses"], "Metric"], [2945, 2947, ["identification", "model"], "Method"], [2954, 2957, ["learning", "rate", "setting"], "Metric"], [3038, 3041, ["ResNet", "-", "50"], "Method"], [3043, 3045, ["pretrained", "model"], "Method"], [3052, 3054, ["identification", "loss"], "Metric"], [3128, 3129, ["identification"], "Task"], [3174, 3176, ["activation", "map"], "Method"], [3273, 3277, ["person", "re", "-", "identification"], "Task"], [3282, 3283, ["Optimization"], "Task"], [3283, 3285, ["Input", "preparation"], "Task"], [3327, 3330, ["ResNet", "-", "50"], "Method"], [3331, 3332, ["VGG16"], "Method"], [3422, 3424, ["Matconvnet", "package"], "Method"], [3429, 3430, ["embedding"], "Method"], [3431, 3432, ["CaffeNet"], "Method"], [3433, 3434, ["VGG16"], "Method"], [3435, 3438, ["ResNet", "-", "50"], "Method"], [3458, 3459, ["VGG16net"], "Method"], [3462, 3463, ["CaffeNet"], "Method"], [3477, 3478, ["CaffeNet"], "Method"], [3481, 3482, ["VGG16"], "Method"], [3483, 3486, ["ResNet", "-", "50"], "Method"], [3488, 3490, ["learning", "rate"], "Metric"], [3508, 3514, ["mini", "-", "batch", "stochastic", "gradient", "descent"], "Method"], [3515, 3516, ["SGD"], "Method"], [3580, 3582, ["identification", "losses"], "Method"], [3588, 3590, ["dropout", "function"], "Method"], [3593, 3595, ["convolutional", "layer"], "Method"], [3628, 3630, ["memory", "consumption"], "Metric"], [3632, 3634, ["pretrained", "model"], "Method"], [3643, 3647, ["fine", "-", "tuned", "model"], "Method"], [3659, 3660, ["CaffeNet"], "Method"], [3666, 3673, ["4", ",", "096", "-", "dim", "pedestrian", "descriptor"], "Method"], [3717, 3718, ["ranking"], "Task"], [3753, 3754, ["Market1501"], "Material"], [3764, 3766, ["network", "structures"], "Method"], [3775, 3776, ["Market1501"], "Material"], [3790, 3793, ["image", "retrieval", "task"], "Task"], [3827, 3828, ["Market1501"], "Material"], [3892, 3895, ["Deformable", "Part", "Model"], "Method"], [3896, 3898, ["DPM", ")"], "Method"], [3901, 3905, ["hand", "-", "drawn", "bboxes"], "Method"], [3943, 3947, ["person", "re", "-", "identification"], "Task"], [3950, 3952, ["realistic", "setting"], "Task"], [3966, 3967, ["Market1501"], "Material"], [4014, 4015, ["retrieval"], "Task"], [4074, 4075, ["DPM"], "Method"], [4207, 4211, ["rank", "-", "1", "accuracy"], "Metric"], [4212, 4215, ["mean", "average", "precision"], "Metric"], [4216, 4217, ["mAP"], "Metric"], [4222, 4223, ["Market1501"], "Material"], [4236, 4237, ["mAP"], "Metric"], [4240, 4246, ["Person", "Re", "-", "i", "d", "Evaluation"], "Task"], [4255, 4257, ["baseline", "networks"], "Method"], [4260, 4264, ["fine", "-", "tuning", "method"], "Method"], [4301, 4305, ["rank", "-", "1", "accuracy"], "Metric"], [4306, 4307, ["CaffeNet"], "Method"], [4308, 4309, ["VGG16"], "Method"], [4310, 4313, ["ResNet", "-", "50"], "Method"], [4345, 4347, ["network", "architectures"], "Method"], [4367, 4368, ["CaffeNet"], "Method"], [4369, 4370, ["VGG16"], "Method"], [4371, 4374, ["ResNet", "-", "50"], "Method"], [4375, 4376, ["Market1501"], "Material"], [4397, 4401, ["single", "-", "shot", "setting"], "Task"], [4471, 4474, ["Cross", "-", "entropy"], "Metric"], [4475, 4477, ["Contrastive", "loss"], "Method"], [4481, 4485, ["cross", "-", "entropy", "loss"], "Method"], [4493, 4495, ["DeepID", "network"], "Method"], [4509, 4512, ["rank", "-", "1"], "Metric"], [4513, 4514, ["mAP"], "Metric"], [4522, 4524, ["contrastive", "loss"], "Method"], [4527, 4531, ["rank", "-", "1", "accuracy"], "Metric"], [4534, 4535, ["mAP"], "Metric"], [4540, 4542, ["contrastive", "loss"], "Method"], [4555, 4556, ["regularization"], "Method"], [4560, 4561, ["verification"], "Task"], [4562, 4566, ["Cross", "-", "entropy", "loss"], "Method"], [4616, 4619, ["mean", "average", "precision"], "Metric"], [4623, 4627, ["rank", "-", "1", "accuracy"], "Metric"], [4628, 4629, ["Market1501"], "Material"], [4639, 4643, ["multiple", "-", "query", "evaluation"], "Task"], [4648, 4649, ["CaffeNet"], "Method"], [4653, 4657, ["rank", "-", "1", "accuracy"], "Metric"], [4660, 4661, ["mAP"], "Metric"], [4673, 4677, ["rank", "-", "1", "accuracy"], "Metric"], [4680, 4681, ["mAP"], "Metric"], [4685, 4688, ["ResNet", "-", "50"], "Method"], [4695, 4699, ["rank", "-", "1", "accuracy"], "Metric"], [4703, 4704, ["mAP"], "Metric"], [4726, 4730, ["single", "-", "shot", "setting"], "Task"], [4775, 4778, ["rank", "-", "1"], "Metric"], [4781, 4782, ["mAP"], "Metric"], [4810, 4814, ["multi", "-", "shot", "setting"], "Task"], [4845, 4847, ["image", "retrieval"], "Task"], [4855, 4858, ["random", "searching", "pool"], "Method"], [4955, 4958, ["rank", "-", "1"], "Metric"], [4961, 4962, ["mAP"], "Metric"], [4976, 4980, ["multi", "-", "shot", "setting"], "Task"], [5010, 5013, ["re", "-", "identification"], "Task"], [5026, 5029, ["camera", "-", "6"], "Method"], [5033, 5035, ["resolution", "camera"], "Method"], [5042, 5044, ["HD", "cameras"], "Method"], [5068, 5069, ["mAP"], "Metric"], [5071, 5074, ["rank", "-", "1"], "Metric"], [5145, 5146, ["embeddings"], "Method"], [5152, 5154, ["clustering", "algorithm"], "Method"], [5176, 5178, ["pedestrian", "descriptor"], "Method"], [5244, 5246, ["image", "retrieval"], "Task"], [5248, 5252, ["re", "-", "ID", "performance"], "Metric"], [5256, 5257, ["ResNet"], "Method"], [5281, 5282, ["accuracy"], "Metric"], [5295, 5297, ["rank1", "accuracy"], "Metric"], [5300, 5301, ["mAP"], "Metric"], [5314, 5315, ["mAP"], "Metric"], [5348, 5350, ["ResNet", "Baseline"], "Method"], [5436, 5437, ["embedding"], "Method"], [5473, 5475, ["Instance", "Retrieval"], "Task"], [5484, 5488, ["generic", "image", "retrieval", "task"], "Task"], [5580, 5582, ["pooling", "layer"], "Method"], [5587, 5589, ["MAC", "layer"], "Method"], [5677, 5678, ["CaffeNet"], "Method"], [5679, 5680, ["VGG16"], "Method"], [5709, 5710, ["CaffeNet"], "Method"], [5711, 5713, ["pretrained", "model"], "Method"], [5725, 5726, ["VGG16"], "Method"], [5756, 5757, ["VGG16"], "Method"], [5763, 5764, ["retrieval"], "Task"], [5803, 5805, ["image", "retrieval"], "Task"], [5867, 5869, ["siamese", "network"], "Method"], [5873, 5875, ["identification", "loss"], "Metric"], [5919, 5923, ["generic", "instance", "retrieval", "task"], "Task"], [5938, 5940, ["car", "recognition"], "Task"], [5941, 5945, ["fine", "-", "grained", "classification"], "Task"], [5955, 5957, ["robust", "descriptor"], "Method"], [5964, 5968, ["person", "re", "-", "identification"], "Task"], [94, 95, ["verification"], "Task"], [128, 129, ["embedding"], "Method"], [144, 147, ["re", "-", "ID"], "Task"], [182, 185, ["re", "-", "ID"], "Task"], [204, 208, ["Person", "Re", "-", "identification"], "Task"], [221, 224, ["re", "-", "ID"], "Task"], [253, 256, ["re", "-", "ID"], "Task"], [310, 311, ["embedding"], "Method"], [364, 365, ["CNN"], "Method"], [369, 370, ["verification"], "Task"], [410, 411, ["embedding"], "Method"], [440, 443, ["re", "-", "ID"], "Task"], [461, 462, ["verification"], "Task"], [502, 503, ["verification"], "Task"], [510, 513, ["re", "-", "ID"], "Task"], [529, 530, ["verification"], "Task"], [557, 560, ["re", "-", "ID"], "Task"], [644, 645, ["CNN"], "Method"], [825, 826, ["verification"], "Task"], [834, 835, ["CNN"], "Method"], [835, 836, ["embedding"], "Method"], [866, 869, ["re", "-", "ID"], "Task"], [920, 921, ["CNN"], "Method"], [945, 949, ["person", "re", "-", "identification"], "Task"], [991, 992, ["verification"], "Task"], [1010, 1011, ["verification"], "Task"], [1018, 1019, ["verification"], "Task"], [1061, 1062, ["verification"], "Task"], [1188, 1189, ["verification"], "Task"], [1366, 1367, ["Market1501"], "Material"], [1409, 1410, ["CNN"], "Method"], [1419, 1420, ["FC"], "Method"], [1508, 1509, ["verification"], "Task"], [1531, 1532, ["verification"], "Task"], [1535, 1536, ["verification"], "Task"], [1608, 1609, ["verification"], "Task"], [1694, 1695, ["embedding"], "Method"], [1732, 1735, ["re", "-", "ID"], "Task"], [1744, 1745, ["verification"], "Task"], [1835, 1836, ["verification"], "Task"], [1839, 1840, ["verification"], "Task"], [2014, 2015, ["verification"], "Task"], [2059, 2060, ["verification"], "Task"], [2098, 2099, ["verification"], "Task"], [2151, 2152, ["CNN"], "Method"], [2175, 2176, ["verification"], "Task"], [2182, 2183, ["CNN"], "Method"], [2240, 2241, ["CNN"], "Method"], [2245, 2246, ["CaffeNet"], "Method"], [2264, 2265, ["verification"], "Task"], [2279, 2282, ["re", "-", "ID"], "Task"], [2336, 2339, ["fully", "-", "connected"], "Method"], [2352, 2353, ["CNN"], "Method"], [2520, 2521, ["CNN"], "Method"], [2556, 2557, ["verification"], "Task"], [2697, 2698, ["verification"], "Task"], [2826, 2829, ["re", "-", "ID"], "Task"], [2948, 2949, ["verification"], "Task"], [3001, 3002, ["CNN"], "Method"], [3055, 3056, ["verification"], "Task"], [3065, 3066, ["Market1501"], "Material"], [3131, 3132, ["verification"], "Task"], [3219, 3220, ["embedding"], "Method"], [3321, 3322, ["CaffeNet"], "Method"], [3452, 3455, ["ResNet", "-", "50"], "Method"], [3569, 3570, ["verification"], "Task"], [3618, 3619, ["CaffeNet"], "Method"], [4249, 4250, ["CNN"], "Method"], [4316, 4317, ["Market1501"], "Material"], [4336, 4337, ["Market1501"], "Material"], [4390, 4393, ["rank", "-", "1"], "Metric"], [4516, 4519, ["ResNet", "-", "50"], "Method"], [4549, 4552, ["re", "-", "ID"], "Task"], [4620, 4621, ["mAP"], "Metric"], [4871, 4874, ["re", "-", "ID"], "Task"], [4999, 5000, ["Market1501"], "Material"], [5046, 5049, ["re", "-", "ID"], "Task"], [5120, 5121, ["embedding"], "Method"], [5219, 5220, ["Market1501"], "Material"], [5365, 5368, ["re", "-", "ID"], "Task"], [5385, 5386, ["Market1501"], "Material"], [5423, 5424, ["Market1501"], "Material"], [5480, 5481, ["verification"], "Task"], [5754, 5755, ["CaffeNet"], "Method"], [5877, 5878, ["verification"], "Task"], [5886, 5887, ["embedding"], "Method"], [5907, 5910, ["re", "-", "ID"], "Task"]]} {"doc_id": "193b518bc3025804c6d587c74cbc154d91478417", "text": ["document", ":", "Learning", "to", "Adapt", "Structured", "Output", "Space", "for", "Semantic", "Segmentation", "Convolutional", "neural", "network", "-", "based", "approaches", "for", "semantic", "segmentation", "rely", "on", "supervision", "with", "pixel", "-", "level", "ground", "truth", ",", "but", "may", "not", "generalize", "well", "to", "unseen", "image", "domains", ".", "As", "the", "labeling", "process", "is", "tedious", "and", "labor", "intensive", ",", "developing", "algorithms", "that", "can", "adapt", "source", "ground", "truth", "labels", "to", "the", "target", "domain", "is", "of", "great", "interest", ".", "In", "this", "paper", ",", "we", "propose", "an", "adversarial", "learning", "method", "for", "domain", "adaptation", "in", "the", "context", "of", "semantic", "segmentation", ".", "Considering", "semantic", "segmentations", "as", "structured", "outputs", "that", "contain", "spatial", "similarities", "between", "the", "source", "and", "target", "domains", ",", "we", "adopt", "adversarial", "learning", "in", "the", "output", "space", ".", "To", "further", "enhance", "the", "adapted", "model", ",", "we", "construct", "a", "multi", "-", "level", "adversarial", "network", "to", "effectively", "perform", "output", "space", "domain", "adaptation", "at", "different", "feature", "levels", ".", "Extensive", "experiments", "and", "ablation", "study", "are", "conducted", "under", "various", "domain", "adaptation", "settings", ",", "including", "synthetic", "-", "to", "-", "real", "and", "cross", "-", "city", "scenarios", ".", "We", "show", "that", "the", "proposed", "method", "performs", "favorably", "against", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "in", "terms", "of", "accuracy", "and", "visual", "quality", ".", "section", ":", "Introduction", "Semantic", "segmentation", "aims", "to", "assign", "each", "pixel", "a", "semantic", "label", ",", "e.g.", ",", "person", ",", "car", ",", "road", "or", "tree", ",", "in", "an", "image", ".", "Recently", ",", "methods", "based", "on", "convolutional", "neural", "networks", "(", "CNNs", ")", "have", "achieved", "significant", "progress", "in", "semantic", "segmentation", "with", "applications", "for", "autonomous", "driving", "and", "image", "editing", ".", "The", "crux", "of", "CNN", "-", "based", "approaches", "is", "to", "annotate", "a", "large", "number", "of", "images", "that", "cover", "possible", "scene", "variations", ".", "However", ",", "this", "trained", "model", "may", "not", "generalize", "well", "to", "unseen", "images", ",", "especially", "when", "there", "is", "a", "domain", "gap", "between", "the", "training", "(", "source", ")", "and", "test", "(", "target", ")", "images", ".", "For", "instance", ",", "the", "distribution", "of", "appearance", "for", "objects", "and", "scenes", "may", "vary", "in", "different", "cities", ",", "and", "even", "weather", "and", "lighting", "conditions", "can", "change", "significantly", "in", "the", "same", "city", ".", "In", "such", "cases", ",", "relying", "only", "on", "the", "supervised", "model", "that", "requires", "re", "-", "annotating", "per", "-", "pixel", "ground", "truths", "in", "different", "scenarios", "would", "entail", "prohibitively", "high", "labor", "cost", ".", "To", "address", "this", "issue", ",", "knowledge", "transfer", "or", "domain", "adaptation", "techniques", "have", "been", "proposed", "to", "close", "the", "gap", "between", "source", "and", "target", "domains", ",", "where", "annotations", "are", "not", "available", "in", "the", "target", "domain", ".", "For", "image", "classification", ",", "one", "effective", "approach", "is", "to", "align", "features", "across", "two", "domains", "such", "that", "the", "adapted", "features", "can", "generalize", "to", "both", "domains", ".", "Similar", "efforts", "have", "been", "made", "for", "semantic", "segmentation", "via", "adversarial", "learning", "in", "the", "feature", "space", ".", "However", ",", "different", "from", "the", "image", "classification", "task", ",", "feature", "adaptation", "for", "semantic", "segmentation", "may", "suffer", "from", "the", "complexity", "of", "high", "-", "dimensional", "features", "that", "needs", "to", "encode", "diverse", "visual", "cues", ",", "including", "appearance", ",", "shape", "and", "context", ".", "This", "motivates", "us", "to", "develop", "an", "effective", "method", "for", "adapting", "pixel", "-", "level", "prediction", "tasks", "rather", "than", "using", "feature", "adaptation", ".", "In", "semantic", "segmentation", ",", "we", "note", "that", "the", "output", "space", "contains", "rich", "information", ",", "both", "spatially", "and", "locally", ".", "For", "instance", ",", "even", "if", "images", "from", "two", "domains", "are", "very", "different", "in", "appearance", ",", "their", "segmentation", "outputs", "share", "a", "significant", "amount", "of", "similarities", ",", "e.g.", ",", "spatial", "layout", "and", "local", "context", "(", "see", "Figure", "[", "reference", "]", ")", ".", "Based", "on", "this", "observation", ",", "we", "address", "the", "pixel", "-", "level", "domain", "adaptation", "problem", "in", "the", "output", "(", "segmentation", ")", "space", ".", "In", "this", "paper", ",", "we", "propose", "an", "end", "-", "to", "-", "end", "CNN", "-", "based", "domain", "adaptation", "algorithm", "for", "semantic", "segmentation", ".", "Our", "formulation", "is", "based", "on", "adversarial", "learning", "in", "the", "output", "space", ",", "where", "the", "intuition", "is", "to", "directly", "make", "the", "predicted", "label", "distributions", "close", "to", "each", "other", "across", "source", "and", "target", "domains", ".", "Based", "on", "the", "generative", "adversarial", "network", "(", "GAN", ")", ",", "the", "proposed", "model", "consists", "of", "two", "parts", ":", "1", ")", "a", "segmentation", "model", "to", "predict", "output", "results", ",", "and", "2", ")", "a", "discriminator", "to", "distinguish", "whether", "the", "input", "is", "from", "the", "source", "or", "target", "segmentation", "output", ".", "With", "an", "adversarial", "loss", ",", "the", "proposed", "segmentation", "model", "aims", "to", "fool", "the", "discriminator", ",", "with", "the", "goal", "of", "generating", "similar", "distributions", "in", "the", "output", "space", "for", "either", "source", "or", "target", "images", ".", "The", "proposed", "method", "also", "adapts", "features", "as", "the", "errors", "are", "back", "-", "propagated", "to", "the", "feature", "level", "from", "the", "output", "labels", ".", "However", ",", "one", "concern", "is", "that", "lower", "-", "level", "features", "may", "not", "be", "adapted", "well", "as", "they", "are", "far", "away", "from", "the", "high", "-", "level", "output", "labels", ".", "To", "address", "this", "issue", ",", "we", "develop", "a", "multi", "-", "level", "strategy", "by", "incorporating", "adversarial", "learning", "at", "different", "feature", "levels", "of", "the", "segmentation", "model", ".", "For", "instance", ",", "we", "can", "use", "both", "conv5", "and", "conv4", "features", "to", "predict", "segmentation", "results", "in", "the", "output", "space", ".", "Then", "two", "discriminators", "can", "be", "connected", "to", "each", "of", "the", "predicted", "output", "for", "multi", "-", "level", "adversarial", "learning", ".", "We", "perform", "one", "-", "stage", "end", "-", "to", "-", "end", "training", "for", "the", "segmentation", "model", "and", "discriminators", "jointly", ",", "without", "using", "any", "prior", "knowledge", "of", "the", "data", "in", "the", "target", "domain", ".", "In", "the", "testing", "phase", ",", "we", "can", "simply", "discard", "discriminators", "and", "use", "the", "adapted", "segmentation", "model", "on", "target", "images", ",", "with", "no", "extra", "computational", "requirements", ".", "Due", "to", "the", "high", "labor", "cost", "of", "annotating", "segmentation", "ground", "truth", ",", "there", "has", "been", "great", "interest", "in", "large", "-", "scale", "synthetic", "datasets", "with", "annotations", ",", "e.g.", ",", "GTA5", "and", "SYNTHIA", ".", "As", "a", "result", ",", "one", "critical", "setting", "is", "to", "adapt", "the", "model", "trained", "on", "synthetic", "data", "to", "real", "-", "world", "datasets", ",", "such", "as", "Cityscapes", ".", "We", "follow", "this", "setting", "and", "conduct", "extensive", "experiments", "to", "validate", "the", "proposed", "domain", "adaptation", "method", ".", "First", ",", "we", "use", "a", "strong", "baseline", "model", "that", "is", "able", "to", "generalize", "to", "different", "domains", ".", "We", "note", "that", "a", "strong", "baseline", "facilitates", "real", "-", "world", "applications", "and", "can", "evaluate", "the", "limitation", "of", "the", "proposed", "adaptation", "approach", ".", "Based", "on", "this", "baseline", "model", ",", "we", "show", "comparisons", "using", "adversarial", "adaptation", "in", "the", "feature", "and", "output", "spaces", ".", "Furthermore", ",", "we", "show", "that", "the", "multi", "-", "level", "adversarial", "learning", "improves", "the", "results", "over", "single", "-", "level", "adaptation", ".", "In", "addition", "to", "the", "synthetic", "-", "to", "-", "real", "setting", ",", "we", "show", "experimental", "results", "on", "the", "Cross", "-", "City", "dataset", ",", "where", "annotations", "are", "provided", "in", "one", "city", "(", "source", ")", ",", "while", "testing", "the", "model", "on", "another", "unseen", "city", "(", "target", ")", ".", "Overall", ",", "our", "method", "performs", "favorably", "against", "state", "-", "of", "-", "the", "-", "art", "algorithms", "on", "numerous", "benchmark", "datasets", "under", "different", "settings", ".", "The", "contributions", "of", "this", "work", "are", "as", "follows", ".", "First", ",", "we", "propose", "a", "domain", "adaptation", "method", "for", "pixel", "-", "level", "semantic", "segmentation", "via", "adversarial", "learning", ".", "Second", ",", "we", "demonstrate", "that", "adaptation", "in", "the", "output", "(", "segmentation", ")", "space", "can", "effectively", "align", "scene", "layout", "and", "local", "context", "between", "source", "and", "target", "images", ".", "Third", ",", "a", "multi", "-", "level", "adversarial", "learning", "scheme", "is", "developed", "to", "adapt", "features", "at", "different", "levels", "of", "the", "segmentation", "model", ",", "which", "leads", "to", "improved", "performance", ".", "section", ":", "Related", "Work", "Semantic", "Segmentation", ".", "State", "-", "of", "-", "the", "-", "art", "semantic", "segmentation", "methods", "are", "mainly", "based", "on", "the", "recent", "advances", "of", "deep", "neural", "networks", ".", "As", "proposed", "by", "Long", ",", "one", "can", "transform", "a", "classification", "CNN", "(", "e.g.", ",", "AlexNet", ",", "VGG", ",", "or", "ResNet", ")", "to", "a", "fully", "-", "convolutional", "network", "(", "FCN", ")", "for", "semantic", "segmentation", ".", "Numerous", "methods", "have", "since", "been", "developed", "to", "improve", "this", "model", "by", "utilizing", "context", "information", "or", "enlarging", "receptive", "fields", ".", "To", "train", "these", "advanced", "networks", ",", "a", "substantial", "amount", "of", "dense", "pixel", "annotations", "must", "be", "collected", "in", "order", "to", "match", "the", "model", "capacity", "of", "deep", "CNNs", ".", "As", "a", "result", ",", "weakly", "and", "semi", "-", "supervised", "approaches", "are", "proposed", "in", "recent", "years", "to", "reduce", "the", "heavy", "labeling", "cost", "of", "collecting", "segmentation", "ground", "truths", ".", "However", ",", "in", "most", "real", "-", "world", "applications", ",", "it", "is", "difficult", "to", "obtain", "weak", "annotations", "and", "the", "trained", "model", "may", "not", "generalize", "well", "to", "unseen", "image", "domains", ".", "Another", "approach", "to", "tackle", "the", "annotation", "problem", "is", "to", "construct", "synthetic", "datasets", "based", "on", "rendering", ",", "e.g.", ",", "GTA5", "and", "SYNTHIA", ".", "While", "the", "data", "collection", "is", "less", "costly", "since", "the", "pixel", "-", "level", "annotation", "can", "be", "done", "with", "a", "partially", "automated", "process", ",", "these", "datasets", "are", "usually", "used", "in", "conjunction", "with", "real", "-", "world", "datasets", "for", "joint", "learning", "to", "improve", "the", "performance", ".", "However", ",", "when", "training", "solely", "on", "the", "synthetic", "dataset", ",", "the", "model", "does", "not", "generalize", "well", "to", "real", "-", "world", "data", ",", "mainly", "due", "to", "the", "large", "domain", "shift", "between", "synthetic", "images", "and", "real", "-", "world", "images", ",", "i.e.", ",", "appearance", "differences", "are", "still", "significant", "with", "current", "rendering", "techniques", ".", "Although", "synthesizing", "more", "realistic", "images", "can", "decrease", "the", "domain", "shift", ",", "it", "is", "necessary", "to", "use", "domain", "adaptation", "to", "narrow", "the", "performance", "gap", ".", "Domain", "Adaptation", ".", "Domain", "adaptation", "methods", "for", "image", "classification", "have", "been", "developed", "to", "address", "the", "domain", "-", "shift", "problem", "between", "the", "source", "and", "target", "domains", ".", "Numerous", "methods", "are", "developed", "based", "on", "CNN", "classifiers", "due", "to", "performance", "gain", ".", "The", "main", "insight", "behind", "these", "approaches", "is", "to", "tackle", "the", "problem", "by", "aligning", "the", "feature", "distribution", "between", "source", "and", "target", "images", ".", "Ganin", "propose", "the", "Domain", "-", "Adversarial", "Neural", "Network", "(", "DANN", ")", "to", "transfer", "the", "feature", "distribution", ".", "A", "number", "of", "variants", "have", "since", "been", "proposed", "with", "different", "loss", "functions", "or", "classifiers", ".", "Recently", ",", "the", "PixelDA", "method", "addresses", "domain", "adaptation", "for", "image", "classification", "by", "transferring", "the", "source", "images", "to", "target", "domain", ",", "thereby", "obtaining", "a", "simulated", "training", "set", "for", "target", "images", ".", "We", "note", "that", "domain", "adaptation", "for", "pixel", "-", "level", "prediction", "tasks", "have", "not", "been", "explored", "widely", ".", "Hoffman", "introduce", "the", "task", "of", "domain", "adaptation", "on", "semantic", "segmentation", "by", "applying", "adversarial", "learning", "(", "i.e.", ",", "DANN", ")", "in", "a", "fully", "-", "convolutional", "way", "on", "feature", "representations", "and", "additional", "category", "constraints", "similar", "to", "the", "constrained", "CNN", ".", "Other", "methods", "focus", "on", "adapting", "synthetic", "-", "to", "-", "real", "or", "cross", "-", "city", "images", "by", "adopting", "class", "-", "wise", "adversarial", "learning", "or", "label", "transfer", ".", "Similar", "to", "the", "PixelDA", "method", ",", "one", "concurrent", "work", ",", "CyCADA", "uses", "the", "CycleGAN", "and", "transfers", "source", "domain", "images", "to", "the", "target", "domain", "with", "pixel", "alignment", ",", "thus", "generating", "extra", "training", "data", "combined", "with", "feature", "space", "adversarial", "learning", ".", "Although", "feature", "space", "adaptation", "has", "been", "successfully", "applied", "to", "image", "classification", ",", "pixel", "-", "level", "tasks", "such", "as", "semantic", "segmentation", "remains", "challenging", "based", "on", "feature", "adaptation", "-", "based", "approaches", ".", "In", "this", "paper", ",", "we", "use", "the", "property", "that", "pixel", "-", "level", "predictions", "are", "structured", "outputs", "that", "contain", "information", "spatially", "and", "locally", ",", "to", "propose", "an", "efficient", "domain", "adaptation", "algorithm", "through", "adversarial", "learning", "in", "the", "output", "space", ".", "section", ":", "Algorithmic", "Overview", "subsection", ":", "Overview", "of", "the", "Proposed", "Model", "Our", "domain", "adaptation", "algorithm", "consists", "of", "two", "modules", ":", "a", "segmentation", "network", "and", "the", "discriminator", ",", "where", "indicates", "the", "level", "of", "a", "discriminator", "in", "the", "multi", "-", "level", "adversarial", "learning", ".", "Two", "sets", "of", "images", "from", "source", "and", "target", "domains", "are", "denoted", "as", "and", ".", "We", "first", "forward", "the", "source", "image", "(", "with", "annotations", ")", "to", "the", "segmentation", "network", "for", "optimizing", ".", "Then", "we", "predict", "the", "segmentation", "softmax", "output", "for", "the", "target", "image", "(", "without", "annotations", ")", ".", "Since", "our", "goal", "is", "to", "make", "segmentation", "predictions", "of", "source", "and", "target", "images", "(", "i.e.", ",", "and", ")", "close", "to", "each", "other", ",", "we", "use", "these", "two", "predictions", "as", "the", "input", "to", "the", "discriminator", "to", "distinguish", "whether", "the", "input", "is", "from", "the", "source", "or", "target", "domain", ".", "With", "an", "adversarial", "loss", "on", "the", "target", "prediction", ",", "the", "network", "propagates", "gradients", "from", "to", ",", "which", "would", "encourage", "to", "generate", "similar", "segmentation", "distributions", "in", "the", "target", "domain", "to", "the", "source", "prediction", ".", "Figure", "[", "reference", "]", "shows", "the", "overview", "of", "the", "proposed", "algorithm", ".", "subsection", ":", "Objective", "Function", "for", "Domain", "Adaptation", "With", "the", "proposed", "network", ",", "we", "formulate", "the", "adaptation", "task", "containing", "two", "loss", "functions", "from", "both", "modules", ":", "where", "is", "the", "cross", "-", "entropy", "loss", "using", "ground", "truth", "annotations", "in", "the", "source", "domain", ",", "and", "is", "the", "adversarial", "loss", "that", "adapts", "predicted", "segmentations", "of", "target", "images", "to", "the", "distribution", "of", "source", "predictions", "(", "see", "Section", "[", "reference", "]", ")", ".", "In", "(", "[", "reference", "]", ")", ",", "is", "the", "weight", "used", "to", "balance", "the", "two", "losses", ".", "section", ":", "Output", "Space", "Adaptation", "Different", "from", "image", "classification", "based", "on", "features", "that", "describe", "the", "global", "visual", "information", "of", "the", "image", ",", "high", "-", "dimensional", "features", "learned", "for", "semantic", "segmentation", "encodes", "complex", "representations", ".", "As", "a", "result", ",", "adaptation", "in", "the", "feature", "space", "may", "not", "be", "the", "best", "choice", "for", "semantic", "segmentation", ".", "On", "the", "other", "hand", ",", "although", "segmentation", "outputs", "are", "in", "the", "low", "-", "dimensional", "space", ",", "they", "contain", "rich", "information", ",", "e.g.", ",", "scene", "layout", "and", "context", ".", "Our", "intuition", "is", "that", "no", "matter", "images", "are", "from", "the", "source", "or", "target", "domain", ",", "their", "segmentations", "should", "share", "strong", "similarities", ",", "spatially", "and", "locally", ".", "Thus", ",", "we", "utilize", "this", "property", "to", "adapt", "low", "-", "dimensional", "softmax", "outputs", "of", "segmentation", "predictions", "via", "an", "adversarial", "learning", "scheme", ".", "subsection", ":", "Single", "-", "level", "Adversarial", "Learning", "Discriminator", "Training", ".", "Before", "introducing", "how", "to", "adapt", "the", "segmentation", "network", "via", "adversarial", "learning", ",", "we", "first", "describe", "the", "training", "objective", "for", "the", "discriminator", ".", "Given", "the", "segmentation", "softmax", "output", ",", "where", "is", "the", "number", "of", "categories", ",", "we", "forward", "to", "a", "fully", "-", "convolutional", "discriminator", "using", "a", "cross", "-", "entropy", "loss", "for", "the", "two", "classes", "(", "i.e.", ",", "source", "and", "target", ")", ".", "The", "loss", "can", "be", "written", "as", ":", "where", "if", "the", "sample", "is", "drawn", "from", "the", "target", "domain", ",", "and", "for", "the", "sample", "from", "the", "source", "domain", ".", "Segmentation", "Network", "Training", ".", "First", ",", "we", "define", "the", "segmentation", "loss", "in", "(", "[", "reference", "]", ")", "as", "the", "cross", "-", "entropy", "loss", "for", "images", "from", "the", "source", "domain", ":", "where", "is", "the", "ground", "truth", "annotations", "for", "source", "images", "and", "is", "the", "segmentation", "output", ".", "Second", ",", "for", "images", "in", "the", "target", "domain", ",", "we", "forward", "them", "to", "and", "obtain", "the", "prediction", ".", "To", "make", "the", "distribution", "of", "closer", "to", ",", "we", "use", "an", "adversarial", "loss", "in", "(", "[", "reference", "]", ")", "as", ":", "This", "loss", "is", "designed", "to", "train", "the", "segmentation", "network", "and", "fool", "the", "discriminator", "by", "maximizing", "the", "probability", "of", "the", "target", "prediction", "being", "considered", "as", "the", "source", "prediction", ".", "subsection", ":", "Multi", "-", "level", "Adversarial", "Learning", "Although", "performing", "adversarial", "learning", "in", "the", "output", "space", "directly", "adapts", "predictions", ",", "low", "-", "level", "features", "may", "not", "be", "adapted", "well", "as", "they", "are", "far", "away", "from", "the", "output", ".", "Similar", "to", "the", "deep", "supervision", "method", "that", "uses", "auxiliary", "loss", "for", "semantic", "segmentation", ",", "we", "incorporate", "additional", "adversarial", "module", "in", "the", "low", "-", "level", "feature", "space", "to", "enhance", "the", "adaptation", ".", "The", "training", "objective", "for", "the", "segmentation", "network", "can", "be", "extended", "from", "(", "[", "reference", "]", ")", "as", ":", "where", "indicates", "the", "level", "used", "for", "predicting", "the", "segmentation", "output", ".", "We", "note", "that", ",", "the", "segmentation", "output", "is", "still", "predicted", "in", "each", "feature", "space", ",", "before", "passing", "through", "individual", "discriminators", "for", "adversarial", "learning", ".", "Hence", ",", "and", "remain", "in", "the", "same", "form", "as", "in", "(", "[", "reference", "]", ")", "and", "(", "[", "reference", "]", ")", ",", "respectively", ".", "Based", "on", "(", "[", "reference", "]", ")", ",", "we", "optimize", "the", "following", "min", "-", "max", "criterion", ":", "The", "ultimate", "goal", "is", "to", "minimize", "the", "segmentation", "loss", "in", "for", "source", "images", ",", "while", "maximizing", "the", "probability", "of", "target", "predictions", "being", "considered", "as", "source", "predictions", ".", "road", "sidewalk", "building", "wall", "fence", "pole", "light", "sign", "veg", "terrain", "sky", "person", "rider", "car", "truck", "bus", "train", "mbike", "bike", "section", ":", "Network", "Architecture", "and", "Training", "Discriminator", ".", "For", "the", "discriminator", ",", "we", "use", "an", "architecture", "similar", "to", "but", "utilize", "all", "fully", "-", "convolutional", "layers", "to", "retain", "the", "spatial", "information", ".", "The", "network", "consists", "of", "5", "convolution", "layers", "with", "kernel", "and", "stride", "of", "2", ",", "where", "the", "channel", "number", "is", "{", "64", ",", "128", ",", "256", ",", "512", ",", "1", "}", ",", "respectively", ".", "Except", "for", "the", "last", "layer", ",", "each", "convolution", "layer", "is", "followed", "by", "a", "leaky", "ReLU", "parameterized", "by", ".", "An", "up", "-", "sampling", "layer", "is", "added", "to", "the", "last", "convolution", "layer", "for", "re", "-", "scaling", "the", "output", "to", "the", "size", "of", "the", "input", ".", "We", "do", "not", "use", "any", "batch", "-", "normalization", "layers", "as", "we", "jointly", "train", "the", "discriminator", "with", "the", "segmentation", "network", "using", "a", "small", "batch", "size", ".", "Segmentation", "Network", ".", "It", "is", "essential", "to", "build", "upon", "a", "good", "baseline", "model", "to", "achieve", "high", "-", "quality", "segmentation", "results", ".", "We", "adopt", "the", "DeepLab", "-", "v2", "framework", "with", "ResNet", "-", "101", "model", "pre", "-", "trained", "on", "ImageNet", "as", "our", "segmentation", "baseline", "network", ".", "However", ",", "we", "do", "not", "use", "the", "multi", "-", "scale", "fusion", "strategy", "due", "to", "the", "memory", "issue", ".", "Similar", "to", "the", "recent", "work", "on", "semantic", "segmentation", ",", "we", "remove", "the", "last", "classification", "layer", "and", "modify", "the", "stride", "of", "the", "last", "two", "convolution", "layers", "from", "2", "to", "1", ",", "making", "the", "resolution", "of", "the", "output", "feature", "maps", "effectively", "times", "the", "input", "image", "size", ".", "To", "enlarge", "the", "receptive", "field", ",", "we", "apply", "dilated", "convolution", "layers", "in", "conv4", "and", "conv5", "layers", "with", "a", "stride", "of", "2", "and", "4", ",", "respectively", ".", "After", "the", "last", "layer", ",", "we", "use", "the", "Atrous", "Spatial", "Pyramid", "Pooling", "(", "ASPP", ")", "as", "the", "final", "classifier", ".", "Finally", ",", "we", "apply", "an", "up", "-", "sampling", "layer", "along", "with", "the", "softmax", "output", "to", "match", "the", "size", "of", "the", "input", "image", ".", "Based", "on", "this", "architecture", ",", "our", "segmentation", "model", "achieves", "65.1", "%", "mean", "intersection", "-", "over", "-", "union", "(", "IoU", ")", "when", "trained", "on", "the", "Cityscapes", "training", "set", "and", "tested", "on", "the", "Cityscapes", "validation", "set", ".", "Multi", "-", "level", "Adaptation", "Model", ".", "We", "construct", "the", "above", "-", "mentioned", "discriminator", "and", "segmentation", "network", "as", "our", "single", "-", "level", "adaptation", "model", ".", "For", "the", "multi", "-", "level", "structure", ",", "we", "extract", "feature", "maps", "from", "the", "conv4", "layer", "and", "add", "an", "ASPP", "module", "as", "the", "auxiliary", "classifier", ".", "Similarly", ",", "a", "discriminator", "with", "the", "same", "architecture", "is", "added", "for", "adversarial", "learning", ".", "Figure", "[", "reference", "]", "shows", "the", "proposed", "multi", "-", "level", "adaptation", "model", ".", "In", "this", "paper", ",", "we", "use", "two", "levels", "due", "to", "the", "balance", "of", "its", "efficiency", "and", "accuracy", ".", "Network", "Training", ".", "To", "train", "the", "proposed", "single", "/", "multi", "-", "level", "adaptation", "model", ",", "we", "find", "that", "jointly", "training", "the", "segmentation", "network", "and", "discriminators", "in", "one", "stage", "is", "effective", ".", "In", "each", "training", "batch", ",", "we", "first", "forward", "the", "source", "image", "to", "optimize", "the", "segmentation", "network", "for", "in", "(", "[", "reference", "]", ")", "and", "generate", "the", "output", ".", "For", "the", "target", "image", ",", "we", "obtain", "the", "segmentation", "output", ",", "and", "pass", "it", "along", "with", "to", "the", "discriminator", "for", "optimizing", "in", "(", "[", "reference", "]", ")", ".", "In", "addition", ",", "we", "compute", "the", "adversarial", "loss", "in", "(", "[", "reference", "]", ")", "for", "the", "target", "prediction", ".", "For", "the", "multi", "-", "level", "training", "objective", "in", "(", "[", "reference", "]", ")", ",", "we", "simply", "repeat", "the", "same", "procedure", "for", "each", "adaptation", "module", ".", "We", "implement", "our", "network", "using", "the", "PyTorch", "toolbox", "on", "a", "single", "Titan", "X", "GPU", "with", "12", "GB", "memory", ".", "To", "train", "the", "segmentation", "network", ",", "we", "use", "the", "Stochastic", "Gradient", "Descent", "(", "SGD", ")", "optimizer", "with", "Nesterov", "acceleration", "where", "the", "momentum", "is", "0.9", "and", "the", "weight", "decay", "is", ".", "The", "initial", "learning", "rate", "is", "set", "as", "and", "is", "decreased", "using", "the", "polynomial", "decay", "with", "power", "of", "0.9", "as", "mentioned", "in", ".", "For", "training", "the", "discriminator", ",", "we", "use", "the", "Adam", "optimizer", "with", "the", "learning", "rate", "as", "and", "the", "same", "polynomial", "decay", "as", "the", "segmentation", "network", ".", "The", "momentum", "is", "set", "as", "0.9", "and", "0.99", ".", "section", ":", "Experimental", "Results", "In", "this", "section", ",", "we", "present", "experimental", "results", "to", "validate", "the", "proposed", "domain", "adaptation", "method", "for", "semantic", "segmentation", "under", "different", "settings", ".", "First", ",", "we", "show", "evaluations", "of", "the", "model", "trained", "on", "synthetic", "datasets", "(", "i.e.", ",", "GTA5", "and", "SYNTHIA", ")", "and", "test", "the", "adapted", "model", "on", "real", "-", "world", "images", "from", "the", "Cityscapes", "dataset", ".", "Extensive", "experiments", "including", "comparisons", "to", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "and", "ablation", "study", "are", "also", "conducted", ",", "e.g.", ",", "adaptation", "in", "the", "feature", "/", "output", "spaces", "and", "single", "/", "multi", "-", "level", "adversarial", "learning", ".", "Second", ",", "we", "carry", "out", "experiments", "on", "the", "Cross", "-", "City", "dataset", ",", "where", "the", "model", "is", "trained", "on", "one", "city", "and", "adapted", "to", "another", "city", "without", "using", "annotations", ".", "In", "all", "the", "experiments", ",", "the", "IoU", "metric", "is", "used", ".", "The", "code", "and", "model", "are", "available", "at", ".", "subsection", ":", "GTA5", "The", "GTA5", "dataset", "consists", "of", "images", "with", "the", "resolution", "of", "synthesized", "from", "the", "video", "game", "based", "on", "the", "city", "of", "Los", "Angeles", ".", "The", "ground", "truth", "annotations", "are", "compatible", "with", "the", "Cityscapes", "dataset", "that", "contains", "19", "categories", ".", "Following", ",", "we", "use", "the", "full", "set", "of", "GTA5", "and", "adapt", "the", "model", "to", "the", "Cityscapes", "training", "set", "with", "2975", "images", ".", "During", "testing", ",", "we", "evaluate", "on", "the", "Cityscapes", "validation", "set", "with", "500", "images", ".", "Overall", "Results", ".", "We", "present", "adaptation", "results", "in", "Table", "[", "reference", "]", "with", "comparisons", "to", "the", "state", "-", "of", "-", "the", "-", "art", "domain", "adaptation", "methods", ".", "For", "these", "approaches", ",", "the", "baseline", "model", "is", "trained", "using", "VGG", "-", "based", "architectures", ".", "To", "fairly", "evaluate", "our", "method", ",", "we", "first", "use", "the", "same", "baseline", "architecture", "(", "VGG", "-", "16", ")", "and", "train", "our", "model", "with", "the", "proposed", "single", "-", "level", "adaptation", "module", ".", "Table", "[", "reference", "]", "shows", "that", "our", "method", "performs", "favorably", "against", "the", "other", "algorithms", ".", "While", "these", "methods", "all", "have", "feature", "adaptation", "modules", ",", "our", "results", "show", "that", "adapting", "the", "model", "in", "the", "output", "space", "achieves", "better", "performance", ".", "We", "note", "that", "CyCADA", "has", "a", "pixel", "adaptation", "module", "by", "transforming", "source", "domain", "images", "to", "the", "target", "domain", "and", "hence", "obtains", "additional", "training", "samples", ".", "Although", "this", "strategy", "achieves", "a", "similar", "performance", "as", "ours", ",", "one", "can", "always", "apply", "pixel", "transformation", "combined", "with", "our", "output", "space", "adaptation", "to", "improve", "the", "results", ".", "On", "the", "other", "hand", ",", "we", "argue", "that", "utilizing", "a", "stronger", "baseline", "model", "is", "critical", "for", "understanding", "the", "importance", "of", "different", "adaptation", "components", "as", "well", "as", "for", "enhancing", "the", "performance", "to", "enable", "real", "-", "world", "applications", ".", "Thus", ",", "we", "use", "the", "ResNet", "-", "101", "based", "network", "introduced", "in", "Section", "[", "reference", "]", "and", "train", "the", "proposed", "adaptation", "model", ".", "Table", "[", "reference", "]", "shows", "the", "baseline", "results", "only", "trained", "on", "source", "images", "without", "adaptation", ",", "with", "comparisons", "to", "our", "adapted", "models", "under", "different", "settings", ",", "including", "feature", "adaptation", "and", "single", "/", "multi", "-", "level", "adversarial", "learning", "in", "the", "output", "space", ".", "Figure", "[", "reference", "]", "presents", "some", "example", "results", "for", "adapted", "segmentation", ".", "We", "note", "that", "for", "small", "objects", "such", "as", "poles", "and", "traffic", "signs", ",", "they", "are", "harder", "to", "adapt", "since", "they", "easily", "get", "merged", "with", "background", "classes", ".", "In", "addition", ",", "another", "factor", "to", "evaluate", "the", "adaptation", "performance", "is", "to", "measure", "how", "much", "gap", "is", "narrowed", "between", "the", "adaptation", "model", "and", "the", "fully", "-", "supervised", "model", ".", "Hence", ",", "we", "train", "the", "model", "using", "annotated", "ground", "truths", "in", "the", "Cityscapes", "dataset", "as", "the", "oracle", "results", ".", "Table", "[", "reference", "]", "shows", "the", "gap", "under", "different", "baseline", "models", ".", "We", "observe", "that", ",", "although", "the", "oracle", "result", "does", "not", "differ", "a", "lot", "between", "VGG", "-", "16", "and", "ResNet", "-", "101", "based", "models", ",", "the", "gap", "is", "larger", "for", "the", "VGG", "one", ".", "It", "suggests", "us", "that", "to", "narrow", "the", "gap", ",", "using", "a", "deeper", "model", "with", "larger", "capacity", "is", "more", "practical", ".", "Parameter", "Analysis", ".", "During", "optimizing", "the", "segmentation", "network", ",", "it", "is", "essential", "to", "balance", "the", "weight", "between", "segmentation", "and", "adversarial", "losses", ".", "We", "first", "consider", "the", "single", "-", "level", "case", "in", "(", "[", "reference", "]", ")", "and", "conduct", "experiments", "to", "observe", "the", "impact", "of", "changing", ".", "Table", "[", "reference", "]", "shows", "that", "a", "smaller", "may", "not", "facilitate", "the", "training", "process", "significantly", ",", "while", "a", "larger", "may", "propagate", "incorrect", "gradients", "to", "the", "network", ".", "We", "empirically", "choose", "as", "0.001", "in", "the", "single", "-", "level", "setting", ".", "Feature", "Level", "v.s", ".", "Output", "Space", "Adaptation", ".", "In", "the", "single", "-", "level", "setting", "in", "(", "[", "reference", "]", ")", ",", "we", "compare", "results", "by", "using", "feature", "-", "level", "or", "output", "space", "adaptation", "via", "adversarial", "learning", ".", "For", "feature", "-", "level", "adaptation", ",", "we", "adopt", "a", "similar", "strategy", "as", "used", "in", "and", "train", "our", "model", "accordingly", ".", "Table", "[", "reference", "]", "shows", "that", "the", "proposed", "adaptation", "method", "in", "the", "output", "space", "performs", "better", "than", "the", "one", "in", "the", "feature", "level", ".", "In", "addition", ",", "Table", "[", "reference", "]", "shows", "that", "adaptation", "in", "the", "feature", "space", "is", "more", "sensitive", "to", ",", "which", "causes", "the", "training", "process", "more", "difficult", ",", "while", "output", "space", "adaptation", "allows", "for", "a", "wider", "range", "of", ".", "One", "reason", "is", "that", "as", "feature", "adaptation", "is", "performed", "in", "the", "high", "-", "dimensional", "space", ",", "the", "problem", "for", "the", "discriminator", "becomes", "easier", ".", "Thus", ",", "such", "an", "adapted", "model", "can", "not", "effectively", "match", "distributions", "between", "source", "and", "target", "domains", "via", "adversarial", "learning", ".", "Single", "-", "level", "v.s", ".", "Multi", "-", "level", "Adversarial", "Learning", ".", "We", "have", "shown", "the", "merits", "of", "adopting", "adversarial", "learning", "in", "the", "output", "space", ".", "In", "addition", ",", "we", "present", "the", "results", "of", "using", "multi", "-", "level", "adversarial", "learning", "in", "Table", "[", "reference", "]", ".", "Here", ",", "we", "utilize", "an", "additional", "adversarial", "module", "(", "see", "Figure", "[", "reference", "]", ")", "and", "jointly", "optimize", "(", "[", "reference", "]", ")", "for", "two", "levels", ".", "To", "properly", "balance", "and", ",", "we", "use", "the", "same", "weight", "as", "in", "the", "single", "-", "level", "setting", "for", "the", "high", "-", "level", "output", "space", "(", "i.e.", ",", "=", "1", "and", "=", "0.001", ")", ".", "Since", "the", "low", "-", "level", "output", "carries", "less", "information", "to", "predict", "the", "segmentation", ",", "we", "use", "smaller", "weights", "for", "both", "the", "segmentation", "and", "adversarial", "loss", "(", "i.e.", ",", "=", "0.1", "and", "=", "0.0002", ")", ".", "Evaluation", "results", "show", "that", "our", "multi", "-", "level", "adversarial", "adaptation", "further", "improves", "the", "segmentation", "accuracy", ".", "More", "results", "and", "analysis", "are", "presented", "in", "the", "supplementary", "material", ".", "road", "sidewalk", "building", "light", "sign", "veg", "sky", "person", "rider", "car", "bus", "mbike", "bike", "subsection", ":", "SYNTHIA", "To", "adapt", "from", "the", "SYNTHIA", "to", "Cityscapes", "datasets", ",", "we", "use", "the", "SYNTHIA", "-", "RAND", "-", "CITYSCAPES", "set", "as", "the", "source", "domain", "which", "contains", "9400", "images", "compatible", "with", "the", "cityscapes", "annotated", "classes", ".", "Similar", "to", ",", "we", "evaluate", "images", "on", "the", "Cityscapes", "validation", "set", "with", "13", "classes", ".", "For", "the", "weight", "in", "(", "[", "reference", "]", ")", "and", "(", "[", "reference", "]", ")", ",", "we", "use", "the", "same", "ones", "as", "in", "the", "case", "of", "GTA5", "dataset", ".", "Table", "[", "reference", "]", "shows", "evaluation", "results", "of", "the", "proposed", "algorithm", "against", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "that", "use", "feature", "adaptation", ".", "Similar", "to", "the", "experiments", "with", "the", "GTA5", "dataset", ",", "we", "first", "utilize", "the", "same", "VGG", "-", "based", "model", "and", "train", "our", "single", "-", "level", "adaptation", "model", "for", "fair", "comparisons", ".", "The", "experimental", "results", "suggest", "that", "adapting", "the", "model", "in", "the", "output", "space", "performs", "better", ".", "Second", ",", "we", "compare", "results", "using", "different", "components", "of", "the", "proposed", "method", "with", "the", "ResNet", "based", "model", ".", "We", "show", "that", "the", "multi", "-", "level", "adaptation", "module", "improves", "the", "results", "over", "the", "baseline", ",", "feature", "space", "adaptation", "and", "single", "-", "level", "adaptation", "models", ".", "In", "addition", ",", "we", "present", "comparisons", "of", "mean", "IoU", "gap", "between", "adapted", "and", "oracle", "results", "in", "Table", "[", "reference", "]", ".", "Our", "method", "achieves", "the", "smallest", "gap", "and", "is", "the", "only", "one", "that", "can", "minimize", "the", "gap", "below", "30", "%", ".", "road", "sidewalk", "building", "light", "sign", "veg", "sky", "person", "rider", "car", "bus", "mbike", "bike", "subsection", ":", "Cross", "-", "City", "Dataset", "In", "addition", "to", "the", "synthetic", "-", "to", "-", "real", "adaptation", "for", "a", "larger", "domain", "gap", ",", "we", "conduct", "experiment", "on", "the", "Cross", "-", "City", "dataset", "with", "smaller", "domain", "gaps", "between", "cities", ".", "The", "dataset", "contains", "four", "different", "cities", ":", "Rio", ",", "Rome", ",", "Tokyo", "and", "Taipei", ",", "in", "which", "each", "city", "has", "3200", "images", "without", "annotations", "and", "100", "images", "with", "pixel", "-", "level", "ground", "truths", "for", "13", "classes", ".", "Similar", "to", ",", "we", "use", "the", "Cityscapes", "training", "set", "as", "the", "source", "domain", "and", "adapt", "it", "to", "each", "target", "city", "using", "3200", "images", ",", "while", "100", "annotated", "images", "are", "used", "for", "evaluation", ".", "Since", "a", "smaller", "domain", "gap", "results", "in", "smaller", "output", "differences", ",", "we", "use", "smaller", "weights", "for", "the", "adversarial", "loss", "(", "i.e.", ",", ")", "when", "training", "our", "models", ",", "while", "the", "weights", "for", "segmentation", "remain", "the", "same", "as", "previous", "experiments", ".", "We", "show", "our", "results", "in", "Table", "[", "reference", "]", "with", "comparisons", "to", "and", "our", "baseline", "models", "under", "different", "settings", ".", "Again", ",", "our", "final", "multi", "-", "level", "model", "achieves", "consistent", "improvement", "for", "different", "cities", ",", "which", "demonstrates", "the", "advantages", "of", "the", "proposed", "adaptation", "method", "in", "the", "output", "space", ".", "Note", "that", "the", "state", "-", "of", "-", "the", "-", "art", "method", "uses", "a", "different", "baseline", "model", ",", "and", "we", "present", "it", "as", "a", "reference", "to", "analyze", "how", "much", "the", "proposed", "algorithm", "can", "improve", ".", "section", ":", "Concluding", "Remarks", "In", "this", "paper", ",", "we", "exploit", "the", "fact", "that", "segmentations", "are", "structured", "outputs", "and", "share", "many", "similarities", "between", "source", "and", "target", "domains", ".", "We", "tackle", "the", "domain", "adaptation", "problem", "for", "semantic", "segmentation", "via", "adversarial", "learning", "in", "the", "output", "space", ".", "To", "further", "enhance", "the", "adapted", "model", ",", "we", "construct", "a", "multi", "-", "level", "adversarial", "network", "to", "effectively", "perform", "output", "space", "domain", "adaptation", "at", "different", "feature", "levels", ".", "Experimental", "results", "show", "that", "the", "proposed", "method", "performs", "favorably", "against", "numerous", "baseline", "models", "and", "the", "state", "-", "of", "-", "the", "-", "art", "algorithms", ".", "We", "hope", "that", "our", "proposed", "method", "can", "be", "a", "generic", "adaptation", "model", "for", "a", "wide", "range", "of", "pixel", "-", "level", "prediction", "tasks", ".", "Acknowledgments", ".", "W.", "-", "C.", "Hung", "is", "supported", "in", "part", "by", "the", "NSF", "CAREER", "Grant", "#", "1149783", ",", "gifts", "from", "Adobe", "and", "NVIDIA", ".", "bibliography", ":", "References"], "entity": [[9, 11, ["Semantic", "Segmentation"], "Task"], [11, 17, ["Convolutional", "neural", "network", "-", "based", "approaches"], "Method"], [18, 20, ["semantic", "segmentation"], "Task"], [42, 44, ["labeling", "process"], "Task"], [75, 78, ["adversarial", "learning", "method"], "Method"], [79, 81, ["domain", "adaptation"], "Method"], [85, 87, ["semantic", "segmentation"], "Task"], [89, 91, ["semantic", "segmentations"], "Task"], [107, 109, ["adversarial", "learning"], "Method"], [124, 129, ["multi", "-", "level", "adversarial", "network"], "Method"], [155, 160, ["synthetic", "-", "to", "-", "real"], "Task"], [161, 165, ["cross", "-", "city", "scenarios"], "Task"], [187, 188, ["accuracy"], "Metric"], [189, 191, ["visual", "quality"], "Metric"], [195, 197, ["Semantic", "segmentation"], "Task"], [225, 228, ["convolutional", "neural", "networks"], "Method"], [229, 230, ["CNNs"], "Method"], [236, 238, ["semantic", "segmentation"], "Task"], [241, 246, ["autonomous", "driving", "and", "image", "editing"], "Task"], [250, 254, ["CNN", "-", "based", "approaches"], "Method"], [340, 342, ["supervised", "model"], "Method"], [359, 361, ["labor", "cost"], "Metric"], [367, 369, ["knowledge", "transfer"], "Method"], [397, 399, ["image", "classification"], "Task"], [427, 429, ["semantic", "segmentation"], "Task"], [430, 432, ["adversarial", "learning"], "Method"], [442, 445, ["image", "classification", "task"], "Task"], [446, 448, ["feature", "adaptation"], "Method"], [449, 451, ["semantic", "segmentation"], "Task"], [485, 491, ["adapting", "pixel", "-", "level", "prediction", "tasks"], "Task"], [494, 496, ["feature", "adaptation"], "Method"], [498, 500, ["semantic", "segmentation"], "Task"], [572, 577, ["output", "(", "segmentation", ")", "space"], "Task"], [597, 599, ["semantic", "segmentation"], "Task"], [605, 607, ["adversarial", "learning"], "Method"], [636, 639, ["generative", "adversarial", "network"], "Method"], [640, 641, ["GAN"], "Method"], [654, 656, ["segmentation", "model"], "Method"], [665, 666, ["discriminator"], "Method"], [682, 684, ["adversarial", "loss"], "Method"], [687, 689, ["segmentation", "model"], "Method"], [693, 694, ["discriminator"], "Method"], [771, 775, ["multi", "-", "level", "strategy"], "Method"], [777, 779, ["adversarial", "learning"], "Method"], [785, 787, ["segmentation", "model"], "Method"], [801, 802, ["segmentation"], "Task"], [821, 826, ["multi", "-", "level", "adversarial", "learning"], "Method"], [832, 838, ["end", "-", "to", "-", "end", "training"], "Task"], [840, 842, ["segmentation", "model"], "Method"], [843, 844, ["discriminators"], "Method"], [868, 869, ["discriminators"], "Method"], [873, 875, ["segmentation", "model"], "Method"], [882, 884, ["computational", "requirements"], "Metric"], [889, 891, ["labor", "cost"], "Metric"], [892, 896, ["annotating", "segmentation", "ground", "truth"], "Task"], [955, 958, ["domain", "adaptation", "method"], "Method"], [983, 987, ["real", "-", "world", "applications"], "Task"], [995, 997, ["adaptation", "approach"], "Method"], [1008, 1010, ["adversarial", "adaptation"], "Method"], [1023, 1028, ["multi", "-", "level", "adversarial", "learning"], "Method"], [1041, 1047, ["synthetic", "-", "to", "-", "real", "setting"], "Task"], [1119, 1122, ["domain", "adaptation", "method"], "Method"], [1123, 1128, ["pixel", "-", "level", "semantic", "segmentation"], "Task"], [1129, 1131, ["adversarial", "learning"], "Method"], [1178, 1180, ["segmentation", "model"], "Method"], [1191, 1193, ["Semantic", "Segmentation"], "Task"], [1201, 1204, ["semantic", "segmentation", "methods"], "Method"], [1212, 1215, ["deep", "neural", "networks"], "Method"], [1225, 1227, ["classification", "CNN"], "Method"], [1230, 1231, ["AlexNet"], "Method"], [1232, 1233, ["VGG"], "Method"], [1235, 1236, ["ResNet"], "Method"], [1239, 1243, ["fully", "-", "convolutional", "network"], "Method"], [1244, 1245, ["FCN"], "Method"], [1247, 1249, ["semantic", "segmentation"], "Task"], [1293, 1295, ["deep", "CNNs"], "Method"], [1300, 1306, ["weakly", "and", "semi", "-", "supervised", "approaches"], "Method"], [1314, 1317, ["heavy", "labeling", "cost"], "Metric"], [1318, 1322, ["collecting", "segmentation", "ground", "truths"], "Task"], [1329, 1331, ["world", "applications"], "Task"], [1357, 1359, ["annotation", "problem"], "Task"], [1366, 1367, ["rendering"], "Method"], [1370, 1371, ["GTA5"], "Method"], [1376, 1378, ["data", "collection"], "Task"], [1383, 1387, ["pixel", "-", "level", "annotation"], "Task"], [1393, 1395, ["automated", "process"], "Method"], [1409, 1411, ["joint", "learning"], "Task"], [1463, 1465, ["rendering", "techniques"], "Method"], [1482, 1484, ["domain", "adaptation"], "Method"], [1490, 1492, ["Domain", "Adaptation"], "Method"], [1493, 1496, ["Domain", "adaptation", "methods"], "Method"], [1497, 1499, ["image", "classification"], "Task"], [1505, 1509, ["domain", "-", "shift", "problem"], "Task"], [1522, 1524, ["CNN", "classifiers"], "Method"], [1551, 1552, ["Ganin"], "Method"], [1554, 1559, ["Domain", "-", "Adversarial", "Neural", "Network"], "Method"], [1560, 1561, ["DANN"], "Method"], [1578, 1580, ["loss", "functions"], "Method"], [1581, 1582, ["classifiers"], "Method"], [1586, 1588, ["PixelDA", "method"], "Method"], [1589, 1591, ["domain", "adaptation"], "Method"], [1592, 1594, ["image", "classification"], "Task"], [1616, 1618, ["domain", "adaptation"], "Method"], [1619, 1624, ["pixel", "-", "level", "prediction", "tasks"], "Task"], [1635, 1637, ["domain", "adaptation"], "Method"], [1638, 1640, ["semantic", "segmentation"], "Task"], [1642, 1644, ["adversarial", "learning"], "Method"], [1647, 1648, ["DANN"], "Method"], [1651, 1655, ["fully", "-", "convolutional", "way"], "Method"], [1656, 1658, ["feature", "representations"], "Method"], [1665, 1667, ["constrained", "CNN"], "Method"], [1673, 1678, ["synthetic", "-", "to", "-", "real"], "Task"], [1679, 1683, ["cross", "-", "city", "images"], "Task"], [1685, 1690, ["class", "-", "wise", "adversarial", "learning"], "Method"], [1691, 1693, ["label", "transfer"], "Method"], [1697, 1699, ["PixelDA", "method"], "Method"], [1704, 1705, ["CyCADA"], "Method"], [1707, 1708, ["CycleGAN"], "Method"], [1728, 1732, ["feature", "space", "adversarial", "learning"], "Method"], [1734, 1737, ["feature", "space", "adaptation"], "Method"], [1742, 1744, ["image", "classification"], "Task"], [1745, 1749, ["pixel", "-", "level", "tasks"], "Task"], [1751, 1753, ["semantic", "segmentation"], "Task"], [1794, 1796, ["adversarial", "learning"], "Method"], [1822, 1824, ["segmentation", "network"], "Method"], [1826, 1827, ["discriminator"], "Method"], [1837, 1842, ["multi", "-", "level", "adversarial", "learning"], "Method"], [1869, 1871, ["segmentation", "network"], "Method"], [1872, 1873, ["optimizing"], "Task"], [1984, 1986, ["Objective", "Function"], "Metric"], [1987, 1989, ["Domain", "Adaptation"], "Method"], [1997, 1999, ["adaptation", "task"], "Task"], [2001, 2003, ["loss", "functions"], "Method"], [2010, 2014, ["cross", "-", "entropy", "loss"], "Metric"], [2026, 2028, ["adversarial", "loss"], "Method"], [2068, 2071, ["Output", "Space", "Adaptation"], "Method"], [2073, 2075, ["image", "classification"], "Task"], [2094, 2096, ["semantic", "segmentation"], "Task"], [2116, 2118, ["semantic", "segmentation"], "Task"], [2191, 2194, ["adversarial", "learning", "scheme"], "Method"], [2197, 2202, ["Single", "-", "level", "Adversarial", "Learning"], "Method"], [2202, 2204, ["Discriminator", "Training"], "Method"], [2211, 2213, ["segmentation", "network"], "Method"], [2214, 2216, ["adversarial", "learning"], "Method"], [2221, 2223, ["training", "objective"], "Metric"], [2225, 2226, ["discriminator"], "Method"], [2244, 2248, ["fully", "-", "convolutional", "discriminator"], "Method"], [2250, 2254, ["cross", "-", "entropy", "loss"], "Metric"], [2293, 2296, ["Segmentation", "Network", "Training"], "Method"], [2312, 2316, ["cross", "-", "entropy", "loss"], "Metric"], [2367, 2369, ["adversarial", "loss"], "Method"], [2384, 2386, ["segmentation", "network"], "Method"], [2389, 2390, ["discriminator"], "Method"], [2407, 2412, ["Multi", "-", "level", "Adversarial", "Learning"], "Method"], [2414, 2416, ["adversarial", "learning"], "Method"], [2445, 2448, ["deep", "supervision", "method"], "Method"], [2453, 2455, ["semantic", "segmentation"], "Task"], [2459, 2461, ["adversarial", "module"], "Method"], [2471, 2472, ["adaptation"], "Task"], [2474, 2476, ["training", "objective"], "Metric"], [2478, 2480, ["segmentation", "network"], "Method"], [2521, 2522, ["discriminators"], "Method"], [2523, 2525, ["adversarial", "learning"], "Method"], [2562, 2566, ["min", "-", "max", "criterion"], "Metric"], [2574, 2576, ["segmentation", "loss"], "Metric"], [2615, 2617, ["Network", "Architecture"], "Method"], [2619, 2620, ["Discriminator"], "Method"], [2623, 2624, ["discriminator"], "Method"], [2634, 2638, ["fully", "-", "convolutional", "layers"], "Method"], [2649, 2651, ["convolution", "layers"], "Method"], [2684, 2686, ["convolution", "layer"], "Method"], [2690, 2693, ["leaky", "ReLU", "parameterized"], "Method"], [2696, 2700, ["up", "-", "sampling", "layer"], "Method"], [2705, 2707, ["convolution", "layer"], "Method"], [2725, 2729, ["batch", "-", "normalization", "layers"], "Method"], [2734, 2735, ["discriminator"], "Method"], [2737, 2739, ["segmentation", "network"], "Method"], [2745, 2747, ["Segmentation", "Network"], "Method"], [2756, 2758, ["baseline", "model"], "Method"], [2763, 2764, ["segmentation"], "Task"], [2769, 2773, ["DeepLab", "-", "v2", "framework"], "Method"], [2774, 2778, ["ResNet", "-", "101", "model"], "Method"], [2782, 2783, ["ImageNet"], "Method"], [2785, 2788, ["segmentation", "baseline", "network"], "Method"], [2796, 2801, ["multi", "-", "scale", "fusion", "strategy"], "Method"], [2813, 2815, ["semantic", "segmentation"], "Task"], [2820, 2822, ["classification", "layer"], "Method"], [2830, 2832, ["convolution", "layers"], "Method"], [2860, 2863, ["dilated", "convolution", "layers"], "Method"], [2864, 2868, ["conv4", "and", "conv5", "layers"], "Method"], [2886, 2890, ["Atrous", "Spatial", "Pyramid", "Pooling"], "Method"], [2891, 2892, ["ASPP"], "Method"], [2896, 2897, ["classifier"], "Method"], [2903, 2907, ["up", "-", "sampling", "layer"], "Method"], [2927, 2929, ["segmentation", "model"], "Method"], [2932, 2938, ["mean", "intersection", "-", "over", "-", "union"], "Metric"], [2939, 2940, ["IoU"], "Metric"], [2956, 2961, ["Multi", "-", "level", "Adaptation", "Model"], "Method"], [2968, 2972, ["discriminator", "and", "segmentation", "network"], "Method"], [2974, 2979, ["single", "-", "level", "adaptation", "model"], "Method"], [2993, 2995, ["conv4", "layer"], "Method"], [2998, 3000, ["ASPP", "module"], "Method"], [3002, 3004, ["auxiliary", "classifier"], "Method"], [3008, 3009, ["discriminator"], "Method"], [3016, 3018, ["adversarial", "learning"], "Task"], [3026, 3031, ["multi", "-", "level", "adaptation", "model"], "Method"], [3048, 3049, ["accuracy"], "Metric"], [3050, 3052, ["Network", "Training"], "Method"], [3057, 3058, ["single"], "Method"], [3059, 3064, ["multi", "-", "level", "adaptation", "model"], "Method"], [3071, 3073, ["segmentation", "network"], "Method"], [3074, 3075, ["discriminators"], "Method"], [3095, 3097, ["segmentation", "network"], "Method"], [3153, 3155, ["target", "prediction"], "Task"], [3158, 3163, ["multi", "-", "level", "training", "objective"], "Task"], [3178, 3180, ["adaptation", "module"], "Method"], [3187, 3189, ["PyTorch", "toolbox"], "Method"], [3192, 3195, ["Titan", "X", "GPU"], "Method"], [3203, 3205, ["segmentation", "network"], "Method"], [3209, 3212, ["Stochastic", "Gradient", "Descent"], "Method"], [3213, 3214, ["SGD"], "Method"], [3232, 3234, ["learning", "rate"], "Metric"], [3255, 3256, ["discriminator"], "Method"], [3260, 3262, ["Adam", "optimizer"], "Method"], [3264, 3266, ["learning", "rate"], "Metric"], [3274, 3276, ["segmentation", "network"], "Method"], [3302, 3305, ["domain", "adaptation", "method"], "Method"], [3306, 3308, ["semantic", "segmentation"], "Task"], [3361, 3363, ["ablation", "study"], "Method"], [3369, 3370, ["adaptation"], "Task"], [3377, 3378, ["single"], "Method"], [3379, 3384, ["multi", "-", "level", "adversarial", "learning"], "Method"], [3534, 3537, ["domain", "adaptation", "methods"], "Method"], [3543, 3545, ["baseline", "model"], "Method"], [3548, 3552, ["VGG", "-", "based", "architectures"], "Method"], [3567, 3570, ["VGG", "-", "16"], "Method"], [3578, 3583, ["single", "-", "level", "adaptation", "module"], "Method"], [3604, 3607, ["feature", "adaptation", "modules"], "Method"], [3626, 3627, ["CyCADA"], "Method"], [3629, 3632, ["pixel", "adaptation", "module"], "Method"], [3662, 3664, ["pixel", "transformation"], "Method"], [3667, 3670, ["output", "space", "adaptation"], "Method"], [3686, 3688, ["baseline", "model"], "Method"], [3696, 3698, ["adaptation", "components"], "Method"], [3707, 3711, ["real", "-", "world", "applications"], "Task"], [3717, 3722, ["ResNet", "-", "101", "based", "network"], "Method"], [3732, 3734, ["adaptation", "model"], "Method"], [3762, 3764, ["feature", "adaptation"], "Method"], [3765, 3766, ["single"], "Method"], [3767, 3772, ["multi", "-", "level", "adversarial", "learning"], "Method"], [3786, 3788, ["adapted", "segmentation"], "Task"], [3824, 3825, ["adaptation"], "Task"], [3836, 3838, ["adaptation", "model"], "Method"], [3840, 3844, ["fully", "-", "supervised", "model"], "Method"], [3890, 3893, ["VGG", "-", "16"], "Method"], [3894, 3899, ["ResNet", "-", "101", "based", "models"], "Method"], [3906, 3908, ["VGG", "one"], "Method"], [3920, 3922, ["deeper", "model"], "Method"], [3929, 3931, ["Parameter", "Analysis"], "Method"], [3935, 3937, ["segmentation", "network"], "Method"], [3987, 3989, ["training", "process"], "Task"], [4014, 4016, ["Feature", "Level"], "Method"], [4018, 4021, ["Output", "Space", "Adaptation"], "Method"], [4040, 4043, ["feature", "-", "level"], "Method"], [4044, 4047, ["output", "space", "adaptation"], "Method"], [4048, 4050, ["adversarial", "learning"], "Method"], [4079, 4081, ["adaptation", "method"], "Method"], [4117, 4119, ["training", "process"], "Task"], [4123, 4126, ["output", "space", "adaptation"], "Method"], [4138, 4140, ["feature", "adaptation"], "Method"], [4153, 4154, ["discriminator"], "Method"], [4161, 4163, ["adapted", "model"], "Method"], [4174, 4176, ["adversarial", "learning"], "Method"], [4182, 4187, ["Multi", "-", "level", "Adversarial", "Learning"], "Method"], [4195, 4197, ["adversarial", "learning"], "Method"], [4211, 4216, ["multi", "-", "level", "adversarial", "learning"], "Method"], [4228, 4230, ["adversarial", "module"], "Method"], [4295, 4296, ["segmentation"], "Task"], [4304, 4305, ["segmentation"], "Task"], [4306, 4308, ["adversarial", "loss"], "Metric"], [4323, 4328, ["multi", "-", "level", "adversarial", "adaptation"], "Method"], [4331, 4333, ["segmentation", "accuracy"], "Metric"], [4461, 4463, ["feature", "adaptation"], "Method"], [4478, 4482, ["VGG", "-", "based", "model"], "Method"], [4485, 4490, ["single", "-", "level", "adaptation", "model"], "Method"], [4523, 4526, ["ResNet", "based", "model"], "Method"], [4531, 4536, ["multi", "-", "level", "adaptation", "module"], "Method"], [4543, 4546, ["feature", "space", "adaptation"], "Method"], [4617, 4623, ["synthetic", "-", "to", "-", "real", "adaptation"], "Task"], [4747, 4748, ["segmentation"], "Task"], [4779, 4783, ["multi", "-", "level", "model"], "Method"], [4797, 4799, ["adaptation", "method"], "Method"], [4872, 4874, ["semantic", "segmentation"], "Task"], [4875, 4877, ["adversarial", "learning"], "Method"], [4892, 4897, ["multi", "-", "level", "adversarial", "network"], "Method"], [4943, 4945, ["adaptation", "model"], "Method"], [4950, 4955, ["pixel", "-", "level", "prediction", "tasks"], "Task"], [134, 136, ["domain", "adaptation"], "Method"], [150, 152, ["domain", "adaptation"], "Method"], [370, 372, ["domain", "adaptation"], "Method"], [567, 569, ["domain", "adaptation"], "Method"], [593, 595, ["domain", "adaptation"], "Method"], [1032, 1033, ["single"], "Method"], [1162, 1167, ["multi", "-", "level", "adversarial", "learning"], "Method"], [1757, 1759, ["feature", "adaptation"], "Method"], [1790, 1792, ["domain", "adaptation"], "Method"], [1813, 1815, ["domain", "adaptation"], "Method"], [3191, 3192, ["single"], "Method"], [3421, 3422, ["IoU"], "Metric"], [3955, 3956, ["single"], "Method"], [4009, 4010, ["single"], "Method"], [4024, 4025, ["single"], "Method"], [4052, 4055, ["feature", "-", "level"], "Method"], [4262, 4263, ["single"], "Method"], [4547, 4548, ["single"], "Method"], [4561, 4562, ["IoU"], "Metric"], [4868, 4870, ["domain", "adaptation"], "Method"], [4902, 4904, ["domain", "adaptation"], "Method"]]} {"doc_id": "1d0dcb458aa4d30b51f7c74b159be687f39120a0", "text": ["document", ":", "Pose", "-", "driven", "Deep", "Convolutional", "Model", "for", "Person", "Re", "-", "identification", "Feature", "extraction", "and", "matching", "are", "two", "crucial", "components", "in", "person", "Re", "-", "Identification", "(", "ReID", ")", ".", "The", "large", "pose", "deformations", "and", "the", "complex", "view", "variations", "exhibited", "by", "the", "captured", "person", "images", "significantly", "increase", "the", "difficulty", "of", "learning", "and", "matching", "of", "the", "features", "from", "person", "images", ".", "To", "overcome", "these", "difficulties", ",", "in", "this", "work", "we", "propose", "a", "Pose", "-", "driven", "Deep", "Convolutional", "(", "PDC", ")", "model", "to", "learn", "improved", "feature", "extraction", "and", "matching", "models", "from", "end", "to", "end", ".", "Our", "deep", "architecture", "explicitly", "leverages", "the", "human", "part", "cues", "to", "alleviate", "the", "pose", "variations", "and", "learn", "robust", "feature", "representations", "from", "both", "the", "global", "image", "and", "different", "local", "parts", ".", "To", "match", "the", "features", "from", "global", "human", "body", "and", "local", "body", "parts", ",", "a", "pose", "driven", "feature", "weighting", "sub", "-", "network", "is", "further", "designed", "to", "learn", "adaptive", "feature", "fusions", ".", "Extensive", "experimental", "analyses", "and", "results", "on", "three", "popular", "datasets", "demonstrate", "significant", "performance", "improvements", "of", "our", "model", "over", "all", "published", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "section", ":", "Introduction", "Person", "Re", "-", "Identification", "(", "ReID", ")", "is", "an", "important", "component", "in", "a", "video", "surveillance", "system", ".", "Here", "person", "ReID", "refers", "to", "the", "process", "of", "identifying", "a", "probe", "person", "from", "a", "gallery", "captured", "by", "different", "cameras", ",", "and", "is", "generally", "deployed", "in", "the", "following", "scenario", ":", "given", "a", "probe", "image", "or", "video", "sequence", "containing", "a", "specific", "person", "under", "a", "certain", "camera", ",", "querying", "the", "images", ",", "locations", ",", "and", "time", "stamps", "of", "this", "person", "from", "other", "cameras", ".", "Despite", "decades", "of", "studies", ",", "the", "person", "ReID", "problem", "is", "still", "far", "from", "being", "solved", ".", "This", "is", "mainly", "because", "of", "challenging", "situations", "like", "complex", "view", "variations", "and", "large", "pose", "deformations", "on", "the", "captured", "person", "images", ".", "Most", "of", "traditional", "works", "try", "to", "address", "these", "challenges", "with", "the", "following", "two", "approaches", ":", "(", "1", ")", "representing", "the", "visual", "appearance", "of", "a", "person", "using", "customized", "local", "invariant", "features", "extracted", "from", "images", "or", "(", "2", ")", "learning", "a", "discriminative", "distance", "metric", "to", "reduce", "the", "distance", "among", "features", "of", "images", "containing", "the", "same", "person", ".", "Because", "the", "human", "poses", "and", "viewpoints", "are", "uncontrollable", "in", "real", "scenarios", ",", "hand", "-", "coded", "features", "may", "be", "not", "robust", "enough", "to", "pose", "and", "viewpoint", "variations", ".", "Distance", "metric", "is", "computed", "for", "each", "pair", "of", "cameras", ",", "making", "distance", "metric", "learning", "based", "person", "ReID", "suffers", "from", "the", "computational", "complexity", ".", "In", "recent", "years", ",", "deep", "learning", "has", "demonstrated", "strong", "model", "capabilities", "and", "obtains", "very", "promising", "performances", "in", "many", "computer", "vision", "tasks", ".", "Meanwhile", ",", "the", "release", "of", "person", "ReID", "datasets", "like", "CUHK", "03", ",", "Market", "-", "1501", ",", "and", "MARS", ",", "both", "of", "which", "contain", "many", "annotated", "person", "images", ",", "makes", "training", "deep", "models", "for", "person", "ReID", "feasible", ".", "Therefore", ",", "many", "researchers", "attempt", "to", "leverage", "deep", "models", "in", "person", "ReID", ".", "Most", "of", "these", "methods", "first", "learn", "a", "pedestrian", "feature", "and", "then", "compute", "Euclidean", "distance", "to", "measure", "the", "similarity", "between", "two", "samples", ".", "More", "specifically", ",", "existing", "deep", "learning", "based", "person", "ReID", "approaches", "can", "be", "summarized", "into", "two", "categories", ":", "1", ")", "use", "Softmax", "Loss", "with", "person", "ID", "labels", "to", "learn", "a", "global", "representation", ",", "and", "2", ")", "first", "learn", "local", "representations", "using", "predefined", "rigid", "body", "parts", ",", "then", "fuse", "the", "local", "and", "global", "representations", "to", "depict", "person", "images", ".", "Deep", "learning", "based", "methods", "have", "demonstrated", "significant", "performance", "improvements", "over", "the", "traditional", "methods", ".", "Although", "these", "approaches", "have", "achieved", "remarkable", "results", "on", "mainstream", "person", "ReID", "datasets", ",", "most", "of", "them", "do", "not", "consider", "pose", "variation", "of", "human", "body", ".", "Because", "pose", "variations", "may", "significantly", "change", "the", "appearance", "of", "a", "person", ",", "considering", "the", "human", "pose", "cues", "is", "potential", "to", "help", "person", "re", "-", "identification", ".", "Although", "there", "are", "several", "methods", "that", "segment", "the", "person", "images", "according", "to", "the", "predefined", "configuration", ",", "such", "simple", "segmentation", "can", "not", "capture", "the", "pose", "cues", "effectively", ".", "Some", "recent", "works", "attempt", "to", "use", "pose", "estimation", "algorithms", "to", "predict", "human", "pose", "and", "then", "train", "deep", "models", "for", "person", "ReID", ".", "However", ",", "they", "use", "manually", "cropped", "human", "body", "parts", "and", "their", "models", "are", "not", "trained", "from", "end", "to", "end", ".", "Therefore", ",", "the", "potential", "of", "pose", "information", "to", "boost", "the", "ReID", "performance", "has", "not", "been", "fully", "explored", ".", "To", "better", "alleviate", "the", "challenges", "from", "pose", "variations", ",", "we", "propose", "a", "Pose", "-", "driven", "Deep", "Convolutional", "(", "PDC", ")", "model", "for", "person", "ReID", ".", "The", "proposed", "PDC", "model", "learns", "the", "global", "representation", "depicting", "the", "whole", "body", "and", "local", "representations", "depicting", "body", "parts", "simultaneously", ".", "The", "global", "representation", "is", "learned", "using", "the", "Softmax", "Loss", "with", "person", "ID", "labels", "on", "the", "whole", "input", "image", ".", "For", "the", "learning", "of", "local", "representations", ",", "a", "novel", "Feature", "Embedding", "sub", "-", "Net", "(", "FEN", ")", "is", "proposed", "to", "learn", "and", "readjust", "human", "parts", "so", "that", "parts", "are", "affine", "transformed", "and", "re", "-", "located", "at", "more", "reasonable", "regions", "which", "can", "be", "easily", "recognizable", "through", "two", "different", "cameras", ".", "In", "Feature", "Embedding", "sub", "-", "Net", ",", "each", "body", "part", "region", "is", "first", "automatically", "cropped", ".", "The", "cropped", "part", "regions", "are", "hence", "transformed", "by", "a", "Pose", "Transformation", "Network", "(", "PTN", ")", "to", "eliminate", "the", "pose", "variations", ".", "The", "local", "representations", "are", "hence", "learned", "on", "the", "transformed", "regions", ".", "We", "further", "propose", "a", "Feature", "Weighting", "sub", "-", "Net", "(", "FWN", ")", "to", "learn", "the", "weights", "of", "global", "representations", "and", "local", "representations", "on", "different", "parts", ".", "Therefore", ",", "more", "reasonable", "feature", "fusion", "is", "conducted", "to", "facilitate", "feature", "similarity", "measurement", ".", "Some", "more", "detailed", "descriptions", "to", "our", "local", "representation", "generation", "are", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "Our", "method", "first", "locates", "the", "key", "body", "joints", "from", "the", "input", "image", ",", ",", "illustrated", "in", "Fig", ".", "[", "reference", "]", "(", "c", ")", ".", "From", "the", "detected", "joints", ",", "six", "body", "parts", "are", "extracted", ",", ",", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "d", ")", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "e", ")", ",", "those", "parts", "are", "extracted", "and", "normalized", "into", "fixed", "sizes", "and", "orientations", ".", "Finally", ",", "they", "are", "fed", "into", "the", "Pose", "Transformation", "Network", "(", "PTN", ")", "to", "further", "eliminate", "the", "pose", "variations", ".", "With", "the", "normalized", "and", "transformed", "part", "regions", ",", ",", "Fig", ".", "[", "reference", "]", "(", "f", ")", ",", "local", "representations", "are", "learned", "by", "training", "the", "deep", "neural", "network", ".", "Different", "parts", "commonly", "convey", "different", "levels", "of", "discriminative", "cues", "to", "identify", "the", "person", ".", "We", "thus", "further", "learn", "weights", "for", "representations", "on", "different", "parts", "with", "a", "sub", "-", "network", ".", "Most", "of", "current", "deep", "learning", "based", "person", "ReID", "works", "do", "not", "consider", "the", "human", "pose", "cues", "and", "the", "weights", "of", "representation", "on", "different", "parts", ".", "This", "paper", "proposes", "a", "novel", "deep", "architecture", "that", "transforms", "body", "parts", "into", "normalized", "and", "homologous", "feature", "representations", "to", "better", "overcome", "the", "pose", "variations", ".", "Moreover", ",", "a", "sub", "-", "network", "is", "proposed", "to", "automatically", "learn", "weights", "for", "different", "parts", "to", "facilitate", "feature", "similarity", "measurement", ".", "Both", "the", "representation", "and", "weighting", "are", "learned", "jointly", "from", "end", "to", "end", ".", "Since", "pose", "estimation", "is", "not", "the", "focus", "of", "this", "paper", ",", "the", "used", "pose", "estimation", "algorithm", ",", ",", "Fully", "Convolutional", "Networks", "(", "FCN", ")", "based", "pose", "estimation", "method", "is", "simple", "and", "trained", "independently", ".", "Once", "the", "FCN", "is", "trained", ",", "it", "is", "incorporated", "in", "our", "framework", ",", "which", "is", "hence", "trained", "in", "an", "end", "-", "to", "-", "end", "manner", ",", ",", "using", "images", "as", "inputs", "and", "person", "ID", "labels", "as", "outputs", ".", "Experimental", "results", "on", "three", "popular", "datasets", "show", "that", "our", "algorithm", "significantly", "outperforms", "many", "state", "-", "of", "-", "the", "-", "art", "ones", ".", "section", ":", "Related", "Work", "Traditional", "algorithms", "perform", "person", "re", "-", "identification", "through", "two", "ways", ":", "(", "a", ")", "acquiring", "robust", "local", "features", "visually", "representing", "a", "person", "’s", "appearance", "and", "then", "encoding", "them", ";", "(", "b", ")", "closing", "the", "gap", "between", "a", "person", "’s", "different", "features", "by", "learning", "a", "discriminative", "distance", "metric", ".", "Some", "recent", "works", "have", "started", "to", "apply", "deep", "learning", "in", "person", "ReID", "and", "achieved", "promising", "performance", ".", "In", "the", "following", ",", "we", "briefly", "review", "recent", "deep", "learning", "based", "person", "ReID", "methods", ".", "Deep", "learning", "is", "commonly", "used", "to", "either", "learn", "a", "person", "’s", "representation", "or", "the", "distance", "metric", ".", "When", "handling", "a", "pair", "of", "person", "images", ",", "existing", "deep", "learning", "methods", "usually", "learn", "feature", "representations", "of", "each", "person", "by", "using", "a", "deep", "matching", "function", "from", "convolutional", "features", "or", "from", "the", "Fully", "Connected", "(", "FC", ")", "features", ".", "Apart", "from", "deep", "metric", "learning", "methods", ",", "some", "algorithms", "first", "learn", "image", "representations", "directly", "with", "the", "Triplet", "Loss", "or", "the", "Siamese", "Contrastive", "Loss", ",", "then", "utilize", "Euclidean", "distance", "for", "comparison", ".", "Wang", "use", "a", "joint", "learning", "framework", "to", "unify", "single", "-", "image", "representation", "and", "cross", "-", "image", "representation", "using", "a", "doublet", "or", "triplet", "CNN", ".", "Shi", "propose", "a", "moderate", "positive", "mining", "method", "to", "use", "deep", "distance", "metric", "learning", "for", "person", "ReID", ".", "Another", "novel", "method", "learns", "deep", "attributes", "feature", "for", "ReID", "with", "semi", "-", "supervised", "learning", ".", "Xiao", "train", "one", "network", "with", "several", "person", "ReID", "datasets", "using", "a", "Domain", "Guided", "Dropout", "algorithm", ".", "Predefined", "rigid", "body", "parts", "are", "also", "used", "by", "many", "deep", "learning", "based", "methods", "for", "the", "purpose", "of", "learning", "local", "pedestrian", "features", ".", "Different", "from", "these", "algorithms", ",", "our", "work", "and", "the", "ones", "in", "use", "more", "accurate", "human", "pose", "estimation", "algorithms", "to", "acquire", "human", "pose", "features", ".", "However", ",", "due", "to", "the", "limited", "accuracy", "of", "pose", "estimation", "algorithms", "as", "well", "as", "reasons", "like", "occlusion", "and", "lighting", "change", ",", "pose", "estimation", "might", "be", "not", "accurate", "enough", ".", "Moreover", ",", "different", "parts", "convey", "different", "levels", "of", "discriminative", "cues", ".", "Therefore", ",", "we", "normalize", "the", "part", "regions", "to", "get", "more", "robust", "feature", "representation", "using", "Feature", "Embedding", "sub", "-", "Net", "(", "FEN", ")", "and", "propose", "a", "Feature", "Weighting", "sub", "-", "Net", "(", "FWN", ")", "to", "learn", "the", "weight", "for", "each", "part", "feature", ".", "In", "this", "way", ",", "the", "part", "with", "high", "discriminative", "power", "can", "be", "identified", "and", "emphasized", ".", "This", "also", "makes", "our", "work", "different", "from", "existing", "ones", ",", "which", "do", "not", "consider", "the", "inaccuracy", "of", "human", "poses", "estimation", "and", "weighting", "on", "different", "parts", "features", ".", "section", ":", "Pose", "-", "driven", "Deep", "ReID", "Model", "In", "this", "section", ",", "we", "describe", "the", "overall", "framework", "of", "the", "proposed", "approach", ",", "where", "we", "mainly", "introduce", "the", "Feature", "Embedding", "sub", "-", "Net", "(", "FEN", ")", "and", "the", "Feature", "Weighting", "sub", "-", "Net", "(", "FWN", ")", ".", "Details", "about", "the", "training", "and", "test", "procedures", "of", "the", "proposed", "approach", "will", "also", "be", "presented", ".", "subsection", ":", "Framework", "Fig", ".", "[", "reference", "]", "shows", "the", "framework", "of", "our", "proposed", "deep", "ReID", "model", ".", "It", "can", "be", "seen", "that", "the", "global", "image", "and", "part", "images", "are", "simultaneously", "considered", "during", "each", "round", "of", "training", ".", "Given", "a", "training", "sample", ",", "we", "use", "an", "human", "pose", "estimation", "algorithm", "to", "acquire", "the", "locations", "of", "human", "pose", "joints", ".", "These", "pose", "joints", "are", "combined", "into", "different", "human", "body", "parts", ".", "The", "part", "regions", "are", "first", "transformed", "using", "our", "Feature", "Embedding", "sub", "-", "Net", "(", "FEN", ")", "and", "then", "are", "combined", "to", "form", "a", "new", "modified", "part", "image", "containing", "the", "normalized", "body", "parts", ".", "The", "global", "image", "and", "the", "new", "modified", "part", "image", "are", "then", "fed", "into", "our", "CNN", "together", ".", "The", "two", "images", "share", "the", "same", "weights", "for", "the", "first", "several", "layers", ",", "then", "have", "their", "own", "network", "weights", "in", "the", "subsequent", "layers", ".", "At", "last", ",", "we", "use", "Feature", "Weighting", "sub", "-", "Net", "(", "FWN", ")", "to", "learn", "the", "weights", "of", "part", "features", "before", "fusing", "them", "with", "the", "global", "features", "for", "final", "Softmax", "Loss", "computation", ".", "Considering", "that", "pedestrian", "images", "form", "different", "datasets", "have", "different", "sizes", ",", "it", "is", "not", "appropriate", "to", "directly", "use", "the", "CNN", "models", "pre", "-", "trained", "on", "the", "ImageNet", "dataset", ".", "We", "thus", "modify", "and", "design", "a", "network", "based", "on", "the", "GoogLeNet", ",", "as", "shown", "in", "the", "Table", "[", "reference", "]", ".", "Layers", "from", "data", "to", "inception", "(", "4e", ")", "in", "Table", "[", "reference", "]", "corresponds", "to", "the", "blue", "CNN", "block", "in", "Fig", ".", "[", "reference", "]", ",", "CNNg", "and", "CNNp", "are", "inception", "(", "5a", ")", "and", "inception", "(", "5b", ")", ",", "respectively", ".", "The", "green", "CONV", "matches", "the", "subsequent", "1", "1", "convolution", ".", "The", "loss", "layers", "are", "not", "shown", "in", "Table", "[", "reference", "]", ".", "The", "Batch", "Normalization", "Layers", "are", "inserted", "before", "every", "ReLU", "Layer", "to", "accelerate", "the", "convergence", ".", "We", "employ", "a", "Convolutional", "Layer", "and", "a", "Global", "Average", "Pooling", "Layer", "(", "GAP", ")", "at", "the", "end", "of", "network", "to", "let", "our", "network", "can", "fit", "different", "sizes", "of", "input", "images", ".", "In", "this", "work", ",", "we", "fix", "input", "image", "size", "as", "512", "256", ".", "subsection", ":", "Feature", "Embedding", "sub", "-", "Net", "The", "Feature", "Embedding", "sub", "-", "Net", "(", "FEN", ")", "is", "divided", "into", "four", "steps", ",", "including", "locating", "the", "joint", ",", "generating", "the", "original", "part", "images", ",", "PTN", ",", "and", "outputting", "the", "final", "modified", "part", "images", ".", "With", "a", "given", "person", "image", ",", "FEN", "first", "locates", "the", "14", "joints", "of", "human", "body", "using", "human", "pose", "estimation", "algorithm", ".", "Fig", ".", "[", "reference", "]", "(", "c", ")", "shows", "an", "example", "of", "the", "14", "joints", "of", "human", "body", ".", "According", "to", "number", ",", "the", "14", "joints", "are", "{", "}", ".", "Then", "we", "propose", "six", "rectangles", "to", "cover", "six", "different", "parts", "of", "human", "body", ",", "including", "the", "head", "region", ",", "the", "upper", "body", ",", "two", "arms", "and", "two", "legs", ".", "For", "each", "human", "joint", ",", "we", "calculate", "a", "response", "feature", "map", ".", "The", "horizontal", "and", "vertical", "dimensions", "of", "the", "feature", "maps", "are", "denoted", "by", "and", ",", "respectively", ".", "With", "the", "feature", "maps", ",", "the", "fourteen", "body", "joints", ",", "can", "be", "located", "by", "finding", "the", "center", "of", "mass", "with", "the", "feature", "values", ":", "where", "in", "Eq", ".", "[", "reference", "]", "are", "the", "coordinates", "of", "joints", ",", "and", "is", "the", "value", "of", "pixels", "in", "response", "feature", "maps", ".", "Different", "from", ",", "we", "do", "not", "use", "complex", "pose", "estimation", "networks", "as", "the", "pre", "-", "trained", "network", ".", "Instead", ",", "we", "use", "a", "standard", "FCN", "trained", "on", "the", "LSP", "dataset", "and", "MPII", "human", "pose", "dataset", ".", "In", "the", "second", "step", ",", "the", "FEN", "uses", "the", "14", "human", "joints", "to", "further", "locate", "six", "sub", "-", "regions", "(", "head", ",", "upper", "body", ",", "left", "arm", ",", "right", "arm", ",", "left", "leg", ",", "and", "right", "leg", ")", "as", "human", "parts", ".", "These", "parts", "are", "normalized", "through", "cropping", ",", "rotating", ",", "and", "resizing", "to", "fixed", "size", "and", "orientation", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "d", ")", ",", "the", "14", "located", "body", "joints", "are", "assigned", "to", "six", "rectangles", "indicating", "six", "parts", ".", "The", "head", "part", ",", "the", "upper", "body", "part", ",", "the", "left", "arm", "part", ",", "the", "right", "arm", "part", ",", "the", "left", "leg", "part", ",", "and", "the", "right", "leg", "part", ",", "respectively", ".", "For", "each", "body", "part", "set", ",", "The", "corresponding", "sub", "-", "region", "bounding", "box", "can", "be", "obtained", "based", "on", "the", "location", "coordinates", "of", "all", "body", "joints", "in", "each", "part", "set", ":", "An", "example", "of", "the", "extracted", "six", "body", "sub", "-", "regions", "are", "visualized", "in", "Fig", ".", "[", "reference", "]", "(", "d", ")", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "e", ")", ",", "these", "body", "sub", "-", "regions", "are", "normalized", "through", "cropping", ",", "rotating", ",", "and", "resizing", "to", "fixed", "sizes", "and", "orientations", ".", "All", "body", "parts", "are", "rotated", "to", "fixed", "vertical", "direction", ".", "Arms", "and", "legs", "are", "resized", "to", "256", "64", ",", "upper", "body", "is", "resized", "to", "256", "128", "and", "head", "is", "resized", "to", "128", "128", ".", "Those", "resized", "and", "rotated", "parts", "are", "combined", "to", "form", "the", "body", "part", "image", ".", "Because", "6", "body", "parts", "have", "different", "sizes", ",", "black", "area", "is", "unavoidable", "in", "body", "part", "image", ".", "Simply", "resizing", "and", "rotation", "can", "not", "overcome", "the", "complex", "pose", "variations", ",", "especially", "if", "the", "pose", "estimations", "are", "inaccurate", ".", "We", "thus", "design", "a", "PTN", "modified", "from", "Spatial", "Transformer", "Networks", "(", "STN", ")", "to", "learn", "the", "angles", "required", "for", "rotating", "the", "five", "body", "parts", ".", "STN", "is", "a", "spatial", "transformer", "module", "which", "can", "be", "inserted", "to", "a", "neural", "network", "to", "provide", "spatial", "transformation", "capabilities", ".", "It", "thus", "is", "potential", "to", "adjust", "the", "localizations", "and", "angles", "of", "parts", ".", "A", "STN", "is", "a", "small", "net", "which", "allows", "for", "end", "-", "to", "-", "end", "training", "with", "standard", "back", "-", "propagation", ",", "therefore", ",", "the", "introduction", "of", "STN", "does", "n’t", "substantially", "increase", "the", "complexity", "of", "training", "procedure", ".", "The", "STN", "consist", "of", "three", "components", ":", "localisation", "network", ",", "parameterised", "sampling", "grid", ",", "and", "differentiable", "image", "sampling", ".", "The", "localisation", "network", "takes", "the", "input", "feature", "map", "and", "outputs", "the", "parameters", "of", "the", "transformation", ".", "For", "our", "net", ",", "we", "choose", "affine", "transformation", "so", "our", "transformation", "parameter", "is", "6", "-", "dimensional", ".", "The", "parameterized", "sampling", "grid", "computes", "each", "output", "pixel", "and", "the", "differentiable", "image", "sampling", "component", "produces", "the", "sampled", "output", "image", ".", "For", "more", "details", "about", "STN", ",", "please", "refer", "to", ".", "As", "discussed", "above", ",", "we", "use", "a", "6", "-", "dimensional", "parameter", "to", "complete", "affine", "transformation", ":", "where", "the", "are", "the", "scale", "and", "rotation", "parameters", ",", "while", "the", "are", "the", "translation", "parameters", ".", "The", "in", "Eq", ".", "[", "reference", "]", "are", "the", "target", "coordinates", "of", "the", "output", "image", "and", "the", "are", "the", "source", "coordinates", "of", "the", "input", "image", ".", "Usually", "the", "STN", "computes", "one", "affine", "transform", "for", "the", "whole", "image", ",", "considering", "a", "pedestrian", "’s", "different", "parts", "have", "various", "orientations", "and", "sizes", "from", "each", "other", ",", "STN", "is", "not", "applicable", "to", "a", "part", "image", ".", "Inspired", "by", "STN", ",", "we", "design", "a", "Pose", "Transformer", "Network", "(", "PTN", ")", "which", "computes", "the", "affine", "transformation", "for", "each", "part", "in", "part", "image", "individually", "and", "combines", "6", "transformed", "parts", "together", ".", "Similar", "to", "STN", ",", "our", "PTN", "is", "also", "a", "small", "net", "and", "does", "n’t", "substantially", "increase", "the", "complexity", "of", "our", "training", "procedure", ".", "As", "a", "consequence", ",", "PTN", "has", "potential", "to", "perform", "better", "than", "STN", "for", "person", "images", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "detailed", "structure", "of", "PTN", ".", "Considering", "a", "pedestrian", "’s", "head", "seldom", "has", "a", "large", "rotation", "angle", ",", "we", "do", "n’t", "insert", "a", "PTN", "net", "for", "the", "pedestrian", "’s", "head", "part", ".", "Therefore", ",", "we", "totally", "have", "5", "independent", "PTN", ",", "namely", ",", ",", ",", ",", ".", "Each", "PTN", "can", "generate", "a", "6", "-", "dimensional", "transformation", "parameter", "and", "use", "to", "adjust", "pedestrian", "’s", "part", ",", "we", "can", "get", "modified", "body", "part", ".", "By", "combining", "the", "five", "transformed", "parts", "and", "a", "head", "part", "together", ",", "we", "obtain", "the", "modified", "part", "image", ".", "subsection", ":", "Feature", "Weighting", "sub", "-", "Net", "The", "generated", "part", "features", "are", "combined", "with", "the", "global", "feature", "to", "generate", "a", "robust", "feature", "representation", "for", "precise", "person", "re", "-", "identification", ".", "As", "the", "poses", "generated", "by", "the", "pose", "detector", "might", "be", "affected", "by", "factors", "like", "occlusions", ",", "pose", "changes", ",", "etc", ".", "Then", "inaccurate", "part", "detection", "results", "could", "be", "obtained", ".", "Examples", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "Therefore", ",", "the", "part", "features", "could", "be", "not", "reliable", "enough", ".", "This", "happens", "frequently", "in", "real", "applications", "with", "unconstrained", "video", "gathering", "environment", ".", "Simply", "fusing", "global", "feature", "and", "the", "part", "feature", "may", "introduces", "noises", ".", "This", "motivates", "us", "to", "introduce", "Feature", "Weighting", "sub", "-", "Net", "(", "FWN", ")", "to", "seek", "a", "more", "optimal", "feature", "fusion", ".", "FWN", "is", "consisted", "with", "a", "Weight", "Layer", "and", "a", "nonlinear", "transformation", ",", "which", "decides", "the", "importance", "of", "each", "dimension", "in", "the", "part", "feature", "vector", ".", "Considering", "that", "a", "single", "linear", "Weight", "Layer", "might", "cause", "excessive", "response", "on", "some", "specific", "dimensions", "of", "the", "part", "vector", ",", "we", "add", "a", "nonlinear", "function", "to", "equalize", "the", "response", "of", "part", "feature", "vector", ",", "and", "the", "fused", "feature", "representation", "is", "where", "the", "and", "the", "are", "the", "global", "and", "part", "feature", "vectors", ".", "The", "and", "in", "Eq", ".", "[", "reference", "]", "are", "the", "weight", "and", "bias", "vectors", "which", "have", "the", "same", "dimensions", "with", ".", "The", "means", "the", "Hadamard", "product", "of", "two", "vectors", ",", "and", "the", "means", "concatenation", "of", "two", "vectors", "together", ".", "The", "imposes", "the", "hyperbolic", "tangent", "nonlinearity", ".", "is", "our", "final", "person", "feature", "generated", "by", "and", ".", "To", "allow", "back", "-", "propagation", "of", "the", "loss", "through", "the", "FWN", ",", "we", "give", "the", "gradient", "formula", ":", "where", ",", ",", ",", ",", ",", "and", "are", "the", "dimensions", "of", "and", ".", "subsection", ":", "ReID", "Feature", "Extraction", "The", "global", "feature", "and", "body", "-", "part", "features", "are", "learned", "by", "training", "the", "Pose", "-", "driven", "Deep", "Convolutional", "model", ".", "These", "two", "types", "of", "features", "are", "then", "fused", "under", "a", "unified", "framework", "for", "multi", "-", "class", "person", "identification", ".", "PDC", "extracts", "the", "global", "feature", "maps", "from", "the", "global", "body", "-", "based", "representation", "and", "learns", "a", "1024", "-", "dimensional", "feature", "embedding", ".", "Similarly", ",", "a", "1024", "-", "dimension", "feature", "is", "acquired", "from", "the", "modified", "part", "image", "after", "the", "FEN", ".", "The", "global", "body", "feature", "and", "the", "local", "body", "part", "features", "are", "compensated", "into", "a", "2048", "-", "dimensional", "feature", "as", "the", "final", "representation", ".", "After", "being", "weighted", "by", "FWN", ",", "the", "final", "representation", "is", "used", "for", "Person", "ReID", "with", "Euclidean", "distance", ".", "section", ":", "Experiment", "subsection", ":", "Datasets", "We", "select", "three", "widely", "used", "person", "ReID", "datasets", "as", "our", "evaluation", "protocols", ",", "including", "the", "CUHK", "03", ",", "Market", "1501", ",", "and", "VIPeR", ".", "Note", "that", ",", "because", "the", "amount", "of", "images", "in", "VIPeR", "is", "not", "enough", "for", "training", "a", "deep", "model", ",", "we", "combine", "the", "training", "sets", "of", "VIPeR", ",", "CUHK", "03", "and", "Market", "1501", "together", "to", "train", "the", "model", "for", "VIPeR", ".", "CUHK", "03", ":", "This", "dataset", "is", "made", "up", "of", "14", ",", "096", "images", "of", "1", ",", "467", "different", "persons", "taken", "by", "six", "campus", "cameras", ".", "Each", "person", "only", "appears", "in", "two", "views", ".", "This", "dataset", "provides", "two", "types", "of", "annotations", ",", "including", "manually", "labelled", "pedestrian", "bounding", "boxes", "and", "bounding", "boxes", "automatically", "detected", "by", "the", "Deformable", "-", "Part", "-", "Model", "(", "DPM", ")", "detector", ".", "We", "denote", "the", "two", "corresponding", "subsets", "as", "labeled", "dataset", "and", "detected", "dataset", ",", "respectively", ".", "The", "dataset", "also", "provides", "20", "test", "sets", ",", "each", "includes", "100", "identities", ".", "We", "select", "the", "first", "set", "and", "use", "100", "identities", "for", "testing", "and", "the", "rest", "1", ",", "367", "identities", "for", "training", ".", "We", "report", "the", "averaged", "performance", "after", "repeating", "the", "experiments", "for", "20", "times", ".", "Market", "1501", ":", "This", "dataset", "is", "made", "up", "of", "32", ",", "368", "pedestrian", "images", "taken", "by", "six", "manually", "configured", "cameras", ".", "It", "has", "1", ",", "501", "different", "persons", "in", "it", ".", "On", "average", ",", "there", "are", "3.6", "images", "for", "each", "person", "captured", "from", "each", "angle", ".", "The", "images", "can", "be", "classified", "into", "two", "types", ",", ",", "cropped", "images", "and", "images", "of", "pedestrians", "automatically", "detected", "by", "the", "DPM", ".", "Because", "Market", "1501", "has", "provided", "the", "training", "set", "and", "testing", "set", ",", "we", "use", "images", "in", "the", "training", "set", "for", "training", "our", "PDC", "network", "and", "follow", "the", "protocol", "to", "report", "the", "ReID", "performance", ".", "VIPeR", ":", "This", "dataset", "is", "made", "up", "of", "632", "person", "images", "captured", "from", "two", "views", ".", "Each", "pair", "of", "images", "depicting", "a", "person", "are", "collected", "by", "different", "cameras", "with", "varying", "viewpoints", "and", "illumination", "conditions", ".", "Because", "the", "amount", "of", "images", "in", "VIPeR", "is", "not", "enough", "to", "train", "the", "deep", "model", ",", "we", "also", "perform", "data", "augmentation", "with", "similar", "methods", "in", "existing", "deep", "learning", "based", "person", "ReID", "works", ".", "For", "each", "training", "image", ",", "we", "generate", "5", "augmented", "images", "around", "the", "image", "center", "by", "performing", "random", "2D", "transformations", ".", "Finally", ",", "we", "combine", "the", "augmented", "training", "images", "of", "VIPeR", ",", "training", "images", "of", "CUHK", "03", "and", "Market", "1501", "together", ",", "as", "the", "final", "training", "set", ".", "subsection", ":", "Implementation", "Details", "The", "pedestrian", "representations", "are", "learned", "through", "multi", "-", "class", "classification", "CNN", ".", "We", "use", "the", "full", "body", "and", "body", "parts", "to", "learn", "the", "representations", "with", "Softmax", "Loss", ",", "respectively", ".", "We", "report", "rank1", ",", "rank5", ",", "rank10", "and", "rank20", "accuracy", "of", "cumulative", "match", "curve", "(", "CMC", ")", "on", "the", "three", "datasets", "to", "evaluate", "the", "ReID", "performance", ".", "As", "for", "Market", "-", "1051", ",", "mean", "Average", "Precision", "(", "mAP", ")", "is", "also", "reported", "as", "an", "additional", "criterion", "to", "evaluate", "the", "performance", ".", "Our", "model", "is", "trained", "and", "fine", "-", "tuned", "on", "Caffe", ".", "Stochastic", "Gradient", "Descent", "(", "SGD", ")", "is", "used", "to", "optimize", "our", "model", ".", "Images", "for", "training", "are", "randomly", "divided", "into", "several", "batches", ",", "each", "of", "which", "includes", "16", "images", ".", "The", "initial", "learning", "rate", "is", "set", "as", "0.01", ",", "and", "is", "gradually", "lowered", "after", "each", "iterations", ".", "It", "should", "be", "noted", "that", ",", "the", "learning", "rate", "in", "part", "localization", "network", "is", "only", "0.1", "%", "of", "that", "in", "feature", "learning", "network", ".", "For", "each", "dataset", ",", "we", "train", "a", "model", "on", "its", "corresponding", "training", "set", "as", "the", "pretrained", "body", "-", "based", "model", ".", "For", "the", "overall", "network", "training", ",", "the", "network", "is", "initialized", "using", "pretrained", "body", "-", "based", "model", ".", "Then", ",", "we", "adopt", "the", "same", "training", "strategy", "as", "described", "above", ".", "We", "implement", "our", "approach", "with", "GTX", "TITAN", "X", "GPU", ",", "Intel", "i7", "CPU", ",", "and", "128", "GB", "memory", ".", "All", "images", "are", "resized", "to", ".", "The", "mean", "value", "is", "subtracted", "from", "each", "channel", "(", "B", ",", "G", ",", "and", "R", ")", "for", "training", "the", "network", ".", "The", "images", "of", "each", "dataset", "are", "randomized", "in", "the", "process", "of", "training", "stage", ".", "subsection", ":", "Evaluation", "of", "Individual", "Components", "We", "evaluate", "five", "variants", "of", "our", "approach", "to", "verify", "the", "validity", "of", "individual", "components", "in", "our", "PDC", ",", ",", "components", "like", "Feature", "Embedding", "sub", "-", "Net", "(", "FEN", ")", "and", "Feature", "Weighting", "sub", "-", "Net", "(", "FWN", ")", ".", "Comparisons", "on", "three", "datasets", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "In", "the", "table", ",", "“", "Global", "Only", "”", "means", "we", "train", "our", "deep", "model", "without", "using", "any", "part", "information", ".", "“", "Global", "+", "Part", "”", "denotes", "CNN", "trained", "through", "two", "streams", "without", "FEN", "and", "FWN", ".", "Based", "on", "“", "Global", "+", "Part", "”", ",", "considering", "FEN", "is", "denoted", "as", "“", "Global", "+", "Part", "+", "FEN", "”", ".", "Similarly", ",", "“", "Global", "+", "Part", "+", "FWN", "”", "means", "considering", "FWN", ".", "In", "addition", ",", "“", "Part", "Only", "”", "denotes", "only", "using", "part", "features", ".", "PDC", "considers", "all", "of", "these", "components", ".", "From", "the", "experimental", "results", ",", "it", "can", "be", "observed", "that", ",", "fusing", "global", "features", "and", "part", "features", "achieves", "better", "performance", "than", "only", "using", "one", "of", "them", ".", "Compared", "with", "“", "Global", "Only", "”", ",", "considering", "extra", "part", "cues", ",", ",", "“", "Global", "+", "Part", "”", ",", "largely", "improves", "the", "ReID", "performance", "and", "achieves", "the", "rank1", "accuracy", "of", "85.07", "%", "and", "76.33", "%", "on", "CUHK", "03", "labeled", "and", "detected", "datasets", ",", "respectively", ".", "Moreover", ",", "using", "FEN", "and", "FWN", "further", "boosts", "the", "rank1", "identification", "rate", ".", "This", "shows", "that", "training", "our", "model", "using", "PTN", "and", "Weight", "Layer", "gets", "more", "competitive", "performance", "on", "three", "datasets", ".", "The", "above", "experiments", "shows", "that", "each", "of", "the", "components", "in", "our", "method", "is", "helpful", "for", "improving", "the", "performance", ".", "By", "considering", "all", "of", "these", "components", ",", "PDC", "exhibits", "the", "best", "performance", ".", "subsection", ":", "Comparison", "with", "Related", "Works", "CUHK", "03", ":", "For", "the", "CUHK", "03", "dataset", ",", "we", "compare", "our", "PDC", "with", "some", "recent", "methods", ",", "including", "distance", "metric", "learning", "methods", ":", "MLAPG", ",", "LOMO", "+", "XQDA", ",", "BoW", "+", "HS", ",", "WARCA", ",", "LDNS", ",", "feature", "extraction", "method", ":", "GOG", "and", "deep", "learning", "based", "methods", ":", "IDLA", ",", "PersonNet", ",", "DGDropout", ",", "SI", "+", "CI", ",", "Gate", "S", "-", "CNN", ",", "LSTM", "S", "-", "CNN", ",", "EDM", ",", "PIE", "and", "Spindle", ".", "We", "conduct", "experiments", "on", "both", "the", "detected", "dataset", "and", "the", "labeled", "dataset", ".", "Experimental", "results", "are", "presented", "in", "Table", "[", "reference", "]", "and", "Table", "[", "reference", "]", ".", "Experimental", "results", "show", "that", "our", "approach", "outperforms", "all", "distance", "metric", "learning", "methods", "by", "a", "large", "margin", ".", "It", "can", "be", "seen", "that", "PIE", ",", "Spindle", "and", "our", "PDC", "which", "all", "use", "the", "human", "pose", "cues", "achieve", "better", "performance", "than", "the", "other", "methods", ".", "This", "shows", "the", "advantages", "of", "considering", "extra", "pose", "cues", "in", "person", "ReID", ".", "It", "is", "also", "clear", "that", ",", "our", "PDC", "achieves", "the", "rank1", "accuracy", "of", "78.29", "and", "88.70", "on", "detected", "and", "labeled", "datasets", ",", "respectively", ".", "This", "leads", "to", "11.19", "and", "0.20", "performance", "gains", "over", "the", "reported", "performance", "of", "PIE", "and", "Spindle", ",", "respectively", ".", "Market", "1501", ":", "On", "Market", "1501", ",", "the", "compared", "works", "that", "learn", "distance", "metrics", "for", "person", "ReID", "include", "LOMO", "+", "XQDA", ",", "BoW", "+", "Kissme", ",", "WARCA", ",", "LDNS", ",", "TMA", "and", "HVIL", ".", "Compared", "works", "based", "on", "deep", "learning", "are", "PersonNet", ",", "Gate", "S", "-", "CNN", ",", "LSTM", "S", "-", "CNN", ",", "PIE", "and", "Spindle", ".", "DGDropout", "does", "not", "report", "performance", "on", "Market1501", ".", "So", "we", "implemented", "DGDroput", "and", "show", "experimental", "results", "in", "Table", "[", "reference", "]", ".", "It", "is", "clear", "that", "our", "method", "outperforms", "these", "compared", "works", "by", "a", "large", "margin", ".", "Specifically", ",", "PDC", "achieves", "rank1", "accuracy", "of", "84.14", "%", ",", "and", "mAP", "of", "63.41", "%", "using", "the", "single", "query", "mode", ".", "They", "are", "higher", "than", "the", "rank1", "accuracy", "and", "mAP", "of", "PIE", ",", "which", "performs", "best", "among", "the", "compared", "works", ".", "This", "is", "because", "our", "PDC", "not", "only", "learns", "pose", "invariant", "features", "with", "FEN", "but", "also", "learns", "better", "fusion", "strategy", "with", "FWN", "to", "emphasize", "the", "more", "discriminative", "features", ".", "VIPeR", ":", "We", "also", "evaluate", "our", "method", "by", "comparing", "it", "with", "several", "existing", "methods", "on", "VIPeR", ".", "The", "compared", "methods", "include", "distance", "metric", "learning", "ones", ":", "MLAPG", ",", "LOMO", "+", "XQDA", ",", "BoW", ",", "WARCA", "and", "LDNS", ",", "and", "deep", "learning", "based", "ones", ":", "IDLA", ",", "DGDropout", ",", "SI", "+", "CI", ",", "Gate", "S", "-", "CNN", ",", "LSTM", "S", "-", "CNN", ",", "MTL", "-", "LORAE", "and", "Spindle", ".", "From", "the", "results", "shown", "in", "Table", "[", "reference", "]", ",", "our", "PDC", "achieves", "the", "rank1", "accuracy", "of", "51.27", "%", ".", "This", "outperforms", "most", "of", "compared", "methods", "except", "Spindle", "which", "also", "considers", "the", "human", "pose", "cues", ".", "We", "assume", "the", "reason", "might", "be", "because", ",", "Spindle", "involves", "more", "training", "sets", "to", "learn", "the", "model", "for", "VIPeR", ".", "Therefore", ",", "the", "training", "set", "of", "Spindle", "is", "larger", "than", "ours", ",", ",", "the", "combination", "of", "Market", "1501", ",", "CUHK03", "and", "VIPeR", ".", "For", "the", "other", "two", "datasets", ",", "our", "PDC", "achieves", "better", "performance", "than", "Spindle", ".", "subsection", ":", "Evaluation", "of", "Feature", "Weighting", "sub", "-", "Net", "To", "test", "the", "effectiveness", "of", "Feature", "Weighting", "sub", "-", "Net", "(", "FWN", ")", ",", "we", "verify", "the", "performance", "of", "five", "variants", "of", "FWN", ",", "which", "are", "denoted", "as", ",", "=", "{", "0", ",", "1", ",", "2", ",", "3", ",", "4", "}", ",", "where", "is", "the", "number", "of", "Weight", "Layers", "in", "FWN", "with", "nonlinear", "transformation", ".", "For", "example", ",", "means", "we", "cascade", "two", "Weight", "Layers", "with", "nonlinear", "transformation", ",", "means", "we", "only", "have", "one", "Weight", "Layer", "without", "nonlinear", "transformation", ".", "The", "experimental", "results", "are", "shown", "in", "Table", "[", "reference", "]", ".", "As", "we", "can", "see", "that", "one", "Weight", "Layer", "with", "nonlinear", "transformation", "gets", "the", "best", "performance", "on", "the", "three", "datasets", ".", "The", "ReID", "performance", "starts", "to", "drop", "as", "we", "increase", "of", "the", "number", "of", "Weight", "Layers", ",", "despite", "more", "computations", "are", "being", "brought", "in", ".", "It", "also", "can", "be", "observed", "that", ",", "using", "one", "layer", "with", "nonlinear", "transformation", "gets", "better", "performance", "than", "one", "layer", "without", "nonlinear", "transformation", ",", ",", ".", "This", "means", "adding", "one", "nonlinear", "transformation", "after", "a", "Weight", "Layer", "learns", "more", "reliable", "weights", "for", "feature", "fusion", "and", "matching", ".", "Based", "on", "the", "above", "observations", ",", "we", "adopt", "as", "our", "final", "model", "in", "this", "paper", ".", "Examples", "of", "features", "before", "and", "after", "FWN", "are", "shown", "Fig", ".", "[", "reference", "]", ".", "section", ":", "Conclusions", "This", "paper", "presents", "a", "pose", "-", "driven", "deep", "convolutional", "model", "for", "the", "person", "ReID", ".", "The", "proposed", "deep", "architecture", "explicitly", "leverages", "the", "human", "part", "cues", "to", "learn", "effective", "feature", "representations", "and", "adaptive", "similarity", "measurements", ".", "For", "the", "feature", "representations", ",", "both", "global", "human", "body", "and", "local", "body", "parts", "are", "transformed", "to", "a", "normalized", "and", "homologous", "state", "for", "better", "feature", "embedding", ".", "For", "similarity", "measurements", ",", "weights", "of", "feature", "representations", "from", "human", "body", "and", "different", "body", "parts", "are", "learned", "to", "adaptively", "chase", "a", "more", "discriminative", "feature", "fusion", ".", "Experimental", "results", "on", "three", "benchmark", "datasets", "demonstrate", "the", "superiority", "of", "the", "proposed", "model", "over", "current", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "Acknowledgments", "This", "work", "is", "partly", "supported", "by", "National", "Science", "Foundation", "of", "China", "under", "Grant", "No", ".", "61572050", ",", "91538111", ",", "61620106009", ",", "61429201", ",", "61672519", ",", "and", "the", "National", "1000", "Youth", "Talents", "Plan", ".", "Dr.", "Qi", "Tian", "is", "supported", "by", "ARO", "grant", "W911NF", "-", "15", "-", "1", "-", "0290", "and", "Faculty", "Research", "Gift", "Awards", "by", "NEC", "Laboratories", "of", "America", "and", "Blippar", ".", "bibliography", ":", "References"], "entity": [[9, 13, ["Person", "Re", "-", "identification"], "Task"], [13, 15, ["Feature", "extraction"], "Task"], [16, 17, ["matching"], "Task"], [22, 26, ["person", "Re", "-", "Identification"], "Task"], [27, 28, ["ReID"], "Task"], [50, 56, ["learning", "and", "matching", "of", "the", "features"], "Task"], [71, 76, ["Pose", "-", "driven", "Deep", "Convolutional"], "Method"], [77, 78, ["PDC"], "Method"], [83, 85, ["feature", "extraction"], "Method"], [86, 88, ["matching", "models"], "Method"], [94, 96, ["deep", "architecture"], "Method"], [109, 112, ["robust", "feature", "representations"], "Method"], [136, 143, ["pose", "driven", "feature", "weighting", "sub", "-", "network"], "Method"], [148, 151, ["adaptive", "feature", "fusions"], "Method"], [183, 187, ["Person", "Re", "-", "Identification"], "Task"], [188, 189, ["ReID"], "Task"], [196, 199, ["video", "surveillance", "system"], "Task"], [210, 212, ["probe", "person"], "Task"], [337, 340, ["discriminative", "distance", "metric"], "Method"], [380, 382, ["Distance", "metric"], "Metric"], [391, 394, ["distance", "metric", "learning"], "Method"], [400, 402, ["computational", "complexity"], "Metric"], [407, 409, ["deep", "learning"], "Method"], [421, 424, ["computer", "vision", "tasks"], "Task"], [434, 436, ["CUHK", "03"], "Material"], [437, 440, ["Market", "-", "1501"], "Material"], [442, 443, ["MARS"], "Material"], [455, 457, ["deep", "models"], "Method"], [469, 471, ["deep", "models"], "Method"], [526, 528, ["global", "representation"], "Method"], [534, 536, ["local", "representations"], "Method"], [545, 549, ["local", "and", "global", "representations"], "Method"], [554, 558, ["Deep", "learning", "based", "methods"], "Method"], [614, 618, ["person", "re", "-", "identification"], "Task"], [652, 655, ["pose", "estimation", "algorithms"], "Method"], [662, 664, ["deep", "models"], "Method"], [698, 699, ["ReID"], "Task"], [718, 723, ["Pose", "-", "driven", "Deep", "Convolutional"], "Method"], [724, 725, ["PDC"], "Method"], [737, 739, ["global", "representation"], "Method"], [744, 746, ["local", "representations"], "Method"], [752, 754, ["global", "representation"], "Method"], [772, 776, ["learning", "of", "local", "representations"], "Task"], [779, 784, ["Feature", "Embedding", "sub", "-", "Net"], "Method"], [785, 786, ["FEN"], "Method"], [820, 825, ["Feature", "Embedding", "sub", "-", "Net"], "Method"], [844, 847, ["Pose", "Transformation", "Network"], "Method"], [848, 850, ["PTN", ")"], "Method"], [857, 859, ["local", "representations"], "Method"], [871, 876, ["Feature", "Weighting", "sub", "-", "Net"], "Method"], [887, 889, ["local", "representations"], "Method"], [897, 899, ["feature", "fusion"], "Method"], [903, 906, ["feature", "similarity", "measurement"], "Task"], [913, 916, ["local", "representation", "generation"], "Method"], [1004, 1007, ["Pose", "Transformation", "Network"], "Method"], [1008, 1010, ["PTN", ")"], "Method"], [1042, 1045, ["deep", "neural", "network"], "Method"], [1072, 1075, ["sub", "-", "network"], "Method"], [1079, 1081, ["deep", "learning"], "Method"], [1082, 1085, ["person", "ReID", "works"], "Task"], [1106, 1108, ["deep", "architecture"], "Method"], [1128, 1131, ["sub", "-", "network"], "Method"], [1142, 1145, ["feature", "similarity", "measurement"], "Task"], [1150, 1151, ["weighting"], "Task"], [1160, 1162, ["pose", "estimation"], "Task"], [1172, 1175, ["pose", "estimation", "algorithm"], "Method"], [1177, 1180, ["Fully", "Convolutional", "Networks"], "Method"], [1181, 1182, ["FCN"], "Method"], [1183, 1187, ["based", "pose", "estimation", "method"], "Method"], [1195, 1196, ["FCN"], "Method"], [1260, 1264, ["person", "re", "-", "identification"], "Task"], [1301, 1304, ["discriminative", "distance", "metric"], "Method"], [1312, 1314, ["deep", "learning"], "Method"], [1337, 1339, ["Deep", "learning"], "Method"], [1346, 1349, ["person", "’s", "representation"], "Method"], [1351, 1353, ["distance", "metric"], "Metric"], [1363, 1366, ["deep", "learning", "methods"], "Method"], [1368, 1370, ["feature", "representations"], "Method"], [1376, 1379, ["deep", "matching", "function"], "Method"], [1385, 1387, ["Fully", "Connected"], "Method"], [1388, 1389, ["FC"], "Method"], [1394, 1398, ["deep", "metric", "learning", "methods"], "Method"], [1426, 1429, ["joint", "learning", "framework"], "Method"], [1431, 1435, ["single", "-", "image", "representation"], "Task"], [1436, 1440, ["cross", "-", "image", "representation"], "Task"], [1442, 1446, ["doublet", "or", "triplet", "CNN"], "Method"], [1450, 1454, ["moderate", "positive", "mining", "method"], "Method"], [1456, 1460, ["deep", "distance", "metric", "learning"], "Method"], [1472, 1473, ["ReID"], "Task"], [1474, 1478, ["semi", "-", "supervised", "learning"], "Method"], [1490, 1494, ["Domain", "Guided", "Dropout", "algorithm"], "Method"], [1504, 1508, ["deep", "learning", "based", "methods"], "Method"], [1512, 1516, ["learning", "local", "pedestrian", "features"], "Task"], [1531, 1535, ["human", "pose", "estimation", "algorithms"], "Method"], [1547, 1548, ["accuracy"], "Metric"], [1549, 1552, ["pose", "estimation", "algorithms"], "Method"], [1562, 1564, ["pose", "estimation"], "Task"], [1592, 1594, ["feature", "representation"], "Method"], [1595, 1600, ["Feature", "Embedding", "sub", "-", "Net"], "Method"], [1601, 1602, ["FEN"], "Method"], [1606, 1611, ["Feature", "Weighting", "sub", "-", "Net"], "Method"], [1612, 1613, ["FWN"], "Method"], [1656, 1659, ["human", "poses", "estimation"], "Task"], [1693, 1698, ["Feature", "Embedding", "sub", "-", "Net"], "Method"], [1699, 1700, ["FEN"], "Method"], [1703, 1708, ["Feature", "Weighting", "sub", "-", "Net"], "Method"], [1709, 1710, ["FWN"], "Method"], [1764, 1765, ["training"], "Task"], [1774, 1778, ["human", "pose", "estimation", "algorithm"], "Method"], [1806, 1811, ["Feature", "Embedding", "sub", "-", "Net"], "Method"], [1812, 1813, ["FEN"], "Method"], [1845, 1846, ["CNN"], "Method"], [1883, 1884, ["FWN"], "Method"], [1901, 1904, ["Softmax", "Loss", "computation"], "Task"], [1924, 1926, ["CNN", "models"], "Method"], [1931, 1933, ["ImageNet", "dataset"], "Material"], [1944, 1945, ["GoogLeNet"], "Method"], [1981, 1982, ["CNNg"], "Method"], [1983, 1984, ["CNNp"], "Method"], [1998, 2000, ["green", "CONV"], "Method"], [2020, 2023, ["Batch", "Normalization", "Layers"], "Method"], [2027, 2029, ["ReLU", "Layer"], "Method"], [2037, 2039, ["Convolutional", "Layer"], "Method"], [2041, 2045, ["Global", "Average", "Pooling", "Layer"], "Method"], [2046, 2047, ["GAP"], "Method"], [2080, 2085, ["Feature", "Embedding", "sub", "-", "Net"], "Method"], [2086, 2091, ["Feature", "Embedding", "sub", "-", "Net"], "Method"], [2092, 2093, ["FEN"], "Method"], [2111, 2112, ["PTN"], "Method"], [2137, 2141, ["human", "pose", "estimation", "algorithm"], "Method"], [2209, 2212, ["response", "feature", "map"], "Method"], [2285, 2288, ["pose", "estimation", "networks"], "Method"], [2290, 2294, ["pre", "-", "trained", "network"], "Method"], [2301, 2302, ["FCN"], "Method"], [2305, 2307, ["LSP", "dataset"], "Material"], [2308, 2312, ["MPII", "human", "pose", "dataset"], "Material"], [2319, 2320, ["FEN"], "Method"], [2360, 2361, ["cropping"], "Method"], [2502, 2503, ["cropping"], "Method"], [2603, 2604, ["PTN"], "Method"], [2606, 2609, ["Spatial", "Transformer", "Networks"], "Method"], [2610, 2611, ["STN"], "Method"], [2624, 2625, ["STN"], "Method"], [2627, 2630, ["spatial", "transformer", "module"], "Method"], [2636, 2638, ["neural", "network"], "Method"], [2658, 2659, ["STN"], "Method"], [2666, 2672, ["end", "-", "to", "-", "end", "training"], "Task"], [2674, 2677, ["back", "-", "propagation"], "Method"], [2683, 2684, ["STN"], "Method"], [2691, 2693, ["training", "procedure"], "Method"], [2695, 2696, ["STN"], "Method"], [2701, 2703, ["localisation", "network"], "Method"], [2704, 2707, ["parameterised", "sampling", "grid"], "Method"], [2709, 2712, ["differentiable", "image", "sampling"], "Method"], [2714, 2716, ["localisation", "network"], "Method"], [2747, 2750, ["parameterized", "sampling", "grid"], "Method"], [2756, 2760, ["differentiable", "image", "sampling", "component"], "Method"], [2770, 2771, ["STN"], "Method"], [2789, 2791, ["affine", "transformation"], "Task"], [2836, 2837, ["STN"], "Method"], [2839, 2841, ["affine", "transform"], "Method"], [2861, 2862, ["STN"], "Method"], [2872, 2873, ["STN"], "Method"], [2877, 2880, ["Pose", "Transformer", "Network"], "Method"], [2881, 2882, ["PTN"], "Method"], [2904, 2905, ["STN"], "Method"], [2907, 2908, ["PTN"], "Method"], [2919, 2920, ["complexity"], "Metric"], [2922, 2924, ["training", "procedure"], "Method"], [2929, 2930, ["PTN"], "Method"], [2936, 2937, ["STN"], "Method"], [2938, 2940, ["person", "images"], "Task"], [2951, 2952, ["PTN"], "Method"], [2970, 2972, ["PTN", "net"], "Method"], [2986, 2987, ["PTN"], "Method"], [2995, 2996, ["PTN"], "Method"], [3040, 3045, ["Feature", "Weighting", "sub", "-", "Net"], "Method"], [3058, 3061, ["robust", "feature", "representation"], "Method"], [3074, 3076, ["pose", "detector"], "Method"], [3091, 3093, ["part", "detection"], "Task"], [3126, 3130, ["unconstrained", "video", "gathering", "environment"], "Task"], [3154, 3155, ["FWN"], "Method"], [3161, 3163, ["feature", "fusion"], "Method"], [3164, 3165, ["FWN"], "Method"], [3169, 3171, ["Weight", "Layer"], "Method"], [3173, 3175, ["nonlinear", "transformation"], "Method"], [3193, 3196, ["linear", "Weight", "Layer"], "Method"], [3212, 3214, ["nonlinear", "function"], "Method"], [3225, 3228, ["fused", "feature", "representation"], "Method"], [3306, 3307, ["FWN"], "Method"], [3311, 3313, ["gradient", "formula"], "Method"], [3362, 3364, ["unified", "framework"], "Method"], [3365, 3370, ["multi", "-", "class", "person", "identification"], "Task"], [3371, 3372, ["PDC"], "Method"], [3379, 3384, ["global", "body", "-", "based", "representation"], "Method"], [3387, 3392, ["1024", "-", "dimensional", "feature", "embedding"], "Method"], [3438, 3439, ["FWN"], "Method"], [3473, 3475, ["CUHK", "03"], "Material"], [3476, 3478, ["Market", "1501"], "Material"], [3480, 3481, ["VIPeR"], "Material"], [3491, 3492, ["VIPeR"], "Material"], [3498, 3500, ["deep", "model"], "Method"], [3507, 3508, ["VIPeR"], "Material"], [3509, 3511, ["CUHK", "03"], "Material"], [3512, 3514, ["Market", "1501"], "Material"], [3520, 3521, ["VIPeR"], "Material"], [3576, 3585, ["Deformable", "-", "Part", "-", "Model", "(", "DPM", ")", "detector"], "Method"], [3648, 3650, ["Market", "1501"], "Material"], [3714, 3715, ["DPM"], "Method"], [3717, 3719, ["Market", "1501"], "Material"], [3747, 3749, ["ReID", "performance"], "Metric"], [3750, 3751, ["VIPeR"], "Material"], [3791, 3792, ["VIPeR"], "Material"], [3798, 3800, ["deep", "model"], "Method"], [3804, 3806, ["data", "augmentation"], "Method"], [3811, 3813, ["deep", "learning"], "Method"], [3834, 3837, ["random", "2D", "transformations"], "Method"], [3847, 3848, ["VIPeR"], "Material"], [3852, 3854, ["CUHK", "03"], "Material"], [3855, 3857, ["Market", "1501"], "Material"], [3870, 3872, ["pedestrian", "representations"], "Method"], [3875, 3880, ["multi", "-", "class", "classification", "CNN"], "Method"], [3901, 3902, ["rank1"], "Metric"], [3903, 3904, ["rank5"], "Metric"], [3905, 3906, ["rank10"], "Metric"], [3907, 3909, ["rank20", "accuracy"], "Metric"], [3910, 3913, ["cumulative", "match", "curve"], "Metric"], [3914, 3915, ["CMC"], "Metric"], [3923, 3925, ["ReID", "performance"], "Metric"], [3928, 3931, ["Market", "-", "1051"], "Material"], [3932, 3935, ["mean", "Average", "Precision"], "Metric"], [3936, 3938, ["mAP", ")"], "Metric"], [3959, 3960, ["Caffe"], "Method"], [3961, 3964, ["Stochastic", "Gradient", "Descent"], "Method"], [3965, 3967, ["SGD", ")"], "Method"], [3993, 3995, ["learning", "rate"], "Metric"], [4015, 4017, ["learning", "rate"], "Metric"], [4018, 4021, ["part", "localization", "network"], "Method"], [4028, 4031, ["feature", "learning", "network"], "Method"], [4047, 4052, ["pretrained", "body", "-", "based", "model"], "Method"], [4056, 4058, ["network", "training"], "Task"], [4064, 4069, ["pretrained", "body", "-", "based", "model"], "Method"], [4076, 4078, ["training", "strategy"], "Method"], [4087, 4091, ["GTX", "TITAN", "X", "GPU"], "Method"], [4164, 4165, ["PDC"], "Method"], [4169, 4174, ["Feature", "Embedding", "sub", "-", "Net"], "Method"], [4175, 4176, ["FEN"], "Method"], [4178, 4183, ["Feature", "Weighting", "sub", "-", "Net"], "Method"], [4184, 4185, ["FWN"], "Method"], [4211, 4213, ["deep", "model"], "Method"], [4220, 4224, ["Global", "+", "Part", "”"], "Method"], [4225, 4226, ["CNN"], "Method"], [4231, 4232, ["FEN"], "Method"], [4233, 4234, ["FWN"], "Method"], [4244, 4245, ["FEN"], "Method"], [4267, 4268, ["FWN"], "Method"], [4282, 4283, ["PDC"], "Method"], [4319, 4322, ["Global", "Only", "”"], "Method"], [4338, 4340, ["ReID", "performance"], "Metric"], [4352, 4354, ["CUHK", "03"], "Material"], [4364, 4365, ["FEN"], "Method"], [4366, 4367, ["FWN"], "Method"], [4370, 4373, ["rank1", "identification", "rate"], "Metric"], [4381, 4385, ["PTN", "and", "Weight", "Layer"], "Method"], [4419, 4420, ["PDC"], "Method"], [4443, 4444, ["PDC"], "Method"], [4450, 4454, ["distance", "metric", "learning", "methods"], "Method"], [4455, 4456, ["MLAPG"], "Method"], [4457, 4460, ["LOMO", "+", "XQDA"], "Method"], [4461, 4464, ["BoW", "+", "HS"], "Method"], [4465, 4466, ["WARCA"], "Method"], [4467, 4468, ["LDNS"], "Method"], [4469, 4472, ["feature", "extraction", "method"], "Method"], [4473, 4479, ["GOG", "and", "deep", "learning", "based", "methods"], "Method"], [4480, 4481, ["IDLA"], "Method"], [4482, 4483, ["PersonNet"], "Method"], [4484, 4485, ["DGDropout"], "Method"], [4486, 4489, ["SI", "+", "CI"], "Method"], [4490, 4494, ["Gate", "S", "-", "CNN"], "Method"], [4495, 4499, ["LSTM", "S", "-", "CNN"], "Method"], [4500, 4501, ["EDM"], "Method"], [4502, 4503, ["PIE"], "Method"], [4504, 4505, ["Spindle"], "Method"], [4542, 4546, ["distance", "metric", "learning", "methods"], "Method"], [4556, 4557, ["PIE"], "Method"], [4558, 4559, ["Spindle"], "Method"], [4561, 4562, ["PDC"], "Method"], [4597, 4598, ["PDC"], "Method"], [4627, 4628, ["PIE"], "Method"], [4633, 4635, ["Market", "1501"], "Material"], [4637, 4639, ["Market", "1501"], "Material"], [4645, 4647, ["distance", "metrics"], "Metric"], [4651, 4654, ["LOMO", "+", "XQDA"], "Method"], [4655, 4658, ["BoW", "+", "Kissme"], "Method"], [4659, 4660, ["WARCA"], "Method"], [4661, 4662, ["LDNS"], "Method"], [4663, 4664, ["TMA"], "Method"], [4665, 4666, ["HVIL"], "Method"], [4671, 4673, ["deep", "learning"], "Method"], [4674, 4675, ["PersonNet"], "Method"], [4676, 4680, ["Gate", "S", "-", "CNN"], "Method"], [4681, 4685, ["LSTM", "S", "-", "CNN"], "Method"], [4686, 4687, ["PIE"], "Method"], [4688, 4689, ["Spindle"], "Method"], [4690, 4691, ["DGDropout"], "Method"], [4696, 4697, ["Market1501"], "Material"], [4701, 4702, ["DGDroput"], "Method"], [4729, 4730, ["PDC"], "Method"], [4738, 4739, ["mAP"], "Metric"], [4744, 4747, ["single", "query", "mode"], "Method"], [4756, 4759, ["mAP", "of", "PIE"], "Metric"], [4772, 4773, ["PDC"], "Method"], [4780, 4781, ["FEN"], "Method"], [4785, 4787, ["fusion", "strategy"], "Method"], [4788, 4789, ["FWN"], "Method"], [4811, 4812, ["VIPeR"], "Material"], [4817, 4821, ["distance", "metric", "learning", "ones"], "Method"], [4822, 4823, ["MLAPG"], "Method"], [4824, 4827, ["LOMO", "+", "XQDA"], "Method"], [4828, 4829, ["BoW"], "Method"], [4830, 4831, ["WARCA"], "Method"], [4832, 4833, ["LDNS"], "Method"], [4835, 4839, ["deep", "learning", "based", "ones"], "Method"], [4840, 4841, ["IDLA"], "Method"], [4842, 4843, ["DGDropout"], "Method"], [4844, 4847, ["SI", "+", "CI"], "Method"], [4848, 4852, ["Gate", "S", "-", "CNN"], "Method"], [4853, 4857, ["LSTM", "S", "-", "CNN"], "Method"], [4858, 4861, ["MTL", "-", "LORAE"], "Method"], [4862, 4863, ["Spindle"], "Method"], [4875, 4876, ["PDC"], "Method"], [4891, 4892, ["Spindle"], "Method"], [4918, 4919, ["VIPeR"], "Material"], [4936, 4938, ["Market", "1501"], "Material"], [4939, 4940, ["CUHK03"], "Material"], [4941, 4942, ["VIPeR"], "Material"], [4950, 4951, ["PDC"], "Method"], [4961, 4966, ["Feature", "Weighting", "sub", "-", "Net"], "Method"], [4971, 4976, ["Feature", "Weighting", "sub", "-", "Net"], "Method"], [4977, 4978, ["FWN"], "Method"], [4988, 4989, ["FWN"], "Method"], [5016, 5017, ["FWN"], "Method"], [5018, 5020, ["nonlinear", "transformation"], "Method"], [5028, 5030, ["Weight", "Layers"], "Method"], [5031, 5033, ["nonlinear", "transformation"], "Method"], [5039, 5041, ["Weight", "Layer"], "Method"], [5042, 5044, ["nonlinear", "transformation"], "Method"], [5062, 5064, ["Weight", "Layer"], "Method"], [5065, 5067, ["nonlinear", "transformation"], "Method"], [5077, 5079, ["ReID", "performance"], "Metric"], [5108, 5110, ["one", "layer"], "Method"], [5111, 5113, ["nonlinear", "transformation"], "Method"], [5120, 5122, ["nonlinear", "transformation"], "Method"], [5133, 5135, ["Weight", "Layer"], "Method"], [5140, 5144, ["feature", "fusion", "and", "matching"], "Task"], [5167, 5168, ["FWN"], "Method"], [5183, 5189, ["pose", "-", "driven", "deep", "convolutional", "model"], "Method"], [5196, 5198, ["deep", "architecture"], "Method"], [5207, 5209, ["feature", "representations"], "Method"], [5216, 5218, ["feature", "representations"], "Method"], [5237, 5239, ["feature", "embedding"], "Method"], [5241, 5243, ["similarity", "measurements"], "Task"], [5246, 5248, ["feature", "representations"], "Method"], [5262, 5265, ["discriminative", "feature", "fusion"], "Method"], [2, 7, ["Pose", "-", "driven", "Deep", "Convolutional"], "Method"], [202, 203, ["ReID"], "Task"], [268, 269, ["ReID"], "Task"], [396, 397, ["ReID"], "Task"], [431, 432, ["ReID"], "Task"], [459, 460, ["ReID"], "Task"], [473, 474, ["ReID"], "Task"], [505, 506, ["ReID"], "Task"], [578, 579, ["ReID"], "Task"], [666, 667, ["ReID"], "Task"], [729, 730, ["ReID"], "Task"], [733, 734, ["PDC"], "Method"], [877, 878, ["FWN"], "Method"], [1316, 1317, ["ReID"], "Task"], [1334, 1335, ["ReID"], "Task"], [1462, 1463, ["ReID"], "Task"], [1486, 1487, ["ReID"], "Task"], [1672, 1673, ["ReID"], "Task"], [1743, 1744, ["ReID"], "Task"], [1877, 1882, ["Feature", "Weighting", "sub", "-", "Net"], "Method"], [2127, 2128, ["FEN"], "Method"], [3063, 3067, ["person", "re", "-", "identification"], "Task"], [3148, 3153, ["Feature", "Weighting", "sub", "-", "Net"], "Method"], [3329, 3330, ["ReID"], "Task"], [3345, 3350, ["Pose", "-", "driven", "Deep", "Convolutional"], "Method"], [3409, 3410, ["FEN"], "Method"], [3447, 3448, ["ReID"], "Task"], [3464, 3465, ["ReID"], "Task"], [3522, 3524, ["CUHK", "03"], "Material"], [3738, 3739, ["PDC"], "Method"], [3815, 3816, ["ReID"], "Task"], [4253, 4254, ["FEN"], "Method"], [4263, 4264, ["FWN"], "Method"], [4343, 4344, ["rank1"], "Metric"], [4431, 4433, ["CUHK", "03"], "Material"], [4436, 4438, ["CUHK", "03"], "Material"], [4588, 4589, ["ReID"], "Task"], [4600, 4601, ["rank1"], "Metric"], [4649, 4650, ["ReID"], "Task"], [4731, 4732, ["rank1"], "Metric"], [4753, 4754, ["rank1"], "Metric"], [4796, 4797, ["VIPeR"], "Material"], [4878, 4879, ["rank1"], "Metric"], [5192, 5193, ["ReID"], "Task"]]} {"doc_id": "1f08598381af9146d0fd9a61b30d0e51a7331689", "text": ["document", ":", "Distributed", "Prioritized", "Experience", "Replay", "We", "propose", "a", "distributed", "architecture", "for", "deep", "reinforcement", "learning", "at", "scale", ",", "that", "enables", "agents", "to", "learn", "effectively", "from", "orders", "of", "magnitude", "more", "data", "than", "previously", "possible", ".", "The", "algorithm", "decouples", "acting", "from", "learning", ":", "the", "actors", "interact", "with", "their", "own", "instances", "of", "the", "environment", "by", "selecting", "actions", "according", "to", "a", "shared", "neural", "network", ",", "and", "accumulate", "the", "resulting", "experience", "in", "a", "shared", "experience", "replay", "memory", ";", "the", "learner", "replays", "samples", "of", "experience", "and", "updates", "the", "neural", "network", ".", "The", "architecture", "relies", "on", "prioritized", "experience", "replay", "to", "focus", "only", "on", "the", "most", "significant", "data", "generated", "by", "the", "actors", ".", "Our", "architecture", "substantially", "improves", "the", "state", "of", "the", "art", "on", "the", "Arcade", "Learning", "Environment", ",", "achieving", "better", "final", "performance", "in", "a", "fraction", "of", "the", "wall", "-", "clock", "training", "time", ".", "section", ":", "Introduction", "A", "broad", "trend", "in", "deep", "learning", "is", "that", "combining", "more", "computation", "Dean:2012:LSD:2999134.2999271", "with", "more", "powerful", "models", "DBLP", ":", "journals", "/", "corr", "/", "KaiserGSVPJU17", "and", "larger", "datasets", "imagenet_cvpr09", "yields", "more", "impressive", "results", ".", "It", "is", "reasonable", "to", "hope", "that", "a", "similar", "principle", "holds", "for", "deep", "reinforcement", "learning", ".", "There", "are", "a", "growing", "number", "of", "examples", "to", "justify", "this", "optimism", ":", "effective", "use", "of", "greater", "computational", "resources", "has", "been", "a", "critical", "factor", "in", "the", "success", "of", "such", "algorithms", "as", "Gorila", "gorila", ",", "A3C", "a3c", ",", "GPU", "Advantage", "Actor", "Critic", "ga3c", ",", "Distributed", "PPO", "heess", ":", "dppo", "and", "AlphaGo", "alphago", ".", "Deep", "learning", "frameworks", "such", "as", "TensorFlow", "tensorflow", "support", "distributed", "training", ",", "making", "large", "scale", "machine", "learning", "systems", "easier", "to", "implement", "and", "deploy", ".", "Despite", "this", ",", "much", "current", "research", "in", "deep", "reinforcement", "learning", "concerns", "itself", "with", "improving", "performance", "within", "the", "computational", "budget", "of", "a", "single", "machine", ",", "and", "the", "question", "of", "how", "to", "best", "harness", "more", "resources", "is", "comparatively", "underexplored", ".", "In", "this", "paper", "we", "describe", "an", "approach", "to", "scaling", "up", "deep", "reinforcement", "learning", "by", "generating", "more", "data", "and", "selecting", "from", "it", "in", "a", "prioritized", "fashion", "prioritized", "-", "replay", ".", "Standard", "approaches", "to", "distributed", "training", "of", "neural", "networks", "focus", "on", "parallelizing", "the", "computation", "of", "gradients", ",", "to", "more", "rapidly", "optimize", "the", "parameters", "Dean:2012:LSD:2999134.2999271", ".", "In", "contrast", ",", "we", "distribute", "the", "generation", "and", "selection", "of", "experience", "data", ",", "and", "find", "that", "this", "alone", "suffices", "to", "improve", "results", ".", "This", "is", "complementary", "to", "distributing", "gradient", "computation", ",", "and", "the", "two", "approaches", "can", "be", "combined", ",", "but", "in", "this", "work", "we", "focus", "purely", "on", "data", "-", "generation", ".", "We", "use", "this", "distributed", "architecture", "to", "scale", "up", "variants", "of", "Deep", "Q", "-", "Networks", "(", "DQN", ")", "and", "Deep", "Deterministic", "Policy", "Gradient", "(", "DDPG", ")", ",", "and", "we", "evaluate", "these", "on", "the", "Arcade", "Learning", "Environment", "benchmark", "bellemare2013arcade", ",", "and", "on", "a", "range", "of", "continuous", "control", "tasks", ".", "Our", "architecture", "achieves", "a", "new", "state", "of", "the", "art", "performance", "on", "Atari", "games", ",", "using", "a", "fraction", "of", "the", "wall", "-", "clock", "time", "compared", "to", "the", "previous", "state", "of", "the", "art", ",", "and", "without", "per", "-", "game", "hyperparameter", "tuning", ".", "We", "empirically", "investigate", "the", "scalability", "of", "our", "framework", ",", "analysing", "how", "prioritization", "affects", "performance", "as", "we", "increase", "the", "number", "of", "data", "-", "generating", "workers", ".", "Our", "experiments", "include", "an", "analysis", "of", "factors", "such", "as", "the", "replay", "capacity", ",", "the", "recency", "of", "the", "experience", ",", "and", "the", "use", "of", "different", "data", "-", "generating", "policies", "for", "different", "workers", ".", "Finally", ",", "we", "discuss", "implications", "for", "deep", "reinforcement", "learning", "agents", "that", "may", "apply", "beyond", "our", "distributed", "framework", ".", "section", ":", "Background", "paragraph", ":", "Distributed", "Stochastic", "Gradient", "Descent", "Distributed", "stochastic", "gradient", "descent", "is", "widely", "used", "in", "supervised", "learning", "to", "speed", "up", "training", "of", "deep", "neural", "networks", ",", "by", "parallelizing", "the", "computation", "of", "the", "gradients", "used", "to", "update", "their", "parameters", ".", "The", "resulting", "parameter", "updates", "may", "be", "applied", "synchronously", "krizhevsky2014one", "or", "asynchronously", "Dean:2012:LSD:2999134.2999271", ".", "Both", "approaches", "have", "proven", "effective", "and", "are", "an", "increasingly", "standard", "part", "of", "the", "deep", "learning", "toolbox", ".", "Inspired", "by", "this", ",", "applied", "distributed", "asynchronous", "parameter", "updates", "and", "distributed", "data", "generation", "to", "deep", "reinforcement", "learning", ".", "Asynchronous", "parameter", "updates", "and", "parallel", "data", "generation", "have", "also", "been", "successfully", "used", "within", "a", "single", "-", "machine", ",", "in", "a", "multi", "-", "threaded", "rather", "than", "a", "distributed", "context", "a3c", ".", "GPU", "Asynchronous", "Actor", "-", "Critic", "[", "GA3C;", "][]", "ga3c", "and", "Parallel", "Advantage", "Actor", "-", "Critic", "[", "PAAC;", "][]", "paac", "adapt", "this", "approach", "to", "make", "efficient", "use", "of", "GPUs", ".", "paragraph", ":", "Distributed", "Importance", "Sampling", "A", "complementary", "family", "of", "techniques", "for", "speeding", "up", "training", "is", "based", "on", "variance", "reduction", "by", "means", "of", "importance", "sampling", "[", "cf.", "][]", "hastings1970monte", ".", "This", "has", "been", "shown", "to", "be", "useful", "in", "the", "context", "of", "neural", "networks", "Hinton2007", "-", "vs", ".", "Sampling", "non", "-", "uniformly", "from", "a", "dataset", "and", "weighting", "updates", "according", "to", "the", "sampling", "probability", "in", "order", "to", "counteract", "the", "bias", "thereby", "introduced", "can", "increase", "the", "speed", "of", "convergence", "by", "reducing", "the", "variance", "of", "the", "gradients", ".", "One", "way", "of", "doing", "this", "is", "to", "select", "samples", "with", "probability", "proportional", "to", "the", "norm", "of", "the", "corresponding", "gradients", ".", "In", "supervised", "learning", ",", "this", "approach", "has", "been", "successfully", "extended", "to", "the", "distributed", "setting", "alain2015variance", ".", "An", "alternative", "is", "to", "rank", "samples", "according", "to", "their", "latest", "known", "loss", "value", "and", "make", "the", "sampling", "probability", "a", "function", "of", "the", "rank", "rather", "than", "of", "the", "loss", "itself", "loshchilov2015online", ".", "paragraph", ":", "Prioritized", "Experience", "Replay", "Experience", "replay", "experience", "-", "replay", "has", "long", "been", "used", "in", "reinforcement", "learning", "to", "improve", "data", "efficiency", ".", "It", "is", "particularly", "useful", "when", "training", "neural", "network", "function", "approximators", "with", "stochastic", "gradient", "descent", "algorithms", ",", "as", "in", "Neural", "Fitted", "Q", "-", "Iteration", "nfq", "and", "Deep", "Q", "-", "Learning", "dqn", ".", "Experience", "replay", "may", "also", "help", "to", "prevent", "overfitting", "by", "allowing", "the", "agent", "to", "learn", "from", "data", "generated", "by", "previous", "versions", "of", "the", "policy", ".", "Prioritized", "experience", "replay", "prioritized", "-", "replay", "extends", "classic", "prioritized", "sweeping", "ideas", "prioritized", "-", "sweeping", "to", "work", "with", "deep", "neural", "network", "function", "approximators", ".", "The", "approach", "is", "strongly", "related", "to", "the", "importance", "sampling", "techniques", "discussed", "in", "the", "previous", "section", ",", "but", "using", "a", "more", "general", "class", "of", "biased", "sampling", "procedures", "that", "focus", "learning", "on", "the", "most", "‘", "surprising", "’", "experiences", ".", "Biased", "sampling", "can", "be", "particularly", "helpful", "in", "reinforcement", "learning", ",", "since", "the", "reward", "signal", "may", "be", "sparse", "and", "the", "data", "distribution", "depends", "on", "the", "agent", "’s", "policy", ".", "As", "a", "result", ",", "prioritized", "experience", "replay", "is", "used", "in", "many", "agents", ",", "such", "as", "Prioritized", "Dueling", "DQN", "dueling", ",", "UNREAL", "unreal", ",", "DQfD", "dqfd", ",", "and", "Rainbow", "rainbow", ".", "In", "an", "ablation", "study", "conducted", "to", "investigate", "the", "relative", "importance", "of", "several", "algorithmic", "ingredients", "rainbow", ",", "prioritization", "was", "found", "to", "be", "the", "most", "important", "ingredient", "contributing", "to", "the", "agent", "’s", "performance", ".", "section", ":", "Our", "Contribution", ":", "Distributed", "Prioritized", "Experience", "Replay", "In", "this", "paper", "we", "extend", "prioritized", "experience", "replay", "to", "the", "distributed", "setting", "and", "show", "that", "this", "is", "a", "highly", "scalable", "approach", "to", "deep", "reinforcement", "learning", ".", "We", "introduce", "a", "few", "key", "modifications", "that", "enable", "this", "scalability", ",", "and", "we", "refer", "to", "our", "approach", "as", "Ape", "-", "X", ".", "As", "in", "Gorila", "gorila", ",", "we", "decompose", "the", "standard", "deep", "reinforcement", "learning", "algorithm", "into", "two", "parts", ",", "which", "run", "concurrently", "with", "no", "high", "-", "level", "synchronization", ".", "The", "first", "part", "consists", "of", "stepping", "through", "an", "environment", ",", "evaluating", "a", "policy", "implemented", "as", "a", "deep", "neural", "network", ",", "and", "storing", "the", "observed", "data", "in", "a", "replay", "memory", ".", "We", "refer", "to", "this", "as", "acting", ".", "The", "second", "part", "consists", "of", "sampling", "batches", "of", "data", "from", "the", "memory", "to", "update", "the", "policy", "parameters", ".", "We", "term", "this", "learning", ".", "[", "width", "=", "trim=0pt", "400pt", "0pt", "0pt", ",", "clip", "]", "images", "/", "apex_architecture.pdf", "[", "t", "!", "]", "Actor", "{", "algorithmic}", "[", "1", "]", "T", "agent", "in", "environment", "instance", ",", "storing", "experiences", ".", "call", "to", "obtain", "latest", "network", "parameters", ".", "initial", "state", "from", "environment", ".", "to", "T", "an", "action", "using", "the", "current", "policy", ".", "the", "action", "in", "the", "environment", ".", "data", "to", "local", "buffer", ".", "a", "background", "thread", ",", "periodically", "send", "data", "to", "replay", ".", "buffered", "data", "(", "e.g.", "batch", "of", "multi", "-", "step", "transitions", ")", ".", "priorities", "for", "experience", "(", "e.g.", "absolute", "TD", "error", ")", ".", "call", "to", "add", "experience", "to", "replay", "memory", ".", "latest", "network", "parameters", ".", "[", "t", "!", "]", "Learner", "{", "algorithmic}", "[", "1", "]", "network", "using", "batches", "sampled", "from", "memory", ".", "to", "T", "the", "parameters", "T", "times", ".", "a", "prioritized", "batch", "of", "transitions", "(", "in", "a", "background", "thread", ")", ".", "learning", "rule", ";", "e.g.", "double", "Q", "-", "learning", "or", "DDPG", "priorities", "for", "experience", ",", "(", "e.g.", "absolute", "TD", "error", ")", ".", "call", "to", "update", "priorities", ".", "old", "experience", "from", "replay", "memory", ".", "In", "principle", ",", "both", "acting", "and", "learning", "may", "be", "distributed", "across", "multiple", "workers", ".", "In", "our", "experiments", ",", "hundreds", "of", "actors", "run", "on", "CPUs", "to", "generate", "data", ",", "and", "a", "single", "learner", "running", "on", "a", "GPU", "samples", "the", "most", "useful", "experiences", "(", "Figure", "[", "reference", "]", ")", ".", "Pseudocode", "for", "the", "actors", "and", "learners", "is", "shown", "in", "Algorithms", "[", "reference", "]", "and", "[", "reference", "]", ".", "Updated", "network", "parameters", "are", "periodically", "communicated", "to", "the", "actors", "from", "the", "learner", ".", "In", "contrast", "to", "gorila", ",", "we", "use", "a", "shared", ",", "centralized", "replay", "memory", ",", "and", "instead", "of", "sampling", "uniformly", ",", "we", "prioritize", ",", "to", "sample", "the", "most", "useful", "data", "more", "often", ".", "Since", "priorities", "are", "shared", ",", "high", "priority", "data", "discovered", "by", "any", "actor", "can", "benefit", "the", "whole", "system", ".", "Priorities", "can", "be", "defined", "in", "various", "ways", ",", "depending", "on", "the", "learning", "algorithm", ";", "two", "instances", "are", "described", "in", "the", "next", "sections", ".", "In", "Prioritized", "DQN", "prioritized", "-", "replay", "priorities", "for", "new", "transitions", "were", "initialized", "to", "the", "maximum", "priority", "seen", "so", "far", ",", "and", "only", "updated", "once", "they", "were", "sampled", ".", "This", "does", "not", "scale", "well", ":", "due", "to", "the", "large", "number", "of", "actors", "in", "our", "architecture", ",", "waiting", "for", "the", "learner", "to", "update", "priorities", "would", "result", "in", "a", "myopic", "focus", "on", "the", "most", "recent", "data", ",", "which", "has", "maximum", "priority", "by", "construction", ".", "Instead", ",", "we", "take", "advantage", "of", "the", "computation", "the", "actors", "in", "Ape", "-", "X", "are", "already", "doing", "to", "evaluate", "their", "local", "copies", "of", "the", "policy", ",", "by", "making", "them", "also", "compute", "suitable", "priorities", "for", "new", "transitions", "online", ".", "This", "ensures", "that", "data", "entering", "the", "replay", "has", "more", "accurate", "priorities", ",", "at", "no", "extra", "cost", ".", "Sharing", "experiences", "has", "certain", "advantages", "compared", "to", "sharing", "gradients", ".", "Low", "latency", "communication", "is", "not", "as", "important", "as", "in", "distributed", "SGD", ",", "because", "experience", "data", "becomes", "outdated", "less", "rapidly", "than", "gradients", ",", "provided", "the", "learning", "algorithm", "is", "robust", "to", "off", "-", "policy", "data", ".", "Across", "the", "system", ",", "we", "take", "advantage", "of", "this", "by", "batching", "all", "communications", "with", "the", "centralized", "replay", ",", "increasing", "the", "efficiency", "and", "throughput", "at", "the", "cost", "of", "some", "latency", ".", "With", "this", "approach", "it", "is", "even", "possible", "for", "actors", "and", "learners", "to", "run", "in", "different", "data", "-", "centers", "without", "limiting", "performance", ".", "Finally", ",", "by", "learning", "off", "-", "policy", "[", "cf.", "][]", "SuttonBarto:1998", ",", "SuttonBarto:2017", ",", "we", "can", "further", "take", "advantage", "of", "Ape", "-", "X", "’s", "ability", "to", "combine", "data", "from", "many", "distributed", "actors", ",", "by", "giving", "the", "different", "actors", "different", "exploration", "policies", ",", "broadening", "the", "diversity", "of", "the", "experience", "they", "jointly", "encounter", ".", "As", "we", "will", "see", "in", "the", "results", ",", "this", "can", "be", "sufficient", "to", "make", "progress", "on", "difficult", "exploration", "problems", ".", "subsection", ":", "Ape", "-", "X", "DQN", "The", "general", "framework", "we", "have", "described", "may", "be", "combined", "with", "different", "learning", "algorithms", ".", "First", ",", "we", "combined", "it", "with", "a", "variant", "of", "DQN", "dqn", "with", "some", "of", "the", "components", "of", "Rainbow", "rainbow", ".", "More", "specifically", ",", "we", "used", "double", "Q", "-", "learning", "doubleq", ",", "deepdoubleqlearning", "with", "multi", "-", "step", "bootstrap", "targets", "[", "cf.", "][]", "Sutton:1988", ",", "SuttonBarto:1998", ",", "SuttonBarto:2017", ",", "a3c", "as", "the", "learning", "algorithm", ",", "and", "a", "dueling", "network", "architecture", "dueling", "as", "the", "function", "approximator", ".", "This", "results", "in", "computing", "for", "all", "elements", "in", "the", "batch", "the", "loss", "with", "where", "is", "a", "time", "index", "for", "an", "experience", "sampled", "from", "the", "replay", "starting", "with", "state", "and", "action", ",", "and", "denotes", "parameters", "of", "the", "target", "network", "dqn", ",", "a", "slow", "moving", "copy", "of", "the", "online", "parameters", ".", "Multi", "-", "step", "returns", "are", "truncated", "if", "the", "episode", "ends", "in", "fewer", "than", "steps", ".", "In", "principle", ",", "Q", "-", "learning", "variants", "are", "off", "-", "policy", "methods", ",", "so", "we", "are", "free", "to", "choose", "the", "policies", "we", "use", "to", "generate", "data", ".", "However", ",", "in", "practice", ",", "the", "choice", "of", "behaviour", "policy", "does", "affect", "both", "exploration", "and", "the", "quality", "of", "function", "approximation", ".", "Furthermore", ",", "we", "are", "using", "a", "multi", "-", "step", "return", "with", "no", "off", "-", "policy", "correction", ",", "which", "in", "theory", "could", "adversely", "affect", "the", "value", "estimation", ".", "Nonetheless", ",", "in", "Ape", "-", "X", "DQN", ",", "each", "actor", "executes", "a", "different", "policy", ",", "and", "this", "allows", "experience", "to", "be", "generated", "from", "a", "variety", "of", "strategies", ",", "relying", "on", "the", "prioritization", "mechanism", "to", "pick", "out", "the", "most", "effective", "experiences", ".", "In", "our", "experiments", ",", "the", "actors", "use", "-", "greedy", "policies", "with", "different", "values", "of", ".", "Low", "policies", "allow", "exploring", "deeper", "in", "the", "environment", ",", "while", "high", "policies", "prevent", "over", "-", "specialization", ".", "subsection", ":", "Ape", "-", "X", "DPG", "To", "test", "the", "generality", "of", "the", "framework", "we", "also", "combined", "it", "with", "a", "continuous", "-", "action", "policy", "gradient", "system", "based", "on", "DDPG", "ddpg", ",", "an", "implementation", "of", "deterministic", "policy", "gradients", "also", "similar", "to", "older", "methods", "adhdp", ",", "acd", ",", "and", "tested", "it", "on", "continuous", "control", "tasks", "from", "the", "DeepMind", "Control", "Suite", "tassa2018suite", ".", "The", "Ape", "-", "X", "DPG", "setup", "is", "similar", "to", "Ape", "-", "X", "DQN", ",", "but", "the", "actor", "’s", "policy", "is", "now", "represented", "explicitly", "by", "a", "separate", "policy", "network", ",", "in", "addition", "to", "the", "Q", "-", "network", ".", "The", "two", "networks", "are", "optimized", "separately", ",", "by", "minimizing", "different", "losses", "on", "the", "sampled", "experience", ".", "We", "denote", "the", "policy", "and", "Q", "-", "network", "parameters", "by", "and", "respectively", ",", "and", "adopt", "the", "same", "convention", "as", "above", "to", "denote", "target", "networks", ".", "The", "Q", "-", "network", "outputs", "an", "action", "-", "value", "estimate", "for", "a", "given", "state", ",", "and", "multi", "-", "dimensional", "action", ".", "It", "is", "updated", "using", "temporal", "-", "difference", "learning", "with", "a", "multi", "-", "step", "bootstrap", "target", ".", "The", "Q", "-", "network", "loss", "can", "be", "written", "as", ",", "where", "The", "policy", "network", "outputs", "an", "action", ".", "The", "policy", "parameters", "are", "updated", "using", "policy", "gradient", "ascent", "on", "the", "estimated", "Q", "-", "value", ",", "using", "gradient", "—", "note", "that", "this", "depends", "on", "the", "policy", "parameters", "only", "through", "the", "action", "that", "is", "input", "to", "the", "critic", "network", ".", "Further", "details", "of", "the", "Ape", "-", "X", "DPG", "algorithm", "are", "available", "in", "the", "appendix", ".", "section", ":", "Experiments", "subsection", ":", "Atari", "[", "width=", "]", "images", "/", "apex_iclr_aggregate_scatter_jan_04", "[", "width=", "]", "images", "/", "apex_iclr_selected_levels_jan_04", "In", "our", "first", "set", "of", "experiments", "we", "evaluate", "Ape", "-", "X", "DQN", "on", "Atari", ",", "and", "show", "state", "of", "the", "art", "results", "on", "this", "standard", "reinforcement", "learning", "benchmark", ".", "We", "use", "360", "actor", "machines", "(", "each", "using", "one", "CPU", "core", ")", "to", "feed", "data", "into", "the", "replay", "memory", "as", "fast", "as", "they", "can", "generate", "it", ";", "approximately", "139", "frames", "per", "second", "(", "FPS", ")", "each", ",", "for", "a", "total", "of", "50", "K", "FPS", ",", "which", "corresponds", "to", "12.5", "K", "transitions", "(", "because", "of", "a", "fixed", "action", "repeat", "of", "4", ")", ".", "The", "actors", "batch", "experience", "data", "locally", "before", "sending", "it", "to", "the", "replay", ":", "up", "to", "100", "transitions", "may", "be", "buffered", "at", "a", "time", ",", "which", "are", "then", "sent", "asynchronously", "in", "batches", "of", ".", "The", "learner", "asynchronously", "prefetches", "up", "to", "16", "batches", "of", "512", "transitions", ",", "and", "computes", "updates", "for", "19", "such", "batches", "each", "second", ",", "meaning", "that", "gradients", "are", "computed", "for", "9.7", "K", "transitions", "per", "second", "on", "average", ".", "To", "reduce", "memory", "and", "bandwidth", "requirements", ",", "observation", "data", "is", "compressed", "using", "a", "PNG", "codec", "when", "sent", "and", "when", "stored", "in", "the", "replay", ".", "The", "learner", "decompresses", "data", "as", "it", "prefetches", "it", ",", "in", "parallel", "with", "computing", "and", "applying", "gradients", ".", "The", "learner", "also", "asynchronously", "handles", "any", "requests", "for", "parameters", "from", "actors", ".", "Actors", "copy", "the", "network", "parameters", "from", "the", "learner", "every", "400", "frames", "(", "2.8", "seconds", ")", ".", "Each", "actor", "executes", "an", "-", "greedy", "policy", "where", "with", ",", ".", "Each", "is", "held", "constant", "throughout", "training", ".", "The", "episode", "length", "is", "limited", "to", "50000", "frames", "during", "training", ".", "The", "capacity", "of", "the", "shared", "experience", "replay", "memory", "is", "soft", "-", "limited", "to", "2", "million", "transitions", ":", "adding", "new", "data", "is", "always", "permitted", ",", "to", "not", "slow", "down", "the", "actors", ",", "but", "every", "100", "learning", "steps", "any", "excess", "data", "above", "this", "capacity", "threshold", "is", "removed", "en", "masse", ",", "in", "FIFO", "order", ".", "The", "median", "actual", "size", "of", "the", "memory", "is", "2035050", ".", "Data", "is", "sampled", "according", "to", "proportional", "prioritization", ",", "with", "a", "priority", "exponent", "of", "0.6", "and", "an", "importance", "sampling", "exponent", "set", "to", "0.4", ".", "In", "Figure", "[", "reference", "]", ",", "on", "the", "left", ",", "we", "compare", "the", "median", "human", "normalized", "score", "across", "all", "57", "games", "to", "several", "baselines", ":", "DQN", ",", "Prioritized", "DQN", ",", "Distributional", "DQN", "distributional", ",", "Rainbow", ",", "and", "Gorila", ".", "In", "all", "cases", "the", "performance", "is", "measured", "at", "the", "end", "of", "training", "under", "the", "no", "-", "op", "starts", "testing", "regime", "dqn", ".", "On", "the", "right", ",", "we", "show", "initial", "learning", "curves", "(", "taken", "from", "the", "greediest", "actor", ")", "for", "a", "selection", "of", "6", "games", "(", "full", "learning", "curves", "for", "all", "games", "are", "in", "the", "appendix", ")", ".", "Given", "that", "Ape", "-", "X", "can", "harness", "substantially", "more", "computation", "than", "most", "baselines", ",", "one", "might", "expect", "it", "to", "train", "faster", ".", "Figure", "[", "reference", "]", "shows", "that", "this", "was", "indeed", "the", "case", ".", "Perhaps", "more", "surprisingly", ",", "our", "agent", "achieved", "a", "substantially", "higher", "final", "performance", ".", "In", "Table", "[", "reference", "]", "we", "compare", "the", "median", "human", "-", "normalized", "performance", "of", "Ape", "-", "X", "DQN", "on", "the", "Atari", "benchmark", "to", "corresponding", "metrics", "as", "reported", "for", "other", "baseline", "agents", "in", "their", "respective", "publications", ".", "Whenever", "available", "we", "report", "results", "both", "for", "no", "-", "op", "starts", "and", "for", "human", "starts", ".", "The", "human", "-", "starts", "regime", "gorila", "corresponds", "to", "a", "more", "challenging", "generalization", "test", ",", "as", "the", "agent", "is", "initialized", "from", "random", "starts", "drawn", "from", "games", "played", "by", "human", "experts", ".", "Ape", "-", "X", "’s", "performance", "is", "higher", "than", "the", "performance", "of", "any", "of", "the", "baselines", "according", "to", "both", "metrics", ".", "subsection", ":", "Continuous", "Control", "In", "a", "second", "set", "of", "experiments", "we", "evaluated", "Ape", "-", "X", "DPG", "on", "four", "continuous", "control", "tasks", ".", "In", "the", "manipulator", "domain", "the", "agent", "must", "learn", "to", "bring", "a", "ball", "to", "a", "specified", "location", ".", "In", "the", "humanoid", "domain", "the", "agent", "must", "learn", "to", "control", "a", "humanoid", "body", "to", "solve", "three", "distinct", "tasks", "of", "increasing", "complexity", ":", "Standing", ",", "Walking", "and", "Running", ".", "Since", "here", "we", "learn", "from", "features", ",", "rather", "than", "from", "pixels", ",", "the", "observation", "space", "is", "much", "smaller", "than", "it", "is", "in", "the", "Atari", "domain", ".", "We", "therefore", "use", "small", ",", "fully", "-", "connected", "networks", "(", "details", "in", "the", "appendix", ")", ".", "With", "64", "actors", "on", "this", "domain", ",", "we", "obtain", "14", "K", "total", "FPS", "(", "the", "same", "number", "of", "transitions", "per", "second", ";", "here", "we", "do", "not", "use", "action", "repeats", ")", ".", "We", "process", "86", "batches", "of", "256", "transitions", "per", "second", ",", "or", "22", "K", "transitions", "processed", "per", "second", ".", "Figure", "[", "reference", "]", "shows", "that", "Ape", "-", "X", "DPG", "achieved", "very", "good", "performance", "on", "all", "four", "tasks", ".", "The", "figure", "shows", "the", "performance", "of", "Ape", "-", "X", "DPG", "for", "different", "numbers", "of", "actors", ":", "as", "the", "number", "of", "actors", "increases", "our", "agent", "becomes", "increasingly", "effective", "at", "solving", "these", "problems", "rapidly", "and", "reliably", ",", "outperforming", "a", "standard", "DDPG", "baseline", "trained", "for", "over", "10", "times", "longer", ".", "A", "parallel", "paper", "d4pg", "builds", "on", "this", "work", "by", "combining", "Ape", "-", "X", "DPG", "with", "distributional", "value", "functions", ",", "and", "the", "resulting", "algorithm", "is", "successfully", "applied", "to", "further", "continuous", "control", "tasks", ".", "[", "width=0.8clip", "]", "images", "/", "apex_iclr_continuous_control", "section", ":", "Analysis", "[", "width=0.85trim=0pt", "0", "0pt", "40pt", ",", "clip", "]", "images", "/", "apex_iclr_num_actors_prioritized_only", "In", "this", "section", "we", "describe", "additional", "Ape", "-", "X", "DQN", "experiments", "on", "Atari", "that", "helped", "improve", "our", "understanding", "of", "the", "framework", ",", "and", "we", "investigate", "the", "contribution", "of", "different", "components", ".", "First", ",", "we", "investigated", "how", "the", "performance", "scales", "with", "the", "number", "of", "actors", ".", "We", "trained", "our", "agent", "with", "different", "numbers", "of", "actors", "(", "8", ",", "16", ",", "32", ",", "64", ",", "128", "and", "256", ")", "for", "35", "hours", "on", "a", "subset", "of", "6", "Atari", "games", ".", "In", "all", "experiments", "we", "kept", "the", "size", "of", "the", "shared", "experience", "replay", "memory", "fixed", "at", "1", "million", "transitions", ".", "Figure", "[", "reference", "]", "shows", "that", "the", "performance", "consistently", "improved", "as", "the", "number", "of", "actors", "increased", ".", "The", "appendix", "contains", "learning", "curves", "for", "additional", "games", ",", "and", "a", "comparison", "of", "the", "scalability", "of", "the", "algorithm", "with", "and", "without", "prioritized", "replay", ".", "It", "is", "perhaps", "surprising", "that", "performance", "improved", "so", "substantially", "purely", "by", "increasing", "the", "number", "of", "actors", ",", "without", "changing", "the", "rate", "at", "which", "the", "network", "parameters", "are", "updated", ",", "the", "structure", "of", "the", "network", ",", "or", "the", "update", "rule", ".", "We", "hypothesize", "that", "the", "proposed", "architecture", "helps", "with", "a", "common", "deep", "reinforcement", "learning", "failure", "mode", ",", "in", "which", "the", "policy", "discovered", "is", "a", "local", "optimum", "in", "the", "parameter", "space", ",", "but", "not", "a", "global", "one", ",", "e.g.", ",", "due", "to", "insufficient", "exploration", ".", "Using", "a", "large", "number", "of", "actors", "with", "varying", "amounts", "of", "exploration", "helps", "to", "discover", "promising", "new", "courses", "of", "action", ",", "and", "prioritized", "replay", "ensures", "that", "when", "this", "happens", ",", "the", "learning", "algorithm", "focuses", "its", "efforts", "on", "this", "important", "information", ".", "Next", ",", "we", "investigated", "varying", "the", "capacity", "of", "the", "replay", "memory", "(", "see", "Figure", "[", "reference", "]", ")", ".", "We", "used", "a", "setup", "with", "256", "actors", ",", "for", "a", "median", "of", "37", "K", "total", "environment", "frames", "per", "second", "(", "approximately", "9", "K", "transitions", ")", ".", "With", "such", "a", "large", "number", "of", "actors", ",", "the", "contents", "of", "the", "memory", "is", "replaced", "much", "faster", "than", "in", "most", "DQN", "-", "like", "agents", ".", "We", "observed", "a", "small", "benefit", "to", "using", "a", "larger", "replay", "capacity", ".", "We", "hypothesize", "this", "is", "due", "to", "the", "value", "of", "keeping", "some", "high", "priority", "experiences", "around", "for", "longer", "and", "replaying", "them", ".", "As", "above", ",", "a", "single", "learner", "machine", "trained", "the", "network", "with", "median", "19", "batches", "per", "second", ",", "each", "of", "512", "transitions", ",", "for", "a", "median", "of", "9.7", "K", "transitions", "processed", "per", "second", ".", "[", "width", "=", "trim=0pt", "0", "0pt", "40pt", ",", "clip", "]", "images", "/", "apex_iclr_vary_replay_size.pdf", "Finally", ",", "we", "ran", "additional", "experiments", "to", "disentangle", "potential", "effects", "of", "two", "confounding", "factors", "in", "our", "scalability", "analysis", ":", "recency", "of", "the", "experience", "data", "in", "the", "replay", "memory", ",", "and", "diversity", "of", "the", "data", "-", "generating", "policies", ".", "The", "full", "description", "of", "these", "experiments", "is", "confined", "to", "the", "appendix", ";", "to", "summarize", ",", "neither", "factor", "alone", "is", "sufficient", "to", "explain", "the", "performance", "we", "see", ".", "We", "therefore", "conclude", "that", "the", "results", "are", "due", "substantially", "to", "the", "positive", "effects", "of", "gathering", "more", "experience", "data", ";", "namely", "better", "exploration", "of", "the", "environment", "and", "better", "avoidance", "of", "overfitting", ".", "section", ":", "Conclusion", "We", "have", "designed", ",", "implemented", ",", "and", "analyzed", "a", "distributed", "framework", "for", "prioritized", "replay", "in", "deep", "reinforcement", "learning", ".", "This", "architecture", "achieved", "state", "of", "the", "art", "results", "in", "a", "wide", "range", "of", "discrete", "and", "continuous", "tasks", ",", "both", "in", "terms", "of", "wall", "-", "clock", "learning", "speed", "and", "final", "performance", ".", "In", "this", "paper", "we", "focused", "on", "applying", "the", "Ape", "-", "X", "framework", "to", "DQN", "and", "DPG", ",", "but", "it", "could", "also", "be", "combined", "with", "any", "other", "off", "-", "policy", "reinforcement", "learning", "update", ".", "For", "methods", "that", "use", "temporally", "extended", "sequences", "[", "e.g.", ",", "]", "[", "]", "a3c", ",", "acer", ",", "the", "Ape", "-", "X", "framework", "may", "be", "adapted", "to", "prioritize", "sequences", "of", "past", "experiences", "instead", "of", "individual", "transitions", ".", "Ape", "-", "X", "is", "designed", "for", "regimes", "in", "which", "it", "is", "possible", "to", "generate", "large", "quantities", "of", "data", "in", "parallel", ".", "This", "includes", "simulated", "environments", "but", "also", "a", "variety", "of", "real", "-", "world", "applications", ",", "such", "as", "robotic", "arm", "farms", ",", "self", "-", "driving", "cars", ",", "online", "recommender", "systems", ",", "or", "other", "multi", "-", "user", "systems", "in", "which", "data", "is", "generated", "by", "many", "instances", "of", "the", "same", "environment", "[", "c.f.", "][]", "concurrent", "-", "rl", ".", "In", "applications", "where", "data", "is", "costly", "to", "obtain", ",", "our", "approach", "will", "not", "be", "directly", "applicable", ".", "With", "powerful", "function", "approximators", ",", "overfitting", "is", "an", "issue", ":", "generating", "more", "training", "data", "is", "the", "simplest", "way", "of", "addressing", "it", ",", "but", "may", "also", "provide", "guidance", "towards", "data", "-", "efficient", "solutions", ".", "Many", "deep", "reinforcement", "learning", "algorithms", "are", "fundamentally", "limited", "by", "their", "ability", "to", "explore", "effectively", "in", "large", "domains", ".", "Ape", "-", "X", "uses", "a", "naive", "yet", "effective", "mechanism", "to", "address", "this", "issue", ":", "generating", "a", "diverse", "set", "of", "experiences", "and", "then", "identifying", "and", "learning", "from", "the", "most", "useful", "events", ".", "The", "success", "of", "this", "approach", "suggests", "that", "simple", "and", "direct", "approaches", "to", "exploration", "may", "be", "feasible", ",", "even", "for", "synchronous", "agents", ".", "Our", "architecture", "illustrates", "that", "distributed", "systems", "are", "now", "practical", "both", "for", "research", "and", ",", "potentially", ",", "large", "-", "scale", "applications", "of", "deep", "reinforcement", "learning", ".", "We", "hope", "that", "the", "algorithms", ",", "architecture", ",", "and", "analysis", "we", "have", "presented", "will", "help", "to", "accelerate", "future", "efforts", "in", "this", "direction", ".", "subsubsection", ":", "Acknowledgments", "We", "would", "like", "to", "acknowledge", "the", "contributions", "of", "our", "colleagues", "at", "DeepMind", ",", "whose", "input", "and", "support", "has", "been", "vital", "to", "the", "success", "of", "this", "work", ".", "Thanks", "in", "particular", "to", "Tom", "Schaul", ",", "Joseph", "Modayil", ",", "Sriram", "Srinivasan", ",", "Georg", "Ostrovski", ",", "Josh", "Abramson", ",", "Todd", "Hester", ",", "Jean", "-", "Baptiste", "Lespiau", ",", "Alban", "Rrustemi", "and", "Dan", "Belov", ".", "bibliography", ":", "References", "appendix", ":", "Recency", "of", "Experience", "[", "width=", "]", "images", "/", "apex_iclr_virtual_actors", "[", "width=", "]", "images", "/", "apex_iclr_varying_the_pool_of_policies.pdf", "In", "our", "main", "experiments", "we", "do", "not", "change", "the", "size", "of", "the", "replay", "memory", "in", "proportion", "to", "the", "number", "of", "actors", ",", "so", "by", "changing", "the", "number", "of", "actors", "we", "also", "increased", "the", "rate", "at", "which", "the", "contents", "of", "the", "replay", "memory", "is", "replaced", ".", "This", "means", "that", "in", "the", "experiments", "with", "more", "actors", ",", "transitions", "in", "the", "replay", "memory", "are", "more", "recent", ":", "they", "are", "generated", "by", "following", "policies", "whose", "parameters", "are", "closer", "to", "version", "of", "the", "parameters", "being", "optimized", "by", "the", "learner", ",", "and", "in", "this", "sense", "they", "are", "more", "on", "-", "policy", ".", "Could", "this", "alone", "be", "sufficient", "to", "explain", "the", "improved", "performance", "?", "If", "so", ",", "we", "might", "be", "able", "to", "recover", "the", "results", "without", "needing", "a", "large", "number", "of", "actor", "machines", ".", "To", "test", "this", ",", "we", "constructed", "an", "experiment", "wherein", "we", "replicate", "the", "rate", "at", "which", "the", "contents", "of", "the", "replay", "memory", "is", "replaced", "in", "the", "256", "-", "actor", "experiments", ",", "but", "instead", "of", "actually", "using", "256", "actors", ",", "we", "use", "32", "actors", "but", "add", "each", "transition", "they", "generate", "to", "the", "replay", "memory", "8", "times", "over", ".", "In", "this", "setup", ",", "the", "contents", "of", "the", "replay", "memory", "is", "similarly", "generated", "by", "policies", "with", "a", "recent", "version", "of", "the", "network", "parameters", ":", "the", "only", "difference", "is", "that", "the", "data", "is", "not", "as", "diverse", "as", "in", "the", "256", "-", "actor", "case", ".", "We", "observe", "(", "see", "Figure", "[", "reference", "]", ")", "that", "this", "does", "not", "recover", "the", "same", "performance", ",", "and", "therefore", "conclude", "that", "the", "recency", "of", "the", "experience", "alone", "is", "not", "sufficient", "to", "explain", "the", "performance", "of", "our", "method", ".", "Indeed", ",", "we", "see", "that", "adding", "the", "same", "data", "multiple", "times", "can", "sometimes", "harm", "performance", ",", "since", "although", "it", "increases", "recency", "this", "comes", "at", "the", "expense", "of", "diversity", ".", "Note", ":", "in", "principle", ",", "duplicating", "the", "added", "data", "in", "this", "fashion", "has", "a", "similar", "effect", "to", "reducing", "the", "capacity", "of", "the", "replay", "memory", ",", "and", "indeed", ",", "our", "results", "with", "a", "smaller", "replay", "memory", "in", "Figure", "[", "reference", "]", "do", "corroborate", "the", "finding", ".", "However", ",", "we", "test", "also", "by", "duplicating", "the", "data", "primarily", "in", "order", "to", "exclude", "any", "effects", "arising", "from", "the", "implementation", ".", "In", "particular", ",", "in", "contrast", "to", "simply", "reducing", "the", "replay", "capacity", ",", "duplicating", "each", "data", "point", "means", "that", "the", "computational", "demands", "on", "the", "replay", "server", "in", "these", "runs", "are", "the", "same", "as", "when", "we", "use", "the", "corresponding", "number", "of", "real", "actors", ".", "appendix", ":", "Varying", "the", "Data", "-", "Generating", "Policies", "Another", "factor", "that", "could", "conceivably", "contribute", "to", "the", "scalability", "of", "our", "algorithm", "is", "the", "fact", "that", "each", "actor", "has", "a", "different", ".", "To", "determine", "the", "extent", "to", "which", "this", "impacts", "upon", "the", "performance", ",", "we", "ran", "an", "experiment", "(", "see", "Figure", "[", "reference", "]", ")", "with", "some", "simple", "variations", "on", "the", "mechanism", "we", "use", "to", "choose", "the", "policies", "that", "generate", "the", "data", "we", "train", "on", ".", "The", "first", "alternative", "we", "tested", "is", "to", "choose", "a", "small", "fixed", "set", "of", "6", "values", "for", ",", "instead", "of", "the", "full", "range", "that", "we", "typically", "use", ".", "In", "this", "test", ",", "we", "use", "prioritized", "replay", "as", "normal", ",", "and", "we", "find", "that", "the", "results", "with", "the", "full", "range", "of", "are", "overall", "slightly", "better", ".", "However", ",", "it", "is", "not", "essential", "for", "achieving", "good", "results", "within", "our", "distributed", "framework", ".", "appendix", ":", "Atari", ":", "Additional", "Details", "The", "frames", "received", "from", "the", "environment", "are", "preprocessed", "on", "the", "actor", "side", "with", "the", "standard", "transformations", "introduced", "by", "DQN", ".", "This", "includes", "greyscaling", ",", "frame", "stacking", ",", "repeating", "actions", "4", "times", ",", "and", "clipping", "rewards", "to", ".", "The", "learner", "waits", "for", "at", "least", "50000", "transitions", "to", "be", "accumulated", "in", "the", "replay", "before", "starting", "learning", ".", "We", "use", "a", "Centered", "RMSProp", "optimizer", "with", "a", "learning", "rate", "of", "0.00025", "/", "4", ",", "decay", "of", "0.95", ",", "epsilon", "of", "1.5e", "-", "7", ",", "and", "no", "momentum", "to", "minimize", "the", "multi", "-", "step", "loss", "(", "with", ")", ".", "Gradient", "norms", "are", "clipped", "to", "40", ".", "The", "target", "network", "used", "in", "the", "loss", "calculation", "is", "copied", "from", "the", "online", "network", "every", "2500", "training", "batches", ".", "We", "use", "the", "same", "network", "as", "in", "the", "Dueling", "DDQN", "agent", ".", "appendix", ":", "Continuous", "Control", ":", "Additional", "Details", "The", "critic", "network", "has", "a", "layer", "with", "400", "units", ",", "followed", "by", "a", "tanh", "activation", ",", "followed", "by", "another", "layer", "of", "300", "units", ".", "The", "actor", "network", "has", "a", "layer", "with", "300", "units", ",", "followed", "by", "a", "tanh", "activation", ",", "followed", "by", "another", "layer", "of", "200", "units", ".", "The", "gradient", "used", "to", "update", "the", "actor", "network", "is", "clipped", "to", ",", "element", "-", "wise", ".", "Training", "uses", "the", "Adam", "optimizer", "(", ")", "with", "learning", "rate", "of", ".", "The", "target", "network", "used", "in", "the", "loss", "calculation", "is", "copied", "from", "the", "online", "network", "every", "100", "training", "batches", ".", "Replay", "sampling", "priorities", "are", "set", "according", "to", "the", "absolute", "TD", "error", "as", "given", "by", "the", "critic", ",", "and", "are", "sampled", "by", "the", "learner", "using", "proportional", "prioritized", "sampling", "(", "see", "appendix", "[", "reference", "]", ")", "with", "priority", "exponent", ".", "To", "maintain", "a", "fixed", "replay", "capacity", "of", ",", "transitions", "are", "periodically", "evicted", "using", "proportional", "prioritized", "sampling", ",", "with", "priority", "exponent", ".", "This", "is", "a", "different", "strategy", "for", "removing", "data", "than", "in", "the", "Atari", "experiments", ",", "which", "simply", "removed", "the", "oldest", "data", "first", "-", "it", "remains", "to", "be", "seen", "which", "is", "superior", ".", "Unlike", "the", "original", "DPG", "algorithm", "which", "applies", "autocorrelated", "noise", "sampled", "from", "a", "Ornstein", "-", "Uhlenbeck", "process", "(", ")", ",", "we", "apply", "exploration", "noise", "to", "each", "action", "sampled", "from", "a", "normal", "distribution", "with", ".", "Evaluation", "is", "performed", "using", "the", "noiseless", "deterministic", "policy", ".", "Hyperparameters", "are", "otherwise", "as", "per", "DQN", ".", "Benchmarking", "was", "performed", "in", "two", "continuous", "control", "domains", "(", "(", "a", ")", "Humanoid", "and", "(", "b", ")", "Manipulator", ",", "see", "Figure", "[", "reference", "]", ")", "implemented", "in", "the", "MuJoCo", "physics", "simulator", "(", ")", ".", "Humanoid", "is", "a", "humanoid", "walker", "with", "action", ",", "state", "and", "observation", "dimensionalities", ",", "and", "respectively", ".", "Three", "Humanoid", "tasks", "were", "considered", ":", "walk", "(", "reward", "for", "exceeding", "a", "minimum", "velocity", ")", ",", "run", "(", "reward", "proportional", "to", "movement", "speed", ")", "and", "stand", "(", "reward", "proportional", "to", "standing", "height", ")", ".", "Manipulator", "is", "a", "2", "-", "dimensional", "planar", "arm", "with", ",", "and", ",", "which", "receives", "reward", "for", "catching", "a", "randomly", "-", "initialized", "moving", "ball", ".", ".5", "[", "width=.6", "]", ".", "/", "images", "/", "humanoid.png", ".5", "[", "width=.6", "]", ".", "/", "images", "/", "manipulator.png", "appendix", ":", "Tuning", "On", "Atari", ",", "we", "performed", "some", "limited", "tuning", "of", "the", "learning", "rate", "and", "batch", "size", ":", "we", "found", "that", "larger", "batch", "sizes", "contribute", "significantly", "to", "performance", ",", "when", "using", "many", "actors", ".", "We", "tried", "batch", "sizes", "from", "{", "32", ",", "128", ",", "256", ",", "512", ",", "1024", "}", ",", "seeing", "clear", "benefits", "up", "to", "512", ".", "We", "attempted", "increasing", "the", "learning", "rate", "to", "0.00025", "with", "the", "larger", "batch", "sizes", "but", "this", "destabilized", "training", "on", "some", "games", ".", "We", "also", "tried", "a", "lower", "learning", "rate", "of", "0.00025", "/", "8", ",", "but", "this", "did", "not", "reliably", "improve", "results", ".", "Likewise", "for", "continuous", "control", ",", "we", "experimented", "with", "batch", "sizes", "{", "32", ",", "128", ",", "256", ",", "512", ",", "1024", "}", "and", "learning", "rates", "from", "to", ".", "We", "also", "experimented", "with", "the", "prioritization", "exponents", "from", "to", ",", "with", "results", "proving", "essentially", "consistent", "within", "the", "range", "[", "0.3", ",", "0.7", "]", "(", "beyond", "0.7", ",", "training", "would", "sometimes", "become", "unstable", "and", "diverge", ")", ".", "For", "the", "experiments", "with", "many", "actors", ",", "we", "set", "the", "period", "for", "updating", "network", "parameters", "on", "the", "actors", "to", "be", "high", "enough", "that", "the", "learner", "was", "not", "overloaded", "with", "requests", ",", "and", "we", "set", "the", "number", "of", "transitions", "that", "are", "locally", "accumulated", "on", "each", "actor", "to", "be", "high", "enough", "that", "the", "replay", "server", "would", "not", "be", "overloaded", "with", "network", "traffic", ",", "but", "we", "did", "not", "otherwise", "tune", "those", "parameters", "and", "have", "not", "observed", "them", "to", "have", "significant", "impact", "on", "the", "learning", "dynamics", ".", "appendix", ":", "Implementation", "The", "following", "section", "makes", "explicit", "some", "of", "the", "more", "practical", "details", "that", "may", "be", "of", "interest", "to", "anyone", "wishing", "to", "implement", "a", "similar", "system", ".", "paragraph", ":", "Data", "Storage", "The", "algorithm", "is", "implemented", "using", "TensorFlow", "tensorflow", ".", "Replay", "data", "is", "kept", "in", "a", "distributed", "in", "-", "memory", "key", "-", "value", "store", "implemented", "using", "custom", "TensorFlow", "ops", ",", "similar", "to", "the", "lookup", "ops", "available", "in", "core", "TensorFlow", ".", "The", "ops", "allow", "adding", ",", "reading", ",", "and", "removing", "batches", "of", "Tensor", "data", "efficiently", ".", "paragraph", ":", "Sampling", "Data", "We", "also", "implemented", "ops", "for", "efficiently", "maintaining", "and", "sampling", "from", "a", "prioritized", "distribution", "over", "the", "keys", ",", "using", "the", "algorithm", "for", "proportional", "prioritization", "described", "in", ".", "The", "probability", "of", "sampling", "a", "transition", "is", "where", "is", "the", "priority", "of", "the", "transition", "with", "key", ".", "The", "exponent", "controls", "the", "amount", "of", "prioritization", ",", "and", "when", "uniform", "sampling", "is", "recovered", ".", "The", "proportional", "variant", "sets", "priority", "where", "is", "the", "TD", "error", "for", "transition", ".", "Whenever", "a", "batch", "of", "data", "is", "added", "to", "or", "removed", "from", "the", "store", ",", "or", "is", "processed", "by", "the", "learner", ",", "this", "distribution", "is", "correspondingly", "updated", ",", "recording", "any", "change", "to", "the", "set", "of", "valid", "keys", "and", "the", "priorities", "associated", "with", "them", ".", "A", "background", "thread", "on", "the", "learner", "fetches", "batches", "of", "sampled", "data", "from", "the", "remote", "replay", "and", "decompresses", "it", "using", "the", "learner", "’s", "CPU", ",", "in", "parallel", "with", "the", "gradients", "being", "computed", "on", "the", "GPU", ".", "The", "fetched", "data", "is", "buffered", "in", "a", "TensorFlow", "queue", ",", "so", "that", "the", "GPU", "always", "has", "data", "available", "to", "train", "on", ".", "paragraph", ":", "Adding", "Data", "In", "order", "to", "efficiently", "construct", "-", "step", "transition", "data", ",", "each", "actor", "maintains", "a", "circular", "buffer", "of", "capacity", "containing", "tuples", ",", "where", "is", "the", "current", "size", "of", "the", "buffer", ".", "With", "each", "step", ",", "the", "new", "data", "is", "appended", "and", "the", "accumulated", "per", "-", "step", "discounts", "and", "partial", "returns", "for", "all", "entries", "in", "the", "buffer", "are", "updated", ".", "If", "the", "buffer", "has", "reached", "its", "capacity", ",", ",", "then", "its", "first", "element", "may", "be", "combined", "with", "the", "latest", "state", "and", "value", "estimates", "to", "produce", "a", "valid", "-", "step", "transition", "(", "with", "accompanying", "Q", "-", "values", ")", ".", "However", ",", "instead", "of", "being", "directly", "added", "to", "the", "remote", "replay", "memory", "on", "each", "step", ",", "the", "constructed", "transitions", "are", "first", "stored", "in", "a", "local", "TensorFlow", "queue", ",", "in", "order", "to", "reduce", "the", "number", "of", "requests", "to", "the", "replay", "server", ".", "The", "queue", "is", "periodically", "flushed", ",", "at", "which", "stage", "the", "absolute", "-", "step", "TD", "-", "errors", "(", "and", "thus", "the", "initial", "priorities", ")", "for", "the", "queued", "transitions", "are", "computed", "in", "batch", ",", "using", "the", "buffered", "Q", "-", "values", "to", "avoid", "recomputation", ".", "The", "Q", "-", "value", "estimates", "from", "which", "the", "initial", "priorities", "are", "derived", "are", "therefore", "based", "on", "the", "actor", "’s", "copy", "of", "the", "network", "parameters", "at", "the", "time", "the", "corresponding", "state", "was", "obtained", "from", "the", "environment", ",", "rather", "than", "the", "latest", "version", "on", "the", "learner", ".", "These", "Q", "-", "values", "need", "not", "be", "stored", "after", "this", ",", "since", "the", "learner", "does", "not", "require", "them", ",", "although", "they", "can", "be", "helpful", "for", "debugging", ".", "A", "unique", "key", "is", "assigned", "to", "each", "transition", ",", "which", "records", "which", "actor", "and", "environment", "step", "it", "came", "from", ",", "and", "the", "dequeued", "transition", "tuples", "are", "stored", "in", "the", "remote", "replay", "memory", ".", "As", "mentioned", "in", "the", "previous", "section", ",", "the", "remote", "sampling", "distribution", "is", "immediately", "updated", "with", "the", "newly", "added", "keys", "and", "the", "corresponding", "initial", "priorities", "computed", "by", "the", "actor", ".", "Note", "that", ",", "since", "we", "store", "both", "the", "start", "and", "the", "end", "state", "with", "each", "transition", ",", "we", "are", "storing", "some", "data", "twice", ":", "this", "costs", "more", "RAM", ",", "but", "simplifies", "the", "code", ".", "paragraph", ":", "Contention", "It", "is", "important", "that", "the", "replay", "server", "be", "able", "to", "handle", "all", "requests", "in", "a", "timely", "fashion", ",", "in", "order", "to", "avoid", "slowing", "down", "the", "whole", "system", ".", "Possible", "bottlenecks", "include", "CPU", ",", "network", "bandwidth", ",", "and", "any", "locks", "protecting", "the", "shared", "data", ".", "In", "our", "experiments", "we", "found", "CPU", "to", "be", "the", "main", "bottleneck", ",", "but", "this", "was", "resolved", "by", "ensuring", "all", "requests", "and", "responses", "use", "sufficiently", "large", "batches", ".", "Nonetheless", ",", "it", "is", "advisable", "to", "consider", "all", "of", "these", "potential", "performance", "concerns", "when", "designing", "such", "systems", ".", "paragraph", ":", "Asynchronicity", "In", "our", "framework", ",", "since", "acting", "and", "learning", "proceed", "with", "no", "synchronization", ",", "and", "performance", "depends", "on", "both", ",", "it", "can", "be", "misleading", "to", "consider", "performance", "with", "reference", "to", "only", "one", "of", "these", ".", "For", "example", ",", "the", "results", "after", "a", "given", "total", "number", "of", "environment", "frames", "have", "been", "experienced", "are", "highly", "dependent", "on", "the", "number", "of", "updates", "the", "learner", "has", "performed", "in", "that", "time", ".", "For", "this", "reason", "it", "is", "important", "to", "monitor", "and", "report", "the", "speeds", "of", "all", "parts", "of", "the", "system", "and", "to", "consider", "them", "when", "analyzing", "results", ".", "paragraph", ":", "Failure", "Tolerance", "In", "distributed", "systems", "with", "many", "workers", ",", "it", "is", "inevitable", "that", "interruptions", "or", "failures", "will", "occur", ",", "either", "due", "to", "occasional", "hardware", "issues", "or", "because", "shared", "resources", "are", "needed", "by", "higher", "priority", "jobs", ".", "All", "stateful", "parts", "of", "the", "system", "therefore", "must", "periodically", "save", "their", "work", "and", "be", "able", "to", "resume", "where", "they", "left", "off", "when", "restarted", ".", "In", "our", "system", ",", "actors", "may", "be", "interrupted", "at", "any", "time", "and", "this", "will", "not", "prevent", "continued", "learning", ",", "albeit", "with", "a", "temporarily", "reduced", "rate", "of", "new", "data", "entering", "the", "replay", "memory", ".", "If", "the", "replay", "server", "is", "interrupted", ",", "the", "data", "it", "contains", "is", "discarded", ",", "and", "upon", "resuming", ",", "the", "memory", "is", "refilled", "quickly", "by", "the", "actors", ".", "In", "this", "event", ",", "to", "avoid", "overfitting", ",", "the", "learner", "will", "pause", "training", "briefly", ",", "until", "the", "minimum", "amount", "of", "data", "has", "once", "again", "been", "accumulated", ".", "If", "the", "learner", "is", "interrupted", ",", "progress", "will", "stall", "until", "it", "resumes", ".", "[", "width=0.99", "]", "images", "/", "apex_iclr_full_sweep_oct_24", "[", "width=0.99", "]", "images", "/", "apex_iclr_full_sweep_oct_24_data_efficiency", "[", "height=4cm", "]", "images", "/", "apex_iclr_scalability_of_data_generation", "[", "width=0.95", "]", "images", "/", "apex_iclr_num_actors"], "entity": [[2, 6, ["Distributed", "Prioritized", "Experience", "Replay"], "Task"], [9, 11, ["distributed", "architecture"], "Method"], [12, 15, ["deep", "reinforcement", "learning"], "Task"], [39, 40, ["learning"], "Task"], [57, 60, ["shared", "neural", "network"], "Method"], [82, 84, ["neural", "network"], "Method"], [89, 92, ["prioritized", "experience", "replay"], "Task"], [116, 119, ["Arcade", "Learning", "Environment"], "Material"], [129, 134, ["wall", "-", "clock", "training", "time"], "Metric"], [142, 144, ["deep", "learning"], "Task"], [181, 184, ["deep", "reinforcement", "learning"], "Task"], [221, 226, ["GPU", "Advantage", "Actor", "Critic", "ga3c"], "Method"], [227, 230, ["Distributed", "PPO", "heess"], "Method"], [231, 232, ["dppo"], "Method"], [233, 235, ["AlphaGo", "alphago"], "Method"], [236, 239, ["Deep", "learning", "frameworks"], "Method"], [241, 243, ["TensorFlow", "tensorflow"], "Method"], [244, 246, ["distributed", "training"], "Task"], [248, 253, ["large", "scale", "machine", "learning", "systems"], "Task"], [266, 269, ["deep", "reinforcement", "learning"], "Task"], [276, 278, ["computational", "budget"], "Metric"], [307, 310, ["deep", "reinforcement", "learning"], "Method"], [329, 334, ["distributed", "training", "of", "neural", "networks"], "Task"], [338, 341, ["computation", "of", "gradients"], "Task"], [356, 362, ["generation", "and", "selection", "of", "experience", "data"], "Task"], [377, 380, ["distributing", "gradient", "computation"], "Task"], [397, 400, ["data", "-", "generation"], "Task"], [404, 406, ["distributed", "architecture"], "Method"], [411, 415, ["Deep", "Q", "-", "Networks"], "Method"], [416, 417, ["DQN"], "Method"], [419, 423, ["Deep", "Deterministic", "Policy", "Gradient"], "Method"], [424, 425, ["DDPG"], "Method"], [433, 437, ["Arcade", "Learning", "Environment", "benchmark"], "Material"], [444, 447, ["continuous", "control", "tasks"], "Task"], [459, 461, ["Atari", "games"], "Task"], [467, 471, ["wall", "-", "clock", "time"], "Metric"], [482, 487, ["per", "-", "game", "hyperparameter", "tuning"], "Method"], [499, 500, ["prioritization"], "Task"], [537, 541, ["data", "-", "generating", "policies"], "Method"], [551, 555, ["deep", "reinforcement", "learning", "agents"], "Method"], [560, 562, ["distributed", "framework"], "Method"], [568, 572, ["Distributed", "Stochastic", "Gradient", "Descent"], "Method"], [572, 576, ["Distributed", "stochastic", "gradient", "descent"], "Method"], [580, 582, ["supervised", "learning"], "Task"], [585, 590, ["training", "of", "deep", "neural", "networks"], "Task"], [630, 633, ["deep", "learning", "toolbox"], "Method"], [639, 643, ["distributed", "asynchronous", "parameter", "updates"], "Method"], [644, 647, ["distributed", "data", "generation"], "Method"], [648, 651, ["deep", "reinforcement", "learning"], "Task"], [652, 655, ["Asynchronous", "parameter", "updates"], "Task"], [656, 659, ["parallel", "data", "generation"], "Task"], [682, 687, ["GPU", "Asynchronous", "Actor", "-", "Critic"], "Method"], [692, 697, ["Parallel", "Advantage", "Actor", "-", "Critic"], "Method"], [709, 710, ["GPUs"], "Method"], [713, 716, ["Distributed", "Importance", "Sampling"], "Method"], [722, 725, ["speeding", "up", "training"], "Task"], [728, 730, ["variance", "reduction"], "Method"], [733, 735, ["importance", "sampling"], "Method"], [751, 756, ["neural", "networks", "Hinton2007", "-", "vs"], "Task"], [785, 786, ["convergence"], "Metric"], [815, 817, ["supervised", "learning"], "Task"], [826, 828, ["distributed", "setting"], "Task"], [863, 866, ["Prioritized", "Experience", "Replay"], "Task"], [866, 871, ["Experience", "replay", "experience", "-", "replay"], "Method"], [876, 878, ["reinforcement", "learning"], "Method"], [880, 882, ["data", "efficiency"], "Task"], [889, 893, ["neural", "network", "function", "approximators"], "Method"], [894, 898, ["stochastic", "gradient", "descent", "algorithms"], "Method"], [901, 907, ["Neural", "Fitted", "Q", "-", "Iteration", "nfq"], "Method"], [908, 913, ["Deep", "Q", "-", "Learning", "dqn"], "Method"], [914, 916, ["Experience", "replay"], "Task"], [938, 944, ["Prioritized", "experience", "replay", "prioritized", "-", "replay"], "Method"], [946, 948, ["prioritized", "sweeping"], "Method"], [949, 952, ["prioritized", "-", "sweeping"], "Method"], [955, 960, ["deep", "neural", "network", "function", "approximators"], "Method"], [968, 971, ["importance", "sampling", "techniques"], "Method"], [984, 987, ["biased", "sampling", "procedures"], "Method"], [998, 1000, ["Biased", "sampling"], "Method"], [1005, 1007, ["reinforcement", "learning"], "Task"], [1022, 1025, ["agent", "’s", "policy"], "Method"], [1030, 1033, ["prioritized", "experience", "replay"], "Task"], [1046, 1048, ["UNREAL", "unreal"], "Method"], [1049, 1051, ["DQfD", "dqfd"], "Method"], [1053, 1055, ["Rainbow", "rainbow"], "Method"], [1093, 1097, ["Distributed", "Prioritized", "Experience", "Replay"], "Task"], [1102, 1105, ["prioritized", "experience", "replay"], "Task"], [1107, 1109, ["distributed", "setting"], "Task"], [1119, 1122, ["deep", "reinforcement", "learning"], "Task"], [1141, 1144, ["Ape", "-", "X"], "Method"], [1154, 1158, ["deep", "reinforcement", "learning", "algorithm"], "Method"], [1184, 1185, ["policy"], "Method"], [1188, 1191, ["deep", "neural", "network"], "Method"], [1297, 1299, ["background", "thread"], "Method"], [1323, 1326, ["absolute", "TD", "error"], "Metric"], [1344, 1345, ["Learner"], "Method"], [1376, 1378, ["learning", "rule"], "Method"], [1380, 1384, ["double", "Q", "-", "learning"], "Method"], [1392, 1395, ["absolute", "TD", "error"], "Metric"], [1414, 1415, ["learning"], "Task"], [1431, 1432, ["CPUs"], "Method"], [1439, 1440, ["learner"], "Method"], [1490, 1491, ["gorila"], "Method"], [1548, 1550, ["learning", "algorithm"], "Method"], [1642, 1645, ["Ape", "-", "X"], "Method"], [1696, 1699, ["Low", "latency", "communication"], "Task"], [1705, 1707, ["distributed", "SGD"], "Method"], [1720, 1722, ["learning", "algorithm"], "Method"], [1745, 1747, ["centralized", "replay"], "Method"], [1752, 1753, ["throughput"], "Metric"], [1786, 1789, ["off", "-", "policy"], "Method"], [1802, 1805, ["Ape", "-", "X"], "Method"], [1851, 1853, ["exploration", "problems"], "Task"], [1856, 1859, ["Ape", "-", "X"], "Method"], [1859, 1860, ["DQN"], "Method"], [1871, 1873, ["learning", "algorithms"], "Method"], [1891, 1893, ["Rainbow", "rainbow"], "Method"], [1899, 1904, ["double", "Q", "-", "learning", "doubleq"], "Method"], [1905, 1906, ["deepdoubleqlearning"], "Method"], [1924, 1926, ["learning", "algorithm"], "Method"], [1929, 1932, ["dueling", "network", "architecture"], "Method"], [1935, 1937, ["function", "approximator"], "Method"], [1975, 1977, ["network", "dqn"], "Method"], [2005, 2009, ["Q", "-", "learning", "variants"], "Method"], [2010, 2014, ["off", "-", "policy", "methods"], "Method"], [2037, 2039, ["behaviour", "policy"], "Method"], [2042, 2043, ["exploration"], "Task"], [2047, 2049, ["function", "approximation"], "Method"], [2056, 2060, ["multi", "-", "step", "return"], "Method"], [2062, 2066, ["off", "-", "policy", "correction"], "Method"], [2074, 2076, ["value", "estimation"], "Task"], [2080, 2083, ["Ape", "-", "X"], "Method"], [2083, 2084, ["DQN"], "Method"], [2108, 2110, ["prioritization", "mechanism"], "Method"], [2125, 2128, ["-", "greedy", "policies"], "Method"], [2155, 2156, ["DPG"], "Method"], [2169, 2175, ["continuous", "-", "action", "policy", "gradient", "system"], "Method"], [2183, 2186, ["deterministic", "policy", "gradients"], "Method"], [2191, 2192, ["adhdp"], "Method"], [2193, 2194, ["acd"], "Method"], [2199, 2202, ["continuous", "control", "tasks"], "Task"], [2204, 2207, ["DeepMind", "Control", "Suite"], "Material"], [2210, 2213, ["Ape", "-", "X"], "Method"], [2213, 2214, ["DPG"], "Method"], [2221, 2222, ["DQN"], "Method"], [2225, 2228, ["actor", "’s", "policy"], "Method"], [2235, 2237, ["policy", "network"], "Method"], [2242, 2245, ["Q", "-", "network"], "Method"], [2288, 2291, ["Q", "-", "network"], "Method"], [2293, 2297, ["action", "-", "value", "estimate"], "Method"], [2312, 2316, ["temporal", "-", "difference", "learning"], "Method"], [2325, 2329, ["Q", "-", "network", "loss"], "Method"], [2336, 2338, ["policy", "network"], "Method"], [2348, 2351, ["policy", "gradient", "ascent"], "Method"], [2378, 2380, ["critic", "network"], "Method"], [2401, 2402, ["Atari"], "Task"], [2427, 2428, ["Atari"], "Task"], [2439, 2442, ["reinforcement", "learning", "benchmark"], "Task"], [2446, 2448, ["actor", "machines"], "Method"], [2587, 2589, ["PNG", "codec"], "Method"], [2647, 2650, ["-", "greedy", "policy"], "Method"], [2676, 2680, ["shared", "experience", "replay", "memory"], "Method"], [2771, 2774, ["human", "normalized", "score"], "Metric"], [2782, 2783, ["DQN"], "Method"], [2791, 2792, ["Rainbow"], "Method"], [2794, 2795, ["Gorila"], "Method"], [2810, 2817, ["no", "-", "op", "starts", "testing", "regime", "dqn"], "Method"], [2831, 2833, ["greediest", "actor"], "Method"], [2855, 2858, ["Ape", "-", "X"], "Method"], [2908, 2913, ["median", "human", "-", "normalized", "performance"], "Metric"], [2914, 2917, ["Ape", "-", "X"], "Method"], [2917, 2918, ["DQN"], "Method"], [2943, 2947, ["no", "-", "op", "starts"], "Task"], [2949, 2951, ["human", "starts"], "Task"], [2953, 2958, ["human", "-", "starts", "regime", "gorila"], "Method"], [2963, 2965, ["generalization", "test"], "Task"], [2982, 2985, ["Ape", "-", "X"], "Method"], [3004, 3006, ["Continuous", "Control"], "Method"], [3020, 3023, ["continuous", "control", "tasks"], "Task"], [3026, 3028, ["manipulator", "domain"], "Task"], [3043, 3045, ["humanoid", "domain"], "Task"], [3065, 3066, ["Walking"], "Task"], [3067, 3068, ["Running"], "Task"], [3241, 3244, ["distributional", "value", "functions"], "Method"], [3254, 3257, ["continuous", "control", "tasks"], "Task"], [3284, 3287, ["Ape", "-", "X"], "Method"], [3287, 3288, ["DQN"], "Method"], [3290, 3291, ["Atari"], "Task"], [3353, 3355, ["Atari", "games"], "Task"], [3413, 3415, ["prioritized", "replay"], "Method"], [3453, 3455, ["update", "rule"], "Method"], [3466, 3471, ["deep", "reinforcement", "learning", "failure", "mode"], "Task"], [3529, 3531, ["learning", "algorithm"], "Method"], [3647, 3649, ["learner", "machine"], "Method"], [3687, 3688, ["apex_iclr_vary_replay_size.pdf"], "Method"], [3704, 3706, ["scalability", "analysis"], "Task"], [3721, 3725, ["data", "-", "generating", "policies"], "Method"], [3796, 3798, ["distributed", "framework"], "Method"], [3799, 3801, ["prioritized", "replay"], "Task"], [3802, 3805, ["deep", "reinforcement", "learning"], "Task"], [3819, 3823, ["discrete", "and", "continuous", "tasks"], "Task"], [3828, 3833, ["wall", "-", "clock", "learning", "speed"], "Metric"], [3850, 3851, ["DQN"], "Method"], [3852, 3853, ["DPG"], "Method"], [3863, 3869, ["off", "-", "policy", "reinforcement", "learning", "update"], "Method"], [3885, 3886, ["acer"], "Method"], [3906, 3909, ["Ape", "-", "X"], "Method"], [3936, 3940, ["real", "-", "world", "applications"], "Task"], [3943, 3946, ["robotic", "arm", "farms"], "Task"], [3947, 3951, ["self", "-", "driving", "cars"], "Task"], [3952, 3955, ["online", "recommender", "systems"], "Task"], [3958, 3962, ["multi", "-", "user", "systems"], "Task"], [3979, 3980, ["rl"], "Method"], [4000, 4002, ["function", "approximators"], "Method"], [4032, 4036, ["deep", "reinforcement", "learning", "algorithms"], "Method"], [4049, 4052, ["Ape", "-", "X"], "Method"], [4089, 4091, ["direct", "approaches"], "Method"], [4092, 4093, ["exploration"], "Task"], [4099, 4101, ["synchronous", "agents"], "Task"], [4106, 4108, ["distributed", "systems"], "Method"], [4118, 4122, ["large", "-", "scale", "applications"], "Task"], [4123, 4126, ["deep", "reinforcement", "learning"], "Task"], [4316, 4317, ["learner"], "Method"], [4430, 4431, ["policies"], "Method"], [4616, 4618, ["replay", "server"], "Method"], [4641, 4643, ["Generating", "Policies"], "Method"], [4775, 4777, ["distributed", "framework"], "Method"], [4780, 4781, ["Atari"], "Task"], [4802, 4803, ["DQN"], "Method"], [4808, 4810, ["frame", "stacking"], "Method"], [4837, 4838, ["learning"], "Task"], [4842, 4845, ["Centered", "RMSProp", "optimizer"], "Method"], [4847, 4849, ["learning", "rate"], "Metric"], [4891, 4893, ["loss", "calculation"], "Task"], [4897, 4899, ["online", "network"], "Method"], [4912, 4915, ["Dueling", "DDQN", "agent"], "Method"], [4918, 4920, ["Continuous", "Control"], "Task"], [4924, 4926, ["critic", "network"], "Method"], [4948, 4950, ["actor", "network"], "Method"], [4960, 4962, ["tanh", "activation"], "Method"], [4977, 4979, ["actor", "network"], "Method"], [4987, 4988, ["Training"], "Method"], [4990, 4992, ["Adam", "optimizer"], "Method"], [4995, 4997, ["learning", "rate"], "Metric"], [5005, 5007, ["loss", "calculation"], "Task"], [5011, 5013, ["online", "network"], "Method"], [5026, 5029, ["absolute", "TD", "error"], "Metric"], [5033, 5034, ["critic"], "Method"], [5040, 5041, ["learner"], "Method"], [5042, 5045, ["proportional", "prioritized", "sampling"], "Method"], [5069, 5072, ["proportional", "prioritized", "sampling"], "Method"], [5120, 5124, ["Ornstein", "-", "Uhlenbeck", "process"], "Method"], [5141, 5142, ["Evaluation"], "Task"], [5146, 5149, ["noiseless", "deterministic", "policy"], "Method"], [5150, 5151, ["Hyperparameters"], "Method"], [5155, 5156, ["DQN"], "Method"], [5169, 5170, ["Humanoid"], "Material"], [5174, 5175, ["Manipulator"], "Method"], [5185, 5188, ["MuJoCo", "physics", "simulator"], "Method"], [5191, 5192, ["Humanoid"], "Material"], [5241, 5242, ["Manipulator"], "Method"], [5287, 5288, ["Atari"], "Task"], [5296, 5298, ["learning", "rate"], "Metric"], [5299, 5301, ["batch", "size"], "Metric"], [5346, 5348, ["learning", "rate"], "Metric"], [5368, 5370, ["learning", "rate"], "Metric"], [5385, 5387, ["continuous", "control"], "Task"], [5405, 5407, ["learning", "rates"], "Metric"], [5497, 5499, ["replay", "server"], "Method"], [5526, 5528, ["learning", "dynamics"], "Task"], [5559, 5561, ["Data", "Storage"], "Task"], [5566, 5568, ["TensorFlow", "tensorflow"], "Method"], [5575, 5583, ["distributed", "in", "-", "memory", "key", "-", "value", "store"], "Method"], [5586, 5588, ["TensorFlow", "ops"], "Method"], [5592, 5594, ["lookup", "ops"], "Method"], [5596, 5598, ["core", "TensorFlow"], "Method"], [5600, 5601, ["ops"], "Method"], [5610, 5612, ["Tensor", "data"], "Material"], [5621, 5622, ["ops"], "Method"], [5639, 5641, ["proportional", "prioritization"], "Task"], [5671, 5673, ["uniform", "sampling"], "Method"], [5677, 5679, ["proportional", "variant"], "Method"], [5927, 5929, ["replay", "server"], "Method"], [5940, 5946, ["absolute", "-", "step", "TD", "-", "errors"], "Metric"], [6042, 6043, ["debugging"], "Task"], [6085, 6088, ["remote", "sampling", "distribution"], "Method"], [6148, 6150, ["replay", "server"], "Method"], [6240, 6243, ["acting", "and", "learning"], "Task"], [6329, 6331, ["Failure", "Tolerance"], "Task"], [6332, 6334, ["distributed", "systems"], "Task"], [6406, 6407, ["learning"], "Task"], [6424, 6426, ["replay", "server"], "Method"], [6458, 6459, ["learner"], "Method"], [1043, 1044, ["DQN"], "Method"], [1385, 1386, ["DDPG"], "Method"], [1562, 1563, ["DQN"], "Method"], [1883, 1884, ["DQN"], "Method"], [2152, 2155, ["Ape", "-", "X"], "Method"], [2177, 2178, ["DDPG"], "Method"], [2218, 2221, ["Ape", "-", "X"], "Method"], [2385, 2388, ["Ape", "-", "X"], "Method"], [2388, 2389, ["DPG"], "Method"], [2422, 2425, ["Ape", "-", "X"], "Method"], [2425, 2426, ["DQN"], "Method"], [2785, 2786, ["DQN"], "Method"], [2788, 2789, ["DQN"], "Method"], [2920, 2921, ["Atari"], "Task"], [3014, 3017, ["Ape", "-", "X"], "Method"], [3017, 3018, ["DPG"], "Method"], [3092, 3093, ["Atari"], "Task"], [3166, 3169, ["Ape", "-", "X"], "Method"], [3169, 3170, ["DPG"], "Method"], [3185, 3188, ["Ape", "-", "X"], "Method"], [3188, 3189, ["DPG"], "Method"], [3217, 3218, ["DDPG"], "Method"], [3236, 3239, ["Ape", "-", "X"], "Method"], [3239, 3240, ["DPG"], "Method"], [3604, 3605, ["DQN"], "Method"], [3845, 3848, ["Ape", "-", "X"], "Method"], [3888, 3891, ["Ape", "-", "X"], "Method"], [5088, 5089, ["Atari"], "Task"], [5111, 5112, ["DPG"], "Method"], [5208, 5209, ["Humanoid"], "Material"]]} {"doc_id": "2138a7127429d67746ec78de46d6820fee0e548e", "text": ["document", ":", "Graph2Seq", ":", "Graph", "to", "Sequence", "Learning", "with", "Attention", "-", "Based", "Neural", "Networks", "The", "celebrated", "Sequence", "to", "Sequence", "learning", "(", "Seq2Seq", ")", "technique", "and", "its", "numerous", "variants", "achieve", "excellent", "performance", "on", "many", "tasks", ".", "However", ",", "many", "machine", "learning", "tasks", "have", "inputs", "naturally", "represented", "as", "graphs", ";", "existing", "Seq2Seq", "models", "face", "a", "significant", "challenge", "in", "achieving", "accurate", "conversion", "from", "graph", "form", "to", "the", "appropriate", "sequence", ".", "To", "address", "this", "challenge", ",", "we", "introduce", "a", "novel", "general", "end", "-", "to", "-", "end", "graph", "-", "to", "-", "sequence", "neural", "encoder", "-", "decoder", "model", "that", "maps", "an", "input", "graph", "to", "a", "sequence", "of", "vectors", "and", "uses", "an", "attention", "-", "based", "LSTM", "method", "to", "decode", "the", "target", "sequence", "from", "these", "vectors", ".", "Our", "method", "first", "generates", "the", "node", "and", "graph", "embeddings", "using", "an", "improved", "graph", "-", "based", "neural", "network", "with", "a", "novel", "aggregation", "strategy", "to", "incorporate", "edge", "direction", "information", "in", "the", "node", "embeddings", ".", "We", "further", "introduce", "an", "attention", "mechanism", "that", "aligns", "node", "embeddings", "and", "the", "decoding", "sequence", "to", "better", "cope", "with", "large", "graphs", ".", "Experimental", "results", "on", "bAbI", ",", "Shortest", "Path", ",", "and", "Natural", "Language", "Generation", "tasks", "demonstrate", "that", "our", "model", "achieves", "state", "-", "of", "-", "the", "-", "art", "performance", "and", "significantly", "outperforms", "existing", "graph", "neural", "networks", ",", "Seq2Seq", ",", "and", "Tree2Seq", "models", ";", "using", "the", "proposed", "bi", "-", "directional", "node", "embedding", "aggregation", "strategy", ",", "the", "model", "can", "converge", "rapidly", "to", "the", "optimal", "performance", ".", "section", ":", "Introduction", "The", "celebrated", "Sequence", "to", "Sequence", "learning", "(", "Seq2Seq", ")", "technique", "and", "its", "numerous", "variants", "achieve", "excellent", "performance", "on", "many", "tasks", "such", "as", "Neural", "Machine", "Translation", "bahdanau2014neural", ",", "gehring2017convolutional", ",", "Natural", "Language", "Generation", "(", "NLG", ")", "DBLP", ":", "conf", "/", "acl", "/", "SongPZWG17", "and", "Speech", "Recognition", "zhang2017very", ".", "Most", "of", "the", "proposed", "Seq2Seq", "models", "can", "be", "viewed", "as", "a", "family", "of", "encoder", "-", "decoders", "DBLP", ":", "conf", "/", "nips", "/", "SutskeverVL14", ",", "cho2014learning", ",", "bahdanau2014neural", ",", "where", "an", "encoder", "reads", "and", "encodes", "a", "source", "input", "in", "the", "form", "of", "sequences", "into", "a", "continuous", "vector", "representation", "of", "fixed", "dimension", ",", "and", "a", "decoder", "takes", "the", "encoded", "vectors", "and", "outputs", "a", "target", "sequence", ".", "Many", "other", "enhancements", "including", "Bidirectional", "Recurrent", "Neural", "Networks", "(", "Bi", "-", "RNN", ")", "schuster1997bidirectional", "or", "Bidirectional", "Long", "Short", "-", "Term", "Memory", "Networks", "(", "Bi", "-", "LSTM", ")", "graves2005framewise", "as", "encoder", ",", "and", "attention", "mechanism", "bahdanau2014neural", ",", "luong2015effective", ",", "have", "been", "proposed", "to", "further", "improve", "its", "practical", "performance", "for", "general", "or", "domain", "-", "specific", "applications", ".", "Despite", "their", "flexibility", "and", "expressive", "power", ",", "a", "significant", "limitation", "with", "the", "Seq2Seq", "models", "is", "that", "they", "can", "only", "be", "applied", "to", "problems", "whose", "inputs", "are", "represented", "as", "sequences", ".", "However", ",", "the", "sequences", "are", "probably", "the", "simplest", "structured", "data", ",", "and", "many", "important", "problems", "are", "best", "expressed", "with", "a", "more", "complex", "structure", "such", "as", "graphs", "that", "have", "more", "capacity", "to", "encode", "complicated", "pair", "-", "wise", "relationships", "in", "the", "data", ".", "For", "example", ",", "one", "task", "in", "NLG", "applications", "is", "to", "translate", "a", "graph", "-", "structured", "semantic", "representation", "such", "as", "Abstract", "Meaning", "Representation", "to", "a", "text", "expressing", "its", "meaning", "banarescu2013abstract", ".", "In", "addition", ",", "path", "planning", "for", "a", "mobile", "robot", "hu2004knowledge", "and", "path", "finding", "for", "question", "answering", "in", "bAbI", "task", "li2015gated", "can", "also", "be", "cast", "as", "graph", "-", "to", "-", "sequence", "problems", ".", "On", "the", "other", "hand", ",", "even", "if", "the", "raw", "inputs", "are", "originally", "expressed", "in", "a", "sequence", "form", ",", "it", "can", "still", "benefit", "from", "the", "enhanced", "inputs", "with", "additional", "information", "(", "to", "formulate", "graph", "inputs", ")", ".", "For", "example", ",", "for", "semantic", "parsing", "tasks", "(", "text", "-", "to", "-", "AMR", "or", "text", "-", "to", "-", "SQL", ")", ",", "they", "have", "been", "shown", "better", "performance", "by", "augmenting", "the", "original", "sentence", "sequences", "with", "other", "structural", "information", "such", "as", "dependency", "parsing", "trees", "pust2015parsing", ".", "Intuitively", ",", "the", "ideal", "solution", "for", "graph", "-", "to", "-", "sequence", "tasks", "is", "to", "build", "a", "more", "powerful", "encoder", "which", "is", "able", "to", "learn", "the", "input", "representation", "regardless", "of", "its", "inherent", "structure", ".", "To", "cope", "with", "graph", "-", "to", "-", "sequence", "problems", ",", "a", "simple", "and", "straightforward", "approach", "is", "to", "directly", "convert", "more", "complex", "structured", "graph", "data", "into", "sequences", "iyer2016summarizing", ",", "gomez2016automatic", ",", "liu2017retrosynthetic", ",", "and", "apply", "sequence", "models", "to", "the", "resulting", "sequences", ".", "However", ",", "the", "Seq2Seq", "model", "often", "fails", "to", "perform", "as", "well", "as", "hoped", "on", "these", "problems", ",", "in", "part", "because", "it", "inevitably", "suffers", "significant", "information", "loss", "due", "to", "the", "conversion", "of", "complex", "structured", "data", "into", "a", "sequence", ",", "especially", "when", "the", "input", "data", "is", "naturally", "represented", "as", "graphs", ".", "Recently", ",", "a", "line", "of", "research", "efforts", "have", "been", "devoted", "to", "incorporate", "additional", "information", "by", "extracting", "syntactic", "information", "such", "as", "the", "phrase", "structure", "of", "a", "source", "sentence", "(", "Tree2seq", ")", "eriguchi2016tree", ",", "by", "utilizing", "attention", "mechanisms", "for", "input", "sets", "(", "Set2seq", ")", "vinyals2015order", ",", "and", "by", "encoding", "sentences", "recursively", "as", "trees", "socher2010learning", ",", "tai2015improved", ".", "Although", "these", "methods", "achieve", "promising", "results", "on", "certain", "classes", "of", "problems", ",", "most", "of", "the", "presented", "techniques", "largely", "depend", "on", "the", "underlying", "application", "and", "may", "not", "be", "able", "to", "generalize", "to", "a", "broad", "class", "of", "problems", "in", "a", "general", "way", ".", "To", "address", "this", "issue", ",", "we", "propose", "Graph2Seq", ",", "a", "novel", "general", "attention", "-", "based", "neural", "network", "model", "for", "graph", "-", "to", "-", "sequence", "learning", ".", "The", "Graph2Seq", "model", "follows", "the", "conventional", "encoder", "-", "decoder", "approach", "with", "two", "main", "components", ",", "a", "graph", "encoder", "and", "a", "sequence", "decoder", ".", "The", "proposed", "graph", "encoder", "aims", "to", "learn", "expressive", "node", "embeddings", "and", "then", "to", "reassemble", "them", "into", "the", "corresponding", "graph", "embeddings", ".", "To", "this", "end", ",", "inspired", "by", "a", "recent", "graph", "representation", "learning", "method", "hamilton2017inductive", ",", "we", "propose", "an", "inductive", "graph", "-", "based", "neural", "network", "to", "learn", "node", "embeddings", "from", "node", "attributes", "through", "aggregation", "of", "neighborhood", "information", "for", "directed", "and", "undirected", "graphs", ",", "which", "explores", "two", "distinct", "aggregators", "on", "each", "node", "to", "yield", "two", "representations", "that", "are", "concatenated", "to", "form", "the", "final", "node", "embedding", ".", "In", "addition", ",", "we", "further", "design", "an", "attention", "-", "based", "RNN", "sequence", "decoder", "that", "takes", "the", "graph", "embedding", "as", "its", "initial", "hidden", "state", "and", "outputs", "a", "target", "prediction", "by", "learning", "to", "align", "and", "translate", "jointly", "based", "on", "the", "context", "vectors", "associated", "with", "the", "corresponding", "nodes", "and", "all", "previous", "predictions", ".", "Our", "code", "and", "data", "are", "available", "at", ".", "Graph2Seq", "is", "simple", "yet", "general", "and", "is", "highly", "extensible", "where", "its", "two", "building", "blocks", ",", "graph", "encoder", "and", "sequence", "decoder", ",", "can", "be", "replaced", "by", "other", "models", "such", "as", "Graph", "Convolutional", "(", "Attention", ")", "Networks", "kipf2016semi", ",", "velickovic2017graph", "or", "their", "extensions", "schlichtkrull2017modeling", ",", "and", "LSTM", "hochreiter1997long", ".", "We", "highlight", "three", "main", "contributions", "of", "this", "paper", "as", "follows", ":", "We", "propose", "a", "novel", "general", "attention", "-", "based", "neural", "networks", "model", "to", "elegantly", "address", "graph", "-", "to", "-", "sequence", "learning", "problems", "that", "learns", "a", "mapping", "between", "graph", "-", "structured", "inputs", "to", "sequence", "outputs", ",", "which", "current", "Seq2Seq", "and", "Tree2Seq", "may", "be", "inadequate", "to", "handle", ".", "We", "propose", "a", "novel", "graph", "encoder", "to", "learn", "a", "bi", "-", "directional", "node", "embeddings", "for", "directed", "and", "undirected", "graphs", "with", "node", "attributes", "by", "employing", "various", "aggregation", "strategies", ",", "and", "to", "learn", "graph", "-", "level", "embedding", "by", "exploiting", "two", "different", "graph", "embedding", "techniques", ".", "Equally", "importantly", ",", "we", "present", "an", "attention", "mechanism", "to", "learn", "the", "alignments", "between", "nodes", "and", "sequence", "elements", "to", "better", "cope", "with", "large", "graphs", ".", "Experimental", "results", "show", "that", "our", "model", "achieves", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "three", "recently", "introduced", "graph", "-", "to", "-", "sequence", "tasks", "and", "significantly", "outperforms", "existing", "graph", "neural", "networks", ",", "Seq2Seq", ",", "and", "Tree2Seq", "models", ".", "section", ":", "Related", "Work", "Our", "model", "draws", "inspiration", "from", "the", "research", "fields", "of", "graph", "representation", "learning", ",", "neural", "networks", "on", "graphs", ",", "and", "neural", "encoder", "-", "decoder", "models", ".", "Graph", "Representation", "Learning", ".", "Graph", "representation", "learning", "has", "been", "proven", "extremely", "useful", "for", "a", "broad", "range", "of", "the", "graph", "-", "based", "analysis", "and", "prediction", "tasks", "hamilton2017representation", ",", "goyal2017graph", ".", "The", "main", "goal", "for", "graph", "representation", "learning", "is", "to", "learn", "a", "mapping", "that", "embeds", "nodes", "as", "points", "in", "a", "low", "-", "dimensional", "vector", "space", ".", "These", "representation", "learning", "approaches", "can", "be", "roughly", "categorized", "into", "two", "classes", "including", "matrix", "factorization", "-", "based", "algorithms", "and", "random", "-", "walk", "based", "methods", ".", "A", "line", "of", "research", "learn", "the", "embeddings", "of", "graph", "nodes", "through", "matrix", "factorization", "roweis2000nonlinear", ",", "belkin2002laplacian", ",", "ahmed2013distributed", ",", "cao2015grarep", ",", "ou2016asymmetric", ".", "These", "methods", "directly", "train", "embeddings", "for", "individual", "nodes", "of", "training", "and", "testing", "data", "jointly", "and", "thus", "inherently", "transductive", ".", "Another", "family", "of", "work", "is", "the", "use", "of", "random", "walk", "-", "based", "methods", "to", "learn", "low", "-", "dimensional", "embeddings", "of", "nodes", "by", "exploring", "neighborhood", "information", "for", "a", "single", "large", "-", "scale", "graph", "duran2017leanring", ",", "hamilton2017inductive", ",", "tang2015line", ",", "grover2016node2vec", ",", "perozzi2014deepwalk", ",", "velickovic2017graph", ".", "GraphSAGE", "hamilton2017inductive", "is", "such", "a", "technique", "that", "learns", "node", "embeddings", "through", "aggregation", "from", "a", "node", "local", "neighborhood", "using", "node", "attributes", "or", "degrees", "for", "inductive", "learning", ",", "which", "has", "better", "capability", "to", "generate", "node", "embeddings", "for", "previously", "unseen", "data", ".", "Our", "graph", "encoder", "is", "an", "extension", "to", "GraphSAGE", "with", "two", "major", "distinctions", ".", "First", ",", "we", "non", "-", "trivially", "generalize", "it", "to", "cope", "with", "both", "directed", "and", "undirected", "graphs", "by", "splitting", "original", "node", "into", "forward", "nodes", "(", "a", "node", "directs", "to", ")", "and", "backward", "nodes", "(", "direct", "to", "a", "node", ")", "according", "to", "edge", "direction", "and", "applying", "two", "distinct", "aggregation", "functions", "to", "these", "types", "of", "nodes", ".", "Second", ",", "we", "exploit", "two", "different", "schemes", "(", "pooling", "-", "based", "and", "supernode", "-", "based", ")", "to", "reassemble", "the", "learned", "node", "embeddings", "to", "generate", "graph", "embedding", ",", "which", "is", "not", "studied", "in", "GraphSAGE", ".", "We", "show", "the", "advantages", "of", "our", "graph", "encoder", "over", "GraphSAGE", "in", "our", "experiments", ".", "Neural", "Networks", "on", "Graphs", ".", "Over", "the", "past", "few", "years", ",", "there", "has", "been", "a", "surge", "of", "approaches", "that", "seek", "to", "learn", "the", "representations", "of", "graph", "nodes", ",", "or", "entire", "(", "sub", ")", "graphs", ",", "based", "on", "Graph", "Neural", "Networks", "(", "GNN", ")", "that", "extend", "well", "-", "known", "network", "architectures", "including", "RNN", "and", "CNN", "to", "graph", "data", "gori2005new", ",", "scarselli2009graph", ",", "li2015gated", ",", "bruna2013spectral", ",", "duvenaud2015convolutional", ",", "niepert2016learning", ",", "defferrard2016convolutional", ",", "yang2016revisiting", ",", "kipf2016semi", ",", "chen2018fastgcn", ".", "A", "line", "of", "research", "is", "the", "neural", "networks", "that", "operate", "on", "graphs", "as", "a", "form", "of", "RNN", "gori2005new", ",", "scarselli2009graph", ",", "and", "recently", "extended", "by", "Li", "et", "al", ".", "li2015gated", "by", "introducing", "modern", "practices", "of", "RNN", "(", "using", "of", "GRU", "updates", ")", "in", "the", "original", "GNN", "framework", ".", "Another", "important", "stream", "of", "work", "that", "has", "recently", "drawn", "fast", "increasing", "interest", "is", "graph", "convolutional", "networks", "(", "GCN", ")", "built", "on", "spectral", "graph", "theory", ",", "introduced", "by", "and", "then", "extended", "by", "with", "fast", "localized", "convolution", ".", "Most", "of", "these", "approaches", "can", "not", "scale", "to", "large", "graphs", ",", "which", "is", "improved", "by", "using", "a", "localized", "first", "-", "order", "approximation", "of", "spectral", "graph", "convolution", "kipf2016semi", "and", "further", "equipping", "with", "important", "sampling", "for", "deriving", "a", "fast", "GCN", "chen2018fastgcn", ".", "The", "closely", "relevant", "work", "to", "our", "graph", "encoder", "is", "GCN", "kipf2016semi", ",", "which", "is", "designed", "for", "semi", "-", "supervised", "learning", "in", "transductive", "setting", "that", "requires", "full", "graph", "Laplacian", "to", "be", "given", "during", "training", "and", "is", "typically", "applicable", "to", "a", "single", "large", "undirected", "graph", ".", "An", "extension", "of", "GCN", "can", "be", "shown", "to", "be", "mathematically", "related", "to", "one", "variant", "of", "our", "graph", "encoder", "on", "undirected", "graphs", ".", "We", "compare", "the", "difference", "between", "our", "graph", "encoder", "and", "GCN", "in", "our", "experiments", ".", "Another", "relevant", "work", "is", "gated", "graph", "sequence", "neural", "networks", "(", "GGS", "-", "NNs", ")", "li2015gated", ".", "Although", "it", "is", "also", "designed", "for", "outputting", "a", "sequence", ",", "it", "is", "essentially", "a", "prediction", "model", "that", "learns", "to", "predict", "a", "sequence", "embedded", "in", "graph", "while", "our", "approach", "is", "a", "generative", "model", "that", "learns", "a", "mapping", "between", "graph", "inputs", "and", "sequence", "outputs", ".", "A", "good", "analogy", "that", "can", "be", "drawn", "between", "our", "proposed", "Graph2Seq", "and", "GGS", "-", "NNs", "is", "the", "relationship", "between", "convolutional", "Seq2Seq", "and", "RNN", ".", "Neural", "Encoder", "-", "Decoder", "Models", ".", "One", "of", "the", "most", "successful", "encoder", "-", "decoder", "architectures", "is", "the", "sequence", "to", "sequence", "learning", "DBLP", ":", "conf", "/", "nips", "/", "SutskeverVL14", ",", "cho2014learning", ",", "bahdanau2014neural", ",", "luong2015effective", ",", "gehring2017convolutional", ",", "which", "are", "originally", "proposed", "for", "machine", "translation", ".", "Recently", ",", "the", "classical", "Seq2Seq", "model", "and", "its", "variants", "have", "been", "applied", "to", "several", "applications", "in", "which", "these", "models", "can", "perform", "mappings", "from", "objects", "to", "sequences", ",", "including", "mapping", "from", "an", "image", "to", "a", "sentence", "vinyals2015show", ",", "models", "for", "computation", "map", "from", "problem", "statements", "of", "a", "python", "program", "to", "their", "solutions", "(", "the", "answers", "to", "the", "program", ")", "zaremba2014learning", ",", "the", "traveling", "salesman", "problem", "for", "the", "set", "of", "points", "vinyals2015pointer", "and", "deep", "generative", "model", "for", "molecules", "generation", "from", "existing", "known", "molecules", "in", "drug", "discovery", ".", "It", "is", "easy", "to", "see", "that", "the", "objects", "that", "are", "mapped", "to", "sequences", "in", "the", "listed", "examples", "are", "often", "naturally", "represented", "in", "graphs", "rather", "than", "sequences", ".", "Recently", ",", "many", "research", "efforts", "and", "the", "key", "contributions", "have", "been", "made", "to", "address", "the", "limitations", "of", "Seq2Seq", "when", "dealing", "with", "more", "complex", "data", ",", "that", "leverage", "external", "information", "using", "specialized", "neural", "models", "attached", "to", "underlying", "targeted", "applications", ",", "including", "Tree2Seq", "eriguchi2016tree", ",", "Set2Seq", "vinyals2015order", ",", "Recursive", "Neural", "Networks", "socher2010learning", ",", "and", "Tree", "-", "Structured", "LSTM", "tai2015improved", ".", "Due", "to", "more", "recent", "advances", "in", "graph", "representations", "and", "graph", "convolutional", "networks", ",", "a", "number", "of", "research", "has", "investigated", "to", "utilize", "various", "GNN", "to", "improve", "the", "performance", "over", "the", "Seq2Seq", "models", "in", "the", "domains", "of", "machine", "translation", "and", "graph", "generation", "bastings2017graph", ",", "beck2018graph", ",", "simonovsky2018graphvae", ",", "li2018learning", ".", "There", "are", "several", "distinctions", "between", "these", "work", "and", "ours", ".", "First", ",", "our", "model", "is", "the", "first", "general", "-", "purpose", "encoder", "-", "decoder", "model", "for", "graph", "-", "to", "-", "sequence", "learning", "that", "is", "applicable", "to", "different", "applications", "while", "the", "aforementioned", "research", "has", "to", "utilize", "domain", "-", "specific", "information", ".", "Second", ",", "we", "design", "our", "own", "graph", "embedding", "techniques", "for", "our", "graph", "decoder", "while", "most", "of", "other", "work", "directly", "apply", "existing", "GNN", "to", "their", "problems", ".", "section", ":", "Graph", "-", "to", "-", "Sequence", "Model", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "our", "graph", "-", "to", "-", "sequence", "model", "includes", "a", "graph", "encoder", ",", "a", "sequence", "decoder", ",", "and", "a", "node", "attention", "mechanism", ".", "Following", "the", "conventional", "encoder", "-", "decoder", "architecture", ",", "the", "graph", "encoder", "first", "generates", "node", "embeddings", ",", "and", "then", "constructs", "graph", "embeddings", "based", "on", "the", "learned", "node", "embeddings", ".", "Finally", ",", "the", "sequence", "decoder", "takes", "both", "the", "graph", "embeddings", "and", "node", "embeddings", "as", "input", "and", "employs", "attention", "over", "the", "node", "embeddings", "whilst", "generating", "sequences", ".", "In", "this", "section", ",", "we", "first", "introduce", "the", "node", "-", "embedding", "generation", "algorithm", "which", "derives", "the", "bi", "-", "directional", "node", "embeddings", "by", "aggregating", "information", "from", "both", "forward", "and", "backward", "neighborhoods", "of", "a", "node", "in", "a", "graph", ".", "Upon", "these", "node", "embeddings", ",", "we", "propose", "two", "methods", "for", "generating", "graph", "embeddings", "capturing", "the", "whole", "-", "graph", "information", ".", "subsection", ":", "Node", "Embedding", "Generation", "Inspired", "by", ",", "we", "design", "a", "new", "inductive", "node", "embedding", "algorithm", "that", "generates", "bi", "-", "directional", "node", "embeddings", "by", "aggregating", "information", "from", "a", "node", "local", "forward", "and", "backward", "neighborhood", "within", "hops", "for", "both", "directed", "and", "undirected", "graphs", ".", "In", "order", "to", "make", "it", "more", "clear", ",", "we", "take", "the", "embedding", "generation", "process", "for", "node", "as", "an", "example", "to", "explain", "our", "node", "embedding", "generation", "algorithm", ":", "We", "first", "transform", "node", "’s", "text", "attribute", "to", "a", "feature", "vector", ",", "av", ",", "by", "looking", "up", "the", "embedding", "matrix", "W", ".", "Note", "that", "for", "some", "tasks", "where", "’s", "text", "attribute", "may", "be", "a", "word", "sequence", ",", "one", "neural", "network", "layer", ",", "such", "as", "an", "LSTM", "layer", ",", "could", "be", "additionally", "used", "to", "generate", "av", ".", "We", "categorize", "the", "neighbors", "of", "into", "forward", "neighbors", ",", ",", "and", "backward", "neighbors", ",", ",", "according", "to", "the", "edge", "direction", ".", "In", "particular", ",", "returns", "the", "nodes", "that", "directs", "to", "and", "returns", "the", "nodes", "that", "direct", "to", ";", "We", "aggregate", "the", "forward", "representations", "of", "’s", "forward", "neighbors", "{", "h⊢u", "-", "k1", ",", "}", "into", "a", "single", "vector", ",", "h⁢N⊢", "(", "v", ")", "k", ",", "where", "is", "the", "iteration", "index", ".", "In", "our", "experiments", ",", "we", "find", "that", "the", "aggregator", "choice", ",", ",", "may", "heavily", "affect", "the", "overall", "performance", "and", "we", "will", "discuss", "it", "later", ".", "Notice", "that", "at", "iteration", ",", "this", "aggregator", "only", "uses", "the", "representations", "generated", "at", ".", "The", "initial", "forward", "representation", "of", "each", "node", "is", "its", "feature", "vector", "calculated", "in", "step", "(", "1", ")", ";", "We", "concatenate", "’s", "current", "forward", "representation", ",", "h⊢v", "-", "k1", ",", "with", "the", "newly", "generated", "neighborhood", "vector", ",", "h⁢N⊢", "(", "v", ")", "k", ".", "This", "concatenated", "vector", "is", "fed", "into", "a", "fully", "connected", "layer", "with", "nonlinear", "activation", "function", ",", "which", "updates", "the", "forward", "representation", "of", ",", "h⊢vk", ",", "to", "be", "used", "at", "the", "next", "iteration", ";", "We", "update", "the", "backward", "representation", "of", ",", "h⊣vk", ",", "using", "the", "similar", "procedure", "as", "introduced", "in", "step", "(", "3", ")", "and", "(", "4", ")", "except", "that", "operating", "on", "the", "backward", "representations", "instead", "of", "the", "forward", "representations", ";", "We", "repeat", "steps", "(", "3", ")", "(", "5", ")", "times", ",", "and", "the", "concatenation", "of", "the", "final", "forward", "and", "backward", "representation", "is", "used", "as", "the", "final", "bi", "-", "directional", "representation", "of", ".", "Since", "the", "neighbor", "information", "from", "different", "hops", "may", "have", "different", "impact", "on", "the", "node", "embedding", ",", "we", "learn", "a", "distinct", "aggregator", "at", "each", "iteration", ".", "Aggregator", "Architectures", ".", "Since", "a", "node", "neighbors", "have", "no", "natural", "ordering", ",", "the", "aggregator", "function", "should", "be", "invariant", "to", "permutations", "of", "its", "inputs", ",", "ensuring", "that", "our", "neural", "network", "model", "can", "be", "trained", "and", "applied", "to", "arbitrarily", "ordered", "node", "-", "neighborhood", "feature", "sets", ".", "In", "practice", ",", "we", "examined", "the", "following", "three", "aggregator", "functions", ":", "Mean", "aggregator", ":", "This", "aggregator", "function", "takes", "the", "element", "-", "wise", "mean", "of", "the", "vectors", "in", "{", "h⊢u", "-", "k1", ",", "}", "and", "{", "h⊣u", "-", "k1", ",", "}", ".", "LSTM", "aggregator", ":", "Similar", "to", "hamilton2017inductive", ",", "we", "also", "examined", "a", "more", "complex", "aggregator", "based", "on", "an", "Long", "Short", "Term", "Memory", "(", "LSTM", ")", "architecture", ".", "Note", "that", "LSTMs", "are", "not", "inherently", "symmetric", "since", "they", "process", "their", "inputs", "sequentially", ".", "We", "use", "LSTMs", "to", "operate", "on", "unordered", "sets", "by", "simply", "applying", "them", "to", "a", "single", "random", "permutation", "of", "the", "node", "neighbors", ".", "Pooling", "aggregator", ":", "In", "this", "aggregator", ",", "each", "neighbor", "’s", "vector", "is", "fed", "through", "a", "fully", "-", "connected", "neural", "network", ",", "and", "an", "element", "-", "wise", "max", "-", "pooling", "operation", "is", "applied", ":", "where", "max", "denotes", "the", "element", "-", "wise", "max", "operator", ",", "and", "is", "a", "nonlinear", "activation", "function", ".", "By", "applying", "max", "-", "pooling", ",", "the", "model", "can", "capture", "different", "information", "across", "the", "neighborhood", "set", ".", "subsection", ":", "Graph", "Embedding", "Generation", "Most", "existing", "works", "of", "graph", "convolution", "neural", "networks", "focus", "more", "on", "node", "embeddings", "rather", "than", "graph", "embeddings", "since", "their", "focus", "is", "on", "the", "node", "-", "wise", "classification", "task", ".", "However", ",", "graph", "embeddings", "that", "convey", "the", "entire", "graph", "information", "are", "essential", "to", "the", "downstream", "decoder", ".", "In", "this", "work", ",", "we", "introduce", "two", "approaches", "(", "i.e.", ",", "Pooling", "-", "based", "and", "Node", "-", "based", ")", "to", "generate", "these", "graph", "embeddings", "from", "the", "node", "embeddings", ".", "Pooling", "-", "based", "Graph", "Embedding", ".", "In", "this", "approach", ",", "we", "investigated", "three", "pooling", "techniques", ":", "max", "-", "pooling", ",", "min", "-", "pooling", "and", "average", "-", "pooling", ".", "In", "our", "experiments", ",", "we", "fed", "the", "node", "embeddings", "to", "a", "fully", "-", "connected", "neural", "network", "and", "applied", "each", "pooling", "method", "element", "-", "wise", ".", "We", "found", "no", "significant", "performance", "difference", "across", "the", "three", "different", "pooling", "approaches", ";", "we", "thus", "adopt", "the", "max", "-", "pooling", "method", "as", "our", "default", "pooling", "approach", ".", "Node", "-", "based", "Graph", "Embedding", ".", "In", "this", "approach", ",", "we", "add", "one", "super", "node", ",", ",", "into", "the", "input", "graph", ",", "and", "all", "other", "nodes", "in", "the", "graph", "direct", "to", ".", "We", "use", "the", "aforementioned", "node", "embedding", "generation", "algorithm", "to", "generate", "the", "embedding", "of", "by", "aggregating", "the", "embeddings", "of", "the", "neighbor", "nodes", ".", "The", "embedding", "of", "that", "captures", "the", "information", "of", "all", "nodes", "is", "regarded", "as", "the", "graph", "embedding", ".", "subsection", ":", "Attention", "Based", "Decoder", "The", "sequence", "decoder", "is", "a", "Recurrent", "Neural", "Network", "(", "RNN", ")", "that", "predicts", "the", "next", "token", ",", "given", "all", "the", "previous", "words", ",", "the", "RNN", "hidden", "state", "for", "time", ",", "and", "a", "context", "vector", "that", "directs", "attention", "to", "the", "encoder", "side", ".", "In", "particular", ",", "the", "context", "vector", "depends", "on", "a", "set", "of", "node", "representations", "(", ",", "…", ",", ")", "which", "the", "graph", "encoder", "maps", "the", "input", "graph", "to", ".", "Each", "node", "representation", "contains", "information", "about", "the", "whole", "graph", "with", "a", "strong", "focus", "on", "the", "parts", "surrounding", "the", "-", "th", "node", "of", "the", "input", "graph", ".", "The", "context", "vector", "is", "computed", "as", "a", "weighted", "sum", "of", "these", "node", "representations", "and", "the", "weight", "of", "each", "node", "representation", "is", "computed", "by", ":", "where", "is", "an", "which", "scores", "how", "well", "the", "input", "node", "around", "position", "and", "the", "output", "at", "position", "match", ".", "The", "score", "is", "based", "on", "the", "RNN", "hidden", "state", "and", "the", "-", "th", "node", "representation", "of", "the", "input", "graph", ".", "We", "parameterize", "the", "alignment", "model", "as", "a", "feed", "-", "forward", "neural", "network", "which", "is", "jointly", "trained", "with", "other", "components", "of", "the", "proposed", "system", ".", "Our", "model", "is", "jointly", "trained", "to", "maximize", "the", "conditional", "log", "-", "probability", "of", "the", "correct", "description", "given", "a", "source", "graph", ".", "In", "the", "inference", "phase", ",", "we", "use", "the", "beam", "search", "to", "generate", "a", "sequence", "with", "the", "beam", "size", "=", "5", ".", "section", ":", "Experiments", "We", "conduct", "experiments", "to", "demonstrate", "the", "effectiveness", "and", "efficiency", "of", "the", "proposed", "method", ".", "Following", "the", "experimental", "settings", "in", "li2015gated", ",", "we", "firstly", "compare", "its", "performance", "with", "classical", "LSTM", ",", "GGS", "-", "NN", ",", "and", "GCN", "based", "methods", "on", "two", "selected", "tasks", "including", "bAbI", "Task", "19", "and", "the", "Shortest", "Path", "Task", ".", "We", "then", "compare", "Graph2Seq", "against", "other", "Seq2Seq", "based", "methods", "on", "a", "real", "-", "world", "application", "-", "Natural", "Language", "Generation", "Task", ".", "Note", "that", "the", "parameters", "of", "all", "baselines", "are", "set", "based", "on", "performance", "on", "the", "development", "set", ".", "Experimental", "Settings", ".", "Our", "proposed", "model", "is", "trained", "using", "the", "Adam", "optimizer", "DBLP", ":", "journals", "/", "corr", "/", "KingmaB14", ",", "with", "mini", "-", "batch", "size", "30", ".", "The", "learning", "rate", "is", "set", "to", "0.001", ".", "We", "apply", "the", "dropout", "strategy", "DBLP", ":", "journals", "/", "jmlr", "/", "SrivastavaHKSS14", "with", "a", "ratio", "of", "0.5", "at", "the", "decoder", "layer", "to", "avoid", "overfitting", ".", "Gradients", "are", "clipped", "when", "their", "norm", "is", "bigger", "than", "20", ".", "For", "the", "graph", "encoder", ",", "the", "default", "hop", "size", "is", "set", "to", "6", ",", "the", "size", "of", "node", "initial", "feature", "vector", "is", "set", "to", "40", ",", "the", "non", "-", "linearity", "function", "is", "ReLU", "DBLP", ":", "journals", "/", "jmlr", "/", "GlorotBB11", ",", "the", "parameters", "of", "aggregators", "are", "randomly", "initialized", ".", "The", "decoder", "has", "1", "layer", "and", "hidden", "state", "size", "is", "80", ".", "Since", "Graph2Seq", "with", "mean", "aggregator", "and", "pooling", "-", "based", "graph", "embeddings", "generally", "performs", "better", "than", "other", "configurations", "(", "we", "defer", "this", "discussion", "to", "Sec", ".", "[", "reference", "]", ")", ",", "we", "use", "this", "setting", "as", "our", "default", "model", "in", "the", "following", "sections", ".", "subsection", ":", "bAbI", "Task", "19", "Setup", ".", "The", "bAbI", "artificial", "intelligence", "(", "AI", ")", "tasks", "DBLP", ":", "journals", "/", "corr", "/", "WestonBCM15", "are", "designed", "to", "test", "reasoning", "capabilities", "that", "an", "AI", "system", "possesses", ".", "Among", "these", "tasks", ",", "Task", "19", "(", "Path", "Finding", ")", "is", "arguably", "the", "most", "challenging", "task", "(", "see", ",", "e.g.", ",", "DBLP", ":", "conf", "/", "nips", "/", "SukhbaatarSWF15", "which", "reports", "an", "accuracy", "of", "less", "than", "20", "%", "for", "all", "methods", "that", "do", "not", "use", "strong", "supervision", ")", ".", "We", "apply", "the", "transformation", "procedure", "introduced", "in", "li2015gated", "to", "transform", "the", "description", "as", "a", "graph", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "The", "left", "part", "shows", "an", "instance", "of", "bAbI", "task", "19", ":", "given", "a", "set", "of", "sentences", "describing", "the", "relative", "geographical", "positions", "for", "a", "pair", "of", "objects", "and", ",", "we", "aim", "to", "find", "the", "geographical", "path", "between", "and", ".", "The", "question", "is", "then", "treated", "as", "finding", "the", "shortest", "path", "between", "two", "nodes", ",", "and", ",", "which", "represent", "and", "in", "the", "graph", ".", "To", "tackle", "this", "problem", "with", "Graph2Seq", ",", "we", "annotate", "with", "text", "attribute", "START", "and", "with", "text", "attribute", "END", ".", "For", "other", "nodes", ",", "we", "assign", "their", "IDs", "in", "the", "graph", "as", "their", "text", "attributes", ".", "It", "is", "worth", "noting", "that", ",", "in", "our", "model", ",", "the", "START", "and", "END", "tokens", "are", "node", "features", "whose", "vector", "representations", "are", "first", "randomly", "initialized", "and", "then", "learned", "by", "the", "model", "later", ".", "In", "contrast", ",", "in", "GGS", "-", "NN", ",", "the", "vector", "representations", "of", "staring", "and", "end", "nodes", "are", "set", "as", "one", "-", "hot", "vectors", ",", "which", "is", "specially", "designed", "for", "the", "shortest", "path", "task", ".", "To", "aggregate", "the", "edge", "information", "into", "the", "node", "embedding", ",", "for", "each", "edge", ",", "we", "additionally", "add", "a", "node", "representing", "this", "edge", "into", "the", "graph", "and", "assign", "the", "edge", "’s", "text", "as", "its", "text", "attribute", ".", "We", "generate", "1000", "training", "examples", ",", "1000", "development", "examples", "and", "1000", "test", "examples", "where", "each", "example", "is", "a", "graph", "-", "path", "pair", ".", "We", "use", "a", "standard", "LSTM", "model", "hochreiter1997long", "and", "GGS", "-", "NN", "li2015gated", "as", "our", "baselines", ".", "Since", "GCN", "kipf2016semi", "itself", "can", "not", "output", "a", "sequence", ",", "we", "also", "create", "a", "baseline", "that", "combines", "GCN", "with", "our", "sequence", "decoder", ".", "figurePath", "Finding", "Example", ".", "Results", ".", "From", "Table", "[", "reference", "]", ",", "we", "can", "see", "that", "the", "LSTM", "model", "fails", "on", "this", "task", "while", "our", "model", "makes", "perfect", "predictions", ",", "which", "underlines", "the", "importance", "of", "the", "use", "of", "graph", "encoder", "to", "directly", "encode", "a", "graph", "instead", "of", "using", "sequence", "model", "on", "the", "converted", "inputs", "from", "a", "graph", ".", "Comparing", "to", "GGS", "-", "NN", "that", "uses", "carefully", "designed", "initial", "embeddings", "for", "different", "types", "of", "nodes", "such", "as", "START", "and", "END", ",", "our", "model", "uses", "a", "purely", "end", "-", "to", "-", "end", "approach", "which", "generates", "the", "initial", "node", "feature", "vectors", "based", "on", "random", "initialization", "of", "the", "embeddings", "for", "words", "in", "text", "attributes", ".", "However", ",", "we", "still", "significantly", "outperform", "GGS", "-", "NN", ",", "demonstrating", "the", "expressive", "power", "of", "our", "graph", "encoder", "that", "considers", "information", "flows", "in", "both", "forward", "and", "backward", "directions", ".", "We", "observe", "similar", "results", "when", "comparing", "our", "whole", "Graph2Seq", "model", "to", "GCN", "with", "our", "decoder", ",", "which", "mainly", "because", "the", "current", "form", "of", "GCN", "kipf2016semi", "is", "designed", "for", "undirected", "graph", "and", "thus", "may", "have", "information", "loss", "when", "converting", "directed", "graph", "to", "undirected", "one", "as", "suggested", "in", "kipf2016semi", ".", "subsection", ":", "Shortest", "Path", "Task", "Setup", ".", "We", "further", "evaluate", "our", "model", "on", "the", "Shortest", "Path", "(", "SP", ")", "Task", "whose", "goal", "is", "to", "find", "the", "shortest", "directed", "path", "between", "two", "nodes", "in", "a", "graph", ",", "introduced", "in", "li2015gated", ".", "For", "this", "task", ",", "we", "created", "datasets", "by", "generating", "random", "graphs", ",", "and", "choosing", "pairs", "random", "nodes", "A", "and", "B", "which", "are", "connected", "by", "a", "unique", "shortest", "directed", "path", ".", "Since", "we", "can", "control", "the", "size", "of", "generated", "graphs", ",", "we", "can", "easily", "test", "the", "performance", "changes", "of", "each", "model", "when", "increasing", "the", "size", "of", "graphs", "as", "well", ".", "Two", "such", "datasets", ",", "SP", "-", "S", "and", "SP", "-", "L", ",", "were", "created", ",", "containing", "S", "mall", "(", "node", "size=5", ")", "and", "L", "arge", "graphs", "(", "node", "size=100", ")", ",", "respectively", ".", "We", "restricted", "the", "length", "of", "the", "generated", "shortest", "paths", "for", "SP", "-", "S", "to", "be", "at", "least", "2", "and", "at", "least", "4", "for", "SP", "-", "L.", "For", "each", "dataset", ",", "we", "used", "1000", "training", "examples", "and", "1000", "development", "examples", "for", "parameter", "tuning", ",", "and", "evaluated", "on", "1000", "test", "examples", ".", "We", "choose", "the", "same", "baselines", "as", "introduced", "in", "the", "previous", "section", ".", "Results", ".", "Table", "[", "reference", "]", "shows", "that", "the", "LSTM", "model", "still", "fails", "on", "both", "of", "these", "two", "datasets", ".", "Our", "Graph2Seq", "model", "achieves", "comparable", "performance", "with", "GGS", "-", "NN", "that", "both", "models", "could", "achieve", "100", "%", "accuracy", "on", "the", "SP", "-", "S", "dataset", "while", "achieves", "much", "better", "on", "larger", "graphs", "on", "the", "SP", "-", "L", "dataset", ".", "This", "is", "because", "our", "graph", "encoder", "is", "more", "expressive", "in", "learning", "the", "graph", "structural", "information", "with", "our", "dual", "-", "direction", "aggregators", ",", "which", "is", "the", "key", "to", "maintaining", "good", "performance", "when", "the", "graph", "size", "grows", "larger", ",", "while", "the", "performance", "of", "GGS", "-", "NN", "significantly", "degrades", "due", "to", "hardness", "of", "capturing", "the", "long", "-", "range", "dependence", "in", "a", "graph", "with", "large", "size", ".", "Compared", "to", "GCN", ",", "it", "achieves", "better", "performance", "than", "GGS", "-", "NN", "but", "still", "much", "lower", "than", "our", "Graph2Seq", ",", "in", "part", "because", "of", "both", "the", "poor", "effectiveness", "of", "graph", "encoder", "and", "incapability", "of", "handling", "with", "directed", "graph", ".", "subsection", ":", "Natural", "Language", "Generation", "Task", "Setup", ".", "We", "finally", "evaluate", "our", "model", "on", "a", "real", "-", "world", "application", "-", "Natural", "Language", "Generation", "(", "NLG", ")", "task", "where", "we", "translate", "a", "structured", "semantic", "representation", "—", "in", "this", "case", "a", "structured", "query", "language", "(", "SQL", ")", "query", "—", "to", "a", "natural", "language", "description", "expressing", "its", "meaning", ".", "As", "indicated", "in", "DBLP", ":", "journals", "/", "is", "/", "SpiliopoulouH92", ",", "the", "structure", "of", "SQL", "query", "is", "essentially", "a", "graph", ".", "Thus", "we", "naturally", "cast", "this", "task", "as", "an", "application", "of", "the", "graph", "-", "to", "-", "sequence", "model", "which", "takes", "a", "graph", "representing", "the", "semantic", "structure", "as", "input", "and", "outputs", "a", "sequence", ".", "Figure", "[", "reference", "]", "illustrates", "the", "process", "of", "translation", "of", "an", "SQL", "query", "to", "a", "corresponding", "natural", "language", "description", "via", "our", "Graph2Seq", "model", ".", "We", "use", "the", "BLEU", "-", "4", "score", "to", "evaluate", "our", "model", "on", "the", "WikiSQL", "dataset", "zhongSeq2SQL2017", ",", "a", "corpus", "of", "87", ",", "726", "hand", "-", "annotated", "instances", "of", "natural", "language", "questions", ",", "SQL", "queries", ",", "and", "SQL", "tables", ".", "WikiSQL", "was", "created", "as", "the", "benchmark", "dataset", "for", "the", "table", "-", "based", "question", "answering", "task", "(", "for", "which", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "is", "82.6", "%", "execution", "accuracy", "yu2018typesql", ")", ";", "here", "we", "reverse", "the", "use", "of", "the", "dataset", ",", "treating", "the", "SQL", "query", "as", "the", "input", "and", "having", "the", "goal", "of", "generating", "the", "correct", "English", "question", ".", "These", "WikiSQL", "SQL", "queries", "are", "split", "into", "training", ",", "development", "and", "test", "sets", ",", "which", "contain", "61297", "queries", ",", "9145", "queries", "and", "17284", "queries", ",", "respectively", ".", "Since", "the", "SQL", "-", "to", "-", "Text", "task", "can", "be", "cast", "as", "”", "machine", "translation", "”", "type", "of", "problems", ",", "we", "implemented", "several", "baselines", "to", "address", "this", "task", ".", "The", "first", "one", "is", "an", "attention", "-", "based", "sequence", "-", "to", "-", "sequence", "(", "Seq2Seq", ")", "model", "proposed", "by", "bahdanau2014neural", ";", "the", "second", "one", "additionally", "introduces", "the", "copy", "mechanism", "in", "the", "decoder", "side", "gu2016incorporating", ";", "the", "third", "one", "is", "a", "tree", "-", "to", "-", "sequence", "(", "Tree2Seq", ")", "model", "proposed", "by", "eriguchi2016tree", "as", "our", "baseline", ";", "the", "fourth", "one", "is", "to", "combine", "a", "GCN", "kipf2016semi", "with", "our", "PGE", "graph", "embeddings", "with", "our", "sequence", "decoder", ";", "the", "fifth", "one", "is", "to", "combine", "a", "GGS", "-", "NN", "li2015gated", "with", "our", "sequence", "decoder", ".", "To", "apply", "these", "baselines", ",", "we", "convert", "an", "SQL", "query", "to", "a", "sequence", "or", "a", "tree", "using", "some", "templates", "which", "we", "discuss", "in", "detail", "in", "the", "Appendix", ".", "Results", ".", "From", "Table", "[", "reference", "]", ",", "we", "can", "see", "that", "our", "Graph2Seq", "model", "performs", "significantly", "better", "than", "the", "Seq2Seq", ",", "Tree2Seq", ",", "and", "Graph2Seq", "baselines", ".", "This", "result", "is", "expected", "since", "the", "structure", "of", "SQL", "query", "is", "essentially", "a", "graph", "despite", "its", "expressions", "in", "sequence", "and", "a", "graph", "encoder", "is", "able", "to", "capture", "much", "more", "information", "directly", "in", "graph", ".", "Among", "all", "Graph2Seq", "models", ",", "our", "Graph2Seq", "model", "performed", "best", ",", "in", "part", "due", "to", "a", "more", "effective", "graph", "encoder", ".", "Tree2Seq", "achieves", "better", "performance", "compared", "to", "Seq2Seq", "since", "its", "tree", "-", "based", "encoder", "explicitly", "takes", "the", "syntactic", "structure", "of", "a", "SQL", "query", "into", "consideration", ".", "Two", "variants", "of", "the", "Graph2Seq", "models", "can", "substantially", "outperform", "Tree2Seq", ",", "which", "demonstrates", "that", "a", "general", "graph", "to", "sequence", "model", "that", "is", "independent", "of", "different", "structural", "information", "in", "complex", "data", "is", "very", "useful", ".", "Interestingly", ",", "we", "also", "observe", "that", "Graph2Seq", "-", "PGE", "(", "pooling", "-", "based", "graph", "embedding", ")", "performs", "better", "than", "Graph2Seq", "-", "NGE", "(", "node", "-", "based", "graph", "embedding", ")", ".", "One", "potential", "reason", "is", "that", "the", "node", "-", "based", "graph", "embedding", "method", "artificially", "added", "a", "super", "node", "in", "graph", "which", "changes", "the", "original", "graph", "topology", "and", "brings", "unnecessary", "noise", "into", "the", "graph", ".", "figureA", "running", "example", "of", "the", "NLG", "task", ".", "subsection", ":", "Impacts", "of", "Aggregator", ",", "Hop", "Size", "and", "Attention", "Mechanism", "on", "Garph2Seq", "Model", "Setup", ".", "We", "now", "investigate", "the", "impact", "of", "the", "aggregator", "and", "the", "hop", "size", "on", "the", "Graph2Seq", "model", ".", "Following", "the", "previous", "SP", "task", ",", "we", "further", "create", "three", "synthetic", "datasets", ":", "i", ")", "SDP⁢DAG", "whose", "graphs", "are", "directed", "acyclic", "graphs", "(", "DAGs", ")", ";", "ii", ")", "SDP⁢DCG", "whose", "graphs", "are", "directed", "cyclic", "graphs", "(", "DCGs", ")", "that", "always", "contain", "cycles", ";", "iii", ")", "SDP⁢SEQ", "whose", "graphs", "are", "essentially", "sequential", "lines", ".", "For", "each", "dataset", ",", "we", "randomly", "generated", "10000", "graphs", "with", "the", "graph", "size", "100", "and", "split", "them", "as", "8000", "/", "1000", "/", "1000", "for", "the", "training", "/", "development", "/", "test", "set", ".", "For", "each", "graph", ",", "we", "generated", "an", "SDP", "query", "by", "choosing", "two", "random", "nodes", "with", "the", "constraints", "that", "there", "should", "be", "a", "unique", "shortest", "path", "connecting", "these", "two", "nodes", ",", "and", "that", "its", "length", "should", "be", "at", "least", "4", ".", "We", "create", "six", "variants", "of", "the", "Graph2Seq", "model", "coupling", "with", "different", "aggregation", "strategies", "in", "the", "node", "embedding", "generation", ".", "The", "first", "three", "(", "Graph2Seq", "-", "MA", ",", "-", "LA", ",", "-", "PA", ")", "use", "the", "M", "ean", "A", "ggregator", ",", "L", "STM", "A", "ggregator", "and", "P", "ooling", "A", "ggregator", "to", "aggregate", "node", "neighbor", "information", ",", "respectively", ".", "Unlike", "these", "three", "models", "that", "aggregate", "the", "information", "of", "both", "forward", "and", "backward", "nodes", ",", "the", "other", "two", "models", "(", "Graph2Seq", "-", "MA", "-", "F", ",", "-", "MA", "-", "B", ")", "only", "consider", "one", "-", "way", "information", "aggregating", "the", "information", "from", "the", "forward", "nodes", "or", "the", "information", "from", "the", "backward", "nodes", "with", "the", "mean", "aggregator", ",", "respectively", ".", "We", "use", "the", "path", "accuracy", "to", "evaluate", "these", "models", ".", "The", "hop", "size", "is", "set", "to", "10", ".", "Impacts", "of", "the", "Aggregator", ".", "Table", "[", "reference", "]", "shows", "that", "on", "the", "SDP", "dataset", ",", "both", "Graph2Seq", "-", "MA", "and", "Graph2Seq", "-", "PA", "achieve", "the", "best", "performance", ".", "On", "more", "complicated", "structured", "data", ",", "such", "as", "SDP", "and", "SDP", ",", "Graph2Seq", "-", "MA", "(", "our", "default", "model", ")", "also", "performs", "better", "than", "other", "variants", ".", "We", "can", "also", "see", "that", "Graph2Seq", "-", "MA", "performs", "better", "than", "Graph2Seq", "-", "MA", "-", "F", "and", "Graph2Seq", "-", "MA", "-", "B", "on", "SDP", "and", "SDP", "since", "it", "captures", "more", "information", "from", "both", "directions", "to", "learn", "better", "node", "embeddings", ".", "However", ",", "Graph2Seq", "-", "MA", "-", "F", "and", "Graph2Seq", "-", "MA", "-", "B", "achieve", "comparable", "performance", "to", "Graph2Seq", "-", "MA", "on", "SDP", ".", "This", "is", "because", "in", "almost", "95", "%", "of", "the", "graphs", ",", "90", "%", "of", "the", "nodes", "could", "reach", "each", "other", "by", "traversing", "the", "graph", "for", "a", "given", "hop", "size", ",", "which", "dramatically", "restores", "its", "information", "loss", ".", "figureTest", "Results", "on", "SDP", ".", "Impact", "of", "Hop", "Size", ".", "To", "study", "the", "impact", "of", "the", "hop", "size", ",", "we", "create", "a", "SDP", "dataset", ",", "SDP", "and", "results", "are", "shown", "in", "Figure", "[", "reference", "]", ".", "We", "see", "that", "the", "performance", "of", "all", "variants", "of", "Graph2Seq", "converges", "to", "its", "optimal", "performance", "when", "increasing", "the", "number", "of", "hop", "size", ".", "Specifically", ",", "Graph2Seq", "-", "MA", "achieves", "significantly", "better", "performance", "than", "its", "counterparts", "considering", "only", "one", "direction", "propagation", ",", "especially", "when", "the", "hop", "size", "is", "small", ".", "As", "the", "hop", "size", "increases", ",", "the", "performance", "differences", "diminish", ".", "This", "is", "the", "desired", "property", "since", "Graph2Seq", "-", "MA", "can", "use", "much", "smaller", "hop", "size", "(", "about", "the", "half", ")", "to", "achieve", "the", "same", "performance", "of", "Graph2Seq", "-", "MA", "-", "F", "or", "Graph2Seq", "-", "MA", "-", "B", "with", "a", "larger", "size", ".", "This", "is", "particularly", "useful", "for", "large", "graphs", "where", "increasing", "hop", "size", "may", "need", "considerable", "computing", "resources", "and", "long", "run", "-", "time", ".", "We", "also", "compare", "Graph2Seq", "with", "GCN", ",", "where", "the", "hop", "size", "means", "the", "number", "of", "layers", "in", "the", "settings", "of", "GCN", ".", "Surprisingly", ",", "even", "Graph2Seq", "-", "MA", "-", "F", "or", "Graph2Seq", "-", "MA", "-", "B", "can", "significantly", "outperform", "GCN", "with", "the", "same", "hope", "size", "despite", "its", "rough", "equivalence", "between", "these", "two", "architectures", ".", "It", "again", "illustrates", "the", "importance", "of", "the", "methods", "that", "could", "take", "into", "account", "both", "directed", "and", "undirected", "graphs", ".", "For", "additional", "experimental", "results", "on", "the", "impact", "of", "hop", "size", "for", "graphs", "of", "different", "sizes", ",", "please", "refer", "to", "the", "Table", "4", "in", "Appendix", "C.", "Impact", "of", "Attention", "Mechanism", ".", "To", "investigate", "the", "impact", "of", "attention", "mechanism", "to", "the", "Graph2Seq", "model", ",", "we", "still", "evaluate", "our", "model", "on", "SDP", ",", "SDP", "and", "SDP", "datasets", "but", "without", "considering", "the", "attention", "strategy", ".", "As", "shown", "in", "Table", "4", ",", "we", "find", "that", "the", "attention", "strategy", "significantly", "improves", "the", "performance", "of", "all", "variants", "of", "Graph2Seq", "by", "at", "least", "14.9", "%", ".", "This", "result", "is", "expected", "since", "for", "larger", "graphs", "it", "is", "more", "difficult", "for", "the", "encoder", "to", "compress", "all", "necessary", "information", "into", "a", "fixed", "-", "length", "vector", ";", "as", "intended", ",", "applying", "the", "attention", "mechanism", "in", "decoding", "enabled", "our", "proposed", "Graph2Seq", "model", "to", "successfully", "handle", "large", "graphs", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "study", "the", "graph", "-", "to", "-", "sequence", "problem", ",", "introducing", "a", "new", "general", "and", "flexible", "Graph2Seq", "model", "that", "follows", "the", "encoder", "-", "decoder", "architecture", ".", "We", "showed", "that", ",", "using", "our", "proposed", "bi", "-", "directional", "node", "embedding", "aggregation", "strategy", ",", "the", "graph", "encoder", "could", "successfully", "learn", "representations", "for", "three", "representative", "classes", "of", "directed", "graph", ",", "i.e.", ",", "directed", "acyclic", "graphs", ",", "directed", "cyclic", "graphs", "and", "sequence", "-", "styled", "graphs", ".", "Experimental", "results", "on", "three", "tasks", "demonstrate", "that", "our", "model", "significantly", "outperforms", "existing", "graph", "neural", "networks", ",", "Seq2Seq", ",", "and", "Tree2Seq", "baselines", "on", "both", "synthetic", "and", "real", "application", "datasets", ".", "We", "also", "showed", "that", "introducing", "an", "attention", "mechanism", "over", "node", "representation", "into", "the", "decoding", "substantially", "enhances", "the", "ability", "of", "our", "model", "to", "produce", "correct", "target", "sequences", "from", "large", "graphs", ".", "Since", "much", "symbolic", "data", "is", "represented", "as", "graphs", "and", "many", "tasks", "express", "their", "desired", "outputs", "as", "sequences", ",", "we", "expect", "Graph2Seq", "to", "be", "broadly", "applicable", "to", "unify", "symbolic", "AI", "and", "beyond", ".", "bibliography", ":", "References", "appendix", ":", "Pseudo", "-", "code", "of", "the", "Graph", "-", "to", "-", "sequence", "Algorithm", "[", "h", "]", "Node", "embedding", "generation", "algorithm", "{", "algorithmic}", "[", "1", "]", "Graph", "⁢G", "(", "V", ",", "E", ")", ";", "node", "initial", "feature", "vector", "av", ",", "∈∀vV", ";", "hops", "K", ";", "weight", "matrices", "Wk", ",", "∈∀k{1", ",", "…", ",", "K", "}", ";", "non", "-", "linearity", "σ", ";", "aggregator", "functions", "AGGREGATEk⊢", ",", "AGGREGATEk⊣", ",", "∈∀k{1", ",", "…", ",", "K", "}", ";", "neighborhood", "functions", "N⊢", ",", "N⊣", "Vector", "representations", "zv", "for", "all", "∈vV", "←", "av", ",", "∈∀vV", "←", "av", ",", "∈∀vV", "←", "AGGREGATEk⊢", "(", "{h⊢u", "-", "k1", ",", "∈∀u⁢N⊢", "(", "v", ")", "}", ")", "←", "σ", "(", "W⋅k", "CONCAT", "(", "h⊢v", "-", "k1", ",", "h⁢N⊢", "(", "v", ")", "k", ")", ")", "h⁢N⊣", "(", "v", ")", "k", "←", "AGGREGATEk⊣", "(", "{h⊣u", "-", "k1", ",", "∈∀u⁢N⊣", "(", "v", ")", "}", ")", "←", "σ", "(", "W⋅k", "CONCAT", "(", "h⊣v", "-", "k1", ",", "h⁢N⊣", "(", "v", ")", "k", ")", ")", "←", "CONCAT", "(", "h⊢vK", ",", "h⊣vK", ")", ",", "∈∀vV", "Algorithm", "[", "reference", "]", "describes", "the", "embedding", "generation", "process", "where", "the", "entire", "graph", "and", "initial", "feature", "vectors", "for", "all", "nodes", "av", ",", ",", "are", "provided", "as", "input", ".", "Here", "denotes", "the", "current", "hop", "in", "the", "outer", "loop", ".", "The", "h⊢vk", "denotes", "node", "’s", "forward", "representation", "which", "aggregates", "the", "information", "of", "nodes", "in", ".", "Similarly", ",", "the", "h⊣vk", "denotes", "node", "’s", "backward", "representation", "which", "is", "generated", "by", "aggregating", "the", "information", "of", "nodes", "in", ".", "Each", "step", "in", "the", "outer", "loop", "of", "Algorithm", "[", "reference", "]", "proceeds", "as", "follows", ".", "First", ",", "each", "node", "in", "a", "graph", "aggregates", "the", "forward", "representations", "of", "the", "nodes", "in", "its", "immediate", "neighborhood", ",", "{", "h⊢u", "-", "k1", ",", "}", ",", "into", "a", "single", "vector", ",", "h⁢N⊢", "(", "v", ")", "k", "(", "line", "5", ")", ".", "Note", "that", "this", "aggregation", "step", "depends", "on", "the", "representations", "generated", "at", "the", "previous", "iteration", "of", "the", "outer", "loop", ",", ",", "and", "the", "forward", "representations", "are", "defined", "as", "the", "input", "node", "feature", "vector", ".", "After", "aggregating", "the", "neighboring", "feature", "vectors", ",", "we", "concatenate", "the", "node", "current", "forward", "representation", ",", "h⊢v", "-", "k1", ",", "with", "the", "aggregated", "neighborhood", "vector", ",", "h⁢N⊢", "(", "v", ")", "k", ".", "Then", "this", "concatenated", "vector", "is", "fed", "through", "a", "fully", "connected", "layer", "with", "nonlinear", "activation", "function", ",", "which", "updates", "the", "forward", "representation", "of", "the", "current", "node", "to", "be", "used", "at", "the", "next", "step", "of", "the", "algorithm", "(", "line", "6", ")", ".", "We", "apply", "similar", "process", "to", "generate", "the", "backward", "representations", "of", "the", "nodes", "(", "line", "7", ",", "8", ")", ".", "Finally", ",", "the", "representation", "of", "each", "node", "zv", "is", "the", "concatenation", "of", "the", "forward", "representation", "(", "i.e.", ",", "h⊢vK", ")", "and", "the", "backward", "representation", "(", "i.e.", ",", "h⊣vK", ")", "at", "the", "last", "iteration", ".", "appendix", ":", "Structured", "Representation", "of", "the", "SQL", "Query", "To", "apply", "Graph2Seq", ",", "Seq2Seq", "and", "Tree2Seq", "models", "on", "the", "natural", "language", "generation", "task", ",", "we", "need", "to", "convert", "the", "SQL", "query", "to", "a", "graph", ",", "sequence", "and", "tree", ",", "respectively", ".", "In", "this", "section", ",", "we", "describe", "these", "representations", "of", "the", "SQL", "query", ".", "subsection", ":", "Sequence", "Representation", "We", "apply", "a", "simple", "template", "to", "construct", "the", "SQL", "query", "sequence", ":", "“", "SELECT", "+", "aggregation", "function", ">", "+", "Split", "Symbol", "+", "selected", "column", ">", "+", "WHERE", "+", "condition0", ">", "+", "Split", "Symbol", "+", "condition1", ">", "+", "…", "”", ".", "subsection", ":", "Tree", "Representation", "We", "apply", "the", "SQL", "Parser", "tool", "to", "convert", "an", "SQL", "query", "to", "a", "tree", "which", "is", "illustrated", "in", "Figure", "[", "reference", "]", ".", "Specifically", ",", "the", "root", "of", "this", "tree", "has", "two", "child", "nodes", ",", "namely", "SELECT", "LIST", "and", "WHERE", "CLAUSE", ".", "The", "child", "nodes", "of", "SELECT", "LIST", "node", "are", "the", "selected", "columns", "in", "the", "SQL", "query", ".", "The", "WHERE", "CLAUSE", "node", "has", "all", "occurred", "logical", "operators", "in", "the", "SQL", "query", "as", "its", "children", ".", "The", "children", "of", "a", "logical", "operator", "node", "are", "the", "columns", "on", "which", "this", "operator", "works", ".", "subsection", ":", "Graph", "Representation", "We", "use", "the", "following", "method", "to", "transform", "the", "SQL", "query", "to", "a", "graph", ":", "SELECT", "Clause", ".", "For", "the", "SELECT", "clause", "such", "as", "“", "SELECT", "company", "”", ",", "we", "first", "create", "a", "node", "assigned", "with", "text", "attribute", "select", ".", "This", "SELECT", "node", "connects", "with", "column", "nodes", "whose", "text", "attributes", "are", "the", "selected", "column", "names", "such", "as", "company", ".", "For", "the", "SQL", "queries", "that", "contain", "aggregation", "functions", "such", "as", "count", "or", "max", ",", "we", "add", "one", "aggregation", "node", "which", "is", "connected", "with", "the", "column", "node", "—", "their", "text", "attributes", "are", "the", "aggregation", "function", "names", ".", "WHERE", "Clause", ".", "The", "WHERE", "clause", "usually", "contains", "more", "than", "one", "condition", ".", "For", "each", "condition", ",", "we", "use", "the", "same", "process", "as", "for", "the", "SELECT", "clause", "to", "create", "nodes", ".", "For", "example", ",", "in", "Figure", "[", "reference", "]", ",", "we", "create", "node", "assets", "and", "for", "the", "first", "condition", ",", "the", "node", "sales", "and", "for", "the", "second", "condition", ".", "We", "then", "integrate", "the", "constraint", "nodes", "that", "have", "the", "same", "text", "attribute", "(", "e.g.", ",", "in", "Figure", "[", "reference", "]", ")", ".", "For", "a", "logical", "operator", "such", "as", "AND", ",", "OR", "and", "NOT", ",", "we", "create", "a", "node", "that", "connects", "with", "all", "column", "nodes", "that", "the", "operator", "works", "on", "(", "e.g.", ",", "AND", "in", "Figure", "[", "reference", "]", ")", ".", "These", "logical", "operator", "nodes", "then", "connect", "with", "SELECT", "node", ".", "appendix", ":", "More", "Results", "on", "the", "Impact", "of", "Hop", "Size", "In", "Algorithm", "[", "reference", "]", ",", "we", "can", "see", "that", "there", "are", "three", "key", "factors", "in", "the", "node", "embedding", "generation", ".", "The", "first", "factor", "is", "the", "aggregator", "choice", "which", "determines", "how", "information", "from", "neighborhood", "nodes", "is", "combined", ".", "The", "other", "two", "are", "the", "hop", "size", "(", ")", "and", "the", "neighborhood", "function", "(", ",", ")", ",", "which", "together", "determine", "which", "neighbor", "nodes", "should", "be", "aggregated", "to", "generate", "each", "node", "embedding", ".", "To", "study", "the", "impact", "of", "the", "hop", "size", "in", "our", "model", ",", "we", "create", "two", "SDP", "datasets", ",", "SDP", "and", "SDP", ",", "where", "each", "graph", "has", "100", "nodes", "or", "1000", "nodes", ",", "respectively", ".", "Both", "of", "these", "two", "datasets", "contain", "8000", "training", "examples", ",", "1000", "dev", "examples", "and", "1000", "test", "examples", ".", "We", "evaluated", "three", "models", ",", "Graph2Seq", "-", "MA", "-", "F", ",", "Graph2Seq", "-", "MA", "-", "B", "and", "Graph2Seq", "-", "MA", ",", "on", "these", "two", "datasets", ";", "results", "are", "listed", "in", "Table", "[", "reference", "]", ".", "We", "see", "that", "Graph2Seq", "-", "MA", "-", "F", "and", "Graph2Seq", "-", "MA", "-", "B", "could", "show", "significant", "performance", "improvements", "with", "increasing", "the", "hop", "size", ".", "Specifically", ",", "on", "the", "SDP", "dataset", ",", "Graph2Seq", "-", "MA", "-", "F", "and", "Graph2Seq", "-", "MA", "-", "B", "achieve", "their", "best", "performance", "when", "the", "hop", "size", "reaches", "7", ";", "further", "increases", "do", "not", "improve", "the", "overall", "performance", ".", "A", "similar", "situation", "is", "also", "observed", "on", "the", "SDP", "dataset", ";", "performance", "converges", "at", "the", "hop", "size", "of", "85", ".", "Interestingly", ",", "the", "average", "diameters", "of", "the", "graphs", "in", "the", "two", "datasets", "are", "6.8", "and", "80.2", ",", "respectively", ",", "suggesting", "that", "the", "ideal", "hop", "size", "for", "best", "Graph2Seq", "-", "MA", "-", "F", "performance", "should", "be", "the", "graph", "diameter", ".", "This", "should", "not", "be", "surprising", ";", "if", "the", "hop", "size", "equals", "the", "graph", "diameter", ",", "each", "node", "is", "guaranteed", "to", "aggregate", "the", "information", "of", "all", "reachable", "nodes", "on", "the", "graph", "within", "its", "embedding", ".", "Note", "that", "in", "the", "experiments", "on", "SDP", ",", "in", "the", "(", "¿", "10", ")", "hop", ",", "we", "always", "use", "the", "aggregator", "in", "the", "10", "-", "th", "hop", ",", "because", "introducing", "too", "many", "aggregators", "(", "i.e.", ",", "parameters", ")", "may", "make", "the", "model", "over", "-", "fitting", ".", "Like", "Graph2Seq", "-", "MA", "-", "F", ",", "Graph2Seq", "-", "MA", "also", "benefited", "from", "increasing", "the", "hop", "size", ".", "However", ",", "on", "both", "datasets", ",", "Graph2Seq", "-", "MA", "could", "reach", "peak", "performance", "at", "a", "smaller", "hop", "size", "than", "Graph2Seq", "-", "MA", "-", "F.", "For", "example", ",", "on", "the", "SDP", "dataset", ",", "Graph2Seq", "-", "MA", "achieves", "99.2", "%", "accuracy", "once", "the", "hop", "size", "is", "greater", "than", "4", "while", "Graph2Seq", "-", "MA", "-", "F", "requires", "a", "hop", "size", "greater", "than", "7", "to", "achieve", "comparable", "accuracy", ";", "similar", "observations", "hold", "for", "the", "SDP", "dataset", ".", "Moreover", ",", "we", "can", "see", "that", "the", "minimum", "required", "hop", "size", "that", "Graph2Seq", "-", "MA", "could", "achieve", "its", "best", "performance", "is", "approximately", "the", "average", "radii", "(", "c.f", ".", "diameter", ")", "of", "the", "graphs", ",", "which", "are", "3.4", "and", "40.1", ",", "respectively", ".", "Recall", "that", "the", "main", "difference", "between", "Graph2Seq", "-", "MA", "and", "Graph2Seq", "-", "MA", "-", "F", "(", "or", "Graph2Seq", "-", "MA", "-", "B", ")", "lies", "in", "whether", "the", "system", "aggregates", "information", "propagated", "from", "backward", "nodes", ";", "the", "performance", "difference", "indicates", "that", "by", "incorporating", "forward", "and", "backward", "nodes", "’", "information", ",", "it", "is", "possible", "for", "the", "model", "to", "achieve", "the", "best", "performance", "by", "traversing", "less", "of", "the", "graph", ".", "This", "is", "useful", "in", "practice", ",", "especially", "for", "large", "graphs", "where", "increasing", "hop", "size", "may", "consume", "considerable", "computing", "resources", "and", "run", "-", "time", ".", "Table", "[", "reference", "]", "also", "makes", "clear", "the", "utility", "of", "the", "attention", "strategy", ";", "the", "performance", "of", "both", "Graph2Seq", "-", "MA", "-", "F", "and", "Graph2Seq", "-", "MA", "decreases", "by", "at", "least", "9.8", "%", "on", "SDP", "and", "14.9", "%", "on", "SDP", ".", "This", "result", "is", "expected", ",", "since", "for", "larger", "graphs", "it", "is", "more", "difficult", "for", "the", "encoder", "to", "compress", "all", "necessary", "information", "into", "a", "fixed", "-", "length", "vector", ";", "as", "intended", ",", "applying", "the", "attention", "mechanism", "in", "decoding", "enabled", "our", "proposed", "Graph2Seq", "model", "to", "handle", "large", "graphs", "successfully", ".", "As", "shown", "in", "Algorithm", "[", "reference", "]", ",", "the", "neighborhood", "function", "takes", "a", "given", "node", "as", "input", "and", "returns", "its", "directly", "connected", "neighbor", "nodes", ",", "which", "are", "then", "fed", "to", "the", "node", "embedding", "generator", ".", "Intuitively", ",", "to", "obtain", "a", "better", "representation", "of", "a", "node", ",", "this", "function", "should", "return", "all", "its", "neighbor", "nodes", "in", "the", "graph", ".", "However", ",", "this", "may", "result", "in", "high", "training", "times", "on", "large", "graphs", ".", "To", "address", "this", ",", "hamilton2017inductive", "proposes", "a", "sampling", "method", "which", "randomly", "selects", "a", "fixed", "number", "of", "neighbor", "nodes", "from", "which", "to", "aggregate", "information", "at", "each", "hop", ".", "We", "use", "this", "sampling", "method", "to", "manage", "the", "neighbor", "node", "size", "at", "each", "aggregation", "step", "."], "entity": [[2, 3, ["Graph2Seq"], "Method"], [4, 8, ["Graph", "to", "Sequence", "Learning"], "Task"], [9, 14, ["Attention", "-", "Based", "Neural", "Networks"], "Method"], [16, 20, ["Sequence", "to", "Sequence", "learning"], "Method"], [38, 41, ["machine", "learning", "tasks"], "Task"], [58, 59, ["conversion"], "Task"], [77, 92, ["end", "-", "to", "-", "end", "graph", "-", "to", "-", "sequence", "neural", "encoder", "-", "decoder", "model"], "Method"], [124, 128, ["node", "and", "graph", "embeddings"], "Method"], [131, 136, ["graph", "-", "based", "neural", "network"], "Method"], [139, 141, ["aggregation", "strategy"], "Method"], [155, 157, ["attention", "mechanism"], "Method"], [175, 176, ["bAbI"], "Task"], [177, 179, ["Shortest", "Path"], "Task"], [181, 185, ["Natural", "Language", "Generation", "tasks"], "Task"], [202, 205, ["graph", "neural", "networks"], "Method"], [206, 207, ["Seq2Seq"], "Method"], [209, 211, ["Tree2Seq", "models"], "Method"], [215, 222, ["bi", "-", "directional", "node", "embedding", "aggregation", "strategy"], "Method"], [238, 242, ["Sequence", "to", "Sequence", "learning"], "Method"], [243, 244, ["Seq2Seq"], "Method"], [258, 261, ["Neural", "Machine", "Translation"], "Task"], [265, 268, ["Natural", "Language", "Generation"], "Task"], [269, 270, ["NLG"], "Task"], [271, 272, ["DBLP"], "Method"], [279, 281, ["Speech", "Recognition"], "Task"], [296, 300, ["encoder", "-", "decoders", "DBLP"], "Method"], [313, 314, ["encoder"], "Method"], [327, 330, ["continuous", "vector", "representation"], "Method"], [336, 337, ["decoder"], "Method"], [351, 356, ["Bidirectional", "Recurrent", "Neural", "Networks", "("], "Method"], [356, 359, ["Bi", "-", "RNN"], "Method"], [362, 369, ["Bidirectional", "Long", "Short", "-", "Term", "Memory", "Networks"], "Method"], [376, 377, ["encoder"], "Method"], [379, 381, ["attention", "mechanism"], "Method"], [395, 401, ["general", "or", "domain", "-", "specific", "applications"], "Task"], [485, 490, ["graph", "-", "structured", "semantic", "representation"], "Method"], [492, 495, ["Abstract", "Meaning", "Representation"], "Method"], [506, 508, ["path", "planning"], "Task"], [510, 512, ["mobile", "robot"], "Task"], [514, 516, ["path", "finding"], "Task"], [517, 519, ["question", "answering"], "Task"], [520, 522, ["bAbI", "task"], "Task"], [528, 534, ["graph", "-", "to", "-", "sequence", "problems"], "Task"], [575, 578, ["semantic", "parsing", "tasks"], "Task"], [579, 584, ["text", "-", "to", "-", "AMR"], "Task"], [585, 590, ["text", "-", "to", "-", "SQL"], "Task"], [610, 612, ["dependency", "parsing"], "Method"], [621, 627, ["graph", "-", "to", "-", "sequence", "tasks"], "Task"], [633, 634, ["encoder"], "Method"], [651, 657, ["graph", "-", "to", "-", "sequence", "problems"], "Task"], [682, 684, ["sequence", "models"], "Method"], [772, 774, ["attention", "mechanisms"], "Method"], [841, 842, ["Graph2Seq"], "Method"], [845, 852, ["general", "attention", "-", "based", "neural", "network", "model"], "Method"], [853, 859, ["graph", "-", "to", "-", "sequence", "learning"], "Task"], [861, 863, ["Graph2Seq", "model"], "Method"], [866, 870, ["encoder", "-", "decoder", "approach"], "Method"], [876, 878, ["graph", "encoder"], "Method"], [880, 882, ["sequence", "decoder"], "Method"], [885, 887, ["graph", "encoder"], "Method"], [912, 917, ["graph", "representation", "learning", "method", "hamilton2017inductive"], "Method"], [921, 927, ["inductive", "graph", "-", "based", "neural", "network"], "Method"], [964, 966, ["node", "embedding"], "Method"], [983, 985, ["graph", "embedding"], "Method"], [1025, 1026, ["Graph2Seq"], "Method"], [1040, 1042, ["graph", "encoder"], "Method"], [1043, 1045, ["sequence", "decoder"], "Method"], [1054, 1056, ["Graph", "Convolutional"], "Method"], [1057, 1060, ["Attention", ")", "Networks"], "Method"], [1069, 1070, ["LSTM"], "Method"], [1087, 1094, ["general", "attention", "-", "based", "neural", "networks", "model"], "Method"], [1097, 1104, ["graph", "-", "to", "-", "sequence", "learning", "problems"], "Task"], [1132, 1134, ["graph", "encoder"], "Method"], [1153, 1155, ["aggregation", "strategies"], "Method"], [1159, 1163, ["graph", "-", "level", "embedding"], "Method"], [1167, 1170, ["graph", "embedding", "techniques"], "Method"], [1177, 1179, ["attention", "mechanism"], "Method"], [1214, 1220, ["graph", "-", "to", "-", "sequence", "tasks"], "Task"], [1224, 1227, ["graph", "neural", "networks"], "Method"], [1228, 1229, ["Seq2Seq"], "Method"], [1231, 1233, ["Tree2Seq", "models"], "Method"], [1247, 1250, ["graph", "representation", "learning"], "Task"], [1251, 1253, ["neural", "networks"], "Method"], [1257, 1262, ["neural", "encoder", "-", "decoder", "models"], "Method"], [1263, 1266, ["Graph", "Representation", "Learning"], "Method"], [1267, 1270, ["Graph", "representation", "learning"], "Method"], [1281, 1288, ["graph", "-", "based", "analysis", "and", "prediction", "tasks"], "Task"], [1296, 1299, ["graph", "representation", "learning"], "Task"], [1318, 1321, ["representation", "learning", "approaches"], "Method"], [1329, 1334, ["matrix", "factorization", "-", "based", "algorithms"], "Method"], [1335, 1340, ["random", "-", "walk", "based", "methods"], "Method"], [1347, 1351, ["embeddings", "of", "graph", "nodes"], "Task"], [1352, 1354, ["matrix", "factorization"], "Method"], [1381, 1382, ["transductive"], "Method"], [1391, 1396, ["random", "walk", "-", "based", "methods"], "Method"], [1398, 1404, ["low", "-", "dimensional", "embeddings", "of", "nodes"], "Task"], [1450, 1452, ["inductive", "learning"], "Method"], [1467, 1469, ["graph", "encoder"], "Method"], [1473, 1474, ["GraphSAGE"], "Material"], [1491, 1495, ["directed", "and", "undirected", "graphs"], "Task"], [1525, 1527, ["aggregation", "functions"], "Method"], [1541, 1548, ["pooling", "-", "based", "and", "supernode", "-", "based"], "Method"], [1557, 1559, ["graph", "embedding"], "Task"], [1565, 1566, ["GraphSAGE"], "Material"], [1573, 1575, ["graph", "encoder"], "Method"], [1576, 1577, ["GraphSAGE"], "Material"], [1581, 1583, ["Neural", "Networks"], "Method"], [1584, 1585, ["Graphs"], "Material"], [1618, 1621, ["Graph", "Neural", "Networks"], "Method"], [1622, 1623, ["GNN"], "Method"], [1629, 1631, ["network", "architectures"], "Method"], [1632, 1633, ["RNN"], "Method"], [1634, 1635, ["CNN"], "Method"], [1664, 1666, ["neural", "networks"], "Method"], [1674, 1675, ["RNN"], "Method"], [1693, 1694, ["RNN"], "Method"], [1697, 1699, ["GRU", "updates"], "Method"], [1703, 1705, ["GNN", "framework"], "Method"], [1719, 1722, ["graph", "convolutional", "networks"], "Method"], [1723, 1724, ["GCN"], "Method"], [1727, 1730, ["spectral", "graph", "theory"], "Method"], [1738, 1741, ["fast", "localized", "convolution"], "Method"], [1759, 1768, ["localized", "first", "-", "order", "approximation", "of", "spectral", "graph", "convolution"], "Method"], [1773, 1775, ["important", "sampling"], "Method"], [1778, 1780, ["fast", "GCN"], "Method"], [1788, 1790, ["graph", "encoder"], "Method"], [1791, 1792, ["GCN"], "Method"], [1798, 1802, ["semi", "-", "supervised", "learning"], "Task"], [1803, 1805, ["transductive", "setting"], "Task"], [1829, 1830, ["GCN"], "Method"], [1842, 1844, ["graph", "encoder"], "Method"], [1854, 1856, ["graph", "encoder"], "Method"], [1857, 1858, ["GCN"], "Method"], [1866, 1871, ["gated", "graph", "sequence", "neural", "networks"], "Method"], [1872, 1875, ["GGS", "-", "NNs"], "Method"], [1892, 1894, ["prediction", "model"], "Method"], [1908, 1910, ["generative", "model"], "Method"], [1931, 1932, ["Graph2Seq"], "Method"], [1933, 1936, ["GGS", "-", "NNs"], "Method"], [1943, 1944, ["RNN"], "Method"], [1945, 1950, ["Neural", "Encoder", "-", "Decoder", "Models"], "Method"], [1956, 1960, ["encoder", "-", "decoder", "architectures"], "Method"], [1962, 1967, ["sequence", "to", "sequence", "learning", "DBLP"], "Method"], [1987, 1989, ["machine", "translation"], "Task"], [2029, 2030, ["computation"], "Task"], [2051, 2054, ["traveling", "salesman", "problem"], "Task"], [2061, 2064, ["deep", "generative", "model"], "Method"], [2065, 2067, ["molecules", "generation"], "Task"], [2072, 2074, ["drug", "discovery"], "Task"], [2119, 2120, ["Seq2Seq"], "Method"], [2133, 2135, ["neural", "models"], "Method"], [2142, 2143, ["Tree2Seq"], "Method"], [2145, 2146, ["Set2Seq"], "Method"], [2148, 2151, ["Recursive", "Neural", "Networks"], "Method"], [2154, 2158, ["Tree", "-", "Structured", "LSTM"], "Method"], [2166, 2168, ["graph", "representations"], "Method"], [2169, 2172, ["graph", "convolutional", "networks"], "Method"], [2182, 2183, ["GNN"], "Method"], [2195, 2197, ["machine", "translation"], "Task"], [2198, 2200, ["graph", "generation"], "Task"], [2225, 2232, ["general", "-", "purpose", "encoder", "-", "decoder", "model"], "Method"], [2233, 2239, ["graph", "-", "to", "-", "sequence", "learning"], "Task"], [2263, 2266, ["graph", "embedding", "techniques"], "Method"], [2268, 2270, ["graph", "decoder"], "Method"], [2278, 2279, ["GNN"], "Method"], [2285, 2291, ["Graph", "-", "to", "-", "Sequence", "Model"], "Method"], [2300, 2306, ["graph", "-", "to", "-", "sequence", "model"], "Method"], [2308, 2310, ["graph", "encoder"], "Method"], [2312, 2314, ["sequence", "decoder"], "Method"], [2317, 2320, ["node", "attention", "mechanism"], "Method"], [2324, 2328, ["encoder", "-", "decoder", "architecture"], "Method"], [2330, 2332, ["graph", "encoder"], "Method"], [2352, 2354, ["sequence", "decoder"], "Method"], [2383, 2388, ["node", "-", "embedding", "generation", "algorithm"], "Method"], [2414, 2416, ["node", "embeddings"], "Method"], [2423, 2425, ["graph", "embeddings"], "Task"], [2434, 2437, ["Node", "Embedding", "Generation"], "Task"], [2444, 2448, ["inductive", "node", "embedding", "algorithm"], "Method"], [2486, 2489, ["embedding", "generation", "process"], "Task"], [2497, 2501, ["node", "embedding", "generation", "algorithm"], "Method"], [2520, 2523, ["embedding", "matrix", "W"], "Method"], [2540, 2543, ["neural", "network", "layer"], "Method"], [2659, 2660, ["aggregator"], "Method"], [2669, 2671, ["forward", "representation"], "Method"], [2689, 2691, ["forward", "representation"], "Method"], [2716, 2719, ["fully", "connected", "layer"], "Method"], [2720, 2723, ["nonlinear", "activation", "function"], "Method"], [2727, 2729, ["forward", "representation"], "Method"], [2744, 2746, ["backward", "representation"], "Method"], [2775, 2777, ["forward", "representations"], "Method"], [2795, 2799, ["forward", "and", "backward", "representation"], "Method"], [2804, 2809, ["bi", "-", "directional", "representation", "of"], "Method"], [2823, 2825, ["node", "embedding"], "Task"], [2835, 2837, ["Aggregator", "Architectures"], "Method"], [2862, 2865, ["neural", "network", "model"], "Method"], [2887, 2889, ["aggregator", "functions"], "Method"], [2890, 2892, ["Mean", "aggregator"], "Method"], [2894, 2896, ["aggregator", "function"], "Method"], [2933, 2934, ["aggregator"], "Method"], [2937, 2941, ["Long", "Short", "Term", "Memory"], "Method"], [2942, 2943, ["LSTM"], "Method"], [2948, 2949, ["LSTMs"], "Method"], [2962, 2963, ["LSTMs"], "Method"], [2982, 2984, ["Pooling", "aggregator"], "Method"], [2997, 3002, ["fully", "-", "connected", "neural", "network"], "Method"], [3005, 3012, ["element", "-", "wise", "max", "-", "pooling", "operation"], "Method"], [3016, 3017, ["max"], "Method"], [3019, 3024, ["element", "-", "wise", "max", "operator"], "Method"], [3034, 3037, ["max", "-", "pooling"], "Method"], [3051, 3054, ["Graph", "Embedding", "Generation"], "Task"], [3058, 3062, ["graph", "convolution", "neural", "networks"], "Method"], [3077, 3082, ["node", "-", "wise", "classification", "task"], "Task"], [3097, 3099, ["downstream", "decoder"], "Task"], [3111, 3118, ["Pooling", "-", "based", "and", "Node", "-", "based"], "Method"], [3129, 3134, ["Pooling", "-", "based", "Graph", "Embedding"], "Method"], [3142, 3144, ["pooling", "techniques"], "Method"], [3145, 3148, ["max", "-", "pooling"], "Method"], [3149, 3152, ["min", "-", "pooling"], "Method"], [3153, 3156, ["average", "-", "pooling"], "Method"], [3168, 3173, ["fully", "-", "connected", "neural", "network"], "Method"], [3176, 3178, ["pooling", "method"], "Method"], [3192, 3194, ["pooling", "approaches"], "Method"], [3199, 3203, ["max", "-", "pooling", "method"], "Method"], [3205, 3208, ["default", "pooling", "approach"], "Method"], [3209, 3214, ["Node", "-", "based", "Graph", "Embedding"], "Method"], [3245, 3249, ["node", "embedding", "generation", "algorithm"], "Method"], [3277, 3279, ["graph", "embedding"], "Method"], [3282, 3285, ["Attention", "Based", "Decoder"], "Method"], [3286, 3288, ["sequence", "decoder"], "Method"], [3290, 3293, ["Recurrent", "Neural", "Network"], "Method"], [3294, 3295, ["RNN"], "Method"], [3338, 3340, ["node", "representations"], "Method"], [3347, 3349, ["graph", "encoder"], "Method"], [3356, 3358, ["node", "representation"], "Method"], [3392, 3394, ["node", "representations"], "Method"], [3399, 3401, ["node", "representation"], "Method"], [3435, 3443, ["-", "th", "node", "representation", "of", "the", "input", "graph"], "Method"], [3447, 3449, ["alignment", "model"], "Method"], [3451, 3456, ["feed", "-", "forward", "neural", "network"], "Method"], [3491, 3493, ["inference", "phase"], "Task"], [3497, 3499, ["beam", "search"], "Method"], [3568, 3569, ["Graph2Seq"], "Method"], [3613, 3622, ["Adam", "optimizer", "DBLP", ":", "journals", "/", "corr", "/", "KingmaB14"], "Method"], [3624, 3628, ["mini", "-", "batch", "size"], "Method"], [3631, 3633, ["learning", "rate"], "Metric"], [3641, 3644, ["dropout", "strategy", "DBLP"], "Method"], [3676, 3678, ["graph", "encoder"], "Method"], [3706, 3708, ["ReLU", "DBLP"], "Method"], [3718, 3719, ["aggregators"], "Method"], [3724, 3725, ["decoder"], "Method"], [3736, 3737, ["Graph2Seq"], "Method"], [3738, 3746, ["mean", "aggregator", "and", "pooling", "-", "based", "graph", "embeddings"], "Method"], [3771, 3773, ["default", "model"], "Method"], [3799, 3800, ["WestonBCM15"], "Method"], [3808, 3810, ["AI", "system"], "Method"], [3819, 3821, ["Path", "Finding"], "Task"], [3843, 3844, ["accuracy"], "Metric"], [3863, 3865, ["transformation", "procedure"], "Method"], [3890, 3892, ["bAbI", "task"], "Task"], [3949, 3950, ["Graph2Seq"], "Method"], [3998, 4000, ["vector", "representations"], "Method"], [4016, 4019, ["GGS", "-", "NN"], "Method"], [4021, 4023, ["vector", "representations"], "Method"], [4042, 4045, ["shortest", "path", "task"], "Task"], [4053, 4055, ["node", "embedding"], "Method"], [4109, 4111, ["LSTM", "model"], "Method"], [4113, 4116, ["GGS", "-", "NN"], "Method"], [4122, 4123, ["GCN"], "Method"], [4138, 4139, ["GCN"], "Method"], [4141, 4143, ["sequence", "decoder"], "Method"], [4144, 4147, ["figurePath", "Finding", "Example"], "Task"], [4161, 4163, ["LSTM", "model"], "Method"], [4182, 4184, ["graph", "encoder"], "Method"], [4192, 4194, ["sequence", "model"], "Method"], [4204, 4207, ["GGS", "-", "NN"], "Method"], [4233, 4235, ["end", "approach"], "Method"], [4244, 4246, ["random", "initialization"], "Method"], [4261, 4264, ["GGS", "-", "NN"], "Method"], [4271, 4273, ["graph", "encoder"], "Method"], [4292, 4294, ["Graph2Seq", "model"], "Method"], [4295, 4296, ["GCN"], "Method"], [4346, 4348, ["Shortest", "Path"], "Task"], [4349, 4350, ["SP"], "Task"], [4504, 4506, ["parameter", "tuning"], "Task"], [4535, 4537, ["LSTM", "model"], "Method"], [4547, 4549, ["Graph2Seq", "model"], "Method"], [4553, 4556, ["GGS", "-", "NN"], "Method"], [4563, 4564, ["accuracy"], "Metric"], [4588, 4590, ["graph", "encoder"], "Method"], [4601, 4605, ["dual", "-", "direction", "aggregators"], "Method"], [4625, 4628, ["GGS", "-", "NN"], "Method"], [4649, 4650, ["GCN"], "Method"], [4656, 4659, ["GGS", "-", "NN"], "Method"], [4665, 4666, ["Graph2Seq"], "Method"], [4676, 4678, ["graph", "encoder"], "Method"], [4706, 4709, ["Natural", "Language", "Generation"], "Task"], [4710, 4711, ["NLG"], "Task"], [4717, 4720, ["structured", "semantic", "representation"], "Method"], [4725, 4728, ["structured", "query", "language"], "Method"], [4729, 4732, ["SQL", ")", "query"], "Method"], [4745, 4746, ["DBLP"], "Material"], [4774, 4780, ["graph", "-", "to", "-", "sequence", "model"], "Method"], [4803, 4804, ["translation"], "Task"], [4806, 4808, ["SQL", "query"], "Task"], [4816, 4818, ["Graph2Seq", "model"], "Method"], [4822, 4826, ["BLEU", "-", "4", "score"], "Metric"], [4832, 4834, ["WikiSQL", "dataset"], "Material"], [4858, 4859, ["WikiSQL"], "Material"], [4867, 4873, ["table", "-", "based", "question", "answering", "task"], "Task"], [4888, 4890, ["execution", "accuracy"], "Metric"], [4917, 4919, ["English", "question"], "Material"], [4949, 4955, ["SQL", "-", "to", "-", "Text", "task"], "Task"], [4960, 4962, ["machine", "translation"], "Task"], [4981, 4989, ["attention", "-", "based", "sequence", "-", "to", "-", "sequence"], "Method"], [4990, 4991, ["Seq2Seq"], "Method"], [5003, 5005, ["copy", "mechanism"], "Method"], [5007, 5009, ["decoder", "side"], "Method"], [5016, 5021, ["tree", "-", "to", "-", "sequence"], "Method"], [5022, 5023, ["Tree2Seq"], "Method"], [5043, 5046, ["PGE", "graph", "embeddings"], "Method"], [5048, 5050, ["sequence", "decoder"], "Method"], [5058, 5061, ["GGS", "-", "NN"], "Method"], [5064, 5066, ["sequence", "decoder"], "Method"], [5108, 5110, ["Graph2Seq", "model"], "Method"], [5115, 5116, ["Seq2Seq"], "Method"], [5117, 5118, ["Tree2Seq"], "Method"], [5120, 5122, ["Graph2Seq", "baselines"], "Method"], [5144, 5146, ["graph", "encoder"], "Method"], [5159, 5161, ["Graph2Seq", "models"], "Method"], [5163, 5165, ["Graph2Seq", "model"], "Method"], [5175, 5177, ["graph", "encoder"], "Method"], [5178, 5179, ["Tree2Seq"], "Method"], [5184, 5185, ["Seq2Seq"], "Method"], [5187, 5191, ["tree", "-", "based", "encoder"], "Method"], [5207, 5209, ["Graph2Seq", "models"], "Method"], [5212, 5213, ["Tree2Seq"], "Method"], [5219, 5223, ["graph", "to", "sequence", "model"], "Method"], [5243, 5246, ["Graph2Seq", "-", "PGE"], "Method"], [5247, 5252, ["pooling", "-", "based", "graph", "embedding"], "Method"], [5256, 5259, ["Graph2Seq", "-", "NGE"], "Method"], [5260, 5265, ["node", "-", "based", "graph", "embedding"], "Method"], [5273, 5279, ["node", "-", "based", "graph", "embedding", "method"], "Method"], [5312, 5313, ["Aggregator"], "Method"], [5314, 5316, ["Hop", "Size"], "Metric"], [5317, 5319, ["Attention", "Mechanism"], "Method"], [5320, 5322, ["Garph2Seq", "Model"], "Method"], [5331, 5332, ["aggregator"], "Method"], [5338, 5340, ["Graph2Seq", "model"], "Method"], [5373, 5376, ["directed", "cyclic", "graphs"], "Method"], [5377, 5378, ["DCGs"], "Method"], [5472, 5475, ["Graph2Seq", "model", "coupling"], "Method"], [5477, 5479, ["aggregation", "strategies"], "Method"], [5481, 5484, ["node", "embedding", "generation"], "Task"], [5489, 5492, ["Graph2Seq", "-", "MA"], "Method"], [5501, 5505, ["M", "ean", "A", "ggregator"], "Method"], [5509, 5510, ["ggregator"], "Method"], [5511, 5513, ["P", "ooling"], "Method"], [5514, 5515, ["ggregator"], "Method"], [5543, 5548, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [5576, 5578, ["mean", "aggregator"], "Method"], [5584, 5586, ["path", "accuracy"], "Metric"], [5602, 5603, ["Aggregator"], "Method"], [5616, 5619, ["Graph2Seq", "-", "MA"], "Method"], [5620, 5623, ["Graph2Seq", "-", "PA"], "Method"], [5636, 5637, ["SDP"], "Material"], [5638, 5639, ["SDP"], "Material"], [5640, 5643, ["Graph2Seq", "-", "MA"], "Method"], [5645, 5647, ["default", "model"], "Method"], [5660, 5663, ["Graph2Seq", "-", "MA"], "Method"], [5666, 5671, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [5672, 5677, ["Graph2Seq", "-", "MA", "-", "B"], "Method"], [5678, 5679, ["SDP"], "Material"], [5680, 5681, ["SDP"], "Material"], [5697, 5702, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [5703, 5708, ["Graph2Seq", "-", "MA", "-", "B"], "Method"], [5712, 5715, ["Graph2Seq", "-", "MA"], "Method"], [5716, 5717, ["SDP"], "Material"], [5758, 5759, ["SDP"], "Material"], [5762, 5764, ["Hop", "Size"], "Metric"], [5780, 5781, ["SDP"], "Material"], [5800, 5801, ["Graph2Seq"], "Method"], [5811, 5813, ["hop", "size"], "Metric"], [5816, 5819, ["Graph2Seq", "-", "MA"], "Method"], [5835, 5837, ["hop", "size"], "Metric"], [5857, 5860, ["Graph2Seq", "-", "MA"], "Method"], [5877, 5882, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [5883, 5888, ["Graph2Seq", "-", "MA", "-", "B"], "Method"], [5918, 5919, ["Graph2Seq"], "Method"], [5920, 5921, ["GCN"], "Method"], [5935, 5936, ["GCN"], "Method"], [5940, 5945, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [5946, 5951, ["Graph2Seq", "-", "MA", "-", "B"], "Method"], [5954, 5955, ["GCN"], "Method"], [5996, 5998, ["hop", "size"], "Metric"], [6015, 6017, ["Attention", "Mechanism"], "Method"], [6023, 6025, ["attention", "mechanism"], "Method"], [6027, 6029, ["Graph2Seq", "model"], "Method"], [6036, 6037, ["SDP"], "Material"], [6038, 6039, ["SDP"], "Material"], [6040, 6042, ["SDP", "datasets"], "Material"], [6046, 6048, ["attention", "strategy"], "Method"], [6059, 6061, ["attention", "strategy"], "Method"], [6069, 6070, ["Graph2Seq"], "Method"], [6090, 6091, ["encoder"], "Method"], [6108, 6110, ["attention", "mechanism"], "Method"], [6111, 6112, ["decoding"], "Task"], [6115, 6117, ["Graph2Seq", "model"], "Method"], [6133, 6139, ["graph", "-", "to", "-", "sequence", "problem"], "Task"], [6151, 6155, ["encoder", "-", "decoder", "architecture"], "Method"], [6163, 6170, ["bi", "-", "directional", "node", "embedding", "aggregation", "strategy"], "Method"], [6172, 6174, ["graph", "encoder"], "Method"], [6213, 6216, ["graph", "neural", "networks"], "Method"], [6217, 6218, ["Seq2Seq"], "Method"], [6236, 6238, ["attention", "mechanism"], "Method"], [6239, 6241, ["node", "representation"], "Method"], [6243, 6244, ["decoding"], "Task"], [6280, 6281, ["Graph2Seq"], "Method"], [6287, 6289, ["symbolic", "AI"], "Task"], [6302, 6308, ["Graph", "-", "to", "-", "sequence", "Algorithm"], "Method"], [6311, 6315, ["Node", "embedding", "generation", "algorithm"], "Method"], [6373, 6375, ["Vector", "representations"], "Method"], [6453, 6456, ["CONCAT", "(", "h⊢vK"], "Method"], [6467, 6470, ["embedding", "generation", "process"], "Task"], [6502, 6506, ["node", "’s", "forward", "representation"], "Method"], [6521, 6523, ["backward", "representation"], "Method"], [6593, 6595, ["aggregation", "step"], "Method"], [6606, 6608, ["outer", "loop"], "Method"], [6612, 6614, ["forward", "representations"], "Method"], [6633, 6637, ["node", "current", "forward", "representation"], "Method"], [6638, 6641, ["h⊢v", "-", "k1"], "Method"], [6662, 6665, ["fully", "connected", "layer"], "Method"], [6666, 6669, ["nonlinear", "activation", "function"], "Method"], [6673, 6675, ["forward", "representation"], "Method"], [6726, 6728, ["forward", "representation"], "Method"], [6735, 6737, ["backward", "representation"], "Method"], [6749, 6751, ["Structured", "Representation"], "Method"], [6753, 6755, ["SQL", "Query"], "Method"], [6757, 6758, ["Graph2Seq"], "Method"], [6759, 6760, ["Seq2Seq"], "Method"], [6761, 6763, ["Tree2Seq", "models"], "Method"], [6765, 6769, ["natural", "language", "generation", "task"], "Task"], [6797, 6799, ["SQL", "query"], "Task"], [6802, 6804, ["Sequence", "Representation"], "Method"], [6846, 6848, ["Tree", "Representation"], "Method"], [6851, 6854, ["SQL", "Parser", "tool"], "Method"], [6941, 6943, ["Graph", "Representation"], "Method"], [7174, 7176, ["Hop", "Size"], "Metric"], [7193, 7196, ["node", "embedding", "generation"], "Task"], [7261, 7263, ["SDP", "datasets"], "Material"], [7264, 7265, ["SDP"], "Material"], [7266, 7267, ["SDP"], "Material"], [7303, 7308, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [7309, 7314, ["Graph2Seq", "-", "MA", "-", "B"], "Method"], [7315, 7318, ["Graph2Seq", "-", "MA"], "Method"], [7336, 7341, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [7342, 7347, ["Graph2Seq", "-", "MA", "-", "B"], "Method"], [7355, 7357, ["hop", "size"], "Metric"], [7365, 7370, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [7371, 7376, ["Graph2Seq", "-", "MA", "-", "B"], "Method"], [7382, 7384, ["hop", "size"], "Metric"], [7411, 7413, ["hop", "size"], "Metric"], [7439, 7441, ["hop", "size"], "Metric"], [7443, 7448, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [7495, 7496, ["SDP"], "Material"], [7536, 7541, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [7542, 7545, ["Graph2Seq", "-", "MA"], "Method"], [7550, 7552, ["hop", "size"], "Metric"], [7559, 7562, ["Graph2Seq", "-", "MA"], "Method"], [7569, 7571, ["hop", "size"], "Metric"], [7572, 7577, ["Graph2Seq", "-", "MA", "-", "F."], "Method"], [7585, 7588, ["Graph2Seq", "-", "MA"], "Method"], [7591, 7592, ["accuracy"], "Metric"], [7594, 7596, ["hop", "size"], "Metric"], [7601, 7606, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [7616, 7617, ["accuracy"], "Metric"], [7633, 7637, ["minimum", "required", "hop", "size"], "Metric"], [7638, 7641, ["Graph2Seq", "-", "MA"], "Method"], [7649, 7651, ["average", "radii"], "Metric"], [7654, 7655, ["diameter"], "Metric"], [7674, 7677, ["Graph2Seq", "-", "MA"], "Method"], [7678, 7683, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [7685, 7690, ["Graph2Seq", "-", "MA", "-", "B"], "Method"], [7755, 7758, ["run", "-", "time"], "Metric"], [7770, 7772, ["attention", "strategy"], "Method"], [7777, 7782, ["Graph2Seq", "-", "MA", "-", "F"], "Method"], [7783, 7786, ["Graph2Seq", "-", "MA"], "Method"], [7793, 7794, ["SDP"], "Material"], [7798, 7799, ["SDP"], "Material"], [7815, 7816, ["encoder"], "Method"], [7833, 7835, ["attention", "mechanism"], "Method"], [7836, 7837, ["decoding"], "Task"], [7840, 7842, ["Graph2Seq", "model"], "Method"], [7844, 7846, ["large", "graphs"], "Task"], [7857, 7859, ["neighborhood", "function"], "Method"], [7879, 7882, ["node", "embedding", "generator"], "Method"], [7913, 7915, ["training", "times"], "Metric"], [7923, 7924, ["hamilton2017inductive"], "Method"], [7926, 7928, ["sampling", "method"], "Method"], [7949, 7951, ["sampling", "method"], "Method"], [21, 22, ["Seq2Seq"], "Method"], [49, 50, ["Seq2Seq"], "Method"], [108, 109, ["LSTM"], "Method"], [287, 288, ["Seq2Seq"], "Method"], [372, 373, ["LSTM"], "Method"], [414, 415, ["Seq2Seq"], "Method"], [479, 480, ["NLG"], "Task"], [692, 693, ["Seq2Seq"], "Method"], [977, 978, ["RNN"], "Method"], [1119, 1120, ["Seq2Seq"], "Method"], [1121, 1122, ["Tree2Seq"], "Method"], [1427, 1428, ["GraphSAGE"], "Material"], [1941, 1942, ["Seq2Seq"], "Method"], [1994, 1995, ["Seq2Seq"], "Method"], [2189, 2190, ["Seq2Seq"], "Method"], [2547, 2548, ["LSTM"], "Method"], [2920, 2921, ["LSTM"], "Method"], [3309, 3310, ["RNN"], "Method"], [3430, 3431, ["RNN"], "Method"], [3541, 3542, ["LSTM"], "Method"], [3543, 3546, ["GGS", "-", "NN"], "Method"], [3548, 3549, ["GCN"], "Method"], [3556, 3557, ["bAbI"], "Task"], [3561, 3563, ["Shortest", "Path"], "Task"], [3571, 3572, ["Seq2Seq"], "Method"], [3581, 3584, ["Natural", "Language", "Generation"], "Task"], [3780, 3781, ["bAbI"], "Task"], [3786, 3787, ["bAbI"], "Task"], [4307, 4308, ["GCN"], "Method"], [4334, 4336, ["Shortest", "Path"], "Task"], [4435, 4436, ["SP"], "Task"], [4439, 4440, ["SP"], "Task"], [4474, 4475, ["SP"], "Task"], [4487, 4488, ["SP"], "Task"], [4566, 4567, ["SP"], "Task"], [4579, 4580, ["SP"], "Task"], [4688, 4691, ["Natural", "Language", "Generation"], "Task"], [4921, 4922, ["WikiSQL"], "Material"], [5039, 5040, ["GCN"], "Method"], [5305, 5306, ["NLG"], "Task"], [5344, 5345, ["SP"], "Task"], [5433, 5434, ["SDP"], "Material"], [5612, 5613, ["SDP"], "Material"], [5777, 5778, ["SDP"], "Material"], [6146, 6147, ["Graph2Seq"], "Method"], [6220, 6221, ["Tree2Seq"], "Method"], [7362, 7363, ["SDP"], "Material"], [7404, 7405, ["SDP"], "Material"], [7582, 7583, ["SDP"], "Material"], [7623, 7624, ["SDP"], "Material"]]} {"doc_id": "21a1654b856cf0c64e60e58258669b374cb05539", "text": ["document", ":", "You", "Only", "Look", "Once", ":", "Unified", ",", "Real", "-", "Time", "Object", "Detection", "We", "present", "YOLO", ",", "a", "new", "approach", "to", "object", "detection", ".", "Prior", "work", "on", "object", "detection", "repurposes", "classifiers", "to", "perform", "detection", ".", "Instead", ",", "we", "frame", "object", "detection", "as", "a", "regression", "problem", "to", "spatially", "separated", "bounding", "boxes", "and", "associated", "class", "probabilities", ".", "A", "single", "neural", "network", "predicts", "bounding", "boxes", "and", "class", "probabilities", "directly", "from", "full", "images", "in", "one", "evaluation", ".", "Since", "the", "whole", "detection", "pipeline", "is", "a", "single", "network", ",", "it", "can", "be", "optimized", "end", "-", "to", "-", "end", "directly", "on", "detection", "performance", ".", "Our", "unified", "architecture", "is", "extremely", "fast", ".", "Our", "base", "YOLO", "model", "processes", "images", "in", "real", "-", "time", "at", "45", "frames", "per", "second", ".", "A", "smaller", "version", "of", "the", "network", ",", "Fast", "YOLO", ",", "processes", "an", "astounding", "155", "frames", "per", "second", "while", "still", "achieving", "double", "the", "mAP", "of", "other", "real", "-", "time", "detectors", ".", "Compared", "to", "state", "-", "of", "-", "the", "-", "art", "detection", "systems", ",", "YOLO", "makes", "more", "localization", "errors", "but", "is", "less", "likely", "to", "predict", "false", "positives", "on", "background", ".", "Finally", ",", "YOLO", "learns", "very", "general", "representations", "of", "objects", ".", "It", "outperforms", "other", "detection", "methods", ",", "including", "DPM", "and", "R", "-", "CNN", ",", "when", "generalizing", "from", "natural", "images", "to", "other", "domains", "like", "artwork", ".", "section", ":", "Introduction", "Humans", "glance", "at", "an", "image", "and", "instantly", "know", "what", "objects", "are", "in", "the", "image", ",", "where", "they", "are", ",", "and", "how", "they", "interact", ".", "The", "human", "visual", "system", "is", "fast", "and", "accurate", ",", "allowing", "us", "to", "perform", "complex", "tasks", "like", "driving", "with", "little", "conscious", "thought", ".", "Fast", ",", "accurate", "algorithms", "for", "object", "detection", "would", "allow", "computers", "to", "drive", "cars", "without", "specialized", "sensors", ",", "enable", "assistive", "devices", "to", "convey", "real", "-", "time", "scene", "information", "to", "human", "users", ",", "and", "unlock", "the", "potential", "for", "general", "purpose", ",", "responsive", "robotic", "systems", ".", "Current", "detection", "systems", "repurpose", "classifiers", "to", "perform", "detection", ".", "To", "detect", "an", "object", ",", "these", "systems", "take", "a", "classifier", "for", "that", "object", "and", "evaluate", "it", "at", "various", "locations", "and", "scales", "in", "a", "test", "image", ".", "Systems", "like", "deformable", "parts", "models", "(", "DPM", ")", "use", "a", "sliding", "window", "approach", "where", "the", "classifier", "is", "run", "at", "evenly", "spaced", "locations", "over", "the", "entire", "image", ".", "More", "recent", "approaches", "like", "R", "-", "CNN", "use", "region", "proposal", "methods", "to", "first", "generate", "potential", "bounding", "boxes", "in", "an", "image", "and", "then", "run", "a", "classifier", "on", "these", "proposed", "boxes", ".", "After", "classification", ",", "post", "-", "processing", "is", "used", "to", "refine", "the", "bounding", "boxes", ",", "eliminate", "duplicate", "detections", ",", "and", "rescore", "the", "boxes", "based", "on", "other", "objects", "in", "the", "scene", ".", "These", "complex", "pipelines", "are", "slow", "and", "hard", "to", "optimize", "because", "each", "individual", "component", "must", "be", "trained", "separately", ".", "We", "reframe", "object", "detection", "as", "a", "single", "regression", "problem", ",", "straight", "from", "image", "pixels", "to", "bounding", "box", "coordinates", "and", "class", "probabilities", ".", "Using", "our", "system", ",", "you", "only", "look", "once", "(", "YOLO", ")", "at", "an", "image", "to", "predict", "what", "objects", "are", "present", "and", "where", "they", "are", ".", "YOLO", "is", "refreshingly", "simple", ":", "see", "Figure", "[", "reference", "]", ".", "A", "single", "convolutional", "network", "simultaneously", "predicts", "multiple", "bounding", "boxes", "and", "class", "probabilities", "for", "those", "boxes", ".", "YOLO", "trains", "on", "full", "images", "and", "directly", "optimizes", "detection", "performance", ".", "This", "unified", "model", "has", "several", "benefits", "over", "traditional", "methods", "of", "object", "detection", ".", "First", ",", "YOLO", "is", "extremely", "fast", ".", "Since", "we", "frame", "detection", "as", "a", "regression", "problem", "we", "do", "n’t", "need", "a", "complex", "pipeline", ".", "We", "simply", "run", "our", "neural", "network", "on", "a", "new", "image", "at", "test", "time", "to", "predict", "detections", ".", "Our", "base", "network", "runs", "at", "45", "frames", "per", "second", "with", "no", "batch", "processing", "on", "a", "Titan", "X", "GPU", "and", "a", "fast", "version", "runs", "at", "more", "than", "150", "fps", ".", "This", "means", "we", "can", "process", "streaming", "video", "in", "real", "-", "time", "with", "less", "than", "25", "milliseconds", "of", "latency", ".", "Furthermore", ",", "YOLO", "achieves", "more", "than", "twice", "the", "mean", "average", "precision", "of", "other", "real", "-", "time", "systems", ".", "For", "a", "demo", "of", "our", "system", "running", "in", "real", "-", "time", "on", "a", "webcam", "please", "see", "our", "project", "webpage", ":", ".", "Second", ",", "YOLO", "reasons", "globally", "about", "the", "image", "when", "making", "predictions", ".", "Unlike", "sliding", "window", "and", "region", "proposal", "-", "based", "techniques", ",", "YOLO", "sees", "the", "entire", "image", "during", "training", "and", "test", "time", "so", "it", "implicitly", "encodes", "contextual", "information", "about", "classes", "as", "well", "as", "their", "appearance", ".", "Fast", "R", "-", "CNN", ",", "a", "top", "detection", "method", ",", "mistakes", "background", "patches", "in", "an", "image", "for", "objects", "because", "it", "ca", "n’t", "see", "the", "larger", "context", ".", "YOLO", "makes", "less", "than", "half", "the", "number", "of", "background", "errors", "compared", "to", "Fast", "R", "-", "CNN", ".", "Third", ",", "YOLO", "learns", "generalizable", "representations", "of", "objects", ".", "When", "trained", "on", "natural", "images", "and", "tested", "on", "artwork", ",", "YOLO", "outperforms", "top", "detection", "methods", "like", "DPM", "and", "R", "-", "CNN", "by", "a", "wide", "margin", ".", "Since", "YOLO", "is", "highly", "generalizable", "it", "is", "less", "likely", "to", "break", "down", "when", "applied", "to", "new", "domains", "or", "unexpected", "inputs", ".", "YOLO", "still", "lags", "behind", "state", "-", "of", "-", "the", "-", "art", "detection", "systems", "in", "accuracy", ".", "While", "it", "can", "quickly", "identify", "objects", "in", "images", "it", "struggles", "to", "precisely", "localize", "some", "objects", ",", "especially", "small", "ones", ".", "We", "examine", "these", "tradeoffs", "further", "in", "our", "experiments", ".", "All", "of", "our", "training", "and", "testing", "code", "is", "open", "source", ".", "A", "variety", "of", "pretrained", "models", "are", "also", "available", "to", "download", ".", "section", ":", "Unified", "Detection", "We", "unify", "the", "separate", "components", "of", "object", "detection", "into", "a", "single", "neural", "network", ".", "Our", "network", "uses", "features", "from", "the", "entire", "image", "to", "predict", "each", "bounding", "box", ".", "It", "also", "predicts", "all", "bounding", "boxes", "across", "all", "classes", "for", "an", "image", "simultaneously", ".", "This", "means", "our", "network", "reasons", "globally", "about", "the", "full", "image", "and", "all", "the", "objects", "in", "the", "image", ".", "The", "YOLO", "design", "enables", "end", "-", "to", "-", "end", "training", "and", "real", "-", "time", "speeds", "while", "maintaining", "high", "average", "precision", ".", "Our", "system", "divides", "the", "input", "image", "into", "an", "grid", ".", "If", "the", "center", "of", "an", "object", "falls", "into", "a", "grid", "cell", ",", "that", "grid", "cell", "is", "responsible", "for", "detecting", "that", "object", ".", "Each", "grid", "cell", "predicts", "bounding", "boxes", "and", "confidence", "scores", "for", "those", "boxes", ".", "These", "confidence", "scores", "reflect", "how", "confident", "the", "model", "is", "that", "the", "box", "contains", "an", "object", "and", "also", "how", "accurate", "it", "thinks", "the", "box", "is", "that", "it", "predicts", ".", "Formally", "we", "define", "confidence", "as", ".", "If", "no", "object", "exists", "in", "that", "cell", ",", "the", "confidence", "scores", "should", "be", "zero", ".", "Otherwise", "we", "want", "the", "confidence", "score", "to", "equal", "the", "intersection", "over", "union", "(", "IOU", ")", "between", "the", "predicted", "box", "and", "the", "ground", "truth", ".", "Each", "bounding", "box", "consists", "of", "5", "predictions", ":", ",", ",", ",", ",", "and", "confidence", ".", "The", "coordinates", "represent", "the", "center", "of", "the", "box", "relative", "to", "the", "bounds", "of", "the", "grid", "cell", ".", "The", "width", "and", "height", "are", "predicted", "relative", "to", "the", "whole", "image", ".", "Finally", "the", "confidence", "prediction", "represents", "the", "IOU", "between", "the", "predicted", "box", "and", "any", "ground", "truth", "box", ".", "Each", "grid", "cell", "also", "predicts", "conditional", "class", "probabilities", ",", ".", "These", "probabilities", "are", "conditioned", "on", "the", "grid", "cell", "containing", "an", "object", ".", "We", "only", "predict", "one", "set", "of", "class", "probabilities", "per", "grid", "cell", ",", "regardless", "of", "the", "number", "of", "boxes", ".", "At", "test", "time", "we", "multiply", "the", "conditional", "class", "probabilities", "and", "the", "individual", "box", "confidence", "predictions", ",", "which", "gives", "us", "class", "-", "specific", "confidence", "scores", "for", "each", "box", ".", "These", "scores", "encode", "both", "the", "probability", "of", "that", "class", "appearing", "in", "the", "box", "and", "how", "well", "the", "predicted", "box", "fits", "the", "object", ".", "For", "evaluating", "YOLO", "on", "Pascal", "VOC", ",", "we", "use", ",", ".", "Pascal", "VOC", "has", "20", "labelled", "classes", "so", ".", "Our", "final", "prediction", "is", "a", "tensor", ".", "subsection", ":", "Network", "Design", "We", "implement", "this", "model", "as", "a", "convolutional", "neural", "network", "and", "evaluate", "it", "on", "the", "Pascal", "VOC", "detection", "dataset", ".", "The", "initial", "convolutional", "layers", "of", "the", "network", "extract", "features", "from", "the", "image", "while", "the", "fully", "connected", "layers", "predict", "the", "output", "probabilities", "and", "coordinates", ".", "Our", "network", "architecture", "is", "inspired", "by", "the", "GoogLeNet", "model", "for", "image", "classification", ".", "Our", "network", "has", "24", "convolutional", "layers", "followed", "by", "2", "fully", "connected", "layers", ".", "Instead", "of", "the", "inception", "modules", "used", "by", "GoogLeNet", ",", "we", "simply", "use", "reduction", "layers", "followed", "by", "convolutional", "layers", ",", "similar", "to", "Lin", "et", "al", ".", "The", "full", "network", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "We", "also", "train", "a", "fast", "version", "of", "YOLO", "designed", "to", "push", "the", "boundaries", "of", "fast", "object", "detection", ".", "Fast", "YOLO", "uses", "a", "neural", "network", "with", "fewer", "convolutional", "layers", "(", "9", "instead", "of", "24", ")", "and", "fewer", "filters", "in", "those", "layers", ".", "Other", "than", "the", "size", "of", "the", "network", ",", "all", "training", "and", "testing", "parameters", "are", "the", "same", "between", "YOLO", "and", "Fast", "YOLO", ".", "The", "final", "output", "of", "our", "network", "is", "the", "tensor", "of", "predictions", ".", "subsection", ":", "Training", "We", "pretrain", "our", "convolutional", "layers", "on", "the", "ImageNet", "1000", "-", "class", "competition", "dataset", ".", "For", "pretraining", "we", "use", "the", "first", "20", "convolutional", "layers", "from", "Figure", "[", "reference", "]", "followed", "by", "a", "average", "-", "pooling", "layer", "and", "a", "fully", "connected", "layer", ".", "We", "train", "this", "network", "for", "approximately", "a", "week", "and", "achieve", "a", "single", "crop", "top", "-", "5", "accuracy", "of", "88", "%", "on", "the", "ImageNet", "2012", "validation", "set", ",", "comparable", "to", "the", "GoogLeNet", "models", "in", "Caffe", "’s", "Model", "Zoo", ".", "We", "use", "the", "Darknet", "framework", "for", "all", "training", "and", "inference", ".", "We", "then", "convert", "the", "model", "to", "perform", "detection", ".", "Ren", "et", "al", ".", "show", "that", "adding", "both", "convolutional", "and", "connected", "layers", "to", "pretrained", "networks", "can", "improve", "performance", ".", "Following", "their", "example", ",", "we", "add", "four", "convolutional", "layers", "and", "two", "fully", "connected", "layers", "with", "randomly", "initialized", "weights", ".", "Detection", "often", "requires", "fine", "-", "grained", "visual", "information", "so", "we", "increase", "the", "input", "resolution", "of", "the", "network", "from", "to", ".", "Our", "final", "layer", "predicts", "both", "class", "probabilities", "and", "bounding", "box", "coordinates", ".", "We", "normalize", "the", "bounding", "box", "width", "and", "height", "by", "the", "image", "width", "and", "height", "so", "that", "they", "fall", "between", "0", "and", "1", ".", "We", "parametrize", "the", "bounding", "box", "and", "coordinates", "to", "be", "offsets", "of", "a", "particular", "grid", "cell", "location", "so", "they", "are", "also", "bounded", "between", "0", "and", "1", ".", "We", "use", "a", "linear", "activation", "function", "for", "the", "final", "layer", "and", "all", "other", "layers", "use", "the", "following", "leaky", "rectified", "linear", "activation", ":", "We", "optimize", "for", "sum", "-", "squared", "error", "in", "the", "output", "of", "our", "model", ".", "We", "use", "sum", "-", "squared", "error", "because", "it", "is", "easy", "to", "optimize", ",", "however", "it", "does", "not", "perfectly", "align", "with", "our", "goal", "of", "maximizing", "average", "precision", ".", "It", "weights", "localization", "error", "equally", "with", "classification", "error", "which", "may", "not", "be", "ideal", ".", "Also", ",", "in", "every", "image", "many", "grid", "cells", "do", "not", "contain", "any", "object", ".", "This", "pushes", "the", "“", "confidence", "”", "scores", "of", "those", "cells", "towards", "zero", ",", "often", "overpowering", "the", "gradient", "from", "cells", "that", "do", "contain", "objects", ".", "This", "can", "lead", "to", "model", "instability", ",", "causing", "training", "to", "diverge", "early", "on", ".", "To", "remedy", "this", ",", "we", "increase", "the", "loss", "from", "bounding", "box", "coordinate", "predictions", "and", "decrease", "the", "loss", "from", "confidence", "predictions", "for", "boxes", "that", "do", "n’t", "contain", "objects", ".", "We", "use", "two", "parameters", ",", "and", "to", "accomplish", "this", ".", "We", "set", "and", ".", "Sum", "-", "squared", "error", "also", "equally", "weights", "errors", "in", "large", "boxes", "and", "small", "boxes", ".", "Our", "error", "metric", "should", "reflect", "that", "small", "deviations", "in", "large", "boxes", "matter", "less", "than", "in", "small", "boxes", ".", "To", "partially", "address", "this", "we", "predict", "the", "square", "root", "of", "the", "bounding", "box", "width", "and", "height", "instead", "of", "the", "width", "and", "height", "directly", ".", "YOLO", "predicts", "multiple", "bounding", "boxes", "per", "grid", "cell", ".", "At", "training", "time", "we", "only", "want", "one", "bounding", "box", "predictor", "to", "be", "responsible", "for", "each", "object", ".", "We", "assign", "one", "predictor", "to", "be", "“", "responsible", "”", "for", "predicting", "an", "object", "based", "on", "which", "prediction", "has", "the", "highest", "current", "IOU", "with", "the", "ground", "truth", ".", "This", "leads", "to", "specialization", "between", "the", "bounding", "box", "predictors", ".", "Each", "predictor", "gets", "better", "at", "predicting", "certain", "sizes", ",", "aspect", "ratios", ",", "or", "classes", "of", "object", ",", "improving", "overall", "recall", ".", "During", "training", "we", "optimize", "the", "following", ",", "multi", "-", "part", "loss", "function", ":", "where", "denotes", "if", "object", "appears", "in", "cell", "and", "denotes", "that", "the", "th", "bounding", "box", "predictor", "in", "cell", "is", "“", "responsible", "”", "for", "that", "prediction", ".", "Note", "that", "the", "loss", "function", "only", "penalizes", "classification", "error", "if", "an", "object", "is", "present", "in", "that", "grid", "cell", "(", "hence", "the", "conditional", "class", "probability", "discussed", "earlier", ")", ".", "It", "also", "only", "penalizes", "bounding", "box", "coordinate", "error", "if", "that", "predictor", "is", "“", "responsible", "”", "for", "the", "ground", "truth", "box", "(", "i.e.", "has", "the", "highest", "IOU", "of", "any", "predictor", "in", "that", "grid", "cell", ")", ".", "We", "train", "the", "network", "for", "about", "135", "epochs", "on", "the", "training", "and", "validation", "data", "sets", "from", "Pascal", "VOC", "2007", "and", "2012", ".", "When", "testing", "on", "2012", "we", "also", "include", "the", "VOC", "2007", "test", "data", "for", "training", ".", "Throughout", "training", "we", "use", "a", "batch", "size", "of", "64", ",", "a", "momentum", "of", "and", "a", "decay", "of", ".", "Our", "learning", "rate", "schedule", "is", "as", "follows", ":", "For", "the", "first", "epochs", "we", "slowly", "raise", "the", "learning", "rate", "from", "to", ".", "If", "we", "start", "at", "a", "high", "learning", "rate", "our", "model", "often", "diverges", "due", "to", "unstable", "gradients", ".", "We", "continue", "training", "with", "for", "75", "epochs", ",", "then", "for", "30", "epochs", ",", "and", "finally", "for", "30", "epochs", ".", "To", "avoid", "overfitting", "we", "use", "dropout", "and", "extensive", "data", "augmentation", ".", "A", "dropout", "layer", "with", "rate", "=", ".5", "after", "the", "first", "connected", "layer", "prevents", "co", "-", "adaptation", "between", "layers", ".", "For", "data", "augmentation", "we", "introduce", "random", "scaling", "and", "translations", "of", "up", "to", "20", "%", "of", "the", "original", "image", "size", ".", "We", "also", "randomly", "adjust", "the", "exposure", "and", "saturation", "of", "the", "image", "by", "up", "to", "a", "factor", "of", "in", "the", "HSV", "color", "space", ".", "subsection", ":", "Inference", "Just", "like", "in", "training", ",", "predicting", "detections", "for", "a", "test", "image", "only", "requires", "one", "network", "evaluation", ".", "On", "Pascal", "VOC", "the", "network", "predicts", "98", "bounding", "boxes", "per", "image", "and", "class", "probabilities", "for", "each", "box", ".", "YOLO", "is", "extremely", "fast", "at", "test", "time", "since", "it", "only", "requires", "a", "single", "network", "evaluation", ",", "unlike", "classifier", "-", "based", "methods", ".", "The", "grid", "design", "enforces", "spatial", "diversity", "in", "the", "bounding", "box", "predictions", ".", "Often", "it", "is", "clear", "which", "grid", "cell", "an", "object", "falls", "in", "to", "and", "the", "network", "only", "predicts", "one", "box", "for", "each", "object", ".", "However", ",", "some", "large", "objects", "or", "objects", "near", "the", "border", "of", "multiple", "cells", "can", "be", "well", "localized", "by", "multiple", "cells", ".", "Non", "-", "maximal", "suppression", "can", "be", "used", "to", "fix", "these", "multiple", "detections", ".", "While", "not", "critical", "to", "performance", "as", "it", "is", "for", "R", "-", "CNN", "or", "DPM", ",", "non", "-", "maximal", "suppression", "adds", "2", "-", "3", "%", "in", "mAP", ".", "subsection", ":", "Limitations", "of", "YOLO", "YOLO", "imposes", "strong", "spatial", "constraints", "on", "bounding", "box", "predictions", "since", "each", "grid", "cell", "only", "predicts", "two", "boxes", "and", "can", "only", "have", "one", "class", ".", "This", "spatial", "constraint", "limits", "the", "number", "of", "nearby", "objects", "that", "our", "model", "can", "predict", ".", "Our", "model", "struggles", "with", "small", "objects", "that", "appear", "in", "groups", ",", "such", "as", "flocks", "of", "birds", ".", "Since", "our", "model", "learns", "to", "predict", "bounding", "boxes", "from", "data", ",", "it", "struggles", "to", "generalize", "to", "objects", "in", "new", "or", "unusual", "aspect", "ratios", "or", "configurations", ".", "Our", "model", "also", "uses", "relatively", "coarse", "features", "for", "predicting", "bounding", "boxes", "since", "our", "architecture", "has", "multiple", "downsampling", "layers", "from", "the", "input", "image", ".", "Finally", ",", "while", "we", "train", "on", "a", "loss", "function", "that", "approximates", "detection", "performance", ",", "our", "loss", "function", "treats", "errors", "the", "same", "in", "small", "bounding", "boxes", "versus", "large", "bounding", "boxes", ".", "A", "small", "error", "in", "a", "large", "box", "is", "generally", "benign", "but", "a", "small", "error", "in", "a", "small", "box", "has", "a", "much", "greater", "effect", "on", "IOU", ".", "Our", "main", "source", "of", "error", "is", "incorrect", "localizations", ".", "section", ":", "Comparison", "to", "Other", "Detection", "Systems", "Object", "detection", "is", "a", "core", "problem", "in", "computer", "vision", ".", "Detection", "pipelines", "generally", "start", "by", "extracting", "a", "set", "of", "robust", "features", "from", "input", "images", "(", "Haar", ",", "SIFT", ",", "HOG", ",", "convolutional", "features", ")", ".", "Then", ",", "classifiers", "or", "localizers", "are", "used", "to", "identify", "objects", "in", "the", "feature", "space", ".", "These", "classifiers", "or", "localizers", "are", "run", "either", "in", "sliding", "window", "fashion", "over", "the", "whole", "image", "or", "on", "some", "subset", "of", "regions", "in", "the", "image", ".", "We", "compare", "the", "YOLO", "detection", "system", "to", "several", "top", "detection", "frameworks", ",", "highlighting", "key", "similarities", "and", "differences", ".", "Deformable", "parts", "models", ".", "Deformable", "parts", "models", "(", "DPM", ")", "use", "a", "sliding", "window", "approach", "to", "object", "detection", ".", "DPM", "uses", "a", "disjoint", "pipeline", "to", "extract", "static", "features", ",", "classify", "regions", ",", "predict", "bounding", "boxes", "for", "high", "scoring", "regions", ",", "etc", ".", "Our", "system", "replaces", "all", "of", "these", "disparate", "parts", "with", "a", "single", "convolutional", "neural", "network", ".", "The", "network", "performs", "feature", "extraction", ",", "bounding", "box", "prediction", ",", "non", "-", "maximal", "suppression", ",", "and", "contextual", "reasoning", "all", "concurrently", ".", "Instead", "of", "static", "features", ",", "the", "network", "trains", "the", "features", "in", "-", "line", "and", "optimizes", "them", "for", "the", "detection", "task", ".", "Our", "unified", "architecture", "leads", "to", "a", "faster", ",", "more", "accurate", "model", "than", "DPM", ".", "R", "-", "CNN", ".", "R", "-", "CNN", "and", "its", "variants", "use", "region", "proposals", "instead", "of", "sliding", "windows", "to", "find", "objects", "in", "images", ".", "Selective", "Search", "generates", "potential", "bounding", "boxes", ",", "a", "convolutional", "network", "extracts", "features", ",", "an", "SVM", "scores", "the", "boxes", ",", "a", "linear", "model", "adjusts", "the", "bounding", "boxes", ",", "and", "non", "-", "max", "suppression", "eliminates", "duplicate", "detections", ".", "Each", "stage", "of", "this", "complex", "pipeline", "must", "be", "precisely", "tuned", "independently", "and", "the", "resulting", "system", "is", "very", "slow", ",", "taking", "more", "than", "40", "seconds", "per", "image", "at", "test", "time", ".", "YOLO", "shares", "some", "similarities", "with", "R", "-", "CNN", ".", "Each", "grid", "cell", "proposes", "potential", "bounding", "boxes", "and", "scores", "those", "boxes", "using", "convolutional", "features", ".", "However", ",", "our", "system", "puts", "spatial", "constraints", "on", "the", "grid", "cell", "proposals", "which", "helps", "mitigate", "multiple", "detections", "of", "the", "same", "object", ".", "Our", "system", "also", "proposes", "far", "fewer", "bounding", "boxes", ",", "only", "98", "per", "image", "compared", "to", "about", "2000", "from", "Selective", "Search", ".", "Finally", ",", "our", "system", "combines", "these", "individual", "components", "into", "a", "single", ",", "jointly", "optimized", "model", ".", "Other", "Fast", "Detectors", "Fast", "and", "Faster", "R", "-", "CNN", "focus", "on", "speeding", "up", "the", "R", "-", "CNN", "framework", "by", "sharing", "computation", "and", "using", "neural", "networks", "to", "propose", "regions", "instead", "of", "Selective", "Search", ".", "While", "they", "offer", "speed", "and", "accuracy", "improvements", "over", "R", "-", "CNN", ",", "both", "still", "fall", "short", "of", "real", "-", "time", "performance", ".", "Many", "research", "efforts", "focus", "on", "speeding", "up", "the", "DPM", "pipeline", ".", "They", "speed", "up", "HOG", "computation", ",", "use", "cascades", ",", "and", "push", "computation", "to", "GPUs", ".", "However", ",", "only", "30Hz", "DPM", "actually", "runs", "in", "real", "-", "time", ".", "Instead", "of", "trying", "to", "optimize", "individual", "components", "of", "a", "large", "detection", "pipeline", ",", "YOLO", "throws", "out", "the", "pipeline", "entirely", "and", "is", "fast", "by", "design", ".", "Detectors", "for", "single", "classes", "like", "faces", "or", "people", "can", "be", "highly", "optimized", "since", "they", "have", "to", "deal", "with", "much", "less", "variation", ".", "YOLO", "is", "a", "general", "purpose", "detector", "that", "learns", "to", "detect", "a", "variety", "of", "objects", "simultaneously", ".", "Deep", "MultiBox", ".", "Unlike", "R", "-", "CNN", ",", "Szegedy", "et", "al", ".", "train", "a", "convolutional", "neural", "network", "to", "predict", "regions", "of", "interest", "instead", "of", "using", "Selective", "Search", ".", "MultiBox", "can", "also", "perform", "single", "object", "detection", "by", "replacing", "the", "confidence", "prediction", "with", "a", "single", "class", "prediction", ".", "However", ",", "MultiBox", "can", "not", "perform", "general", "object", "detection", "and", "is", "still", "just", "a", "piece", "in", "a", "larger", "detection", "pipeline", ",", "requiring", "further", "image", "patch", "classification", ".", "Both", "YOLO", "and", "MultiBox", "use", "a", "convolutional", "network", "to", "predict", "bounding", "boxes", "in", "an", "image", "but", "YOLO", "is", "a", "complete", "detection", "system", ".", "OverFeat", ".", "Sermanet", "et", "al", ".", "train", "a", "convolutional", "neural", "network", "to", "perform", "localization", "and", "adapt", "that", "localizer", "to", "perform", "detection", ".", "OverFeat", "efficiently", "performs", "sliding", "window", "detection", "but", "it", "is", "still", "a", "disjoint", "system", ".", "OverFeat", "optimizes", "for", "localization", ",", "not", "detection", "performance", ".", "Like", "DPM", ",", "the", "localizer", "only", "sees", "local", "information", "when", "making", "a", "prediction", ".", "OverFeat", "can", "not", "reason", "about", "global", "context", "and", "thus", "requires", "significant", "post", "-", "processing", "to", "produce", "coherent", "detections", ".", "MultiGrasp", ".", "Our", "work", "is", "similar", "in", "design", "to", "work", "on", "grasp", "detection", "by", "Redmon", "et", "al", ".", "Our", "grid", "approach", "to", "bounding", "box", "prediction", "is", "based", "on", "the", "MultiGrasp", "system", "for", "regression", "to", "grasps", ".", "However", ",", "grasp", "detection", "is", "a", "much", "simpler", "task", "than", "object", "detection", ".", "MultiGrasp", "only", "needs", "to", "predict", "a", "single", "graspable", "region", "for", "an", "image", "containing", "one", "object", ".", "It", "does", "n’t", "have", "to", "estimate", "the", "size", ",", "location", ",", "or", "boundaries", "of", "the", "object", "or", "predict", "it", "’s", "class", ",", "only", "find", "a", "region", "suitable", "for", "grasping", ".", "YOLO", "predicts", "both", "bounding", "boxes", "and", "class", "probabilities", "for", "multiple", "objects", "of", "multiple", "classes", "in", "an", "image", ".", "section", ":", "Experiments", "First", "we", "compare", "YOLO", "with", "other", "real", "-", "time", "detection", "systems", "on", "Pascal", "VOC", "2007", ".", "To", "understand", "the", "differences", "between", "YOLO", "and", "R", "-", "CNN", "variants", "we", "explore", "the", "errors", "on", "VOC", "2007", "made", "by", "YOLO", "and", "Fast", "R", "-", "CNN", ",", "one", "of", "the", "highest", "performing", "versions", "of", "R", "-", "CNN", ".", "Based", "on", "the", "different", "error", "profiles", "we", "show", "that", "YOLO", "can", "be", "used", "to", "rescore", "Fast", "R", "-", "CNN", "detections", "and", "reduce", "the", "errors", "from", "background", "false", "positives", ",", "giving", "a", "significant", "performance", "boost", ".", "We", "also", "present", "VOC", "2012", "results", "and", "compare", "mAP", "to", "current", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "Finally", ",", "we", "show", "that", "YOLO", "generalizes", "to", "new", "domains", "better", "than", "other", "detectors", "on", "two", "artwork", "datasets", ".", "subsection", ":", "Comparison", "to", "Other", "Real", "-", "Time", "Systems", "Many", "research", "efforts", "in", "object", "detection", "focus", "on", "making", "standard", "detection", "pipelines", "fast", ".", "However", ",", "only", "Sadeghi", "et", "al", ".", "actually", "produce", "a", "detection", "system", "that", "runs", "in", "real", "-", "time", "(", "30", "frames", "per", "second", "or", "better", ")", ".", "We", "compare", "YOLO", "to", "their", "GPU", "implementation", "of", "DPM", "which", "runs", "either", "at", "30Hz", "or", "100Hz", ".", "While", "the", "other", "efforts", "do", "n’t", "reach", "the", "real", "-", "time", "milestone", "we", "also", "compare", "their", "relative", "mAP", "and", "speed", "to", "examine", "the", "accuracy", "-", "performance", "tradeoffs", "available", "in", "object", "detection", "systems", ".", "Fast", "YOLO", "is", "the", "fastest", "object", "detection", "method", "on", "Pascal", ";", "as", "far", "as", "we", "know", ",", "it", "is", "the", "fastest", "extant", "object", "detector", ".", "With", "mAP", ",", "it", "is", "more", "than", "twice", "as", "accurate", "as", "prior", "work", "on", "real", "-", "time", "detection", ".", "YOLO", "pushes", "mAP", "to", "while", "still", "maintaining", "real", "-", "time", "performance", ".", "We", "also", "train", "YOLO", "using", "VGG", "-", "16", ".", "This", "model", "is", "more", "accurate", "but", "also", "significantly", "slower", "than", "YOLO", ".", "It", "is", "useful", "for", "comparison", "to", "other", "detection", "systems", "that", "rely", "on", "VGG", "-", "16", "but", "since", "it", "is", "slower", "than", "real", "-", "time", "the", "rest", "of", "the", "paper", "focuses", "on", "our", "faster", "models", ".", "Fastest", "DPM", "effectively", "speeds", "up", "DPM", "without", "sacrificing", "much", "mAP", "but", "it", "still", "misses", "real", "-", "time", "performance", "by", "a", "factor", "of", "2", ".", "It", "also", "is", "limited", "by", "DPM", "’s", "relatively", "low", "accuracy", "on", "detection", "compared", "to", "neural", "network", "approaches", ".", "R", "-", "CNN", "minus", "R", "replaces", "Selective", "Search", "with", "static", "bounding", "box", "proposals", ".", "While", "it", "is", "much", "faster", "than", "R", "-", "CNN", ",", "it", "still", "falls", "short", "of", "real", "-", "time", "and", "takes", "a", "significant", "accuracy", "hit", "from", "not", "having", "good", "proposals", ".", "Fast", "R", "-", "CNN", "speeds", "up", "the", "classification", "stage", "of", "R", "-", "CNN", "but", "it", "still", "relies", "on", "selective", "search", "which", "can", "take", "around", "2", "seconds", "per", "image", "to", "generate", "bounding", "box", "proposals", ".", "Thus", "it", "has", "high", "mAP", "but", "at", "fps", "it", "is", "still", "far", "from", "real", "-", "time", ".", "The", "recent", "Faster", "R", "-", "CNN", "replaces", "selective", "search", "with", "a", "neural", "network", "to", "propose", "bounding", "boxes", ",", "similar", "to", "Szegedy", "et", "al", ".", "In", "our", "tests", ",", "their", "most", "accurate", "model", "achieves", "7", "fps", "while", "a", "smaller", ",", "less", "accurate", "one", "runs", "at", "18", "fps", ".", "The", "VGG", "-", "16", "version", "of", "Faster", "R", "-", "CNN", "is", "10", "mAP", "higher", "but", "is", "also", "6", "times", "slower", "than", "YOLO", ".", "The", "Zeiler", "-", "Fergus", "Faster", "R", "-", "CNN", "is", "only", "2.5", "times", "slower", "than", "YOLO", "but", "is", "also", "less", "accurate", ".", "subsection", ":", "VOC", "2007", "Error", "Analysis", "To", "further", "examine", "the", "differences", "between", "YOLO", "and", "state", "-", "of", "-", "the", "-", "art", "detectors", ",", "we", "look", "at", "a", "detailed", "breakdown", "of", "results", "on", "VOC", "2007", ".", "We", "compare", "YOLO", "to", "Fast", "R", "-", "CNN", "since", "Fast", "R", "-", "CNN", "is", "one", "of", "the", "highest", "performing", "detectors", "on", "Pascal", "and", "it", "’s", "detections", "are", "publicly", "available", ".", "We", "use", "the", "methodology", "and", "tools", "of", "Hoiem", "et", "al", ".", "For", "each", "category", "at", "test", "time", "we", "look", "at", "the", "top", "N", "predictions", "for", "that", "category", ".", "Each", "prediction", "is", "either", "correct", "or", "it", "is", "classified", "based", "on", "the", "type", "of", "error", ":", "Correct", ":", "correct", "class", "and", "Localization", ":", "correct", "class", ",", "Similar", ":", "class", "is", "similar", ",", "Other", ":", "class", "is", "wrong", ",", "Background", ":", "for", "any", "object", "Figure", "[", "reference", "]", "shows", "the", "breakdown", "of", "each", "error", "type", "averaged", "across", "all", "20", "classes", ".", "YOLO", "struggles", "to", "localize", "objects", "correctly", ".", "Localization", "errors", "account", "for", "more", "of", "YOLO", "’s", "errors", "than", "all", "other", "sources", "combined", ".", "Fast", "R", "-", "CNN", "makes", "much", "fewer", "localization", "errors", "but", "far", "more", "background", "errors", ".", "13.6", "%", "of", "it", "’s", "top", "detections", "are", "false", "positives", "that", "do", "n’t", "contain", "any", "objects", ".", "Fast", "R", "-", "CNN", "is", "almost", "3x", "more", "likely", "to", "predict", "background", "detections", "than", "YOLO", ".", "subsection", ":", "Combining", "Fast", "R", "-", "CNN", "and", "YOLO", "YOLO", "makes", "far", "fewer", "background", "mistakes", "than", "Fast", "R", "-", "CNN", ".", "By", "using", "YOLO", "to", "eliminate", "background", "detections", "from", "Fast", "R", "-", "CNN", "we", "get", "a", "significant", "boost", "in", "performance", ".", "For", "every", "bounding", "box", "that", "R", "-", "CNN", "predicts", "we", "check", "to", "see", "if", "YOLO", "predicts", "a", "similar", "box", ".", "If", "it", "does", ",", "we", "give", "that", "prediction", "a", "boost", "based", "on", "the", "probability", "predicted", "by", "YOLO", "and", "the", "overlap", "between", "the", "two", "boxes", ".", "The", "best", "Fast", "R", "-", "CNN", "model", "achieves", "a", "mAP", "of", "71.8", "%", "on", "the", "VOC", "2007", "test", "set", ".", "When", "combined", "with", "YOLO", ",", "its", "mAP", "increases", "by", "3.2", "%", "to", "75.0", "%", ".", "We", "also", "tried", "combining", "the", "top", "Fast", "R", "-", "CNN", "model", "with", "several", "other", "versions", "of", "Fast", "R", "-", "CNN", ".", "Those", "ensembles", "produced", "small", "increases", "in", "mAP", "between", ".3", "and", ".6", "%", ",", "see", "Table", "[", "reference", "]", "for", "details", ".", "The", "boost", "from", "YOLO", "is", "not", "simply", "a", "byproduct", "of", "model", "ensembling", "since", "there", "is", "little", "benefit", "from", "combining", "different", "versions", "of", "Fast", "R", "-", "CNN", ".", "Rather", ",", "it", "is", "precisely", "because", "YOLO", "makes", "different", "kinds", "of", "mistakes", "at", "test", "time", "that", "it", "is", "so", "effective", "at", "boosting", "Fast", "R", "-", "CNN", "’s", "performance", ".", "Unfortunately", ",", "this", "combination", "does", "n’t", "benefit", "from", "the", "speed", "of", "YOLO", "since", "we", "run", "each", "model", "seperately", "and", "then", "combine", "the", "results", ".", "However", ",", "since", "YOLO", "is", "so", "fast", "it", "does", "n’t", "add", "any", "significant", "computational", "time", "compared", "to", "Fast", "R", "-", "CNN", ".", "subsection", ":", "VOC", "2012", "Results", "On", "the", "VOC", "2012", "test", "set", ",", "YOLO", "scores", "57.9", "%", "mAP", ".", "This", "is", "lower", "than", "the", "current", "state", "of", "the", "art", ",", "closer", "to", "the", "original", "R", "-", "CNN", "using", "VGG", "-", "16", ",", "see", "Table", "[", "reference", "]", ".", "Our", "system", "struggles", "with", "small", "objects", "compared", "to", "its", "closest", "competitors", ".", "On", "categories", "like", "bottle", ",", "sheep", ",", "and", "tv", "/", "monitor", "YOLO", "scores", "8", "-", "10", "%", "lower", "than", "R", "-", "CNN", "or", "Feature", "Edit", ".", "However", ",", "on", "other", "categories", "like", "cat", "and", "train", "YOLO", "achieves", "higher", "performance", ".", "Our", "combined", "Fast", "R", "-", "CNN", "+", "YOLO", "model", "is", "one", "of", "the", "highest", "performing", "detection", "methods", ".", "Fast", "R", "-", "CNN", "gets", "a", "2.3", "%", "improvement", "from", "the", "combination", "with", "YOLO", ",", "boosting", "it", "5", "spots", "up", "on", "the", "public", "leaderboard", ".", "subsection", ":", "Generalizability", ":", "Person", "Detection", "in", "Artwork", "[", "b", "]", ".45", "[", "b", "]", ".55", "Academic", "datasets", "for", "object", "detection", "draw", "the", "training", "and", "testing", "data", "from", "the", "same", "distribution", ".", "In", "real", "-", "world", "applications", "it", "is", "hard", "to", "predict", "all", "possible", "use", "cases", "and", "the", "test", "data", "can", "diverge", "from", "what", "the", "system", "has", "seen", "before", ".", "We", "compare", "YOLO", "to", "other", "detection", "systems", "on", "the", "Picasso", "Dataset", "and", "the", "People", "-", "Art", "Dataset", ",", "two", "datasets", "for", "testing", "person", "detection", "on", "artwork", ".", "Figure", "[", "reference", "]", "shows", "comparative", "performance", "between", "YOLO", "and", "other", "detection", "methods", ".", "For", "reference", ",", "we", "give", "VOC", "2007", "detection", "AP", "on", "person", "where", "all", "models", "are", "trained", "only", "on", "VOC", "2007", "data", ".", "On", "Picasso", "models", "are", "trained", "on", "VOC", "2012", "while", "on", "People", "-", "Art", "they", "are", "trained", "on", "VOC", "2010", ".", "R", "-", "CNN", "has", "high", "AP", "on", "VOC", "2007", ".", "However", ",", "R", "-", "CNN", "drops", "off", "considerably", "when", "applied", "to", "artwork", ".", "R", "-", "CNN", "uses", "Selective", "Search", "for", "bounding", "box", "proposals", "which", "is", "tuned", "for", "natural", "images", ".", "The", "classifier", "step", "in", "R", "-", "CNN", "only", "sees", "small", "regions", "and", "needs", "good", "proposals", ".", "DPM", "maintains", "its", "AP", "well", "when", "applied", "to", "artwork", ".", "Prior", "work", "theorizes", "that", "DPM", "performs", "well", "because", "it", "has", "strong", "spatial", "models", "of", "the", "shape", "and", "layout", "of", "objects", ".", "Though", "DPM", "does", "n’t", "degrade", "as", "much", "as", "R", "-", "CNN", ",", "it", "starts", "from", "a", "lower", "AP", ".", "YOLO", "has", "good", "performance", "on", "VOC", "2007", "and", "its", "AP", "degrades", "less", "than", "other", "methods", "when", "applied", "to", "artwork", ".", "Like", "DPM", ",", "YOLO", "models", "the", "size", "and", "shape", "of", "objects", ",", "as", "well", "as", "relationships", "between", "objects", "and", "where", "objects", "commonly", "appear", ".", "Artwork", "and", "natural", "images", "are", "very", "different", "on", "a", "pixel", "level", "but", "they", "are", "similar", "in", "terms", "of", "the", "size", "and", "shape", "of", "objects", ",", "thus", "YOLO", "can", "still", "predict", "good", "bounding", "boxes", "and", "detections", ".", "section", ":", "Real", "-", "Time", "Detection", "In", "The", "Wild", "YOLO", "is", "a", "fast", ",", "accurate", "object", "detector", ",", "making", "it", "ideal", "for", "computer", "vision", "applications", ".", "We", "connect", "YOLO", "to", "a", "webcam", "and", "verify", "that", "it", "maintains", "real", "-", "time", "performance", ",", "including", "the", "time", "to", "fetch", "images", "from", "the", "camera", "and", "display", "the", "detections", ".", "The", "resulting", "system", "is", "interactive", "and", "engaging", ".", "While", "YOLO", "processes", "images", "individually", ",", "when", "attached", "to", "a", "webcam", "it", "functions", "like", "a", "tracking", "system", ",", "detecting", "objects", "as", "they", "move", "around", "and", "change", "in", "appearance", ".", "A", "demo", "of", "the", "system", "and", "the", "source", "code", "can", "be", "found", "on", "our", "project", "website", ":", ".", "section", ":", "Conclusion", "We", "introduce", "YOLO", ",", "a", "unified", "model", "for", "object", "detection", ".", "Our", "model", "is", "simple", "to", "construct", "and", "can", "be", "trained", "directly", "on", "full", "images", ".", "Unlike", "classifier", "-", "based", "approaches", ",", "YOLO", "is", "trained", "on", "a", "loss", "function", "that", "directly", "corresponds", "to", "detection", "performance", "and", "the", "entire", "model", "is", "trained", "jointly", ".", "Fast", "YOLO", "is", "the", "fastest", "general", "-", "purpose", "object", "detector", "in", "the", "literature", "and", "YOLO", "pushes", "the", "state", "-", "of", "-", "the", "-", "art", "in", "real", "-", "time", "object", "detection", ".", "YOLO", "also", "generalizes", "well", "to", "new", "domains", "making", "it", "ideal", "for", "applications", "that", "rely", "on", "fast", ",", "robust", "object", "detection", ".", "Acknowledgements", ":", "This", "work", "is", "partially", "supported", "by", "ONR", "N00014", "-", "13", "-", "1", "-", "0720", ",", "NSF", "IIS", "-", "1338054", ",", "and", "The", "Allen", "Distinguished", "Investigator", "Award", ".", "bibliography", ":", "References"], "entity": [[9, 14, ["Real", "-", "Time", "Object", "Detection"], "Task"], [16, 17, ["YOLO"], "Method"], [22, 24, ["object", "detection"], "Task"], [28, 30, ["object", "detection"], "Task"], [31, 32, ["classifiers"], "Method"], [34, 35, ["detection"], "Task"], [40, 42, ["object", "detection"], "Task"], [44, 46, ["regression", "problem"], "Task"], [58, 60, ["neural", "network"], "Method"], [77, 79, ["detection", "pipeline"], "Method"], [95, 96, ["detection"], "Task"], [99, 101, ["unified", "architecture"], "Method"], [129, 130, ["YOLO"], "Method"], [146, 150, ["real", "-", "time", "detectors"], "Method"], [160, 162, ["detection", "systems"], "Method"], [163, 164, ["YOLO"], "Method"], [181, 182, ["YOLO"], "Method"], [185, 188, ["representations", "of", "objects"], "Method"], [192, 194, ["detection", "methods"], "Method"], [196, 197, ["DPM"], "Method"], [198, 201, ["R", "-", "CNN"], "Method"], [241, 244, ["human", "visual", "system"], "Method"], [256, 257, ["driving"], "Task"], [267, 269, ["object", "detection"], "Task"], [301, 304, ["responsive", "robotic", "systems"], "Task"], [306, 308, ["detection", "systems"], "Method"], [309, 310, ["classifiers"], "Method"], [312, 313, ["detection"], "Task"], [323, 324, ["classifier"], "Method"], [342, 345, ["deformable", "parts", "models"], "Method"], [346, 347, ["DPM"], "Method"], [350, 353, ["sliding", "window", "approach"], "Method"], [355, 356, ["classifier"], "Method"], [371, 374, ["R", "-", "CNN"], "Method"], [375, 378, ["region", "proposal", "methods"], "Method"], [391, 392, ["classifier"], "Method"], [398, 399, ["classification"], "Task"], [400, 403, ["post", "-", "processing"], "Task"], [447, 449, ["object", "detection"], "Task"], [452, 454, ["regression", "problem"], "Task"], [476, 477, ["YOLO"], "Method"], [492, 493, ["YOLO"], "Method"], [505, 507, ["convolutional", "network"], "Method"], [519, 520, ["YOLO"], "Method"], [527, 528, ["detection"], "Task"], [540, 542, ["object", "detection"], "Task"], [545, 546, ["YOLO"], "Method"], [552, 554, ["frame", "detection"], "Task"], [556, 558, ["regression", "problem"], "Task"], [570, 572, ["neural", "network"], "Method"], [594, 596, ["batch", "processing"], "Method"], [610, 611, ["fps"], "Metric"], [633, 634, ["YOLO"], "Method"], [639, 642, ["mean", "average", "precision"], "Metric"], [644, 648, ["real", "-", "time", "systems"], "Method"], [672, 673, ["YOLO"], "Method"], [680, 681, ["predictions"], "Task"], [683, 691, ["sliding", "window", "and", "region", "proposal", "-", "based", "techniques"], "Method"], [692, 693, ["YOLO"], "Method"], [722, 725, ["top", "detection", "method"], "Method"], [743, 744, ["YOLO"], "Method"], [751, 753, ["background", "errors"], "Metric"], [762, 763, ["YOLO"], "Method"], [764, 768, ["generalizable", "representations", "of", "objects"], "Method"], [779, 780, ["YOLO"], "Method"], [781, 784, ["top", "detection", "methods"], "Method"], [785, 786, ["DPM"], "Method"], [787, 790, ["R", "-", "CNN"], "Method"], [796, 797, ["YOLO"], "Method"], [816, 817, ["YOLO"], "Method"], [827, 829, ["detection", "systems"], "Method"], [830, 831, ["accuracy"], "Metric"], [875, 877, ["pretrained", "models"], "Method"], [885, 887, ["Unified", "Detection"], "Task"], [893, 895, ["object", "detection"], "Task"], [898, 900, ["neural", "network"], "Method"], [955, 957, ["end", "training"], "Task"], [958, 962, ["real", "-", "time", "speeds"], "Metric"], [965, 967, ["average", "precision"], "Metric"], [1014, 1016, ["confidence", "scores"], "Metric"], [1066, 1068, ["confidence", "score"], "Metric"], [1132, 1134, ["confidence", "prediction"], "Task"], [1241, 1242, ["YOLO"], "Method"], [1243, 1245, ["Pascal", "VOC"], "Material"], [1250, 1252, ["Pascal", "VOC"], "Material"], [1267, 1269, ["Network", "Design"], "Method"], [1275, 1278, ["convolutional", "neural", "network"], "Method"], [1290, 1292, ["convolutional", "layers"], "Method"], [1302, 1305, ["fully", "connected", "layers"], "Method"], [1313, 1315, ["network", "architecture"], "Method"], [1319, 1321, ["GoogLeNet", "model"], "Method"], [1322, 1324, ["image", "classification"], "Task"], [1329, 1331, ["convolutional", "layers"], "Method"], [1334, 1337, ["fully", "connected", "layers"], "Method"], [1341, 1343, ["inception", "modules"], "Method"], [1345, 1346, ["GoogLeNet"], "Method"], [1350, 1352, ["reduction", "layers"], "Method"], [1354, 1356, ["convolutional", "layers"], "Method"], [1381, 1382, ["YOLO"], "Method"], [1392, 1394, ["Fast", "YOLO"], "Method"], [1396, 1398, ["neural", "network"], "Method"], [1400, 1402, ["convolutional", "layers"], "Method"], [1424, 1428, ["training", "and", "testing", "parameters"], "Metric"], [1432, 1433, ["YOLO"], "Method"], [1434, 1436, ["Fast", "YOLO"], "Method"], [1455, 1457, ["convolutional", "layers"], "Method"], [1459, 1465, ["ImageNet", "1000", "-", "class", "competition", "dataset"], "Material"], [1467, 1468, ["pretraining"], "Task"], [1473, 1475, ["convolutional", "layers"], "Method"], [1483, 1487, ["average", "-", "pooling", "layer"], "Method"], [1489, 1492, ["fully", "connected", "layer"], "Method"], [1515, 1519, ["ImageNet", "2012", "validation", "set"], "Material"], [1523, 1525, ["GoogLeNet", "models"], "Method"], [1526, 1530, ["Caffe", "’s", "Model", "Zoo"], "Method"], [1534, 1536, ["Darknet", "framework"], "Method"], [1538, 1541, ["training", "and", "inference"], "Task"], [1549, 1550, ["detection"], "Task"], [1559, 1563, ["convolutional", "and", "connected", "layers"], "Method"], [1564, 1566, ["pretrained", "networks"], "Method"], [1577, 1579, ["convolutional", "layers"], "Method"], [1581, 1584, ["fully", "connected", "layers"], "Method"], [1589, 1590, ["Detection"], "Task"], [1673, 1676, ["linear", "activation", "function"], "Method"], [1687, 1691, ["leaky", "rectified", "linear", "activation"], "Method"], [1695, 1699, ["sum", "-", "squared", "error"], "Metric"], [1708, 1712, ["sum", "-", "squared", "error"], "Metric"], [1729, 1732, ["maximizing", "average", "precision"], "Metric"], [1735, 1737, ["localization", "error"], "Metric"], [1739, 1741, ["classification", "error"], "Metric"], [1793, 1794, ["training"], "Task"], [1808, 1812, ["bounding", "box", "coordinate", "predictions"], "Method"], [1841, 1845, ["Sum", "-", "squared", "error"], "Metric"], [1857, 1859, ["error", "metric"], "Metric"], [1898, 1899, ["YOLO"], "Method"], [1914, 1917, ["bounding", "box", "predictor"], "Method"], [1957, 1960, ["bounding", "box", "predictors"], "Method"], [1980, 1981, ["recall"], "Metric"], [1983, 1984, ["training"], "Task"], [2007, 2010, ["bounding", "box", "predictor"], "Method"], [2023, 2025, ["loss", "function"], "Method"], [2027, 2029, ["classification", "error"], "Metric"], [2113, 2117, ["VOC", "2007", "test", "data"], "Material"], [2139, 2142, ["learning", "rate", "schedule"], "Metric"], [2154, 2156, ["learning", "rate"], "Metric"], [2165, 2167, ["learning", "rate"], "Metric"], [2200, 2201, ["dropout"], "Method"], [2203, 2205, ["data", "augmentation"], "Method"], [2207, 2209, ["dropout", "layer"], "Method"], [2216, 2218, ["connected", "layer"], "Method"], [2226, 2228, ["data", "augmentation"], "Task"], [2270, 2271, ["Inference"], "Task"], [2276, 2278, ["predicting", "detections"], "Task"], [2285, 2287, ["network", "evaluation"], "Method"], [2289, 2291, ["Pascal", "VOC"], "Material"], [2306, 2307, ["YOLO"], "Method"], [2319, 2321, ["network", "evaluation"], "Method"], [2323, 2327, ["classifier", "-", "based", "methods"], "Method"], [2329, 2331, ["grid", "design"], "Method"], [2336, 2339, ["bounding", "box", "predictions"], "Task"], [2384, 2388, ["Non", "-", "maximal", "suppression"], "Method"], [2406, 2409, ["R", "-", "CNN"], "Method"], [2410, 2411, ["DPM"], "Method"], [2412, 2416, ["non", "-", "maximal", "suppression"], "Method"], [2422, 2423, ["mAP"], "Metric"], [2428, 2429, ["YOLO"], "Method"], [2429, 2430, ["YOLO"], "Method"], [2435, 2438, ["bounding", "box", "predictions"], "Task"], [2519, 2522, ["predicting", "bounding", "boxes"], "Task"], [2541, 2543, ["loss", "function"], "Method"], [2545, 2546, ["detection"], "Task"], [2549, 2551, ["loss", "function"], "Method"], [2588, 2589, ["IOU"], "Task"], [2596, 2598, ["incorrect", "localizations"], "Task"], [2604, 2606, ["Detection", "Systems"], "Method"], [2606, 2608, ["Object", "detection"], "Task"], [2613, 2615, ["computer", "vision"], "Task"], [2616, 2618, ["Detection", "pipelines"], "Method"], [2643, 2644, ["classifiers"], "Method"], [2645, 2646, ["localizers"], "Method"], [2657, 2658, ["classifiers"], "Method"], [2659, 2660, ["localizers"], "Method"], [2664, 2667, ["sliding", "window", "fashion"], "Method"], [2689, 2692, ["top", "detection", "frameworks"], "Method"], [2699, 2702, ["Deformable", "parts", "models"], "Method"], [2703, 2706, ["Deformable", "parts", "models"], "Method"], [2707, 2708, ["DPM"], "Method"], [2711, 2714, ["sliding", "window", "approach"], "Method"], [2715, 2717, ["object", "detection"], "Task"], [2718, 2719, ["DPM"], "Method"], [2721, 2723, ["disjoint", "pipeline"], "Method"], [2752, 2755, ["convolutional", "neural", "network"], "Method"], [2759, 2761, ["feature", "extraction"], "Task"], [2762, 2765, ["bounding", "box", "prediction"], "Task"], [2766, 2770, ["non", "-", "maximal", "suppression"], "Task"], [2772, 2774, ["contextual", "reasoning"], "Task"], [2795, 2797, ["detection", "task"], "Task"], [2799, 2801, ["unified", "architecture"], "Method"], [2810, 2811, ["DPM"], "Method"], [2812, 2815, ["R", "-", "CNN"], "Method"], [2816, 2819, ["R", "-", "CNN"], "Method"], [2823, 2825, ["region", "proposals"], "Method"], [2835, 2837, ["Selective", "Search"], "Task"], [2843, 2845, ["convolutional", "network"], "Method"], [2849, 2850, ["SVM"], "Method"], [2855, 2857, ["linear", "model"], "Method"], [2863, 2867, ["non", "-", "max", "suppression"], "Method"], [2901, 2902, ["YOLO"], "Method"], [2906, 2909, ["R", "-", "CNN"], "Method"], [2922, 2924, ["convolutional", "features"], "Method"], [2965, 2967, ["Selective", "Search"], "Method"], [2980, 2983, ["jointly", "optimized", "model"], "Method"], [2985, 2987, ["Fast", "Detectors"], "Method"], [2990, 2993, ["R", "-", "CNN"], "Method"], [3004, 3005, ["computation"], "Method"], [3007, 3009, ["neural", "networks"], "Method"], [3014, 3016, ["Selective", "Search"], "Method"], [3020, 3021, ["speed"], "Metric"], [3022, 3023, ["accuracy"], "Metric"], [3025, 3028, ["R", "-", "CNN"], "Method"], [3053, 3055, ["HOG", "computation"], "Task"], [3057, 3058, ["cascades"], "Method"], [3060, 3062, ["push", "computation"], "Task"], [3069, 3070, ["DPM"], "Method"], [3087, 3089, ["detection", "pipeline"], "Method"], [3090, 3091, ["YOLO"], "Method"], [3102, 3103, ["Detectors"], "Method"], [3124, 3125, ["YOLO"], "Method"], [3127, 3130, ["general", "purpose", "detector"], "Method"], [3140, 3142, ["Deep", "MultiBox"], "Task"], [3144, 3147, ["R", "-", "CNN"], "Method"], [3154, 3157, ["convolutional", "neural", "network"], "Method"], [3165, 3167, ["Selective", "Search"], "Method"], [3168, 3169, ["MultiBox"], "Method"], [3178, 3180, ["confidence", "prediction"], "Method"], [3183, 3185, ["class", "prediction"], "Method"], [3188, 3189, ["MultiBox"], "Method"], [3204, 3206, ["detection", "pipeline"], "Task"], [3209, 3212, ["image", "patch", "classification"], "Task"], [3214, 3215, ["YOLO"], "Method"], [3216, 3217, ["MultiBox"], "Method"], [3219, 3221, ["convolutional", "network"], "Method"], [3223, 3225, ["bounding", "boxes"], "Task"], [3229, 3230, ["YOLO"], "Method"], [3233, 3235, ["detection", "system"], "Method"], [3244, 3247, ["convolutional", "neural", "network"], "Method"], [3249, 3250, ["localization"], "Task"], [3253, 3254, ["localizer"], "Method"], [3256, 3257, ["detection"], "Task"], [3258, 3259, ["OverFeat"], "Method"], [3261, 3264, ["sliding", "window", "detection"], "Method"], [3272, 3273, ["OverFeat"], "Method"], [3275, 3276, ["localization"], "Task"], [3278, 3279, ["detection"], "Task"], [3282, 3283, ["DPM"], "Method"], [3285, 3286, ["localizer"], "Method"], [3293, 3294, ["prediction"], "Task"], [3295, 3296, ["OverFeat"], "Method"], [3306, 3309, ["post", "-", "processing"], "Method"], [3314, 3315, ["MultiGrasp"], "Method"], [3325, 3327, ["grasp", "detection"], "Task"], [3333, 3335, ["grid", "approach"], "Method"], [3336, 3339, ["bounding", "box", "prediction"], "Task"], [3343, 3345, ["MultiGrasp", "system"], "Method"], [3346, 3349, ["regression", "to", "grasps"], "Task"], [3352, 3354, ["grasp", "detection"], "Task"], [3360, 3362, ["object", "detection"], "Task"], [3363, 3364, ["MultiGrasp"], "Method"], [3407, 3408, ["grasping"], "Task"], [3409, 3410, ["YOLO"], "Method"], [3433, 3434, ["YOLO"], "Method"], [3436, 3441, ["real", "-", "time", "detection", "systems"], "Method"], [3451, 3452, ["YOLO"], "Method"], [3453, 3457, ["R", "-", "CNN", "variants"], "Method"], [3462, 3464, ["VOC", "2007"], "Material"], [3466, 3467, ["YOLO"], "Method"], [3480, 3483, ["R", "-", "CNN"], "Method"], [3493, 3494, ["YOLO"], "Method"], [3522, 3524, ["VOC", "2012"], "Material"], [3527, 3528, ["mAP"], "Metric"], [3544, 3545, ["YOLO"], "Method"], [3563, 3567, ["Real", "-", "Time", "Systems"], "Task"], [3571, 3573, ["object", "detection"], "Task"], [3577, 3579, ["detection", "pipelines"], "Method"], [3591, 3593, ["detection", "system"], "Method"], [3610, 3611, ["YOLO"], "Method"], [3613, 3615, ["GPU", "implementation"], "Method"], [3616, 3617, ["DPM"], "Method"], [3642, 3643, ["mAP"], "Metric"], [3644, 3645, ["speed"], "Metric"], [3658, 3660, ["Fast", "YOLO"], "Method"], [3667, 3668, ["Pascal"], "Material"], [3680, 3682, ["object", "detector"], "Method"], [3684, 3685, ["mAP"], "Metric"], [3697, 3701, ["real", "-", "time", "detection"], "Task"], [3717, 3718, ["YOLO"], "Method"], [3719, 3722, ["VGG", "-", "16"], "Method"], [3733, 3734, ["YOLO"], "Method"], [3742, 3744, ["detection", "systems"], "Method"], [3747, 3750, ["VGG", "-", "16"], "Method"], [3771, 3772, ["DPM"], "Method"], [3775, 3776, ["DPM"], "Method"], [3784, 3788, ["real", "-", "time", "performance"], "Metric"], [3799, 3800, ["DPM"], "Method"], [3803, 3804, ["accuracy"], "Metric"], [3805, 3806, ["detection"], "Task"], [3808, 3811, ["neural", "network", "approaches"], "Method"], [3818, 3820, ["Selective", "Search"], "Method"], [3821, 3825, ["static", "bounding", "box", "proposals"], "Method"], [3832, 3835, ["R", "-", "CNN"], "Method"], [3848, 3849, ["accuracy"], "Metric"], [3857, 3860, ["R", "-", "CNN"], "Method"], [3863, 3865, ["classification", "stage"], "Method"], [3866, 3869, ["R", "-", "CNN"], "Method"], [3874, 3876, ["selective", "search"], "Method"], [3894, 3895, ["mAP"], "Metric"], [3897, 3898, ["fps"], "Metric"], [3910, 3913, ["R", "-", "CNN"], "Method"], [3914, 3916, ["selective", "search"], "Method"], [3918, 3920, ["neural", "network"], "Method"], [3941, 3942, ["fps"], "Metric"], [3952, 3953, ["fps"], "Metric"], [3975, 3976, ["YOLO"], "Method"], [3991, 3992, ["YOLO"], "Method"], [4000, 4004, ["VOC", "2007", "Error", "Analysis"], "Task"], [4010, 4011, ["YOLO"], "Method"], [4030, 4032, ["VOC", "2007"], "Material"], [4035, 4036, ["YOLO"], "Method"], [4054, 4055, ["Pascal"], "Material"], [4112, 4113, ["Localization"], "Task"], [4151, 4152, ["YOLO"], "Method"], [4158, 4160, ["Localization", "errors"], "Task"], [4180, 4182, ["localization", "errors"], "Task"], [4216, 4218, ["background", "detections"], "Task"], [4229, 4230, ["YOLO"], "Method"], [4230, 4231, ["YOLO"], "Method"], [4238, 4241, ["R", "-", "CNN"], "Method"], [4244, 4245, ["YOLO"], "Method"], [4251, 4254, ["R", "-", "CNN"], "Method"], [4316, 4317, ["mAP"], "Metric"], [4322, 4326, ["VOC", "2007", "test", "set"], "Material"], [4330, 4331, ["YOLO"], "Method"], [4333, 4334, ["mAP"], "Metric"], [4369, 4370, ["mAP"], "Metric"], [4387, 4388, ["YOLO"], "Method"], [4394, 4396, ["model", "ensembling"], "Method"], [4417, 4418, ["YOLO"], "Method"], [4434, 4437, ["R", "-", "CNN"], "Method"], [4451, 4452, ["YOLO"], "Method"], [4467, 4468, ["YOLO"], "Method"], [4477, 4479, ["computational", "time"], "Metric"], [4488, 4490, ["VOC", "2012"], "Material"], [4493, 4497, ["VOC", "2012", "test", "set"], "Material"], [4498, 4499, ["YOLO"], "Method"], [4502, 4503, ["mAP"], "Metric"], [4519, 4522, ["R", "-", "CNN"], "Method"], [4523, 4526, ["VGG", "-", "16"], "Method"], [4564, 4567, ["R", "-", "CNN"], "Method"], [4568, 4570, ["Feature", "Edit"], "Method"], [4580, 4581, ["YOLO"], "Method"], [4600, 4602, ["detection", "methods"], "Method"], [4616, 4617, ["YOLO"], "Method"], [4632, 4634, ["Person", "Detection"], "Task"], [4635, 4636, ["Artwork"], "Task"], [4644, 4646, ["Academic", "datasets"], "Material"], [4647, 4649, ["object", "detection"], "Task"], [4661, 4665, ["real", "-", "world", "applications"], "Task"], [4690, 4691, ["YOLO"], "Method"], [4693, 4695, ["detection", "systems"], "Method"], [4697, 4699, ["Picasso", "Dataset"], "Material"], [4701, 4705, ["People", "-", "Art", "Dataset"], "Material"], [4710, 4712, ["person", "detection"], "Task"], [4723, 4724, ["YOLO"], "Method"], [4726, 4728, ["detection", "methods"], "Method"], [4747, 4750, ["VOC", "2007", "data"], "Material"], [4752, 4754, ["Picasso", "models"], "Method"], [4757, 4759, ["VOC", "2012"], "Material"], [4761, 4764, ["People", "-", "Art"], "Method"], [4768, 4770, ["VOC", "2010"], "Material"], [4771, 4774, ["R", "-", "CNN"], "Method"], [4776, 4777, ["AP"], "Metric"], [4778, 4780, ["VOC", "2007"], "Material"], [4783, 4786, ["R", "-", "CNN"], "Method"], [4792, 4793, ["artwork"], "Material"], [4794, 4797, ["R", "-", "CNN"], "Method"], [4798, 4800, ["Selective", "Search"], "Method"], [4801, 4804, ["bounding", "box", "proposals"], "Method"], [4808, 4810, ["natural", "images"], "Material"], [4812, 4814, ["classifier", "step"], "Method"], [4815, 4818, ["R", "-", "CNN"], "Method"], [4827, 4828, ["DPM"], "Method"], [4835, 4836, ["artwork"], "Material"], [4841, 4842, ["DPM"], "Method"], [4848, 4850, ["spatial", "models"], "Method"], [4859, 4860, ["DPM"], "Method"], [4866, 4869, ["R", "-", "CNN"], "Method"], [4877, 4878, ["YOLO"], "Method"], [4882, 4884, ["VOC", "2007"], "Material"], [4898, 4899, ["DPM"], "Method"], [4900, 4901, ["YOLO"], "Method"], [4947, 4948, ["YOLO"], "Method"], [4959, 4966, ["Real", "-", "Time", "Detection", "In", "The", "Wild"], "Task"], [4966, 4967, ["YOLO"], "Method"], [4972, 4974, ["object", "detector"], "Method"], [4979, 4982, ["computer", "vision", "applications"], "Task"], [4985, 4986, ["YOLO"], "Method"], [5031, 5033, ["webcam", "it"], "Method"], [5036, 5038, ["tracking", "system"], "Method"], [5073, 5074, ["YOLO"], "Method"], [5076, 5078, ["unified", "model"], "Method"], [5079, 5081, ["object", "detection"], "Task"], [5098, 5102, ["classifier", "-", "based", "approaches"], "Method"], [5103, 5104, ["YOLO"], "Method"], [5108, 5110, ["loss", "function"], "Method"], [5114, 5115, ["detection"], "Task"], [5124, 5126, ["Fast", "YOLO"], "Method"], [5129, 5134, ["general", "-", "purpose", "object", "detector"], "Method"], [5138, 5139, ["YOLO"], "Method"], [5155, 5156, ["YOLO"], "Method"], [107, 108, ["YOLO"], "Method"], [143, 144, ["mAP"], "Metric"], [717, 720, ["R", "-", "CNN"], "Method"], [756, 759, ["R", "-", "CNN"], "Method"], [948, 949, ["YOLO"], "Method"], [1283, 1284, ["Pascal"], "Material"], [1389, 1391, ["object", "detection"], "Task"], [1509, 1510, ["accuracy"], "Metric"], [2099, 2100, ["Pascal"], "Material"], [2684, 2685, ["YOLO"], "Method"], [2998, 3001, ["R", "-", "CNN"], "Method"], [3047, 3048, ["DPM"], "Method"], [3051, 3052, ["speed"], "Metric"], [3173, 3175, ["object", "detection"], "Task"], [3193, 3195, ["object", "detection"], "Task"], [3442, 3443, ["Pascal"], "Material"], [3469, 3472, ["R", "-", "CNN"], "Method"], [3500, 3503, ["R", "-", "CNN"], "Method"], [3648, 3649, ["accuracy"], "Metric"], [3654, 3656, ["object", "detection"], "Task"], [3663, 3665, ["object", "detection"], "Task"], [3702, 3703, ["YOLO"], "Method"], [3704, 3705, ["mAP"], "Metric"], [3779, 3780, ["mAP"], "Metric"], [3812, 3815, ["R", "-", "CNN"], "Method"], [3961, 3964, ["R", "-", "CNN"], "Method"], [3966, 3967, ["mAP"], "Metric"], [3982, 3985, ["R", "-", "CNN"], "Method"], [4038, 4041, ["R", "-", "CNN"], "Method"], [4043, 4046, ["R", "-", "CNN"], "Method"], [4164, 4165, ["YOLO"], "Method"], [4174, 4177, ["R", "-", "CNN"], "Method"], [4206, 4209, ["R", "-", "CNN"], "Method"], [4219, 4220, ["YOLO"], "Method"], [4225, 4228, ["R", "-", "CNN"], "Method"], [4267, 4270, ["R", "-", "CNN"], "Method"], [4276, 4277, ["YOLO"], "Method"], [4298, 4299, ["YOLO"], "Method"], [4310, 4313, ["R", "-", "CNN"], "Method"], [4349, 4352, ["R", "-", "CNN"], "Method"], [4359, 4362, ["R", "-", "CNN"], "Method"], [4407, 4410, ["R", "-", "CNN"], "Method"], [4449, 4450, ["speed"], "Metric"], [4482, 4485, ["R", "-", "CNN"], "Method"], [4556, 4557, ["YOLO"], "Method"], [4588, 4591, ["R", "-", "CNN"], "Method"], [4592, 4593, ["YOLO"], "Method"], [4604, 4607, ["R", "-", "CNN"], "Method"], [4737, 4738, ["AP"], "Metric"], [4830, 4831, ["AP"], "Metric"], [4875, 4876, ["AP"], "Metric"], [4886, 4887, ["AP"], "Metric"], [5022, 5023, ["YOLO"], "Method"], [5152, 5154, ["object", "detection"], "Task"], [5173, 5175, ["object", "detection"], "Task"]]} {"doc_id": "232b43584b2236669c0a53702ad89ab10c3886ea", "text": ["In", "this", "work", ",", "we", "build", "on", "recent", "advances", "in", "distributional", "reinforcement", "learning", "to", "give", "a", "generally", "applicable", ",", "flexible", ",", "and", "state", "-", "of", "-", "the", "-", "art", "distributional", "variant", "of", "DQN", ".", "We", "achieve", "this", "by", "using", "quantile", "regression", "to", "approximate", "the", "full", "quantile", "function", "for", "the", "state", "-", "action", "return", "distribution", ".", "By", "reparameterizing", "a", "distribution", "over", "the", "sample", "space", ",", "this", "yields", "an", "implicitly", "defined", "return", "distribution", "and", "gives", "rise", "to", "a", "large", "class", "of", "risk", "-", "sensitive", "policies", ".", "We", "demonstrate", "improved", "performance", "on", "the", "57", "Atari", "2600", "games", "in", "the", "ALE", ",", "and", "use", "our", "algorithm", "’s", "implicitly", "defined", "distributions", "to", "study", "the", "effects", "of", "risk", "-", "sensitive", "policies", "in", "Atari", "games", ".", "ImplicitQuantileNetworksforDistributionalReinforcementLearning", "section", ":", "Introduction", "Distributional", "reinforcement", "learning", "focuses", "on", "the", "intrinsic", "randomness", "of", "returns", "within", "the", "reinforcement", "learning", "(", "RL", ")", "framework", ".", "As", "the", "agent", "interacts", "with", "the", "environment", ",", "irreducible", "randomness", "seeps", "in", "through", "the", "stochasticity", "of", "these", "interactions", ",", "the", "approximations", "in", "the", "agent", "’s", "representation", ",", "and", "even", "the", "inherently", "chaotic", "nature", "of", "physical", "interaction", ".", "Distributional", "RL", "aims", "to", "model", "the", "distribution", "over", "returns", ",", "whose", "mean", "is", "the", "traditional", "value", "function", ",", "and", "to", "use", "these", "distributions", "to", "evaluate", "and", "optimize", "a", "policy", ".", "Any", "distributional", "RL", "algorithm", "is", "characterized", "by", "two", "aspects", ":", "the", "parameterization", "of", "the", "return", "distribution", ",", "and", "the", "distance", "metric", "or", "loss", "function", "being", "optimized", ".", "Together", ",", "these", "choices", "control", "assumptions", "about", "the", "random", "returns", "and", "how", "approximations", "will", "be", "traded", "off", ".", "Categorical", "DQN", "[", "C51", "]", "c51", "combines", "a", "categorical", "distribution", "and", "the", "cross", "-", "entropy", "loss", "with", "the", "Cramér", "-", "minimizing", "projection", ".", "For", "this", ",", "it", "assumes", "returns", "are", "bounded", "in", "a", "known", "range", "and", "trades", "off", "mean", "-", "preservation", "at", "the", "cost", "of", "overestimating", "variance", ".", "C51", "outperformed", "all", "previous", "improvements", "to", "DQN", "on", "a", "set", "of", "57", "Atari", "2600", "games", "in", "the", "Arcade", "Learning", "Environment", ",", "which", "we", "refer", "to", "as", "the", "Atari", "-", "57", "benchmark", ".", "Subsequently", ",", "several", "papers", "have", "built", "upon", "this", "successful", "combination", "to", "achieve", "significant", "improvements", "to", "the", "state", "-", "of", "-", "the", "-", "art", "in", "Atari", "-", "57", ",", "and", "challenging", "continuous", "control", "tasks", ".", "These", "algorithms", "are", "restricted", "to", "assigning", "probabilities", "to", "an", "a", "priori", "fixed", ",", "discrete", "set", "of", "possible", "returns", ".", "dabney2017qr", "propose", "an", "alternate", "pair", "of", "choices", ",", "parameterizing", "the", "distribution", "by", "a", "uniform", "mixture", "of", "Diracs", "whose", "locations", "are", "adjusted", "using", "quantile", "regression", ".", "Their", "algorithm", ",", "QR", "-", "DQN", ",", "while", "restricted", "to", "a", "discrete", "set", "of", "quantiles", ",", "automatically", "adapts", "return", "quantiles", "to", "minimize", "the", "Wasserstein", "distance", "between", "the", "Bellman", "updated", "and", "current", "return", "distributions", ".", "This", "flexibility", "allows", "QR", "-", "DQN", "to", "significantly", "improve", "on", "C51", "’s", "Atari", "-", "57", "performance", ".", "In", "this", "paper", ",", "we", "extend", "the", "approach", "of", "dabney2017qr", ",", "from", "learning", "a", "discrete", "set", "of", "quantiles", "to", "learning", "the", "full", "quantile", "function", ",", "a", "continuous", "map", "from", "probabilities", "to", "returns", ".", "When", "combined", "with", "a", "base", "distribution", ",", "such", "as", ",", "this", "forms", "an", "implicit", "distribution", "capable", "of", "approximating", "any", "distribution", "over", "returns", "given", "sufficient", "network", "capacity", ".", "Our", "approach", ",", "implicit", "quantile", "networks", "(", "IQN", ")", ",", "is", "best", "viewed", "as", "a", "simple", "distributional", "generalization", "of", "the", "DQN", "algorithm", ",", "and", "provides", "several", "benefits", "over", "QR", "-", "DQN", ".", "First", ",", "the", "approximation", "error", "for", "the", "distribution", "is", "no", "longer", "controlled", "by", "the", "number", "of", "quantiles", "output", "by", "the", "network", ",", "but", "by", "the", "size", "of", "the", "network", "itself", ",", "and", "the", "amount", "of", "training", ".", "Second", ",", "IQN", "can", "be", "used", "with", "as", "few", ",", "or", "as", "many", ",", "samples", "per", "update", "as", "desired", ",", "providing", "improved", "data", "efficiency", "with", "increasing", "number", "of", "samples", "per", "training", "update", ".", "Third", ",", "the", "implicit", "representation", "of", "the", "return", "distribution", "allows", "us", "to", "expand", "the", "class", "of", "policies", "to", "more", "fully", "take", "advantage", "of", "the", "learned", "distribution", ".", "Specifically", ",", "by", "taking", "the", "base", "distribution", "to", "be", "non", "-", "uniform", ",", "we", "expand", "the", "class", "of", "policies", "to", "-", "greedy", "policies", "on", "arbitrary", "distortion", "risk", "measures", ".", "We", "begin", "by", "reviewing", "distributional", "reinforcement", "learning", ",", "related", "work", ",", "and", "introducing", "the", "concepts", "surrounding", "risk", "-", "sensitive", "RL", ".", "In", "subsequent", "sections", ",", "we", "introduce", "our", "proposed", "algorithm", ",", "IQN", ",", "and", "present", "a", "series", "of", "experiments", "using", "the", "Atari", "-", "57", "benchmark", ",", "investigating", "the", "robustness", "and", "performance", "of", "IQN", ".", "Despite", "being", "a", "simple", "distributional", "extension", "to", "DQN", ",", "and", "forgoing", "any", "other", "improvements", ",", "IQN", "significantly", "outperforms", "QR", "-", "DQN", "and", "nearly", "matches", "the", "performance", "of", "Rainbow", ",", "which", "combines", "many", "orthogonal", "advances", ".", "In", "fact", ",", "in", "human", "-", "starts", "as", "well", "as", "in", "the", "hardest", "Atari", "games", "(", "where", "current", "RL", "agents", "still", "underperform", "human", "players", ")", "IQN", "improves", "over", "Rainbow", ".", "section", ":", "Background", "/", "Related", "Work", "We", "consider", "the", "standard", "RL", "setting", ",", "in", "which", "the", "interaction", "of", "an", "agent", "and", "an", "environment", "is", "modeled", "as", "a", "Markov", "Decision", "Process", ",", "where", "and", "denote", "the", "state", "and", "action", "spaces", ",", "the", "(", "state", "-", "and", "action", "-", "dependent", ")", "reward", "function", ",", "the", "transition", "kernel", ",", "and", "a", "discount", "factor", ".", "A", "policy", "maps", "a", "state", "to", "a", "distribution", "over", "actions", ".", "For", "an", "agent", "following", "policy", ",", "the", "discounted", "sum", "of", "future", "rewards", "is", "denoted", "by", "the", "random", "variable", ",", "where", ",", ",", ",", "and", ".", "The", "action", "-", "value", "function", "is", "defined", "as", ",", "and", "can", "be", "characterized", "by", "the", "Bellman", "equation", "The", "objective", "in", "RL", "is", "to", "find", "an", "optimal", "policy", ",", "which", "maximizes", ",", "i.e.", "for", "all", "and", "all", ".", "One", "approach", "is", "to", "find", "the", "unique", "fixed", "point", "of", "the", "Bellman", "optimality", "operator", ":", "To", "this", "end", ",", "Q", "-", "learning", "iteratively", "improves", "an", "estimate", ",", ",", "of", "the", "optimal", "action", "-", "value", "function", ",", ",", "by", "repeatedly", "applying", "the", "Bellman", "update", ":", "The", "action", "-", "value", "function", "can", "be", "approximated", "by", "a", "parameterized", "function", "(", "e.g.", "a", "neural", "network", ")", ",", "and", "trained", "by", "minimizing", "the", "squared", "temporal", "difference", "(", "TD", ")", "error", ",", "over", "samples", "observed", "while", "following", "an", "-", "greedy", "policy", "over", ".", "This", "policy", "acts", "greedily", "with", "respect", "to", "with", "probability", "and", "uniformly", "at", "random", "otherwise", ".", "DQN", "uses", "a", "convolutional", "neural", "network", "to", "parameterize", "and", "the", "Q", "-", "learning", "algorithm", "to", "achieve", "human", "-", "level", "play", "on", "the", "Atari", "-", "57", "benchmark", ".", "subsection", ":", "Distributional", "RL", "In", "distributional", "RL", ",", "the", "distribution", "over", "returns", "(", "the", "law", "of", ")", "is", "considered", "instead", "of", "the", "scalar", "value", "function", "that", "is", "its", "expectation", ".", "This", "change", "in", "perspective", "has", "yielded", "new", "insights", "into", "the", "dynamics", "of", "RL", ",", "and", "been", "a", "useful", "tool", "for", "analysis", ".", "Empirically", ",", "distributional", "RL", "algorithms", "show", "improved", "sample", "complexity", "and", "final", "performance", ",", "as", "well", "as", "increased", "robustness", "to", "hyperparameter", "variation", ".", "An", "analogous", "distributional", "Bellman", "equation", "of", "the", "form", "can", "be", "derived", ",", "where", "denotes", "that", "two", "random", "variables", "and", "have", "equal", "probability", "laws", ",", "and", "the", "random", "variables", "and", "are", "distributed", "according", "to", "and", ",", "respectively", ".", "morimura10parametric", "defined", "the", "distributional", "Bellman", "operator", "explicitly", "in", "terms", "of", "conditional", "probabilities", ",", "parameterized", "by", "the", "mean", "and", "scale", "of", "a", "Gaussian", "or", "Laplace", "distribution", ",", "and", "minimized", "the", "Kullback", "-", "Leibler", "(", "KL", ")", "divergence", "between", "the", "Bellman", "target", "and", "the", "current", "estimated", "return", "distribution", ".", "However", ",", "the", "distributional", "Bellman", "operator", "is", "not", "a", "contraction", "in", "the", "KL", ".", "As", "with", "the", "scalar", "setting", ",", "a", "distributional", "Bellman", "optimality", "operator", "can", "be", "defined", "by", "with", "distributed", "according", "to", ".", "While", "the", "distributional", "Bellman", "operator", "for", "policy", "evaluation", "is", "a", "contraction", "in", "the", "-", "Wasserstein", "distance", ",", "this", "no", "longer", "holds", "for", "the", "control", "case", ".", "Convergence", "to", "the", "optimal", "policy", "can", "still", "be", "established", ",", "but", "requires", "a", "more", "involved", "argument", ".", "c51", "parameterize", "the", "return", "distribution", "as", "a", "categorical", "distribution", "over", "a", "fixed", "set", "of", "equidistant", "points", "and", "minimize", "the", "KL", "divergence", "to", "the", "projected", "distributional", "Bellman", "target", ".", "Their", "algorithm", ",", "C51", ",", "outperformed", "previous", "DQN", "variants", "on", "the", "Atari", "-", "57", "benchmark", ".", "Subsequently", ",", "hessel2018rainbow", "combined", "C51", "with", "enhancements", "such", "as", "prioritized", "experience", "replay", ",", "-", "step", "updates", ",", "and", "the", "dueling", "architecture", ",", "leading", "to", "the", "Rainbow", "agent", ",", "current", "state", "-", "of", "-", "the", "-", "art", "in", "Atari", "-", "57", ".", "The", "categorical", "parameterization", ",", "using", "the", "projected", "KL", "loss", ",", "has", "also", "been", "used", "in", "recent", "work", "to", "improve", "the", "critic", "of", "a", "policy", "gradient", "algorithm", ",", "D4PG", ",", "achieving", "significantly", "improved", "robustness", "and", "state", "-", "of", "-", "the", "-", "art", "performance", "across", "a", "variety", "of", "continuous", "control", "tasks", ".", "subsection", ":", "-", "Wasserstein", "Metric", "The", "-", "Wasserstein", "metric", ",", "for", ",", "plays", "a", "key", "role", "in", "recent", "results", "in", "distributional", "RL", ".", "It", "has", "also", "been", "a", "topic", "of", "increasing", "interest", "in", "generative", "modeling", ",", "because", "unlike", "the", "KL", "divergence", ",", "the", "Wasserstein", "metric", "inherently", "trades", "off", "approximate", "solutions", "with", "likelihoods", ".", "The", "-", "Wasserstein", "distance", "is", "the", "metric", "on", "inverse", "cumulative", "distribution", "functions", "(", "c.d.f", ".", ")", ",", "also", "known", "as", "quantile", "functions", ".", "For", "random", "variables", "and", "with", "quantile", "functions", "and", ",", "respectively", ",", "the", "-", "Wasserstein", "distance", "is", "given", "by", "The", "class", "of", "optimal", "transport", "metrics", "express", "distances", "between", "distributions", "in", "terms", "of", "the", "minimal", "cost", "for", "transporting", "mass", "to", "make", "the", "two", "distributions", "identical", ".", "This", "cost", "is", "given", "in", "terms", "of", "some", "metric", ",", ",", "on", "the", "underlying", "space", ".", "The", "-", "Wasserstein", "metric", "corresponds", "to", ".", "We", "are", "particularly", "interested", "in", "the", "Wasserstein", "metrics", "due", "to", "the", "predominant", "use", "of", "spaces", "in", "mean", "-", "value", "reinforcement", "learning", ".", "subsection", ":", "Quantile", "Regression", "for", "Distributional", "RL", "c51", "showed", "that", "the", "distributional", "Bellman", "operator", "is", "a", "contraction", "in", "the", "-", "Wasserstein", "metric", ",", "but", "as", "the", "proposed", "algorithm", "did", "not", "itself", "minimize", "the", "Wasserstein", "metric", ",", "this", "left", "a", "theory", "-", "practice", "gap", "for", "distributional", "RL", ".", "Recently", ",", "this", "gap", "was", "closed", ",", "in", "both", "directions", ".", "First", "and", "most", "relevant", "to", "this", "work", ",", "dabney2017qr", "proposed", "the", "use", "of", "quantile", "regression", "for", "distributional", "RL", "and", "showed", "that", "by", "choosing", "the", "quantile", "targets", "suitably", "the", "resulting", "projected", "distributional", "Bellman", "operator", "is", "a", "contraction", "in", "the", "-", "Wasserstein", "metric", ".", "Concurrently", ",", "rowland2018analysis", "showed", "the", "original", "class", "of", "categorical", "algorithms", "are", "a", "contraction", "in", "the", "Cramér", "distance", ",", "the", "metric", "on", "cumulative", "distribution", "functions", ".", "By", "estimating", "the", "quantile", "function", "at", "precisely", "chosen", "points", ",", "QR", "-", "DQN", "minimizes", "the", "Wasserstein", "distance", "to", "the", "distributional", "Bellman", "target", ".", "This", "estimation", "uses", "quantile", "regression", ",", "which", "has", "been", "shown", "to", "converge", "to", "the", "true", "quantile", "function", "value", "when", "minimized", "using", "stochastic", "approximation", ".", "In", "QR", "-", "DQN", ",", "the", "random", "return", "is", "approximated", "by", "a", "uniform", "mixture", "of", "Diracs", ",", "with", "each", "assigned", "a", "fixed", "quantile", "target", ",", "for", ",", "where", ".", "These", "quantile", "estimates", "are", "trained", "using", "the", "huber1964robust", "quantile", "regression", "loss", ",", "with", "threshold", ",", "on", "the", "pairwise", "TD", "-", "errors", "At", "the", "time", "of", "this", "writing", ",", "QR", "-", "DQN", "achieves", "the", "best", "performance", "on", "Atari", "-", "57", ",", "human", "-", "normalized", "mean", "and", "median", ",", "of", "all", "agents", "that", "do", "not", "combine", "distributional", "RL", ",", "prioritized", "replay", ",", "and", "-", "step", "updates", ".", "subsection", ":", "Risk", "in", "Reinforcement", "Learning", "Distributional", "RL", "algorithms", "have", "been", "theoretically", "justified", "for", "the", "Wasserstein", "and", "Cramér", "metrics", ",", "and", "learning", "the", "distribution", "over", "returns", ",", "in", "and", "of", "itself", ",", "empirically", "results", "in", "significant", "improvements", "to", "data", "efficiency", ",", "final", "performance", ",", "and", "stability", ".", "However", ",", "in", "each", "of", "these", "recent", "works", "the", "policy", "used", "was", "based", "entirely", "on", "the", "mean", "of", "the", "return", "distribution", ",", "just", "as", "in", "standard", "reinforcement", "learning", ".", "A", "natural", "question", "arises", ":", "can", "we", "expand", "the", "class", "of", "policies", "using", "information", "provided", "by", "the", "distribution", "over", "returns", "(", "i.e.", "to", "the", "class", "of", "risk", "-", "sensitive", "policies", ")", "?", "Furthermore", ",", "when", "would", "this", "larger", "policy", "class", "be", "beneficial", "?", "Here", ",", "‘", "risk", "’", "refers", "to", "the", "uncertainty", "over", "possible", "outcomes", ",", "and", "risk", "-", "sensitive", "policies", "are", "those", "which", "depend", "upon", "more", "than", "the", "mean", "of", "the", "outcomes", ".", "At", "this", "point", ",", "it", "is", "important", "to", "highlight", "the", "difference", "between", "intrinsic", "uncertainty", ",", "captured", "by", "the", "distribution", "over", "returns", ",", "and", "parametric", "uncertainty", ",", "the", "uncertainty", "over", "the", "value", "estimate", "typically", "associated", "with", "Bayesian", "approaches", "such", "as", "PSRL", "and", "Kalman", "TD", ".", "Distributional", "RL", "seeks", "to", "capture", "the", "former", ",", "which", "classic", "approaches", "to", "risk", "are", "built", "upon", ".", "Expected", "utility", "theory", "states", "that", "if", "a", "decision", "policy", "is", "consistent", "with", "a", "particular", "set", "of", "four", "axioms", "regarding", "its", "choices", "then", "the", "decision", "policy", "behaves", "as", "though", "it", "is", "maximizing", "the", "expected", "value", "of", "some", "utility", "function", ",", "This", "is", "perhaps", "the", "most", "pervasive", "notion", "of", "risk", "-", "sensitivity", ".", "A", "policy", "maximizing", "a", "linear", "utility", "function", "is", "called", "risk", "-", "neutral", ",", "whereas", "concave", "or", "convex", "utility", "functions", "give", "rise", "to", "risk", "-", "averse", "or", "risk", "-", "seeking", "policies", ",", "respectively", ".", "Many", "previous", "studies", "on", "risk", "-", "sensitive", "RL", "adopt", "the", "utility", "function", "approach", ".", "A", "crucial", "axiom", "of", "expected", "utility", "is", "independence", ":", "given", "random", "variables", ",", "and", ",", "such", "that", "(", "preferred", "over", ")", ",", "any", "mixture", "between", "and", "is", "preferred", "to", "the", "same", "mixture", "between", "and", ".", "Stated", "in", "terms", "of", "the", "cumulative", "probability", "functions", ",", ".", "This", "axiom", "in", "particular", "has", "troubled", "many", "researchers", "because", "it", "is", "consistently", "violated", "by", "human", "behavior", ".", "The", "Allais", "paradox", "is", "a", "frequently", "used", "example", "of", "a", "decision", "problem", "where", "people", "violate", "the", "independence", "axiom", "of", "expected", "utility", "theory", ".", "However", ",", "as", "yaari1987dual", "showed", ",", "this", "axiom", "can", "be", "replaced", "by", "one", "in", "terms", "of", "convex", "combinations", "of", "outcome", "values", ",", "instead", "of", "mixtures", "of", "distributions", ".", "Specifically", ",", "if", "as", "before", ",", "then", "for", "any", "and", "random", "variable", ",", ".", "This", "leads", "to", "an", "alternate", ",", "dual", ",", "theory", "of", "choice", "than", "that", "of", "expected", "utility", ".", "Under", "these", "axioms", "the", "decision", "policy", "behaves", "as", "though", "it", "is", "maximizing", "a", "distorted", "expectation", ",", "for", "some", "continuous", "monotonic", "function", ":", "Such", "a", "function", "is", "known", "as", "a", "distortion", "risk", "measure", ",", "as", "it", "distorts", "the", "cumulative", "probabilities", "of", "the", "random", "variable", ".", "That", "is", ",", "we", "have", "two", "fundamentally", "equivalent", "approaches", "to", "risk", "-", "sensitivity", ".", "Either", ",", "we", "choose", "a", "utility", "function", "and", "follow", "the", "expectation", "of", "this", "utility", ".", "Or", ",", "we", "choose", "a", "reweighting", "of", "the", "distribution", "and", "compute", "expectation", "under", "this", "distortion", "measure", ".", "Indeed", ",", "yaari1987dual", "further", "showed", "that", "these", "two", "functions", "are", "inverses", "of", "each", "other", ".", "The", "choice", "between", "them", "amounts", "to", "a", "choice", "over", "whether", "the", "behavior", "should", "be", "invariant", "to", "mixing", "with", "random", "events", "or", "to", "convex", "combinations", "of", "outcomes", ".", "Distortion", "risk", "measures", "include", ",", "as", "special", "cases", ",", "cumulative", "probability", "weighting", "used", "in", "cumulative", "prospect", "theory", ",", "conditional", "value", "at", "risk", ",", "and", "many", "other", "methods", ".", "Recently", "majumdar2017should", "argued", "for", "the", "use", "of", "distortion", "risk", "measures", "in", "robotics", ".", "section", ":", "Implicit", "Quantile", "Networks", "We", "now", "introduce", "the", "implicit", "quantile", "network", "(", "IQN", ")", ",", "a", "deterministic", "parametric", "function", "trained", "to", "reparameterize", "samples", "from", "a", "base", "distribution", ",", "e.g.", ",", "to", "the", "respective", "quantile", "values", "of", "a", "target", "distribution", ".", "IQN", "provides", "an", "effective", "way", "to", "learn", "an", "implicit", "representation", "of", "the", "return", "distribution", ",", "yielding", "a", "powerful", "function", "approximator", "for", "a", "new", "DQN", "-", "like", "agent", ".", "Let", "be", "the", "quantile", "function", "at", "for", "the", "random", "variable", ".", "For", "notational", "simplicity", "we", "write", ",", "thus", "for", "the", "resulting", "state", "-", "action", "return", "distribution", "sample", "is", ".", "We", "propose", "to", "model", "the", "state", "-", "action", "quantile", "function", "as", "a", "mapping", "from", "state", "-", "actions", "and", "samples", "from", "some", "base", "distribution", ",", "typically", ",", "to", ",", "viewed", "as", "samples", "from", "the", "implicitly", "defined", "return", "distribution", ".", "Let", "be", "a", "distortion", "risk", "measure", ",", "with", "identity", "corresponding", "to", "risk", "-", "neutrality", ".", "Then", ",", "the", "distorted", "expectation", "of", "under", "is", "given", "by", "Notice", "that", "the", "distorted", "expectation", "is", "equal", "to", "the", "expected", "value", "of", "weighted", "by", ",", "that", "is", ",", ".", "The", "immediate", "implication", "of", "this", "is", "that", "for", "any", ",", "there", "exists", "a", "sampling", "distribution", "for", "such", "that", "the", "mean", "of", "is", "equal", "to", "the", "distorted", "expectation", "of", "under", ",", "that", "is", ",", "any", "distorted", "expectation", "can", "be", "represented", "as", "a", "weighted", "sum", "over", "the", "quantiles", ".", "Denote", "by", "the", "risk", "-", "sensitive", "greedy", "policy", "For", "two", "samples", ",", "and", "policy", ",", "the", "sampled", "temporal", "difference", "(", "TD", ")", "error", "at", "step", "is", "Then", ",", "the", "IQN", "loss", "function", "is", "given", "by", "where", "and", "denote", "the", "respective", "number", "of", "iid", "samples", "used", "to", "estimate", "the", "loss", ".", "A", "corresponding", "sample", "-", "based", "risk", "-", "sensitive", "policy", "is", "obtained", "by", "approximating", "in", "Equation", "[", "reference", "]", "by", "samples", "of", ":", "Implicit", "quantile", "networks", "differ", "from", "the", "approach", "of", "dabney2017qr", "in", "two", "ways", ".", "First", ",", "instead", "of", "approximating", "the", "quantile", "function", "at", "fixed", "values", "of", "we", "approximate", "it", "with", "for", "some", "differentiable", "functions", ",", ",", "and", ".", "If", "we", "ignore", "the", "distributional", "interpretation", "for", "a", "moment", "and", "view", "each", "as", "a", "separate", "action", "-", "value", "function", ",", "this", "highlights", "that", "implicit", "quantile", "networks", "are", "a", "type", "of", "universal", "value", "function", "approximator", "(", "UVFA", ")", ".", "There", "may", "be", "additional", "benefits", "to", "implicit", "quantile", "networks", "beyond", "the", "obvious", "increase", "in", "representational", "fidelity", ".", "As", "with", "UVFAs", ",", "we", "might", "hope", "that", "training", "over", "many", "different", "’s", "(", "goals", "in", "the", "case", "of", "the", "UVFA", ")", "leads", "to", "better", "generalization", "between", "values", "and", "improved", "sample", "complexity", "than", "attempting", "to", "train", "each", "separately", ".", "Second", ",", ",", ",", "and", "are", "sampled", "from", "continuous", ",", "independent", ",", "distributions", ".", "Besides", ",", "we", "also", "explore", "risk", "-", "sentive", "policies", ",", "with", "non", "-", "linear", ".", "The", "independent", "sampling", "of", "each", ",", "results", "in", "the", "sample", "TD", "errors", "being", "decorrelated", ",", "and", "the", "estimated", "action", "-", "values", "go", "from", "being", "the", "true", "mean", "of", "a", "mixture", "of", "Diracs", "to", "a", "sample", "mean", "of", "the", "implicit", "distribution", "defined", "by", "reparameterizing", "the", "sampling", "distribution", "via", "the", "learned", "quantile", "function", ".", "subsection", ":", "Implementation", "Consider", "the", "neural", "network", "structure", "used", "by", "the", "DQN", "agent", ".", "Let", "be", "the", "function", "computed", "by", "the", "convolutional", "layers", "and", "the", "subsequent", "fully", "-", "connected", "layers", "mapping", "to", "the", "estimated", "action", "-", "values", ",", "such", "that", ".", "For", "our", "network", "we", "use", "the", "same", "functions", "and", "as", "in", "DQN", ",", "but", "include", "an", "additional", "function", "computing", "an", "embedding", "for", "the", "sample", "point", ".", "We", "combine", "these", "to", "form", "the", "approximation", ",", "where", "denotes", "the", "element", "-", "wise", "(", "Hadamard", ")", "product", ".", "As", "the", "network", "for", "is", "not", "particularly", "deep", ",", "we", "use", "the", "multiplicative", "form", ",", ",", "to", "force", "interaction", "between", "the", "convolutional", "features", "and", "the", "sample", "embedding", ".", "Alternative", "functional", "forms", ",", "e.g.", "concatenation", "or", "a", "‘", "residual", "’", "function", ",", "are", "conceivable", ",", "and", "can", "be", "parameterized", "in", "different", "ways", ".", "To", "investigate", "these", ",", "we", "compared", "performance", "across", "a", "number", "of", "architectural", "variants", "on", "six", "Atari", "2600", "games", "(", "Asterix", ",", "Assault", ",", "Breakout", ",", "Ms", ".", "Pacman", ",", "QBert", ",", "Space", "Invaders", ")", ".", "Full", "results", "are", "given", "in", "the", "Appendix", ".", "Despite", "minor", "variation", "in", "performance", ",", "we", "found", "the", "general", "approach", "to", "be", "robust", "to", "the", "various", "choices", ".", "Based", "upon", "the", "results", "we", "used", "the", "following", "function", "in", "our", "later", "experiments", ",", "for", "embedding", "dimension", ":", "After", "settling", "on", "a", "network", "architecture", ",", "we", "study", "the", "effect", "of", "the", "number", "of", "samples", ",", "and", ",", "used", "in", "the", "estimate", "terms", "of", "Equation", "[", "reference", "]", ".", "We", "hypothesized", "that", ",", "the", "number", "of", "samples", "of", ",", "would", "affect", "the", "sample", "complexity", "of", "IQN", ",", "with", "larger", "values", "leading", "to", "faster", "learning", ",", "and", "that", "with", "one", "would", "potentially", "approach", "the", "performance", "of", "DQN", ".", "This", "would", "support", "the", "hypothesis", "that", "the", "improved", "performance", "of", "many", "distributional", "RL", "algorithms", "rests", "on", "their", "effect", "as", "auxiliary", "loss", "functions", ",", "which", "would", "vanish", "in", "the", "case", "of", ".", "Furthermore", ",", "we", "believed", "that", ",", "the", "number", "of", "samples", "of", ",", "would", "affect", "the", "variance", "of", "the", "gradient", "estimates", "much", "like", "a", "mini", "-", "batch", "size", "hyperparameter", ".", "Our", "prediction", "was", "that", "would", "have", "the", "greatest", "effect", "on", "variance", "of", "the", "long", "-", "term", "performance", "of", "the", "agent", ".", "We", "used", "the", "same", "set", "of", "six", "games", "as", "before", ",", "with", "our", "chosen", "architecture", ",", "and", "varied", ".", "In", "Figure", "[", "reference", "]", "we", "report", "the", "average", "human", "-", "normalized", "scores", "on", "the", "six", "games", "for", "each", "configuration", ".", "Figure", "[", "reference", "]", "(", "left", ")", "shows", "the", "average", "performance", "over", "the", "first", "ten", "million", "frames", ",", "while", "(", "right", ")", "shows", "the", "average", "performance", "over", "the", "last", "ten", "million", "(", "from", "190", "M", "to", "200", "M", ")", ".", "As", "expected", ",", "we", "found", "that", "has", "a", "dramatic", "effect", "on", "early", "performance", ",", "shown", "by", "the", "continual", "improvement", "in", "score", "as", "the", "value", "increases", ".", "Additionally", ",", "we", "observed", "that", "affected", "performance", "very", "differently", "than", "expected", ":", "it", "had", "a", "strong", "effect", "on", "early", "performance", ",", "but", "minimal", "impact", "on", "long", "-", "term", "performance", "past", ".", "Overall", ",", "while", "using", "more", "samples", "for", "both", "distributions", "is", "generally", "favorable", ",", "appears", "to", "be", "sufficient", "to", "achieve", "the", "majority", "of", "improvements", "offered", "by", "IQN", "for", "long", "-", "term", "performance", ",", "with", "variation", "past", "this", "point", "largely", "insignificant", ".", "To", "our", "surprise", "we", "found", "that", "even", "for", ",", "which", "is", "comparable", "to", "DQN", "in", "the", "number", "of", "loss", "components", ",", "the", "longer", "term", "performance", "is", "still", "quite", "strong", "(", "DQN", ")", ".", "In", "an", "informal", "evaluation", ",", "we", "did", "not", "find", "IQN", "to", "be", "sensitive", "to", ",", "the", "number", "of", "samples", "used", "for", "the", "policy", ",", "and", "have", "fixed", "it", "at", "for", "all", "experiments", ".", "section", ":", "Risk", "-", "Sensitive", "Reinforcement", "Learning", "In", "this", "section", ",", "we", "explore", "the", "effects", "of", "varying", "the", "distortion", "risk", "measure", ",", ",", "away", "from", "identity", ".", "This", "only", "affects", "the", "policy", ",", ",", "used", "both", "in", "Equation", "[", "reference", "]", "and", "for", "acting", "in", "the", "environment", ".", "As", "we", "have", "argued", ",", "evaluating", "under", "different", "distortion", "risk", "measures", "is", "equivalent", "to", "changing", "the", "sampling", "distribution", "for", ",", "allowing", "us", "to", "achieve", "various", "forms", "of", "risk", "-", "sensitive", "policies", ".", "We", "focus", "on", "a", "handful", "of", "sampling", "distributions", "and", "their", "corresponding", "distortion", "measures", ".", "The", "first", "one", "is", "the", "cumulative", "probability", "weighting", "parameterization", "proposed", "in", "cumulative", "prospect", "theory", ":", "In", "particular", ",", "we", "use", "the", "parameter", "value", "found", "by", "wu1996curvature", "to", "most", "closely", "match", "human", "subjects", ".", "This", "choice", "is", "interesting", "as", ",", "unlike", "the", "others", "we", "consider", ",", "it", "is", "neither", "globally", "convex", "nor", "concave", ".", "For", "small", "values", "of", "it", "is", "locally", "concave", "and", "for", "larger", "values", "of", "it", "becomes", "locally", "convex", ".", "Recall", "that", "concavity", "corresponds", "to", "risk", "-", "averse", "and", "convexity", "to", "risk", "-", "seeking", "policies", ".", "Second", ",", "we", "consider", "the", "distortion", "risk", "measure", "proposed", "by", "wang2000class", ",", "where", "and", "are", "taken", "to", "be", "the", "standard", "Normal", "cumulative", "distribution", "function", "and", "its", "inverse", ":", "For", ",", "this", "produces", "risk", "-", "averse", "policies", "and", "we", "include", "it", "due", "to", "its", "simple", "interpretation", "and", "ability", "to", "switch", "between", "risk", "-", "averse", "and", "risk", "-", "seeking", "distortions", ".", "Third", ",", "we", "consider", "a", "simple", "power", "formula", "for", "risk", "-", "averse", "(", ")", "or", "risk", "-", "seeking", "(", ")", "policies", ":", "Finally", ",", "we", "consider", "conditional", "value", "-", "at", "-", "risk", "(", "CVaR", ")", ":", "CVaR", "has", "been", "widely", "studied", "in", "and", "out", "of", "reinforcement", "learning", ".", "Its", "implementation", "as", "a", "modification", "to", "the", "sampling", "distribution", "of", "is", "particularly", "simple", ",", "as", "it", "changes", "to", ".", "Another", "interesting", "sampling", "distribution", ",", "not", "included", "in", "our", "experiments", ",", "is", "denoted", "and", "corresponds", "to", "sampled", "by", "averaging", "samples", "from", ".", "In", "Figure", "[", "reference", "]", "(", "right", ")", "we", "give", "an", "example", "of", "a", "distribution", "(", "Neutral", ")", "and", "how", "each", "of", "these", "distortion", "measures", "affects", "the", "implied", "distribution", "due", "to", "changing", "the", "sampling", "distribution", "of", ".", "and", "reduce", "the", "impact", "of", "the", "tails", "of", "the", "distribution", ",", "while", "and", "heavily", "shift", "the", "distribution", "mass", "towards", "the", "tails", ",", "creating", "a", "risk", "-", "averse", "or", "risk", "-", "seeking", "preference", ".", "Additionally", ",", "while", "CVaR", "entirely", "ignores", "all", "values", "corresponding", "to", ",", "gives", "these", "non", "-", "zero", ",", "but", "vanishingly", "small", ",", "probability", ".", "By", "using", "these", "sampling", "distributions", "we", "can", "induce", "various", "risk", "-", "sensitive", "policies", "in", "IQN", ".", "We", "evaluate", "these", "on", "the", "same", "set", "of", "six", "Atari", "2600", "games", "previously", "used", ".", "Our", "algorithm", "simply", "changes", "the", "policy", "to", "maximize", "the", "distorted", "expectations", "instead", "of", "the", "usual", "sample", "mean", ".", "Figure", "[", "reference", "]", "(", "left", ")", "shows", "our", "results", "in", "this", "experiment", ",", "with", "average", "scores", "reported", "under", "the", "usual", ",", "risk", "-", "neutral", ",", "evaluation", "criterion", ".", "Intuitively", ",", "we", "expected", "to", "see", "a", "qualitative", "effect", "from", "risk", "-", "sensitive", "training", ",", "e.g.", "strengthened", "exploration", "from", "a", "risk", "-", "seeking", "objective", ".", "Although", "we", "did", "see", "qualitative", "differences", ",", "these", "did", "not", "always", "match", "our", "expectations", ".", "For", "two", "of", "the", "games", ",", "Asterix", "and", "Assault", ",", "there", "is", "a", "very", "significant", "advantage", "to", "the", "risk", "-", "averse", "policies", ".", "Although", "tends", "to", "perform", "almost", "identically", "to", "the", "standard", "risk", "-", "neutral", "policy", ",", "and", "the", "risk", "-", "seeking", "performs", "as", "well", "or", "worse", "than", "risk", "-", "neutral", ",", "we", "find", "that", "both", "risk", "-", "averse", "policies", "improve", "performance", "over", "standard", "IQN", ".", "However", ",", "we", "also", "observe", "that", "the", "more", "risk", "-", "averse", "of", "the", "two", ",", ",", "suffers", "some", "loss", "in", "performance", "on", "two", "other", "games", "(", "QBert", "and", "Space", "Invaders", ")", ".", "Additionally", ",", "we", "note", "that", "the", "risk", "-", "seeking", "policy", "significantly", "underperforms", "the", "risk", "-", "neutral", "policy", "on", "three", "of", "the", "six", "games", ".", "It", "remains", "an", "open", "question", "as", "to", "exactly", "why", "we", "see", "improved", "performance", "for", "risk", "-", "averse", "policies", ".", "There", "are", "many", "possible", "explanations", "for", "this", "phenomenon", ",", "e.g.", "that", "risk", "-", "aversion", "encodes", "a", "heuristic", "to", "stay", "alive", "longer", ",", "which", "in", "many", "games", "is", "correlated", "with", "increased", "rewards", ".", "section", ":", "Full", "Atari", "-", "57", "Results", "Finally", ",", "we", "evaluate", "IQN", "on", "the", "full", "Atari", "-", "57", "benchmark", ",", "comparing", "with", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "of", "Rainbow", ",", "a", "distributional", "RL", "agent", "that", "combines", "several", "advances", "in", "deep", "RL", ",", "the", "closely", "related", "algorithm", "QR", "-", "DQN", ",", "prioritized", "experience", "replay", "DQN", ",", "and", "the", "original", "DQN", "agent", ".", "Note", "that", "in", "this", "section", "we", "use", "the", "risk", "-", "neutral", "variant", "of", "the", "IQN", ",", "that", "is", ",", "the", "policy", "of", "the", "IQN", "agent", "is", "the", "regular", "-", "greedy", "policy", "with", "respect", "to", "the", "mean", "of", "the", "state", "-", "action", "return", "distribution", ".", "It", "is", "important", "to", "remember", "that", "Rainbow", "builds", "upon", "the", "distributional", "RL", "algorithm", "C51", ",", "but", "also", "includes", "prioritized", "experience", "replay", ",", "Double", "DQN", ",", "Dueling", "Network", "architecture", ",", "Noisy", "Networks", ",", "and", "multi", "-", "step", "updates", ".", "In", "particular", ",", "besides", "the", "distributional", "update", ",", "-", "step", "updates", "and", "prioritized", "experience", "replay", "were", "found", "to", "have", "significant", "impact", "on", "the", "performance", "of", "Rainbow", ".", "Our", "other", "competitive", "baseline", "is", "QR", "-", "DQN", ",", "which", "is", "currently", "state", "-", "of", "-", "the", "-", "art", "for", "agents", "that", "do", "not", "combine", "distributional", "updates", ",", "-", "step", "updates", ",", "and", "prioritized", "replay", ".", "Thus", ",", "between", "QR", "-", "DQN", "and", "the", "much", "more", "complex", "Rainbow", "we", "compare", "to", "the", "two", "most", "closely", "related", ",", "and", "best", "performing", ",", "agents", "in", "published", "work", ".", "In", "particular", ",", "we", "would", "expect", "that", "IQN", "would", "benefit", "from", "the", "additional", "enhancements", "in", "Rainbow", ",", "just", "as", "Rainbow", "improved", "significantly", "over", "C51", ".", "Figure", "[", "reference", "]", "shows", "the", "mean", "(", "left", ")", "and", "median", "(", "right", ")", "human", "-", "normalized", "scores", "during", "training", "over", "the", "Atari", "-", "57", "benchmark", ".", "IQN", "dramatically", "improves", "over", "QR", "-", "DQN", ",", "which", "itself", "improves", "on", "many", "previously", "published", "results", ".", "At", "100", "million", "frames", "IQN", "has", "reached", "the", "same", "level", "of", "performance", "as", "QR", "-", "DQN", "at", "200", "million", "frames", ".", "Table", "[", "reference", "]", "gives", "a", "comparison", "between", "the", "same", "methods", "in", "terms", "of", "their", "best", ",", "human", "-", "normalized", ",", "scores", "per", "game", "under", "the", "30", "random", "no", "-", "op", "start", "condition", ".", "These", "are", "averages", "over", "the", "given", "number", "of", "seeds", ".", "Additionally", ",", "using", "human", "-", "starts", ",", "IQN", "achieves", "median", "human", "-", "normalized", "score", ",", "whereas", "Rainbow", "reaches", ",", "see", "Table", "[", "reference", "]", ".", "Finally", ",", "we", "took", "a", "closer", "look", "at", "the", "games", "in", "which", "each", "algorithm", "continues", "to", "underperform", "humans", ",", "and", "computed", ",", "on", "average", ",", "how", "far", "below", "human", "-", "level", "they", "perform", ".", "We", "refer", "to", "this", "value", "as", "the", "human", "-", "gapThanks", "to", "Joseph", "Modayil", "for", "proposing", "this", "metric", ".", "metric", "and", "give", "results", "in", "Table", "[", "reference", "]", ".", "Interestingly", ",", "C51", "outperforms", "QR", "-", "DQN", "in", "this", "metric", ",", "and", "IQN", "outperforms", "all", "others", ".", "This", "shows", "that", "the", "remaining", "gap", "between", "Rainbow", "and", "IQN", "is", "entirely", "from", "games", "on", "which", "both", "algorithms", "are", "already", "super", "-", "human", ".", "The", "games", "where", "the", "most", "progress", "in", "RL", "is", "needed", "happen", "to", "be", "the", "games", "where", "IQN", "shows", "the", "greatest", "improvement", "over", "QR", "-", "DQN", "and", "Rainbow", ".", "section", ":", "Discussion", "and", "Conclusions", "We", "have", "proposed", "a", "generalization", "of", "recent", "work", "based", "around", "using", "quantile", "regression", "to", "learn", "the", "distribution", "over", "returns", "of", "the", "current", "policy", ".", "Our", "generalization", "leads", "to", "a", "simple", "change", "to", "the", "DQN", "agent", "to", "enable", "distributional", "RL", ",", "the", "natural", "integration", "of", "risk", "-", "sensitive", "policies", ",", "and", "significantly", "improved", "performance", "over", "existing", "methods", ".", "The", "IQN", "algorithm", "provides", ",", "for", "the", "first", "time", ",", "a", "fully", "integrated", "distributional", "RL", "agent", "without", "prior", "assumptions", "on", "the", "parameterization", "of", "the", "return", "distribution", ".", "IQN", "can", "be", "trained", "with", "as", "little", "as", "a", "single", "sample", "from", "each", "state", "-", "action", "value", "distribution", ",", "or", "as", "many", "as", "computational", "limits", "allow", "to", "improve", "the", "algorithm", "’s", "data", "efficiency", ".", "Furthermore", ",", "IQN", "allows", "us", "to", "expand", "the", "class", "of", "control", "policies", "to", "a", "large", "class", "of", "risk", "-", "sensitive", "policies", "connected", "to", "distortion", "risk", "measures", ".", "Finally", ",", "we", "show", "substantial", "gains", "on", "the", "Atari", "-", "57", "benchmark", "over", "QR", "-", "DQN", ",", "and", "even", "halving", "the", "distance", "between", "QR", "-", "DQN", "and", "Rainbow", ".", "Despite", "the", "significant", "empirical", "successes", "in", "this", "paper", "there", "are", "many", "areas", "in", "need", "of", "additional", "theoretical", "analysis", ".", "We", "highlight", "a", "few", "particularly", "relevant", "open", "questions", "we", "were", "unable", "to", "address", "in", "the", "present", "work", ".", "First", ",", "sample", "-", "based", "convergence", "results", "have", "been", "recently", "shown", "for", "a", "class", "of", "categorical", "distributional", "RL", "algorithms", ".", "Could", "existing", "sample", "-", "based", "RL", "convergence", "results", "be", "extended", "to", "the", "QR", "-", "based", "algorithms", "?", "Second", ",", "can", "the", "contraction", "mapping", "results", "for", "a", "fixed", "grid", "of", "quantiles", "given", "by", "dabney2017qr", "be", "extended", "to", "the", "more", "general", "class", "of", "approximate", "quantile", "functions", "studied", "in", "this", "work", "?", "Finally", ",", "and", "particularly", "salient", "to", "our", "experiments", "with", "distortion", "risk", "measures", ",", "theoretical", "guarantees", "for", "risk", "-", "sensitive", "RL", "have", "been", "building", "over", "recent", "years", ",", "but", "have", "been", "largely", "limited", "to", "special", "cases", "and", "restricted", "classes", "of", "risk", "-", "sensitive", "policies", ".", "Can", "the", "convergence", "of", "the", "distribution", "of", "returns", "under", "the", "Bellman", "operator", "be", "leveraged", "to", "show", "convergence", "to", "a", "fixed", "-", "point", "in", "distorted", "expectations", "?", "In", "particular", ",", "can", "the", "control", "results", "of", "c51", "be", "expanded", "to", "cover", "some", "class", "of", "risk", "-", "sensitive", "policies", "?", "There", "remain", "many", "intriguing", "directions", "for", "future", "research", "into", "distributional", "RL", ",", "even", "on", "purely", "empirical", "fronts", ".", "hessel2018rainbow", "recently", "showed", "that", "distributional", "RL", "agents", "can", "be", "significantly", "improved", ",", "when", "combined", "with", "other", "techniques", ".", "Creating", "a", "Rainbow", "-", "IQN", "agent", "could", "yield", "even", "greater", "improvements", "on", "Atari", "-", "57", ".", "We", "also", "recall", "the", "surprisingly", "rich", "return", "distributions", "found", "by", "barthmaron2018d4pg", ",", "and", "hypothesize", "that", "the", "continuous", "control", "setting", "may", "be", "a", "particularly", "fruitful", "area", "for", "the", "application", "of", "distributional", "RL", "in", "general", ",", "and", "IQN", "in", "particular", ".", "bibliography", ":", "References", "section", ":", "Appendix", "subsection", ":", "Architecture", "and", "Hyperparameters", "We", "considered", "multiple", "architectural", "variants", "for", "parameterizing", "an", "IQN", ".", "All", "of", "these", "build", "on", "the", "Q", "-", "network", "of", "a", "regular", "DQN", ",", "which", "can", "be", "seen", "as", "the", "composition", "of", "a", "convolutional", "stack", "and", "an", "MLP", ",", "and", "extend", "it", "by", "an", "embedding", "of", "the", "sample", "point", ",", ",", "and", "a", "merging", "function", ",", "resulting", "in", "the", "function", "For", "the", "embedding", ",", "we", "considered", "a", "number", "of", "variants", ":", "a", "learned", "linear", "embedding", ",", "a", "learned", "MLP", "embedding", "with", "a", "single", "hidden", "layer", "of", "size", ",", "and", "a", "learned", "linear", "function", "of", "cosine", "basis", "functions", "of", "the", "form", ".", "Each", "of", "those", "was", "followed", "by", "either", "a", "ReLU", "or", "sigmoid", "nonlinearity", ".", "For", "the", "merging", "function", ",", "the", "simplest", "choice", "would", "be", "a", "simple", "vector", "concatenation", "of", "and", ".", "Note", "however", ",", "that", "the", "MLP", "which", "takes", "in", "the", "output", "of", "and", "outputs", "the", "action", "-", "value", "quantiles", ",", "only", "has", "a", "single", "hidden", "layer", "in", "the", "DQN", "network", ".", "Therefore", ",", "to", "force", "a", "sufficiently", "early", "interaction", "between", "the", "two", "representations", ",", "we", "also", "considered", "a", "multiplicative", "function", ",", "where", "denotes", "the", "element", "-", "wise", "(", "Hadamard", ")", "product", "of", "two", "vectors", ",", "as", "well", "as", "a", "‘", "residual", "’", "function", ".", "Early", "experiments", "showed", "that", "a", "simple", "linear", "embedding", "of", "was", "insufficient", "to", "achieve", "good", "performance", ",", "and", "the", "residual", "version", "of", "did", "n’t", "show", "any", "marked", "difference", "to", "the", "multiplicative", "variant", ",", "so", "we", "do", "not", "include", "results", "for", "these", "here", ".", "For", "the", "other", "configurations", ",", "Figure", "[", "reference", "]", "shows", "pairwise", "comparisons", "between", "1", ")", "a", "cosine", "basis", "function", "embedding", "and", "a", "completely", "learned", "MLP", "embedding", ",", "2", ")", "an", "embedding", "size", "(", "hidden", "layer", "size", "or", "number", "of", "cosine", "basis", "elements", ")", "32", "and", "64", ",", "3", ")", "ReLU", "and", "sigmoid", "nonlinearity", "following", "the", "embedding", ",", "and", "4", ")", "concatenation", "and", "a", "multiplicative", "interaction", "between", "and", ".", "Each", "comparison", "‘", "violin", "plot", "’", "can", "be", "understood", "as", "a", "marginalization", "over", "the", "other", "variants", "of", "the", "architecture", ",", "with", "the", "human", "-", "normalized", "performance", "at", "the", "end", "of", "training", ",", "averaged", "across", "six", "Atari", "2600", "games", ",", "on", "the", "y", "-", "axis", ".", "Each", "white", "dot", "corresponds", "to", "a", "configuration", "(", "each", "represented", "by", "two", "seeds", ")", ",", "the", "black", "dots", "show", "the", "position", "of", "our", "preferred", "configuration", ".", "The", "width", "of", "the", "colored", "regions", "corresponds", "to", "a", "kernel", "density", "estimate", "of", "the", "number", "of", "configurations", "at", "each", "performance", "level", ".", "Our", "final", "choice", "is", "a", "multiplicative", "interaction", "with", "a", "linear", "function", "of", "a", "cosine", "embedding", ",", "with", "and", "a", "ReLU", "nonlinearity", "(", "see", "Equation", "[", "reference", "]", ")", ",", "as", "this", "configuration", "yielded", "the", "highest", "performance", "consistently", "over", "multiple", "seeds", ".", "Also", "noteworthy", "is", "the", "overall", "robustness", "of", "the", "approach", "to", "these", "variations", ":", "most", "of", "the", "configurations", "consistently", "outperform", "the", "QR", "-", "DQN", "baseline", "shown", "as", "a", "grey", "horizontal", "line", "for", "comparison", ".", "We", "give", "pseudo", "-", "code", "for", "the", "IQN", "loss", "in", "Algorithm", "[", "reference", "]", ".", "All", "other", "hyperparameters", "for", "this", "agent", "correspond", "to", "the", "ones", "used", "by", "dabney2017qr", ".", "In", "particular", ",", "the", "Bellman", "target", "is", "computed", "using", "a", "target", "network", ".", "Notice", "that", "IQN", "will", "generally", "be", "more", "computationally", "expensive", "per", "-", "sample", "than", "QR", "-", "DQN", ".", "However", ",", "in", "practice", "IQN", "requires", "many", "fewer", "samples", "per", "update", "than", "QR", "-", "DQN", "so", "that", "the", "actual", "running", "times", "are", "comparable", ".", "[", "ht", "]", "Implicit", "Quantile", "Network", "Loss", "and", "functions", ",", "#", "Compute", "greedy", "next", "action", "#", "Sample", "quantile", "thresholds", "#", "Compute", "distributional", "temporal", "differences", "#", "Compute", "Huber", "quantile", "loss", "subsection", ":", "Evaluation", "The", "human", "-", "normalized", "scores", "reported", "in", "this", "paper", "are", "given", "by", "the", "formula", "where", ",", "and", "are", "the", "per", "-", "game", "raw", "scores", "(", "undiscounted", "returns", ")", "for", "the", "given", "agent", ",", "a", "reference", "human", "player", ",", "and", "random", "agent", "baseline", ".", "The", "‘", "human", "-", "gap", "’", "metric", "referred", "to", "at", "the", "end", "of", "Section", "[", "reference", "]", "builds", "on", "the", "human", "-", "normalized", "score", ",", "but", "emphasizes", "the", "remaining", "improvement", "for", "the", "agent", "to", "reach", "super", "-", "human", "performance", ".", "It", "is", "given", "by", ",", "with", "a", "value", "of", "corresponding", "to", "random", "play", ",", "and", "a", "value", "of", "corresponding", "to", "super", "-", "human", "level", "of", "performance", ".", "To", "avoid", "degeneracies", "in", "the", "case", "of", ",", "the", "quantity", "is", "being", "clipped", "above", "at", "."], "entity": [[29, 31, ["distributional", "variant"], "Method"], [32, 33, ["DQN"], "Method"], [39, 41, ["quantile", "regression"], "Method"], [79, 83, ["risk", "-", "sensitive", "policies"], "Method"], [91, 94, ["Atari", "2600", "games"], "Material"], [111, 115, ["risk", "-", "sensitive", "policies"], "Task"], [116, 118, ["Atari", "games"], "Task"], [135, 137, ["reinforcement", "learning"], "Method"], [138, 139, ["RL"], "Method"], [165, 168, ["agent", "’s", "representation"], "Method"], [207, 208, ["policy"], "Task"], [220, 225, ["parameterization", "of", "the", "return", "distribution"], "Method"], [228, 230, ["distance", "metric"], "Metric"], [254, 256, ["Categorical", "DQN"], "Method"], [266, 270, ["cross", "-", "entropy", "loss"], "Method"], [272, 276, ["Cramér", "-", "minimizing", "projection"], "Method"], [292, 295, ["mean", "-", "preservation"], "Metric"], [302, 303, ["C51"], "Method"], [308, 309, ["DQN"], "Method"], [314, 317, ["Atari", "2600", "games"], "Material"], [319, 322, ["Arcade", "Learning", "Environment"], "Task"], [358, 361, ["Atari", "-", "57"], "Task"], [364, 367, ["continuous", "control", "tasks"], "Task"], [400, 404, ["uniform", "mixture", "of", "Diracs"], "Method"], [409, 411, ["quantile", "regression"], "Method"], [415, 418, ["QR", "-", "DQN"], "Method"], [449, 452, ["QR", "-", "DQN"], "Method"], [458, 461, ["Atari", "-", "57"], "Task"], [500, 502, ["base", "distribution"], "Method"], [509, 511, ["implicit", "distribution"], "Method"], [526, 529, ["implicit", "quantile", "networks"], "Method"], [539, 545, ["distributional", "generalization", "of", "the", "DQN", "algorithm"], "Method"], [551, 554, ["QR", "-", "DQN"], "Method"], [558, 560, ["approximation", "error"], "Metric"], [594, 595, ["IQN"], "Method"], [614, 616, ["data", "efficiency"], "Metric"], [628, 634, ["implicit", "representation", "of", "the", "return", "distribution"], "Method"], [677, 680, ["distortion", "risk", "measures"], "Metric"], [712, 713, ["IQN"], "Method"], [729, 730, ["robustness"], "Metric"], [733, 734, ["IQN"], "Method"], [739, 741, ["distributional", "extension"], "Method"], [742, 743, ["DQN"], "Method"], [750, 751, ["IQN"], "Method"], [753, 756, ["QR", "-", "DQN"], "Method"], [762, 763, ["Rainbow"], "Method"], [767, 769, ["orthogonal", "advances"], "Method"], [783, 785, ["Atari", "games"], "Material"], [795, 796, ["IQN"], "Method"], [827, 830, ["Markov", "Decision", "Process"], "Method"], [862, 863, ["policy"], "Method"], [874, 877, ["agent", "following", "policy"], "Task"], [912, 914, ["Bellman", "equation"], "Method"], [917, 918, ["RL"], "Method"], [922, 924, ["optimal", "policy"], "Method"], [945, 948, ["Bellman", "optimality", "operator"], "Method"], [953, 956, ["Q", "-", "learning"], "Method"], [975, 977, ["Bellman", "update"], "Method"], [979, 983, ["action", "-", "value", "function"], "Method"], [988, 990, ["parameterized", "function"], "Method"], [993, 995, ["neural", "network"], "Method"], [1002, 1005, ["squared", "temporal", "difference"], "Metric"], [1006, 1007, ["TD"], "Metric"], [1016, 1019, ["-", "greedy", "policy"], "Method"], [1036, 1037, ["DQN"], "Method"], [1039, 1042, ["convolutional", "neural", "network"], "Method"], [1046, 1050, ["Q", "-", "learning", "algorithm"], "Method"], [1105, 1106, ["RL"], "Method"], [1122, 1124, ["sample", "complexity"], "Metric"], [1125, 1127, ["final", "performance"], "Metric"], [1132, 1133, ["robustness"], "Metric"], [1139, 1142, ["distributional", "Bellman", "equation"], "Method"], [1217, 1220, ["estimated", "return", "distribution"], "Method"], [1224, 1227, ["distributional", "Bellman", "operator"], "Method"], [1238, 1240, ["scalar", "setting"], "Task"], [1242, 1246, ["distributional", "Bellman", "optimality", "operator"], "Method"], [1257, 1260, ["distributional", "Bellman", "operator"], "Method"], [1261, 1263, ["policy", "evaluation"], "Task"], [1278, 1280, ["control", "case"], "Task"], [1281, 1282, ["Convergence"], "Task"], [1284, 1286, ["optimal", "policy"], "Method"], [1305, 1307, ["categorical", "distribution"], "Method"], [1329, 1330, ["C51"], "Method"], [1333, 1335, ["DQN", "variants"], "Method"], [1344, 1345, ["hessel2018rainbow"], "Method"], [1346, 1347, ["C51"], "Method"], [1361, 1363, ["dueling", "architecture"], "Method"], [1367, 1369, ["Rainbow", "agent"], "Method"], [1379, 1382, ["Atari", "-", "57"], "Task"], [1384, 1386, ["categorical", "parameterization"], "Method"], [1389, 1392, ["projected", "KL", "loss"], "Method"], [1406, 1409, ["policy", "gradient", "algorithm"], "Method"], [1410, 1411, ["D4PG"], "Method"], [1415, 1416, ["robustness"], "Metric"], [1429, 1432, ["continuous", "control", "tasks"], "Task"], [1439, 1442, ["-", "Wasserstein", "metric"], "Metric"], [1466, 1468, ["generative", "modeling"], "Task"], [1476, 1478, ["Wasserstein", "metric"], "Method"], [1481, 1483, ["approximate", "solutions"], "Method"], [1487, 1490, ["-", "Wasserstein", "distance"], "Method"], [1494, 1498, ["inverse", "cumulative", "distribution", "functions"], "Method"], [1530, 1533, ["optimal", "transport", "metrics"], "Metric"], [1541, 1543, ["minimal", "cost"], "Metric"], [1582, 1584, ["Wasserstein", "metrics"], "Metric"], [1600, 1602, ["Quantile", "Regression"], "Method"], [1609, 1612, ["distributional", "Bellman", "operator"], "Method"], [1614, 1615, ["contraction"], "Method"], [1669, 1671, ["quantile", "regression"], "Method"], [1685, 1689, ["projected", "distributional", "Bellman", "operator"], "Method"], [1694, 1697, ["-", "Wasserstein", "metric"], "Metric"], [1706, 1708, ["categorical", "algorithms"], "Method"], [1710, 1711, ["contraction"], "Method"], [1713, 1715, ["Cramér", "distance"], "Metric"], [1719, 1722, ["cumulative", "distribution", "functions"], "Method"], [1733, 1736, ["QR", "-", "DQN"], "Method"], [1749, 1751, ["quantile", "regression"], "Method"], [1767, 1769, ["stochastic", "approximation"], "Method"], [1771, 1774, ["QR", "-", "DQN"], "Method"], [1782, 1786, ["uniform", "mixture", "of", "Diracs"], "Method"], [1800, 1802, ["quantile", "estimates"], "Method"], [1807, 1810, ["quantile", "regression", "loss"], "Method"], [1827, 1830, ["QR", "-", "DQN"], "Method"], [1839, 1845, ["human", "-", "normalized", "mean", "and", "median"], "Metric"], [1856, 1858, ["prioritized", "replay"], "Method"], [1866, 1867, ["Risk"], "Task"], [1868, 1870, ["Reinforcement", "Learning"], "Task"], [1879, 1883, ["Wasserstein", "and", "Cramér", "metrics"], "Metric"], [1902, 1904, ["data", "efficiency"], "Metric"], [1905, 1907, ["final", "performance"], "Metric"], [1909, 1910, ["stability"], "Metric"], [1930, 1932, ["return", "distribution"], "Method"], [1937, 1939, ["reinforcement", "learning"], "Method"], [1966, 1970, ["risk", "-", "sensitive", "policies"], "Method"], [2044, 2046, ["value", "estimate"], "Method"], [2049, 2051, ["Bayesian", "approaches"], "Method"], [2053, 2054, ["PSRL"], "Method"], [2070, 2071, ["risk"], "Task"], [2075, 2078, ["Expected", "utility", "theory"], "Method"], [2082, 2084, ["decision", "policy"], "Method"], [2098, 2100, ["decision", "policy"], "Method"], [2122, 2125, ["risk", "-", "sensitivity"], "Task"], [2152, 2156, ["risk", "-", "seeking", "policies"], "Task"], [2169, 2172, ["utility", "function", "approach"], "Method"], [2236, 2238, ["Allais", "paradox"], "Material"], [2245, 2247, ["decision", "problem"], "Task"], [2254, 2257, ["expected", "utility", "theory"], "Method"], [2310, 2311, ["choice"], "Task"], [2321, 2323, ["decision", "policy"], "Method"], [2346, 2349, ["distortion", "risk", "measure"], "Metric"], [2371, 2374, ["risk", "-", "sensitivity"], "Task"], [2404, 2406, ["distortion", "measure"], "Metric"], [2449, 2452, ["Distortion", "risk", "measures"], "Metric"], [2458, 2461, ["cumulative", "probability", "weighting"], "Method"], [2463, 2466, ["cumulative", "prospect", "theory"], "Method"], [2467, 2471, ["conditional", "value", "at", "risk"], "Metric"], [2484, 2487, ["distortion", "risk", "measures"], "Metric"], [2488, 2489, ["robotics"], "Task"], [2492, 2495, ["Implicit", "Quantile", "Networks"], "Method"], [2499, 2502, ["implicit", "quantile", "network"], "Method"], [2503, 2504, ["IQN"], "Method"], [2507, 2510, ["deterministic", "parametric", "function"], "Method"], [2531, 2532, ["IQN"], "Method"], [2539, 2545, ["implicit", "representation", "of", "the", "return", "distribution"], "Method"], [2549, 2551, ["function", "approximator"], "Method"], [2554, 2558, ["DQN", "-", "like", "agent"], "Method"], [2580, 2586, ["state", "-", "action", "return", "distribution", "sample"], "Method"], [2593, 2598, ["state", "-", "action", "quantile", "function"], "Method"], [2629, 2632, ["distortion", "risk", "measure"], "Metric"], [2683, 2685, ["sampling", "distribution"], "Method"], [2720, 2725, ["risk", "-", "sensitive", "greedy", "policy"], "Method"], [2746, 2747, ["IQN"], "Method"], [2765, 2766, ["loss"], "Task"], [2769, 2776, ["sample", "-", "based", "risk", "-", "sensitive", "policy"], "Method"], [2789, 2792, ["Implicit", "quantile", "networks"], "Method"], [2849, 2852, ["implicit", "quantile", "networks"], "Method"], [2856, 2860, ["universal", "value", "function", "approximator"], "Method"], [2861, 2862, ["UVFA"], "Method"], [2870, 2873, ["implicit", "quantile", "networks"], "Method"], [2878, 2880, ["representational", "fidelity"], "Metric"], [2883, 2884, ["UVFAs"], "Method"], [2901, 2902, ["UVFA"], "Method"], [2911, 2913, ["sample", "complexity"], "Metric"], [2939, 2943, ["risk", "-", "sentive", "policies"], "Method"], [2998, 3000, ["quantile", "function"], "Method"], [3006, 3009, ["neural", "network", "structure"], "Method"], [3012, 3014, ["DQN", "agent"], "Method"], [3022, 3024, ["convolutional", "layers"], "Method"], [3027, 3031, ["fully", "-", "connected", "layers"], "Method"], [3053, 3054, ["DQN"], "Method"], [3074, 3075, ["approximation"], "Method"], [3079, 3086, ["element", "-", "wise", "(", "Hadamard", ")", "product"], "Method"], [3112, 3114, ["sample", "embedding"], "Method"], [3150, 3152, ["architectural", "variants"], "Method"], [3154, 3157, ["Atari", "2600", "games"], "Material"], [3158, 3159, ["Asterix"], "Material"], [3160, 3161, ["Assault"], "Material"], [3162, 3163, ["Breakout"], "Material"], [3168, 3169, ["QBert"], "Material"], [3170, 3172, ["Space", "Invaders"], "Material"], [3216, 3218, ["embedding", "dimension"], "Metric"], [3223, 3225, ["network", "architecture"], "Method"], [3262, 3264, ["sample", "complexity"], "Metric"], [3265, 3266, ["IQN"], "Method"], [3273, 3274, ["learning"], "Task"], [3285, 3286, ["DQN"], "Method"], [3341, 3346, ["mini", "-", "batch", "size", "hyperparameter"], "Method"], [3530, 3531, ["IQN"], "Method"], [3558, 3559, ["DQN"], "Method"], [3563, 3565, ["loss", "components"], "Method"], [3575, 3576, ["DQN"], "Method"], [3587, 3588, ["IQN"], "Method"], [3613, 3618, ["Risk", "-", "Sensitive", "Reinforcement", "Learning"], "Method"], [3629, 3632, ["distortion", "risk", "measure"], "Metric"], [3667, 3670, ["distortion", "risk", "measures"], "Metric"], [3686, 3690, ["risk", "-", "sensitive", "policies"], "Task"], [3697, 3699, ["sampling", "distributions"], "Method"], [3702, 3704, ["distortion", "measures"], "Metric"], [3710, 3714, ["cumulative", "probability", "weighting", "parameterization"], "Method"], [3716, 3719, ["cumulative", "prospect", "theory"], "Method"], [3787, 3791, ["risk", "-", "seeking", "policies"], "Task"], [3797, 3800, ["distortion", "risk", "measure"], "Metric"], [3812, 3816, ["Normal", "cumulative", "distribution", "function"], "Method"], [3857, 3859, ["power", "formula"], "Method"], [3860, 3872, ["risk", "-", "averse", "(", ")", "or", "risk", "-", "seeking", "(", ")", "policies"], "Task"], [3877, 3883, ["conditional", "value", "-", "at", "-", "risk"], "Method"], [3884, 3885, ["CVaR"], "Method"], [3887, 3888, ["CVaR"], "Method"], [3896, 3898, ["reinforcement", "learning"], "Method"], [3920, 3922, ["sampling", "distribution"], "Method"], [3963, 3965, ["distortion", "measures"], "Metric"], [4013, 4014, ["CVaR"], "Method"], [4036, 4038, ["sampling", "distributions"], "Method"], [4042, 4046, ["risk", "-", "sensitive", "policies"], "Task"], [4047, 4048, ["IQN"], "Method"], [4058, 4061, ["Atari", "2600", "games"], "Material"], [4108, 4110, ["evaluation", "criterion"], "Metric"], [4121, 4125, ["risk", "-", "sensitive", "training"], "Method"], [4128, 4129, ["exploration"], "Task"], [4157, 4158, ["Asterix"], "Material"], [4159, 4160, ["Assault"], "Material"], [4169, 4173, ["risk", "-", "averse", "policies"], "Method"], [4183, 4187, ["risk", "-", "neutral", "policy"], "Method"], [4190, 4193, ["risk", "-", "seeking"], "Method"], [4207, 4211, ["risk", "-", "averse", "policies"], "Method"], [4215, 4216, ["IQN"], "Method"], [4243, 4244, ["QBert"], "Material"], [4245, 4247, ["Space", "Invaders"], "Material"], [4255, 4259, ["risk", "-", "seeking", "policy"], "Method"], [4262, 4266, ["risk", "-", "neutral", "policy"], "Method"], [4287, 4291, ["risk", "-", "averse", "policies"], "Method"], [4335, 4336, ["IQN"], "Method"], [4356, 4357, ["Rainbow"], "Method"], [4374, 4377, ["QR", "-", "DQN"], "Method"], [4378, 4382, ["prioritized", "experience", "replay", "DQN"], "Method"], [4386, 4388, ["DQN", "agent"], "Method"], [4409, 4410, ["policy"], "Method"], [4416, 4420, ["regular", "-", "greedy", "policy"], "Method"], [4439, 4440, ["Rainbow"], "Method"], [4451, 4454, ["prioritized", "experience", "replay"], "Method"], [4455, 4457, ["Double", "DQN"], "Method"], [4458, 4461, ["Dueling", "Network", "architecture"], "Method"], [4462, 4464, ["Noisy", "Networks"], "Method"], [4466, 4470, ["multi", "-", "step", "updates"], "Method"], [4496, 4497, ["Rainbow"], "Task"], [4503, 4506, ["QR", "-", "DQN"], "Method"], [4537, 4540, ["QR", "-", "DQN"], "Method"], [4571, 4572, ["IQN"], "Method"], [4583, 4584, ["Rainbow"], "Method"], [4604, 4608, ["human", "-", "normalized", "scores"], "Metric"], [4617, 4618, ["IQN"], "Method"], [4621, 4624, ["QR", "-", "DQN"], "Method"], [4638, 4639, ["IQN"], "Method"], [4647, 4650, ["QR", "-", "DQN"], "Method"], [4672, 4677, ["human", "-", "normalized", ",", "scores"], "Metric"], [4706, 4707, ["IQN"], "Method"], [4709, 4713, ["human", "-", "normalized", "score"], "Metric"], [4715, 4716, ["Rainbow"], "Method"], [4788, 4789, ["C51"], "Method"], [4790, 4793, ["QR", "-", "DQN"], "Method"], [4798, 4799, ["IQN"], "Method"], [4810, 4811, ["Rainbow"], "Method"], [4812, 4813, ["IQN"], "Method"], [4834, 4835, ["RL"], "Method"], [4843, 4844, ["IQN"], "Method"], [4849, 4852, ["QR", "-", "DQN"], "Method"], [4853, 4854, ["Rainbow"], "Method"], [4871, 4873, ["quantile", "regression"], "Method"], [4893, 4895, ["DQN", "agent"], "Method"], [4901, 4908, ["natural", "integration", "of", "risk", "-", "sensitive", "policies"], "Task"], [4944, 4945, ["IQN"], "Method"], [4975, 4977, ["data", "efficiency"], "Metric"], [4980, 4981, ["IQN"], "Method"], [4988, 4990, ["control", "policies"], "Method"], [4995, 4999, ["risk", "-", "sensitive", "policies"], "Method"], [5001, 5004, ["distortion", "risk", "measures"], "Metric"], [5018, 5021, ["QR", "-", "DQN"], "Method"], [5026, 5027, ["distance"], "Metric"], [5028, 5031, ["QR", "-", "DQN"], "Method"], [5032, 5033, ["Rainbow"], "Method"], [5103, 5107, ["QR", "-", "based", "algorithms"], "Method"], [5112, 5114, ["contraction", "mapping"], "Method"], [5149, 5152, ["distortion", "risk", "measures"], "Metric"], [5179, 5183, ["risk", "-", "sensitive", "policies"], "Method"], [5194, 5196, ["Bellman", "operator"], "Method"], [5218, 5219, ["c51"], "Method"], [5226, 5230, ["risk", "-", "sensitive", "policies"], "Task"], [5279, 5282, ["Atari", "-", "57"], "Task"], [5299, 5302, ["continuous", "control", "setting"], "Task"], [5318, 5319, ["IQN"], "Method"], [5336, 5338, ["architectural", "variants"], "Method"], [5341, 5342, ["IQN"], "Method"], [5349, 5352, ["Q", "-", "network"], "Method"], [5354, 5356, ["regular", "DQN"], "Method"], [5366, 5368, ["convolutional", "stack"], "Method"], [5370, 5371, ["MLP"], "Method"], [5395, 5396, ["embedding"], "Task"], [5405, 5408, ["learned", "linear", "embedding"], "Method"], [5411, 5413, ["MLP", "embedding"], "Method"], [5442, 5446, ["ReLU", "or", "sigmoid", "nonlinearity"], "Method"], [5449, 5451, ["merging", "function"], "Method"], [5459, 5463, ["vector", "concatenation", "of", "and"], "Method"], [5469, 5470, ["MLP"], "Method"], [5488, 5490, ["hidden", "layer"], "Method"], [5492, 5494, ["DQN", "network"], "Method"], [5544, 5547, ["linear", "embedding", "of"], "Method"], [5556, 5558, ["residual", "version"], "Method"], [5567, 5569, ["multiplicative", "variant"], "Method"], [5596, 5600, ["cosine", "basis", "function", "embedding"], "Method"], [5604, 5606, ["MLP", "embedding"], "Method"], [5683, 5686, ["Atari", "2600", "games"], "Material"], [5728, 5731, ["kernel", "density", "estimate"], "Method"], [5754, 5756, ["cosine", "embedding"], "Method"], [5802, 5806, ["QR", "-", "DQN", "baseline"], "Method"], [5822, 5824, ["IQN", "loss"], "Method"], [5854, 5856, ["target", "network"], "Method"], [5859, 5860, ["IQN"], "Method"], [5870, 5873, ["QR", "-", "DQN"], "Method"], [5878, 5879, ["IQN"], "Method"], [5886, 5889, ["QR", "-", "DQN"], "Method"], [5893, 5895, ["running", "times"], "Metric"], [5902, 5907, ["Quantile", "Network", "Loss", "and", "functions"], "Method"], [5910, 5913, ["greedy", "next", "action"], "Method"], [5924, 5927, ["Huber", "quantile", "loss"], "Method"], [5931, 5935, ["human", "-", "normalized", "scores"], "Metric"], [5969, 5972, ["random", "agent", "baseline"], "Method"], [5975, 5980, ["human", "-", "gap", "’", "metric"], "Metric"], [5993, 5997, ["human", "-", "normalized", "score"], "Metric"], [11, 13, ["reinforcement", "learning"], "Method"], [124, 126, ["reinforcement", "learning"], "Method"], [180, 181, ["RL"], "Method"], [211, 212, ["RL"], "Method"], [329, 332, ["Atari", "-", "57"], "Task"], [530, 531, ["IQN"], "Method"], [686, 688, ["reinforcement", "learning"], "Method"], [700, 701, ["RL"], "Method"], [722, 725, ["Atari", "-", "57"], "Task"], [788, 789, ["RL"], "Method"], [810, 811, ["RL"], "Method"], [1058, 1061, ["Atari", "-", "57"], "Task"], [1066, 1067, ["RL"], "Method"], [1069, 1070, ["RL"], "Method"], [1118, 1119, ["RL"], "Method"], [1337, 1340, ["Atari", "-", "57"], "Task"], [1454, 1455, ["RL"], "Method"], [1595, 1597, ["reinforcement", "learning"], "Method"], [1604, 1605, ["RL"], "Method"], [1643, 1644, ["RL"], "Method"], [1673, 1674, ["RL"], "Method"], [1817, 1818, ["TD"], "Metric"], [1835, 1838, ["Atari", "-", "57"], "Task"], [1854, 1855, ["RL"], "Method"], [1871, 1872, ["RL"], "Method"], [2056, 2057, ["TD"], "Metric"], [2059, 2060, ["RL"], "Method"], [2166, 2167, ["RL"], "Method"], [2737, 2738, ["TD"], "Metric"], [2959, 2960, ["TD"], "Metric"], [3299, 3300, ["RL"], "Method"], [4327, 4330, ["Atari", "-", "57"], "Task"], [4339, 4342, ["Atari", "-", "57"], "Task"], [4360, 4361, ["RL"], "Method"], [4368, 4369, ["RL"], "Method"], [4403, 4404, ["IQN"], "Method"], [4412, 4413, ["IQN"], "Method"], [4444, 4445, ["RL"], "Method"], [4612, 4615, ["Atari", "-", "57"], "Task"], [4898, 4899, ["RL"], "Method"], [4918, 4919, ["IQN"], "Method"], [4931, 4932, ["RL"], "Method"], [5013, 5016, ["Atari", "-", "57"], "Task"], [5088, 5089, ["RL"], "Method"], [5096, 5097, ["RL"], "Method"], [5159, 5160, ["RL"], "Method"], [5241, 5242, ["RL"], "Method"], [5254, 5255, ["RL"], "Method"], [5271, 5272, ["IQN"], "Method"], [5313, 5314, ["RL"], "Method"]]} {"doc_id": "23d2d3a6ffebfecaa8930307fdcf451c147757c8", "text": ["document", ":", "SeqGAN", ":", "Sequence", "Generative", "Adversarial", "Nets", "with", "Policy", "Gradient", "As", "a", "new", "way", "of", "training", "generative", "models", ",", "Generative", "Adversarial", "Net", "(", "GAN", ")", "that", "uses", "a", "discriminative", "model", "to", "guide", "the", "training", "of", "the", "generative", "model", "has", "enjoyed", "considerable", "success", "in", "generating", "real", "-", "valued", "data", ".", "However", ",", "it", "has", "limitations", "when", "the", "goal", "is", "for", "generating", "sequences", "of", "discrete", "tokens", ".", "A", "major", "reason", "lies", "in", "that", "the", "discrete", "outputs", "from", "the", "generative", "model", "make", "it", "difficult", "to", "pass", "the", "gradient", "update", "from", "the", "discriminative", "model", "to", "the", "generative", "model", ".", "Also", ",", "the", "discriminative", "model", "can", "only", "assess", "a", "complete", "sequence", ",", "while", "for", "a", "partially", "generated", "sequence", ",", "it", "is", "non", "-", "trivial", "to", "balance", "its", "current", "score", "and", "the", "future", "one", "once", "the", "entire", "sequence", "has", "been", "generated", ".", "In", "this", "paper", ",", "we", "propose", "a", "sequence", "generation", "framework", ",", "called", "SeqGAN", ",", "to", "solve", "the", "problems", ".", "Modeling", "the", "data", "generator", "as", "a", "stochastic", "policy", "in", "reinforcement", "learning", "(", "RL", ")", ",", "SeqGAN", "bypasses", "the", "generator", "differentiation", "problem", "by", "directly", "performing", "gradient", "policy", "update", ".", "The", "RL", "reward", "signal", "comes", "from", "the", "GAN", "discriminator", "judged", "on", "a", "complete", "sequence", ",", "and", "is", "passed", "back", "to", "the", "intermediate", "state", "-", "action", "steps", "using", "Monte", "Carlo", "search", ".", "Extensive", "experiments", "on", "synthetic", "data", "and", "real", "-", "world", "tasks", "demonstrate", "significant", "improvements", "over", "strong", "baselines", ".", "section", ":", "Introduction", "Generating", "sequential", "synthetic", "data", "that", "mimics", "the", "real", "one", "is", "an", "important", "problem", "in", "unsupervised", "learning", ".", "Recently", ",", "recurrent", "neural", "networks", "(", "RNNs", ")", "with", "long", "short", "-", "term", "memory", "(", "LSTM", ")", "cells", "have", "shown", "excellent", "performance", "ranging", "from", "natural", "language", "generation", "to", "handwriting", "generation", ".", "The", "most", "common", "approach", "to", "training", "an", "RNN", "is", "to", "maximize", "the", "log", "predictive", "likelihood", "of", "each", "true", "token", "in", "the", "training", "sequence", "given", "the", "previous", "observed", "tokens", ".", "However", ",", "as", "argued", "in", ",", "the", "maximum", "likelihood", "approaches", "suffer", "from", "so", "-", "called", "exposure", "bias", "in", "the", "inference", "stage", ":", "the", "model", "generates", "a", "sequence", "iteratively", "and", "predicts", "next", "token", "conditioned", "on", "its", "previously", "predicted", "ones", "that", "may", "be", "never", "observed", "in", "the", "training", "data", ".", "Such", "a", "discrepancy", "between", "training", "and", "inference", "can", "incur", "accumulatively", "along", "with", "the", "sequence", "and", "will", "become", "prominent", "as", "the", "length", "of", "sequence", "increases", ".", "To", "address", "this", "problem", ",", "proposed", "a", "training", "strategy", "called", "scheduled", "sampling", "(", "SS", ")", ",", "where", "the", "generative", "model", "is", "partially", "fed", "with", "its", "own", "synthetic", "data", "as", "prefix", "(", "observed", "tokens", ")", "rather", "than", "the", "true", "data", "when", "deciding", "the", "next", "token", "in", "the", "training", "stage", ".", "Nevertheless", ",", "showed", "that", "SS", "is", "an", "inconsistent", "training", "strategy", "and", "fails", "to", "address", "the", "problem", "fundamentally", ".", "Another", "possible", "solution", "of", "the", "training", "/", "inference", "discrepancy", "problem", "is", "to", "build", "the", "loss", "function", "on", "the", "entire", "generated", "sequence", "instead", "of", "each", "transition", ".", "For", "instance", ",", "in", "the", "application", "of", "machine", "translation", ",", "a", "task", "specific", "sequence", "score", "/", "loss", ",", "bilingual", "evaluation", "understudy", "(", "BLEU", ")", ",", "can", "be", "adopted", "to", "guide", "the", "sequence", "generation", ".", "However", ",", "in", "many", "other", "practical", "applications", ",", "such", "as", "poem", "generation", "and", "chatbot", ",", "a", "task", "specific", "loss", "may", "not", "be", "directly", "available", "to", "score", "a", "generated", "sequence", "accurately", ".", "General", "adversarial", "net", "(", "GAN", ")", "proposed", "by", "is", "a", "promising", "framework", "for", "alleviating", "the", "above", "problem", ".", "Specifically", ",", "in", "GAN", "a", "discriminative", "net", "learns", "to", "distinguish", "whether", "a", "given", "data", "instance", "is", "real", "or", "not", ",", "and", "a", "generative", "net", "learns", "to", "confuse", "by", "generating", "high", "quality", "data", ".", "This", "approach", "has", "been", "successful", "and", "been", "mostly", "applied", "in", "computer", "vision", "tasks", "of", "generating", "samples", "of", "natural", "images", ".", "Unfortunately", ",", "applying", "GAN", "to", "generating", "sequences", "has", "two", "problems", ".", "Firstly", ",", "GAN", "is", "designed", "for", "generating", "real", "-", "valued", ",", "continuous", "data", "but", "has", "difficulties", "in", "directly", "generating", "sequences", "of", "discrete", "tokens", ",", "such", "as", "texts", ".", "The", "reason", "is", "that", "in", "GANs", ",", "the", "generator", "starts", "with", "random", "sampling", "first", "and", "then", "a", "determistic", "transform", ",", "govermented", "by", "the", "model", "parameters", ".", "As", "such", ",", "the", "gradient", "of", "the", "loss", "from", "w.r.t", ".", "the", "outputs", "by", "is", "used", "to", "guide", "the", "generative", "model", "(", "paramters", ")", "to", "slightly", "change", "the", "generated", "value", "to", "make", "it", "more", "realistic", ".", "If", "the", "generated", "data", "is", "based", "on", "discrete", "tokens", ",", "the", "“", "slight", "change", "”", "guidance", "from", "the", "discriminative", "net", "makes", "little", "sense", "because", "there", "is", "probably", "no", "corresponding", "token", "for", "such", "slight", "change", "in", "the", "limited", "dictionary", "space", ".", "Secondly", ",", "GAN", "can", "only", "give", "the", "score", "/", "loss", "for", "an", "entire", "sequence", "when", "it", "has", "been", "generated", ";", "for", "a", "partially", "generated", "sequence", ",", "it", "is", "non", "-", "trivial", "to", "balance", "how", "good", "as", "it", "is", "now", "and", "the", "future", "score", "as", "the", "entire", "sequence", ".", "In", "this", "paper", ",", "to", "address", "the", "above", "two", "issues", ",", "we", "follow", "and", "consider", "the", "sequence", "generation", "procedure", "as", "a", "sequential", "decision", "making", "process", ".", "The", "generative", "model", "is", "treated", "as", "an", "agent", "of", "reinforcement", "learning", "(", "RL", ")", ";", "the", "state", "is", "the", "generated", "tokens", "so", "far", "and", "the", "action", "is", "the", "next", "token", "to", "be", "generated", ".", "Unlike", "the", "work", "in", "that", "requires", "a", "task", "-", "specific", "sequence", "score", ",", "such", "as", "BLEU", "in", "machine", "translation", ",", "to", "give", "the", "reward", ",", "we", "employ", "a", "discriminator", "to", "evaluate", "the", "sequence", "and", "feedback", "the", "evaluation", "to", "guide", "the", "learning", "of", "the", "generative", "model", ".", "To", "solve", "the", "problem", "that", "the", "gradient", "can", "not", "pass", "back", "to", "the", "generative", "model", "when", "the", "output", "is", "discrete", ",", "we", "regard", "the", "generative", "model", "as", "a", "stochastic", "parametrized", "policy", ".", "In", "our", "policy", "gradient", ",", "we", "employ", "Monte", "Carlo", "(", "MC", ")", "search", "to", "approximate", "the", "state", "-", "action", "value", ".", "We", "directly", "train", "the", "policy", "(", "generative", "model", ")", "via", "policy", "gradient", ",", "which", "naturally", "avoids", "the", "differentiation", "difficulty", "for", "discrete", "data", "in", "a", "conventional", "GAN", ".", "Extensive", "experiments", "based", "on", "synthetic", "and", "real", "data", "are", "conducted", "to", "investigate", "the", "efficacy", "and", "properties", "of", "the", "proposed", "SeqGAN", ".", "In", "our", "synthetic", "data", "environment", ",", "SeqGAN", "significantly", "outperforms", "the", "maximum", "likelihood", "methods", ",", "scheduled", "sampling", "and", "PG", "-", "BLEU", ".", "In", "three", "real", "-", "world", "tasks", ",", "i.e.", "poem", "generation", ",", "speech", "language", "generation", "and", "music", "generation", ",", "SeqGAN", "significantly", "outperforms", "the", "compared", "baselines", "in", "various", "metrics", "including", "human", "expert", "judgement", ".", "section", ":", "Related", "Work", "Deep", "generative", "models", "have", "recently", "drawn", "significant", "attention", ",", "and", "the", "ability", "of", "learning", "over", "large", "(", "unlabeled", ")", "data", "endows", "them", "with", "more", "potential", "and", "vitality", ".", "first", "proposed", "to", "use", "the", "contrastive", "divergence", "algorithm", "to", "efficiently", "training", "deep", "belief", "nets", "(", "DBN", ")", ".", "proposed", "denoising", "autoencoder", "(", "DAE", ")", "that", "learns", "the", "data", "distribution", "in", "a", "supervised", "learning", "fashion", ".", "Both", "DBN", "and", "DAE", "learn", "a", "low", "dimensional", "representation", "(", "encoding", ")", "for", "each", "data", "instance", "and", "generate", "it", "from", "a", "decoding", "network", ".", "Recently", ",", "variational", "autoencoder", "(", "VAE", ")", "that", "combines", "deep", "learning", "with", "statistical", "inference", "intended", "to", "represent", "a", "data", "instance", "in", "a", "latent", "hidden", "space", ",", "while", "still", "utilizing", "(", "deep", ")", "neural", "networks", "for", "non", "-", "linear", "mapping", ".", "The", "inference", "is", "done", "via", "variational", "methods", ".", "All", "these", "generative", "models", "are", "trained", "by", "maximizing", "(", "the", "lower", "bound", "of", ")", "training", "data", "likelihood", ",", "which", ",", "as", "mentioned", "by", ",", "suffers", "from", "the", "difficulty", "of", "approximating", "intractable", "probabilistic", "computations", ".", "proposed", "an", "alternative", "training", "methodology", "to", "generative", "models", ",", "i.e.", "GANs", ",", "where", "the", "training", "procedure", "is", "a", "minimax", "game", "between", "a", "generative", "model", "and", "a", "discriminative", "model", ".", "This", "framework", "bypasses", "the", "difficulty", "of", "maximum", "likelihood", "learning", "and", "has", "gained", "striking", "successes", "in", "natural", "image", "generation", ".", "However", ",", "little", "progress", "has", "been", "made", "in", "applying", "GANs", "to", "sequence", "discrete", "data", "generation", "problems", ",", "e.g.", "natural", "language", "generation", ".", "This", "is", "due", "to", "the", "generator", "network", "in", "GAN", "is", "designed", "to", "be", "able", "to", "adjust", "the", "output", "continuously", ",", "which", "does", "not", "work", "on", "discrete", "data", "generation", ".", "On", "the", "other", "hand", ",", "a", "lot", "of", "efforts", "have", "been", "made", "to", "generate", "structured", "sequences", ".", "Recurrent", "neural", "networks", "can", "be", "trained", "to", "produce", "sequences", "of", "tokens", "in", "many", "applications", "such", "as", "machine", "translation", ".", "The", "most", "popular", "way", "of", "training", "RNNs", "is", "to", "maximize", "the", "likelihood", "of", "each", "token", "in", "the", "training", "data", "whereas", "pointed", "out", "that", "the", "discrepancy", "between", "training", "and", "generating", "makes", "the", "maximum", "likelihood", "estimation", "suboptimal", "and", "proposed", "scheduled", "sampling", "strategy", "(", "SS", ")", ".", "Later", "theorized", "that", "the", "objective", "function", "underneath", "SS", "is", "improper", "and", "explained", "the", "reason", "why", "GANs", "tend", "to", "generate", "natural", "-", "looking", "samples", "in", "theory", ".", "Consequently", ",", "the", "GANs", "have", "great", "potential", "but", "are", "not", "practically", "feasible", "to", "discrete", "probabilistic", "models", "currently", ".", "As", "pointed", "out", "by", ",", "the", "sequence", "data", "generation", "can", "be", "formulated", "as", "a", "sequential", "decision", "making", "process", ",", "which", "can", "be", "potentially", "be", "solved", "by", "reinforcement", "learning", "techniques", ".", "Modeling", "the", "sequence", "generator", "as", "a", "policy", "of", "picking", "the", "next", "token", ",", "policy", "gradient", "methods", "can", "be", "adopted", "to", "optimize", "the", "generator", "once", "there", "is", "an", "(", "implicit", ")", "reward", "function", "to", "guide", "the", "policy", ".", "For", "most", "practical", "sequence", "generation", "tasks", ",", "e.g.", "machine", "translation", ",", "the", "reward", "signal", "is", "meaningful", "only", "for", "the", "entire", "sequence", ",", "for", "instance", "in", "the", "game", "of", "Go", ",", "the", "reward", "signal", "is", "only", "set", "at", "the", "end", "of", "the", "game", ".", "In", "those", "cases", ",", "state", "-", "action", "evaluation", "methods", "such", "as", "Monte", "Carlo", "(", "tree", ")", "search", "have", "been", "adopted", ".", "By", "contract", ",", "our", "proposed", "SeqGAN", "extends", "GANs", "with", "the", "RL", "-", "based", "generator", "to", "solve", "the", "sequence", "generation", "problem", ",", "where", "a", "reward", "signal", "is", "provided", "by", "the", "discriminator", "at", "the", "end", "of", "each", "episode", "via", "Monte", "Carlo", "approach", ",", "and", "the", "generator", "picks", "the", "action", "and", "learns", "the", "policy", "using", "estimated", "overall", "rewards", ".", "section", ":", "Sequence", "Generative", "Adversarial", "Nets", "The", "sequence", "generation", "problem", "is", "denoted", "as", "follows", ".", "Given", "a", "dataset", "of", "real", "-", "world", "structured", "sequences", ",", "train", "a", "-", "parameterized", "generative", "model", "to", "produce", "a", "sequence", ",", "where", "is", "the", "vocabulary", "of", "candidate", "tokens", ".", "We", "interpret", "this", "problem", "based", "on", "reinforcement", "learning", ".", "In", "timestep", ",", "the", "state", "is", "the", "current", "produced", "tokens", "and", "the", "action", "is", "the", "next", "token", "to", "select", ".", "Thus", "the", "policy", "model", "is", "stochastic", ",", "whereas", "the", "state", "transition", "is", "deterministic", "after", "an", "action", "has", "been", "chosen", ",", "i.e.", "for", "the", "next", "state", "if", "the", "current", "state", "and", "the", "action", ";", "for", "other", "next", "states", ",", ".", "Additionally", ",", "we", "also", "train", "a", "-", "parameterized", "discriminative", "model", "to", "provide", "a", "guidance", "for", "improving", "generator", ".", "is", "a", "probability", "indicating", "how", "likely", "a", "sequence", "is", "from", "real", "sequence", "data", "or", "not", ".", "As", "illustrated", "in", "Figure", "[", "reference", "]", ",", "the", "discriminative", "model", "is", "trained", "by", "providing", "positive", "examples", "from", "the", "real", "sequence", "data", "and", "negative", "examples", "from", "the", "synthetic", "sequences", "generated", "from", "the", "generative", "model", ".", "At", "the", "same", "time", ",", "the", "generative", "model", "is", "updated", "by", "employing", "a", "policy", "gradient", "and", "MC", "search", "on", "the", "basis", "of", "the", "expected", "end", "reward", "received", "from", "the", "discriminative", "model", ".", "The", "reward", "is", "estimated", "by", "the", "likelihood", "that", "it", "would", "fool", "the", "discriminative", "model", ".", "The", "specific", "formulation", "is", "given", "in", "the", "next", "subsection", ".", "subsection", ":", "SeqGAN", "via", "Policy", "Gradient", "Following", ",", "when", "there", "is", "no", "intermediate", "reward", ",", "the", "objective", "of", "the", "generator", "model", "(", "policy", ")", "is", "to", "generate", "a", "sequence", "from", "the", "start", "state", "to", "maximize", "its", "expected", "end", "reward", ":", "where", "is", "the", "reward", "for", "a", "complete", "sequence", ".", "Note", "that", "the", "reward", "is", "from", "the", "discriminator", ",", "which", "we", "will", "discuss", "later", ".", "is", "the", "action", "-", "value", "function", "of", "a", "sequence", ",", "i.e.", "the", "expected", "accumulative", "reward", "starting", "from", "state", ",", "taking", "action", ",", "and", "then", "following", "policy", ".", "The", "rational", "of", "the", "objective", "function", "for", "a", "sequence", "is", "that", "starting", "from", "a", "given", "initial", "state", ",", "the", "goal", "of", "the", "generator", "is", "to", "generate", "a", "sequence", "which", "would", "make", "the", "discriminator", "consider", "it", "is", "real", ".", "The", "next", "question", "is", "how", "to", "estimate", "the", "action", "-", "value", "function", ".", "In", "this", "paper", ",", "we", "use", "the", "REINFORCE", "algorithm", "and", "consider", "the", "estimated", "probability", "of", "being", "real", "by", "the", "discriminator", "as", "the", "reward", ".", "Formally", ",", "we", "have", ":", "However", ",", "the", "discriminator", "only", "provides", "a", "reward", "value", "for", "a", "finished", "sequence", ".", "Since", "we", "actually", "care", "about", "the", "long", "-", "term", "reward", ",", "at", "every", "timestep", ",", "we", "should", "not", "only", "consider", "the", "fitness", "of", "previous", "tokens", "(", "prefix", ")", "but", "also", "the", "resulted", "future", "outcome", ".", "This", "is", "similar", "to", "playing", "the", "games", "such", "as", "Go", "or", "Chess", "where", "players", "sometimes", "would", "give", "up", "the", "immediate", "interests", "for", "the", "long", "-", "term", "victory", ".", "Thus", ",", "to", "evaluate", "the", "action", "-", "value", "for", "an", "intermediate", "state", ",", "we", "apply", "Monte", "Carlo", "search", "with", "a", "roll", "-", "out", "policy", "to", "sample", "the", "unknown", "last", "tokens", ".", "We", "represent", "an", "-", "time", "Monte", "Carlo", "search", "as", "where", "and", "is", "sampled", "based", "on", "the", "roll", "-", "out", "policy", "and", "the", "current", "state", ".", "In", "our", "experiment", ",", "is", "set", "the", "same", "as", "the", "generator", ",", "but", "one", "can", "use", "a", "simplified", "version", "if", "the", "speed", "is", "the", "priority", ".", "To", "reduce", "the", "variance", "and", "get", "more", "accurate", "assessment", "of", "the", "action", "value", ",", "we", "run", "the", "roll", "-", "out", "policy", "starting", "from", "current", "state", "till", "the", "end", "of", "the", "sequence", "for", "times", "to", "get", "a", "batch", "of", "output", "samples", ".", "Thus", ",", "we", "have", ":", "where", ",", "we", "see", "that", "when", "no", "intermediate", "reward", ",", "the", "function", "is", "iteratively", "defined", "as", "the", "next", "-", "state", "value", "starting", "from", "state", "and", "rolling", "out", "to", "the", "end", ".", "A", "benefit", "of", "using", "the", "discriminator", "as", "a", "reward", "function", "is", "that", "it", "can", "be", "dynamically", "updated", "to", "further", "improve", "the", "generative", "model", "iteratively", ".", "Once", "we", "have", "a", "set", "of", "more", "realistic", "generated", "sequences", ",", "we", "shall", "re", "-", "train", "the", "discriminator", "model", "as", "follows", ":", "Each", "time", "when", "a", "new", "discriminator", "model", "has", "been", "obtained", ",", "we", "are", "ready", "to", "update", "the", "generator", ".", "The", "proposed", "policy", "based", "method", "relies", "upon", "optimizing", "a", "parametrized", "policy", "to", "directly", "maximize", "the", "long", "-", "term", "reward", ".", "Following", ",", "the", "gradient", "of", "the", "objective", "function", "w.r.t", ".", "the", "generator", "’s", "parameters", "can", "be", "derived", "as", "The", "above", "form", "is", "due", "to", "the", "deterministic", "state", "transition", "and", "zero", "intermediate", "rewards", ".", "The", "detailed", "derivation", "is", "provided", "in", "the", "appendix", ".", "Using", "likelihood", "ratios", ",", "we", "build", "an", "unbiased", "estimation", "for", "Eq", ".", "(", "[", "reference", "]", ")", "(", "on", "one", "episode", ")", ":", "where", "is", "the", "observed", "intermediate", "state", "sampled", "from", ".", "Since", "the", "expectation", "can", "be", "approximated", "by", "sampling", "methods", ",", "we", "then", "update", "the", "generator", "’s", "parameters", "as", ":", "where", "denotes", "the", "corresponding", "learning", "rate", "at", "-", "th", "step", ".", "Also", "the", "advanced", "gradient", "algorithms", "such", "as", "Adam", "and", "RMSprop", "can", "be", "adopted", "here", ".", "In", "summary", ",", "Algorithm", "[", "reference", "]", "shows", "full", "details", "of", "the", "proposed", "SeqGAN", ".", "At", "the", "beginning", "of", "the", "training", ",", "we", "use", "the", "maximum", "likelihood", "estimation", "(", "MLE", ")", "to", "pre", "-", "train", "on", "training", "set", ".", "We", "found", "the", "supervised", "signal", "from", "the", "pre", "-", "trained", "discriminator", "is", "informative", "to", "help", "adjust", "the", "generator", "efficiently", ".", "After", "the", "pre", "-", "training", ",", "the", "generator", "and", "discriminator", "are", "trained", "alternatively", ".", "As", "the", "generator", "gets", "progressed", "via", "training", "on", "g", "-", "steps", "updates", ",", "the", "discriminator", "needs", "to", "be", "re", "-", "trained", "periodically", "to", "keeps", "a", "good", "pace", "with", "the", "generator", ".", "When", "training", "the", "discriminator", ",", "positive", "examples", "are", "from", "the", "given", "dataset", ",", "whereas", "negative", "examples", "are", "generated", "from", "our", "generator", ".", "In", "order", "to", "keep", "the", "balance", ",", "the", "number", "of", "negative", "examples", "we", "generate", "for", "each", "d", "-", "step", "is", "the", "same", "as", "the", "positive", "examples", ".", "And", "to", "reduce", "the", "variability", "of", "the", "estimation", ",", "we", "use", "different", "sets", "of", "negative", "samples", "combined", "with", "positive", "ones", ",", "which", "is", "similar", "to", "bootstrapping", ".", "[", "t", "]", "Sequence", "Generative", "Adversarial", "Nets", "[", "1", "]", "generator", "policy", "Gθ", ";", "roll", "-", "out", "policy", "Gβ", ";", "discriminator", "Dϕ", ";", "a", "sequence", "dataset", "=", "S{X:1", "T", "}", "Initialize", "Gθ", ",", "Dϕ", "with", "random", "weights", "θ", ",", "ϕ.", "Gθ", "using", "MLE", "on", "S", "negative", "samples", "using", "Gθ", "for", "training", "Dϕ", "Dϕ", "via", "minimizing", "the", "cross", "entropy", "a", "sequence", "Y:1T=", "(", "y1", ",", "…", ",", "yT", ")", "∼Gθ", "in", ":1", "T", "Q", "(", "a", "=", "yt;s", "=", "Y:1", "-", "t1", ")", "by", "Eq", ".", "(", ")", "generator", "parameters", "via", "policy", "gradient", "Eq", ".", "(", ")", "current", "Gθ", "to", "generate", "negative", "examples", "and", "combine", "with", "given", "positive", "examples", "S", "discriminator", "Dϕ", "for", "k", "epochs", "by", "Eq", ".", "(", ")", "converges", "subsection", ":", "The", "Generative", "Model", "for", "Sequences", "We", "use", "recurrent", "neural", "networks", "(", "RNNs", ")", "as", "the", "generative", "model", ".", "An", "RNN", "maps", "the", "input", "embedding", "representations", "of", "the", "sequence", "into", "a", "sequence", "of", "hidden", "states", "by", "using", "the", "update", "function", "recursively", ".", "Moreover", ",", "a", "softmax", "output", "layer", "maps", "the", "hidden", "states", "into", "the", "output", "token", "distribution", "where", "the", "parameters", "are", "a", "bias", "vector", "and", "a", "weight", "matrix", ".", "To", "deal", "with", "the", "common", "vanishing", "and", "exploding", "gradient", "problem", "of", "the", "backpropagation", "through", "time", ",", "we", "leverage", "the", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "cells", "to", "implement", "the", "update", "function", "in", "Eq", ".", "(", "[", "reference", "]", ")", ".", "It", "is", "worth", "noticing", "that", "most", "of", "the", "RNN", "variants", ",", "such", "as", "the", "gated", "recurrent", "unit", "(", "GRU", ")", "and", "soft", "attention", "mechanism", ",", "can", "be", "used", "as", "a", "generator", "in", "SeqGAN", ".", "subsection", ":", "The", "Discriminative", "Model", "for", "Sequences", "Deep", "discriminative", "models", "such", "as", "deep", "neural", "network", "(", "DNN", ")", ",", "convolutional", "neural", "network", "(", "CNN", ")", "and", "recurrent", "convolutional", "neural", "network", "(", "RCNN", ")", "have", "shown", "a", "high", "performance", "in", "complicated", "sequence", "classification", "tasks", ".", "In", "this", "paper", ",", "we", "choose", "the", "CNN", "as", "our", "discriminator", "as", "CNN", "has", "recently", "been", "shown", "of", "great", "effectiveness", "in", "text", "(", "token", "sequence", ")", "classification", ".", "Most", "discriminative", "models", "can", "only", "perform", "classification", "well", "for", "an", "entire", "sequence", "rather", "than", "the", "unfinished", "one", ".", "In", "this", "paper", ",", "we", "also", "focus", "on", "the", "situation", "where", "the", "discriminator", "predicts", "the", "probability", "that", "a", "finished", "sequence", "is", "real", ".", "We", "first", "represent", "an", "input", "sequence", "as", ":", "where", "is", "the", "-", "dimensional", "token", "embedding", "and", "is", "the", "concatenation", "operator", "to", "build", "the", "matrix", ".", "Then", "a", "kernel", "applies", "a", "convolutional", "operation", "to", "a", "window", "size", "of", "words", "to", "produce", "a", "new", "feature", "map", ":", "where", "operator", "is", "the", "summation", "of", "elementwise", "production", ",", "is", "a", "bias", "term", "and", "is", "a", "non", "-", "linear", "function", ".", "We", "can", "use", "various", "numbers", "of", "kernels", "with", "different", "window", "sizes", "to", "extract", "different", "features", ".", "Finally", "we", "apply", "a", "max", "-", "over", "-", "time", "pooling", "operation", "over", "the", "feature", "maps", ".", "To", "enhance", "the", "performance", ",", "we", "also", "add", "the", "highway", "architecture", "based", "on", "the", "pooled", "feature", "maps", ".", "Finally", ",", "a", "fully", "connected", "layer", "with", "sigmoid", "activation", "is", "used", "to", "output", "the", "probability", "that", "the", "input", "sequence", "is", "real", ".", "The", "optimization", "target", "is", "to", "minimize", "the", "cross", "entropy", "between", "the", "ground", "truth", "label", "and", "the", "predicted", "probability", "as", "formulated", "in", "Eq", ".", "(", "[", "reference", "]", ")", ".", "Detailed", "implementations", "of", "the", "generative", "and", "discriminative", "models", "are", "provided", "in", "the", "appendix", ".", "section", ":", "Synthetic", "Data", "Experiments", "To", "test", "the", "efficacy", "and", "add", "our", "understanding", "of", "SeqGAN", ",", "we", "conduct", "a", "simulated", "test", "with", "synthetic", "data", ".", "To", "simulate", "the", "real", "-", "world", "structured", "sequences", ",", "we", "consider", "a", "language", "model", "to", "capture", "the", "dependency", "of", "the", "tokens", ".", "We", "use", "a", "randomly", "initialized", "LSTM", "as", "the", "true", "model", ",", "aka", ",", "the", "oracle", ",", "to", "generate", "the", "real", "data", "distribution", "for", "the", "following", "experiments", ".", "subsection", ":", "Evaluation", "Metric", "The", "benefit", "of", "having", "such", "oracle", "is", "that", "firstly", ",", "it", "provides", "the", "training", "dataset", "and", "secondly", "evaluates", "the", "exact", "performance", "of", "the", "generative", "models", ",", "which", "will", "not", "be", "possible", "with", "real", "data", ".", "We", "know", "that", "MLE", "is", "trying", "to", "minimize", "the", "cross", "-", "entropy", "between", "the", "true", "data", "distribution", "and", "our", "approximation", ",", "i.e.", ".", "However", ",", "the", "most", "accurate", "way", "of", "evaluating", "generative", "models", "is", "that", "we", "draw", "some", "samples", "from", "it", "and", "let", "human", "observers", "review", "them", "based", "on", "their", "prior", "knowledge", ".", "We", "assume", "that", "the", "human", "observer", "has", "learned", "an", "accurate", "model", "of", "the", "natural", "distribution", ".", "Then", "in", "order", "to", "increase", "the", "chance", "of", "passing", "Turing", "Test", ",", "we", "actually", "need", "to", "minimize", "the", "exact", "opposite", "average", "negative", "log", "-", "likelihood", ",", "with", "the", "role", "of", "and", "exchanged", ".", "In", "our", "synthetic", "data", "experiments", ",", "we", "can", "consider", "the", "oracle", "to", "be", "the", "human", "observer", "for", "real", "-", "world", "problems", ",", "thus", "a", "perfect", "evaluation", "metric", "should", "be", "where", "and", "denote", "our", "generative", "model", "and", "the", "oracle", "respectively", ".", "At", "the", "test", "stage", ",", "we", "use", "to", "generate", "100", ",", "000", "sequence", "samples", "and", "calculate", "for", "each", "sample", "by", "and", "their", "average", "score", ".", "Also", "significance", "tests", "are", "performed", "to", "compare", "the", "statistical", "properties", "of", "the", "generation", "performance", "between", "the", "baselines", "and", "SeqGAN", ".", "subsection", ":", "Training", "Setting", "To", "set", "up", "the", "synthetic", "data", "experiments", ",", "we", "first", "initialize", "the", "parameters", "of", "an", "LSTM", "network", "following", "the", "normal", "distribution", "as", "the", "oracle", "describing", "the", "real", "data", "distribution", ".", "Then", "we", "use", "it", "to", "generate", "10", ",", "000", "sequences", "of", "length", "20", "as", "the", "training", "set", "for", "the", "generative", "models", ".", "In", "SeqGAN", "algorithm", ",", "the", "training", "set", "for", "the", "discriminator", "is", "comprised", "by", "the", "generated", "examples", "with", "the", "label", "0", "and", "the", "instances", "from", "with", "the", "label", "1", ".", "For", "different", "tasks", ",", "one", "should", "design", "specific", "structure", "for", "the", "convolutional", "layer", "and", "in", "our", "synthetic", "data", "experiments", ",", "the", "kernel", "size", "is", "from", "1", "to", "and", "the", "number", "of", "each", "kernel", "size", "is", "between", "100", "to", "200", ".", "Dropout", "and", "L2", "regularization", "are", "used", "to", "avoid", "over", "-", "fitting", ".", "Four", "generative", "models", "are", "compared", "with", "SeqGAN", ".", "The", "first", "model", "is", "a", "random", "token", "generation", ".", "The", "second", "one", "is", "the", "MLE", "trained", "LSTM", ".", "The", "third", "one", "is", "scheduled", "sampling", ".", "The", "fourth", "one", "is", "the", "Policy", "Gradient", "with", "BLEU", "(", "PG", "-", "BLEU", ")", ".", "In", "the", "scheduled", "sampling", ",", "the", "training", "process", "gradually", "changes", "from", "a", "fully", "guided", "scheme", "feeding", "the", "true", "previous", "tokens", "into", "LSTM", ",", "towards", "a", "less", "guided", "scheme", "which", "mostly", "feeds", "the", "LSTM", "with", "its", "generated", "tokens", ".", "A", "curriculum", "rate", "is", "used", "to", "control", "the", "probability", "of", "replacing", "the", "true", "tokens", "with", "the", "generated", "ones", ".", "To", "get", "a", "good", "and", "stable", "performance", ",", "we", "decrease", "by", "0.002", "for", "every", "training", "epoch", ".", "In", "the", "PG", "-", "BLEU", "algorithm", ",", "we", "use", "BLEU", ",", "a", "metric", "measuring", "the", "similarity", "between", "a", "generated", "sequence", "and", "references", "(", "training", "data", ")", ",", "to", "score", "the", "finished", "samples", "from", "Monte", "Carlo", "search", ".", "subsection", ":", "Results", "The", "performance", "of", "generating", "sequences", "from", "the", "compared", "policies", "is", "provided", "in", "Table", "[", "reference", "]", ".", "Since", "the", "evaluation", "metric", "is", "fundamentally", "instructive", ",", "we", "can", "see", "the", "impact", "of", "SeqGAN", ",", "which", "outperforms", "other", "baselines", "significantly", ".", "A", "significance", "T", "-", "test", "on", "the", "score", "distribution", "of", "the", "generated", "sequences", "from", "the", "compared", "models", "is", "also", "performed", ",", "which", "demonstrates", "the", "significant", "improvement", "of", "SeqGAN", "over", "all", "compared", "models", ".", "The", "learning", "curves", "shown", "in", "Figure", "[", "reference", "]", "illustrate", "the", "superiority", "of", "SeqGAN", "explicitly", ".", "After", "about", "150", "training", "epochs", ",", "both", "the", "maximum", "likelihood", "estimation", "and", "the", "schedule", "sampling", "methods", "converge", "to", "a", "relatively", "high", "score", ",", "whereas", "SeqGAN", "can", "improve", "the", "limit", "of", "the", "generator", "with", "the", "same", "structure", "as", "the", "baselines", "significantly", ".", "This", "indicates", "the", "prospect", "of", "applying", "adversarial", "training", "strategies", "to", "discrete", "sequence", "generative", "models", "to", "breakthrough", "the", "limitations", "of", "MLE", ".", "Additionally", ",", "SeqGAN", "outperforms", "PG", "-", "BLEU", ",", "which", "means", "the", "discriminative", "signal", "in", "GAN", "is", "more", "general", "and", "effective", "than", "a", "predefined", "score", "(", "e.g.", "BLEU", ")", "to", "guide", "the", "generative", "policy", "to", "capture", "the", "underlying", "distribution", "of", "the", "sequence", "data", ".", "subsection", ":", "Discussion", "In", "our", "synthetic", "data", "experiments", ",", "we", "find", "that", "the", "stability", "of", "SeqGAN", "depends", "on", "the", "training", "strategy", ".", "More", "specifically", ",", "the", "g", "-", "steps", ",", "d", "-", "steps", "and", "parameters", "in", "Algorithm", "[", "reference", "]", "have", "a", "large", "effect", "on", "the", "convergence", "and", "performance", "of", "SeqGAN", ".", "Figure", "[", "reference", "]", "shows", "the", "effect", "of", "these", "parameters", ".", "In", "Figure", "[", "reference", "]", ",", "the", "g", "-", "steps", "is", "much", "larger", "than", "the", "d", "-", "steps", "and", "epoch", "number", ",", "which", "means", "we", "train", "the", "generator", "for", "many", "times", "until", "we", "update", "the", "discriminator", ".", "This", "strategy", "leads", "to", "a", "fast", "convergence", "but", "as", "the", "generator", "improves", "quickly", ",", "the", "discriminator", "can", "not", "get", "fully", "trained", "and", "thus", "will", "provide", "a", "misleading", "signal", "gradually", ".", "In", "Figure", "[", "reference", "]", ",", "with", "more", "discriminator", "training", "epochs", ",", "the", "unstable", "training", "process", "is", "alleviated", ".", "In", "Figure", "[", "reference", "]", ",", "we", "train", "the", "generator", "for", "only", "one", "epoch", "and", "then", "before", "the", "discriminator", "gets", "fooled", ",", "we", "update", "it", "immediately", "based", "on", "the", "more", "realistic", "negative", "examples", ".", "In", "such", "a", "case", ",", "SeqGAN", "learns", "stably", ".", "The", "d", "-", "steps", "in", "all", "three", "training", "strategies", "described", "above", "is", "set", "to", "1", ",", "which", "means", "we", "only", "generate", "one", "set", "of", "negative", "examples", "with", "the", "same", "number", "as", "the", "given", "dataset", ",", "and", "then", "train", "the", "discriminator", "on", "it", "for", "various", "epochs", ".", "But", "actually", "we", "can", "utilize", "the", "potentially", "unlimited", "number", "of", "negative", "examples", "to", "improve", "the", "discriminator", ".", "This", "trick", "can", "be", "considered", "as", "a", "type", "of", "bootstrapping", ",", "where", "we", "combine", "the", "fixed", "positive", "examples", "with", "different", "negative", "examples", "to", "obtain", "multiple", "training", "sets", ".", "Figure", "[", "reference", "]", "shows", "this", "technique", "can", "improve", "the", "overall", "performance", "with", "good", "stability", ",", "since", "the", "discriminator", "is", "shown", "more", "negative", "examples", "and", "each", "time", "the", "positive", "examples", "are", "emphasized", ",", "which", "will", "lead", "to", "a", "more", "comprehensive", "guidance", "for", "training", "generator", ".", "This", "is", "in", "line", "with", "the", "theorem", "in", ".", "When", "analyzing", "the", "convergence", "of", "generative", "adversarial", "nets", ",", "an", "important", "assumption", "is", "that", "the", "discriminator", "is", "allowed", "to", "reach", "its", "optimum", "given", ".", "Only", "if", "the", "discriminator", "is", "capable", "of", "differentiating", "real", "data", "from", "unnatural", "data", "consistently", ",", "the", "supervised", "signal", "from", "it", "can", "be", "meaningful", "and", "the", "whole", "adversarial", "training", "process", "can", "be", "stable", "and", "effective", ".", "section", ":", "Real", "-", "world", "Scenarios", "To", "complement", "the", "previous", "experiments", ",", "we", "also", "test", "SeqGAN", "on", "several", "real", "-", "world", "tasks", ",", "i.e.", "poem", "composition", ",", "speech", "language", "generation", "and", "music", "generation", ".", "subsection", ":", "Text", "Generation", "For", "text", "generation", "scenarios", ",", "we", "apply", "the", "proposed", "SeqGAN", "to", "generate", "Chinese", "poems", "and", "Barack", "Obama", "political", "speeches", ".", "In", "the", "poem", "composition", "task", ",", "we", "use", "a", "corpus", "of", "16", ",", "394", "Chinese", "quatrains", ",", "each", "containing", "four", "lines", "of", "twenty", "characters", "in", "total", ".", "To", "focus", "on", "a", "fully", "automatic", "solution", "and", "stay", "general", ",", "we", "did", "not", "use", "any", "prior", "knowledge", "of", "special", "structure", "rules", "in", "Chinese", "poems", "such", "as", "specific", "phonological", "rules", ".", "In", "the", "Obama", "political", "speech", "generation", "task", ",", "we", "use", "a", "corpus", ",", "which", "is", "a", "collection", "of", "11", ",", "092", "paragraphs", "from", "Obama", "’s", "political", "speeches", ".", "We", "use", "BLEU", "score", "as", "an", "evaluation", "metric", "to", "measure", "the", "similarity", "degree", "between", "the", "generated", "texts", "and", "the", "human", "-", "created", "texts", ".", "BLEU", "is", "originally", "designed", "to", "automatically", "judge", "the", "machine", "translation", "quality", ".", "The", "key", "point", "is", "to", "compare", "the", "similarity", "between", "the", "results", "created", "by", "machine", "and", "the", "references", "provided", "by", "human", ".", "Specifically", ",", "for", "poem", "evaluation", ",", "we", "set", "n", "-", "gram", "to", "be", "2", "(", "BLEU", "-", "2", ")", "since", "most", "words", "(", "dependency", ")", "in", "classical", "Chinese", "poems", "consist", "of", "one", "or", "two", "characters", "and", "for", "the", "similar", "reason", ",", "we", "use", "BLEU", "-", "3", "and", "BLEU", "-", "4", "to", "evaluate", "Obama", "speech", "generation", "performance", ".", "In", "our", "work", ",", "we", "use", "the", "whole", "test", "set", "as", "the", "references", "instead", "of", "trying", "to", "find", "some", "references", "for", "the", "following", "line", "given", "the", "previous", "line", ".", "The", "reason", "is", "in", "generation", "tasks", "we", "only", "provide", "some", "positive", "examples", "and", "then", "let", "the", "model", "catch", "the", "patterns", "of", "them", "and", "generate", "new", "ones", ".", "In", "addition", "to", "BLEU", ",", "we", "also", "choose", "poem", "generation", "as", "a", "case", "for", "human", "judgement", "since", "a", "poem", "is", "a", "creative", "text", "construction", "and", "human", "evaluation", "is", "ideal", ".", "Specifically", ",", "we", "mix", "the", "20", "real", "poems", "and", "20", "each", "generated", "from", "SeqGAN", "and", "MLE", ".", "Then", "70", "experts", "on", "Chinese", "poems", "are", "invited", "to", "judge", "whether", "each", "of", "the", "60", "poem", "is", "created", "by", "human", "or", "machines", ".", "Once", "regarded", "to", "be", "real", ",", "it", "gets", "+", "1", "score", ",", "otherwise", "0", ".", "Finally", ",", "the", "average", "score", "for", "each", "algorithm", "is", "calculated", ".", "The", "experiment", "results", "are", "shown", "in", "Tables", "[", "reference", "]", "and", "[", "reference", "]", ",", "from", "which", "we", "can", "see", "the", "significant", "advantage", "of", "SeqGAN", "over", "the", "MLE", "in", "text", "generation", ".", "Particularly", ",", "for", "poem", "composition", ",", "SeqGAN", "performs", "comparably", "to", "real", "human", "data", ".", "subsection", ":", "Music", "Generation", "For", "music", "composition", ",", "we", "use", "Nottingham", "dataset", "as", "our", "training", "data", ",", "which", "is", "a", "collection", "of", "695", "music", "of", "folk", "tunes", "in", "midi", "file", "format", ".", "We", "study", "the", "solo", "track", "of", "each", "music", ".", "In", "our", "work", ",", "we", "use", "88", "numbers", "to", "represent", "88", "pitches", ",", "which", "correspond", "to", "the", "88", "keys", "on", "the", "piano", ".", "With", "the", "pitch", "sampling", "for", "every", "0.4s", ",", "we", "transform", "the", "midi", "files", "into", "sequences", "of", "numbers", "from", "1", "to", "88", "with", "the", "length", "32", ".", "To", "model", "the", "fitness", "of", "the", "discrete", "piano", "key", "patterns", ",", "BLEU", "is", "used", "as", "the", "evaluation", "metric", ".", "To", "model", "the", "fitness", "of", "the", "continuous", "pitch", "data", "patterns", ",", "the", "mean", "squared", "error", "(", "MSE", ")", "is", "used", "for", "evaluation", ".", "From", "Table", "[", "reference", "]", ",", "we", "see", "that", "SeqGAN", "outperforms", "the", "MLE", "significantly", "in", "both", "metrics", "in", "the", "music", "generation", "task", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "proposed", "a", "sequence", "generation", "method", ",", "SeqGAN", ",", "to", "effectively", "train", "generative", "adversarial", "nets", "for", "structured", "sequences", "generation", "via", "policy", "gradient", ".", "To", "our", "best", "knowledge", ",", "this", "is", "the", "first", "work", "extending", "GANs", "to", "generate", "sequences", "of", "discrete", "tokens", ".", "In", "our", "synthetic", "data", "experiments", ",", "we", "used", "an", "oracle", "evaluation", "mechanism", "to", "explicitly", "illustrate", "the", "superiority", "of", "SeqGAN", "over", "strong", "baselines", ".", "For", "three", "real", "-", "world", "scenarios", ",", "i.e.", ",", "poems", ",", "speech", "language", "and", "music", "generation", ",", "SeqGAN", "showed", "excellent", "performance", "on", "generating", "the", "creative", "sequences", ".", "We", "also", "performed", "a", "set", "of", "experiments", "to", "investigate", "the", "robustness", "and", "stability", "of", "training", "SeqGAN", ".", "For", "future", "work", ",", "we", "plan", "to", "build", "Monte", "Carlo", "tree", "search", "and", "value", "network", "to", "improve", "action", "decision", "making", "for", "large", "scale", "data", "and", "in", "the", "case", "of", "longer", "-", "term", "planning", ".", "section", ":", "Acknowledgments", "We", "sincerely", "thank", "Tianxing", "He", "for", "many", "helpful", "discussions", "and", "comments", "on", "the", "manuscript", ".", "bibliography", ":", "References", "appendix", ":", "Appendix", "In", "Section", "1", ",", "we", "present", "the", "step", "-", "by", "-", "step", "derivation", "of", "Eq", ".", "(", "6", ")", "in", "the", "paper", ".", "In", "Section", "2", ",", "the", "detailed", "realization", "of", "the", "generative", "model", "and", "the", "discriminative", "model", "is", "discussed", ",", "including", "the", "model", "parameter", "settings", ".", "In", "Section", "3", ",", "an", "interesting", "ablation", "study", "is", "provided", ",", "which", "is", "a", "supplementary", "to", "the", "discussions", "of", "the", "synthetic", "data", "experiments", ".", "subsection", ":", "Proof", "for", "Eq", ".", "(", "6", ")", "For", "readability", ",", "we", "provide", "the", "detailed", "derivation", "of", "Eq", ".", "(", "6", ")", "here", "by", "following", ".", "As", "mentioned", "in", "Sequence", "Generative", "Adversarial", "Nets", "section", ",", "the", "state", "transition", "is", "deterministic", "after", "an", "action", "has", "been", "chosen", ",", "i.e.", "for", "the", "next", "state", "if", "the", "current", "state", "and", "the", "action", ";", "for", "other", "next", "states", ",", ".", "In", "addition", ",", "the", "intermediate", "reward", "is", "0", ".", "We", "re", "-", "write", "the", "action", "value", "and", "state", "value", "as", "follows", ":", "For", "the", "start", "state", ",", "the", "value", "is", "calculated", "as", "which", "is", "the", "objective", "function", "to", "maximize", "in", "Eq", ".", "(", "1", ")", "of", "the", "paper", ".", "Then", "we", "can", "obtain", "the", "gradient", "of", "the", "objective", "function", ",", "defined", "in", "Eq", ".", "(", "1", ")", ",", "w.r.t", ".", "the", "generator", "’s", "parameters", ":", "which", "is", "the", "result", "in", "Eq", ".", "(", "6", ")", "of", "the", "paper", ".", "subsection", ":", "Model", "Implementations", "In", "this", "section", ",", "we", "present", "a", "full", "version", "of", "the", "discussed", "generative", "model", "and", "discriminative", "model", "in", "our", "paper", "submission", ".", "subsubsection", ":", "The", "Generative", "Model", "for", "Sequences", "We", "use", "recurrent", "neural", "networks", "(", "RNNs", ")", "as", "the", "generative", "model", ".", "An", "RNN", "maps", "the", "input", "embedding", "representations", "of", "the", "sequence", "into", "a", "sequence", "of", "hidden", "states", "by", "using", "the", "update", "function", "recursively", ".", "Moreover", ",", "a", "softmax", "output", "layer", "maps", "the", "hidden", "states", "into", "the", "output", "token", "distribution", "where", "the", "parameters", "are", "a", "bias", "vector", "and", "a", "weight", "matrix", ".", "The", "vanishing", "and", "exploding", "gradient", "problem", "in", "backpropagation", "through", "time", "(", "BPTT", ")", "issues", "a", "challenge", "of", "learning", "long", "-", "term", "dependencies", "to", "recurrent", "neural", "network", ".", "To", "address", "such", "problems", ",", "gated", "RNNs", "have", "been", "designed", "based", "on", "the", "basic", "idea", "of", "creating", "paths", "through", "time", "that", "have", "derivatives", "that", "neither", "vanish", "nor", "explode", ".", "Among", "various", "gated", "RNNs", ",", "we", "choose", "the", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "to", "be", "our", "generative", "networks", "with", "the", "update", "equations", ":", "where", "is", "the", "vector", "concatenation", "and", "is", "the", "elementwise", "product", ".", "For", "simplicity", ",", "we", "use", "the", "standard", "LSTM", "as", "the", "generator", ",", "while", "it", "is", "worth", "noticing", "that", "most", "of", "the", "RNN", "variants", ",", "such", "as", "the", "gated", "recurrent", "unit", "(", "GRU", ")", "and", "soft", "attention", "mechanism", ",", "can", "be", "used", "as", "a", "generator", "in", "SeqGAN", ".", "The", "standard", "way", "of", "training", "an", "RNN", "is", "the", "maximum", "likelihood", "estimation", "(", "MLE", ")", ",", "which", "involves", "minimizing", "the", "negative", "log", "-", "likelihood", "for", "a", "generated", "sequence", "given", "input", ".", "However", ",", "when", "applying", "MLE", "to", "generative", "models", ",", "there", "is", "a", "discrepancy", "between", "training", "and", "generating", ",", "which", "motivates", "our", "work", ".", "subsubsection", ":", "The", "Discriminative", "Model", "for", "Sequences", "Deep", "discriminative", "models", "such", "as", "deep", "neural", "network", "(", "DNN", ")", ",", "convolutional", "neural", "network", "(", "CNN", ")", "and", "recurrent", "convolutional", "neural", "network", "(", "RCNN", ")", "have", "shown", "a", "high", "performance", "in", "complicated", "sequence", "classification", "tasks", ".", "In", "this", "paper", ",", "we", "choose", "the", "CNN", "as", "our", "discriminator", "as", "CNN", "has", "recently", "been", "shown", "of", "great", "effectiveness", "in", "text", "(", "token", "sequence", ")", "classification", ".", "As", "far", "as", "we", "know", ",", "except", "for", "some", "specific", "tasks", ",", "most", "discriminative", "models", "can", "only", "perform", "classification", "well", "for", "a", "whole", "sequence", "rather", "than", "the", "unfinished", "one", ".", "In", "case", "of", "some", "specific", "tasks", ",", "one", "may", "design", "a", "classifier", "to", "provide", "intermediate", "reward", "signal", "to", "enhance", "the", "performance", "of", "our", "framework", ".", "But", "to", "make", "it", "more", "general", ",", "we", "focus", "on", "the", "situation", "where", "discriminator", "can", "only", "provide", "final", "reward", ",", "i.e.", ",", "the", "probability", "that", "a", "finished", "sequence", "was", "real", ".", "We", "first", "represent", "an", "input", "sequence", "as", ":", "where", "is", "the", "-", "dimensional", "token", "embedding", "and", "is", "the", "vertical", "concatenation", "operator", "to", "build", "the", "matrix", ".", "Then", "a", "kernel", "applies", "a", "convolutional", "operation", "to", "a", "window", "size", "of", "words", "to", "produce", "a", "new", "feature", "map", ":", "where", "operator", "is", "the", "summation", "of", "elementwise", "production", ",", "is", "a", "bias", "term", "and", "is", "a", "non", "-", "linear", "function", ".", "We", "can", "use", "various", "numbers", "of", "kernels", "with", "different", "window", "sizes", "to", "extract", "different", "features", ".", "Specifically", ",", "a", "kernel", "with", "window", "size", "applied", "to", "the", "concatenated", "embeddings", "of", "input", "sequence", "will", "produce", "a", "feature", "map", "Finally", "we", "apply", "a", "max", "-", "over", "-", "time", "pooling", "operation", "over", "the", "feature", "map", "and", "pass", "all", "pooled", "features", "from", "different", "kernels", "to", "a", "fully", "connected", "softmax", "layer", "to", "get", "the", "probability", "that", "a", "given", "sequence", "is", "real", ".", "We", "perform", "an", "empirical", "experiment", "to", "choose", "the", "kernel", "window", "sizes", "and", "numbers", "as", "shown", "in", "Table", "[", "reference", "]", ".", "For", "different", "tasks", ",", "one", "should", "design", "specific", "structures", "for", "the", "discriminator", ".", "To", "enhance", "the", "performance", ",", "we", "also", "add", "the", "highway", "architecture", "before", "the", "final", "fully", "connected", "layer", ":", "where", ",", "and", "are", "highway", "layer", "weights", ",", "denotes", "an", "affine", "transform", "followed", "by", "a", "non", "-", "linear", "activation", "function", "such", "as", "a", "rectified", "linear", "unit", "(", "ReLU", ")", "and", "is", "the", "“", "transform", "gate", "”", "with", "the", "same", "dimensionality", "as", "and", ".", "Finally", ",", "we", "apply", "a", "sigmoid", "transformation", "to", "get", "the", "probability", "that", "a", "given", "sequence", "is", "real", ":", "where", "and", "is", "the", "output", "layer", "weight", "and", "bias", ".", "When", "optimizing", "discriminative", "models", ",", "supervised", "training", "is", "applied", "to", "minimize", "the", "cross", "entropy", ",", "which", "is", "widely", "used", "as", "the", "objective", "function", "for", "classification", "and", "prediction", "tasks", ":", "where", "is", "the", "ground", "truth", "label", "of", "the", "input", "sequence", "and", "is", "the", "predicted", "probability", "from", "the", "discriminative", "models", ".", "subsection", ":", "More", "Ablation", "Study", "In", "the", "Discussion", "subsection", "of", "Synthetic", "Data", "Experiments", "section", "of", "our", "paper", ",", "we", "discussed", "the", "ablation", "study", "of", "three", "hyperparameters", "of", "SeqGAN", ",", "i.e.", ",", "g", "-", "steps", ",", "d", "-", "steps", "and", "epoch", "number", ".", "Here", "we", "provide", "another", "ablation", "study", "which", "is", "instructive", "for", "the", "better", "training", "of", "SeqGAN", ".", "As", "described", "in", "our", "paper", ",", "we", "start", "the", "adversarial", "training", "process", "after", "the", "convergence", "of", "MLE", "supervised", "pre", "-", "training", ".", "Here", "we", "further", "conduct", "experiments", "to", "investigate", "the", "performance", "of", "SeqGAN", "when", "the", "supervised", "pre", "-", "training", "is", "insufficient", ".", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "if", "we", "pre", "-", "train", "the", "generative", "model", "with", "conventional", "MLE", "methods", "for", "only", "20", "epochs", ",", "which", "is", "far", "from", "convergence", ",", "then", "the", "adversarial", "training", "process", "improves", "the", "generator", "quite", "slowly", "and", "unstably", ".", "The", "reason", "is", "that", "in", "SeqGAN", ",", "the", "discriminative", "model", "provides", "reward", "guidance", "when", "training", "the", "generator", "and", "if", "the", "generator", "acts", "almost", "randomly", ",", "the", "discriminator", "will", "identify", "the", "generated", "sequence", "to", "be", "unreal", "with", "high", "confidence", "and", "almost", "every", "action", "the", "generator", "takes", "receives", "a", "low", "(", "unified", ")", "reward", ",", "which", "does", "not", "guide", "the", "generator", "towards", "a", "good", "improvement", "direction", ",", "resulting", "in", "an", "ineffective", "training", "procedure", ".", "This", "indicates", "that", "in", "order", "to", "apply", "adversarial", "training", "strategies", "to", "sequence", "generative", "models", ",", "a", "sufficient", "pre", "-", "training", "is", "necessary", "."], "entity": [[2, 3, ["SeqGAN"], "Method"], [4, 8, ["Sequence", "Generative", "Adversarial", "Nets"], "Method"], [9, 11, ["Policy", "Gradient"], "Method"], [17, 19, ["generative", "models"], "Method"], [20, 23, ["Generative", "Adversarial", "Net"], "Method"], [24, 25, ["GAN"], "Method"], [29, 31, ["discriminative", "model"], "Method"], [37, 39, ["generative", "model"], "Method"], [60, 65, ["generating", "sequences", "of", "discrete", "tokens"], "Task"], [77, 79, ["generative", "model"], "Method"], [89, 91, ["discriminative", "model"], "Method"], [93, 95, ["generative", "model"], "Method"], [99, 101, ["discriminative", "model"], "Method"], [144, 147, ["sequence", "generation", "framework"], "Method"], [149, 150, ["SeqGAN"], "Method"], [158, 160, ["data", "generator"], "Method"], [162, 164, ["stochastic", "policy"], "Method"], [165, 167, ["reinforcement", "learning"], "Method"], [168, 169, ["RL"], "Method"], [171, 172, ["SeqGAN"], "Method"], [174, 177, ["generator", "differentiation", "problem"], "Task"], [180, 183, ["gradient", "policy", "update"], "Method"], [206, 210, ["state", "-", "action", "steps"], "Method"], [211, 214, ["Monte", "Carlo", "search"], "Method"], [249, 251, ["unsupervised", "learning"], "Task"], [254, 257, ["recurrent", "neural", "networks"], "Method"], [258, 259, ["RNNs"], "Method"], [261, 266, ["long", "short", "-", "term", "memory"], "Method"], [267, 268, ["LSTM"], "Method"], [276, 279, ["natural", "language", "generation"], "Task"], [280, 282, ["handwriting", "generation"], "Task"], [290, 291, ["RNN"], "Method"], [319, 322, ["maximum", "likelihood", "approaches"], "Method"], [331, 333, ["inference", "stage"], "Task"], [366, 367, ["inference"], "Task"], [392, 394, ["training", "strategy"], "Method"], [395, 397, ["scheduled", "sampling"], "Method"], [398, 399, ["SS"], "Method"], [403, 405, ["generative", "model"], "Method"], [438, 439, ["SS"], "Method"], [441, 444, ["inconsistent", "training", "strategy"], "Method"], [457, 462, ["training", "/", "inference", "discrepancy", "problem"], "Task"], [466, 468, ["loss", "function"], "Method"], [485, 487, ["machine", "translation"], "Task"], [489, 495, ["task", "specific", "sequence", "score", "/", "loss"], "Metric"], [496, 499, ["bilingual", "evaluation", "understudy"], "Metric"], [500, 502, ["BLEU", ")"], "Metric"], [509, 511, ["sequence", "generation"], "Task"], [522, 524, ["poem", "generation"], "Task"], [525, 526, ["chatbot"], "Task"], [543, 546, ["General", "adversarial", "net"], "Method"], [547, 548, ["GAN"], "Method"], [564, 565, ["GAN"], "Method"], [566, 568, ["discriminative", "net"], "Method"], [583, 585, ["generative", "net"], "Method"], [604, 607, ["computer", "vision", "tasks"], "Task"], [608, 613, ["generating", "samples", "of", "natural", "images"], "Task"], [617, 618, ["GAN"], "Method"], [619, 621, ["generating", "sequences"], "Task"], [627, 628, ["GAN"], "Method"], [643, 648, ["generating", "sequences", "of", "discrete", "tokens"], "Task"], [658, 659, ["GANs"], "Method"], [661, 662, ["generator"], "Method"], [664, 666, ["random", "sampling"], "Method"], [670, 672, ["determistic", "transform"], "Method"], [698, 700, ["generative", "model"], "Method"], [701, 702, ["paramters"], "Method"], [733, 735, ["discriminative", "net"], "Method"], [757, 758, ["GAN"], "Method"], [819, 822, ["sequence", "generation", "procedure"], "Method"], [824, 828, ["sequential", "decision", "making", "process"], "Task"], [830, 832, ["generative", "model"], "Method"], [841, 842, ["RL"], "Method"], [878, 879, ["BLEU"], "Metric"], [880, 882, ["machine", "translation"], "Task"], [891, 892, ["discriminator"], "Method"], [906, 908, ["generative", "model"], "Method"], [922, 924, ["generative", "model"], "Method"], [933, 935, ["generative", "model"], "Method"], [937, 940, ["stochastic", "parametrized", "policy"], "Method"], [943, 945, ["policy", "gradient"], "Method"], [948, 954, ["Monte", "Carlo", "(", "MC", ")", "search"], "Method"], [966, 970, ["policy", "(", "generative", "model"], "Method"], [972, 974, ["policy", "gradient"], "Method"], [987, 988, ["GAN"], "Method"], [1008, 1009, ["SeqGAN"], "Method"], [1016, 1017, ["SeqGAN"], "Method"], [1020, 1023, ["maximum", "likelihood", "methods"], "Method"], [1024, 1026, ["scheduled", "sampling"], "Method"], [1027, 1030, ["PG", "-", "BLEU"], "Method"], [1033, 1037, ["real", "-", "world", "tasks"], "Task"], [1039, 1041, ["poem", "generation"], "Task"], [1046, 1048, ["music", "generation"], "Task"], [1049, 1050, ["SeqGAN"], "Method"], [1057, 1058, ["metrics"], "Metric"], [1059, 1062, ["human", "expert", "judgement"], "Metric"], [1067, 1070, ["Deep", "generative", "models"], "Method"], [1100, 1103, ["contrastive", "divergence", "algorithm"], "Method"], [1106, 1109, ["deep", "belief", "nets"], "Method"], [1110, 1112, ["DBN", ")"], "Method"], [1114, 1116, ["denoising", "autoencoder"], "Method"], [1117, 1118, ["DAE"], "Method"], [1126, 1129, ["supervised", "learning", "fashion"], "Method"], [1131, 1132, ["DBN"], "Method"], [1133, 1134, ["DAE"], "Method"], [1136, 1139, ["low", "dimensional", "representation"], "Method"], [1140, 1141, ["encoding"], "Method"], [1151, 1153, ["decoding", "network"], "Method"], [1156, 1158, ["variational", "autoencoder"], "Method"], [1159, 1160, ["VAE"], "Method"], [1163, 1165, ["deep", "learning"], "Method"], [1166, 1168, ["statistical", "inference"], "Method"], [1184, 1188, ["deep", ")", "neural", "networks"], "Method"], [1189, 1193, ["non", "-", "linear", "mapping"], "Task"], [1195, 1196, ["inference"], "Task"], [1199, 1201, ["variational", "methods"], "Method"], [1204, 1206, ["generative", "models"], "Method"], [1231, 1235, ["approximating", "intractable", "probabilistic", "computations"], "Task"], [1239, 1241, ["training", "methodology"], "Method"], [1242, 1244, ["generative", "models"], "Method"], [1246, 1247, ["GANs"], "Method"], [1250, 1252, ["training", "procedure"], "Method"], [1254, 1256, ["minimax", "game"], "Method"], [1258, 1260, ["generative", "model"], "Method"], [1262, 1264, ["discriminative", "model"], "Method"], [1271, 1274, ["maximum", "likelihood", "learning"], "Method"], [1280, 1283, ["natural", "image", "generation"], "Task"], [1293, 1294, ["GANs"], "Method"], [1295, 1300, ["sequence", "discrete", "data", "generation", "problems"], "Task"], [1302, 1305, ["natural", "language", "generation"], "Task"], [1311, 1313, ["generator", "network"], "Method"], [1314, 1315, ["GAN"], "Method"], [1331, 1334, ["discrete", "data", "generation"], "Task"], [1352, 1355, ["Recurrent", "neural", "networks"], "Method"], [1368, 1370, ["machine", "translation"], "Task"], [1377, 1378, ["RNNs"], "Method"], [1402, 1405, ["maximum", "likelihood", "estimation"], "Method"], [1412, 1413, ["SS"], "Method"], [1430, 1431, ["GANs"], "Method"], [1444, 1445, ["GANs"], "Method"], [1454, 1457, ["discrete", "probabilistic", "models"], "Method"], [1465, 1468, ["sequence", "data", "generation"], "Task"], [1473, 1477, ["sequential", "decision", "making", "process"], "Task"], [1491, 1493, ["sequence", "generator"], "Method"], [1502, 1505, ["policy", "gradient", "methods"], "Method"], [1529, 1532, ["sequence", "generation", "tasks"], "Task"], [1534, 1536, ["machine", "translation"], "Task"], [1554, 1555, ["Go"], "Task"], [1573, 1578, ["state", "-", "action", "evaluation", "methods"], "Method"], [1580, 1586, ["Monte", "Carlo", "(", "tree", ")", "search"], "Method"], [1595, 1596, ["SeqGAN"], "Method"], [1597, 1598, ["GANs"], "Method"], [1607, 1610, ["sequence", "generation", "problem"], "Task"], [1627, 1630, ["Monte", "Carlo", "approach"], "Method"], [1648, 1652, ["Sequence", "Generative", "Adversarial", "Nets"], "Method"], [1653, 1656, ["sequence", "generation", "problem"], "Task"], [1673, 1677, ["-", "parameterized", "generative", "model"], "Method"], [1696, 1698, ["reinforcement", "learning"], "Method"], [1721, 1723, ["policy", "model"], "Method"], [1764, 1768, ["-", "parameterized", "discriminative", "model"], "Method"], [1774, 1775, ["generator"], "Task"], [1801, 1803, ["discriminative", "model"], "Method"], [1824, 1826, ["generative", "model"], "Method"], [1833, 1835, ["generative", "model"], "Method"], [1840, 1842, ["policy", "gradient"], "Method"], [1843, 1845, ["MC", "search"], "Method"], [1856, 1858, ["discriminative", "model"], "Method"], [1871, 1873, ["discriminative", "model"], "Method"], [1886, 1887, ["SeqGAN"], "Method"], [1888, 1890, ["Policy", "Gradient"], "Method"], [1903, 1905, ["generator", "model"], "Method"], [1906, 1907, ["policy"], "Method"], [1940, 1941, ["discriminator"], "Method"], [1973, 1974, ["policy"], "Method"], [2033, 2035, ["REINFORCE", "algorithm"], "Method"], [2045, 2046, ["discriminator"], "Method"], [2058, 2059, ["discriminator"], "Method"], [2115, 2116, ["Chess"], "Task"], [2147, 2150, ["Monte", "Carlo", "search"], "Method"], [2152, 2156, ["roll", "-", "out", "policy"], "Method"], [2166, 2171, ["-", "time", "Monte", "Carlo", "search"], "Method"], [2179, 2183, ["roll", "-", "out", "policy"], "Method"], [2231, 2235, ["roll", "-", "out", "policy"], "Method"], [2296, 2297, ["discriminator"], "Method"], [2312, 2314, ["generative", "model"], "Method"], [2333, 2335, ["discriminator", "model"], "Method"], [2343, 2345, ["discriminator", "model"], "Method"], [2355, 2356, ["generator"], "Method"], [2359, 2362, ["policy", "based", "method"], "Method"], [2366, 2368, ["parametrized", "policy"], "Method"], [2426, 2428, ["unbiased", "estimation"], "Method"], [2429, 2430, ["Eq"], "Task"], [2458, 2460, ["sampling", "methods"], "Method"], [2484, 2486, ["gradient", "algorithms"], "Method"], [2488, 2489, ["Adam"], "Method"], [2490, 2491, ["RMSprop"], "Method"], [2509, 2510, ["SeqGAN"], "Method"], [2521, 2524, ["maximum", "likelihood", "estimation"], "Method"], [2525, 2527, ["MLE", ")"], "Method"], [2545, 2546, ["discriminator"], "Method"], [2552, 2553, ["generator"], "Method"], [2557, 2560, ["pre", "-", "training"], "Task"], [2562, 2565, ["generator", "and", "discriminator"], "Method"], [2583, 2584, ["discriminator"], "Method"], [2603, 2604, ["discriminator"], "Method"], [2638, 2641, ["d", "-", "step"], "Method"], [2656, 2657, ["estimation"], "Task"], [2674, 2675, ["bootstrapping"], "Method"], [2679, 2683, ["Sequence", "Generative", "Adversarial", "Nets"], "Method"], [2686, 2689, ["generator", "policy", "Gθ"], "Method"], [2690, 2695, ["roll", "-", "out", "policy", "Gβ"], "Method"], [2696, 2698, ["discriminator", "Dϕ"], "Method"], [2707, 2708, ["Gθ"], "Method"], [2718, 2719, ["MLE"], "Method"], [2732, 2734, ["cross", "entropy"], "Metric"], [2766, 2769, ["policy", "gradient", "Eq"], "Method"], [2799, 2801, ["Generative", "Model"], "Method"], [2802, 2803, ["Sequences"], "Task"], [2805, 2808, ["recurrent", "neural", "networks"], "Method"], [2809, 2810, ["RNNs"], "Method"], [2813, 2815, ["generative", "model"], "Method"], [2817, 2818, ["RNN"], "Method"], [2842, 2845, ["softmax", "output", "layer"], "Method"], [2871, 2876, ["vanishing", "and", "exploding", "gradient", "problem"], "Task"], [2878, 2881, ["backpropagation", "through", "time"], "Task"], [2885, 2890, ["Long", "Short", "-", "Term", "Memory"], "Method"], [2891, 2892, ["LSTM"], "Method"], [2916, 2918, ["RNN", "variants"], "Method"], [2922, 2925, ["gated", "recurrent", "unit"], "Method"], [2926, 2927, ["GRU"], "Method"], [2929, 2932, ["soft", "attention", "mechanism"], "Method"], [2938, 2939, ["generator"], "Method"], [2940, 2941, ["SeqGAN"], "Method"], [2945, 2947, ["Discriminative", "Model"], "Method"], [2948, 2949, ["Sequences"], "Task"], [2949, 2952, ["Deep", "discriminative", "models"], "Method"], [2954, 2957, ["deep", "neural", "network"], "Method"], [2958, 2959, ["DNN"], "Method"], [2961, 2964, ["convolutional", "neural", "network"], "Method"], [2965, 2966, ["CNN"], "Method"], [2973, 2974, ["RCNN"], "Method"], [2981, 2985, ["complicated", "sequence", "classification", "tasks"], "Task"], [2993, 2994, ["CNN"], "Method"], [2996, 2997, ["discriminator"], "Method"], [2998, 2999, ["CNN"], "Method"], [3007, 3013, ["text", "(", "token", "sequence", ")", "classification"], "Task"], [3015, 3017, ["discriminative", "models"], "Method"], [3020, 3021, ["classification"], "Task"], [3044, 3045, ["discriminator"], "Method"], [3066, 3070, ["-", "dimensional", "token", "embedding"], "Method"], [3082, 3083, ["kernel"], "Method"], [3085, 3087, ["convolutional", "operation"], "Method"], [3104, 3108, ["summation", "of", "elementwise", "production"], "Method"], [3116, 3120, ["non", "-", "linear", "function"], "Method"], [3141, 3148, ["max", "-", "over", "-", "time", "pooling", "operation"], "Method"], [3162, 3164, ["highway", "architecture"], "Method"], [3174, 3177, ["fully", "connected", "layer"], "Method"], [3194, 3195, ["optimization"], "Task"], [3226, 3230, ["generative", "and", "discriminative", "models"], "Method"], [3250, 3251, ["SeqGAN"], "Method"], [3273, 3275, ["language", "model"], "Method"], [3291, 3293, ["true", "model"], "Method"], [3294, 3295, ["aka"], "Method"], [3297, 3298, ["oracle"], "Method"], [3312, 3314, ["Evaluation", "Metric"], "Metric"], [3337, 3339, ["generative", "models"], "Method"], [3352, 3353, ["MLE"], "Method"], [3358, 3361, ["cross", "-", "entropy"], "Metric"], [3380, 3382, ["generative", "models"], "Method"], [3406, 3408, ["human", "observer"], "Method"], [3468, 3472, ["real", "-", "world", "problems"], "Task"], [3476, 3478, ["evaluation", "metric"], "Metric"], [3484, 3486, ["generative", "model"], "Method"], [3524, 3526, ["statistical", "properties"], "Metric"], [3528, 3530, ["generation", "performance"], "Metric"], [3534, 3535, ["SeqGAN"], "Method"], [3559, 3561, ["normal", "distribution"], "Method"], [3589, 3591, ["generative", "models"], "Method"], [3601, 3602, ["discriminator"], "Method"], [3632, 3634, ["convolutional", "layer"], "Method"], [3661, 3662, ["Dropout"], "Method"], [3663, 3665, ["L2", "regularization"], "Method"], [3669, 3672, ["over", "-", "fitting"], "Task"], [3674, 3676, ["generative", "models"], "Method"], [3679, 3680, ["SeqGAN"], "Method"], [3686, 3689, ["random", "token", "generation"], "Method"], [3703, 3705, ["scheduled", "sampling"], "Method"], [3711, 3715, ["Policy", "Gradient", "with", "BLEU"], "Method"], [3716, 3719, ["PG", "-", "BLEU"], "Method"], [3723, 3725, ["scheduled", "sampling"], "Method"], [3733, 3736, ["fully", "guided", "scheme"], "Method"], [3742, 3743, ["LSTM"], "Method"], [3747, 3749, ["guided", "scheme"], "Method"], [3753, 3754, ["LSTM"], "Method"], [3760, 3762, ["curriculum", "rate"], "Metric"], [3804, 3805, ["BLEU"], "Metric"], [3810, 3811, ["similarity"], "Metric"], [3828, 3831, ["Monte", "Carlo", "search"], "Method"], [3854, 3856, ["evaluation", "metric"], "Metric"], [3866, 3867, ["SeqGAN"], "Method"], [3901, 3902, ["SeqGAN"], "Method"], [3920, 3921, ["SeqGAN"], "Method"], [3931, 3934, ["maximum", "likelihood", "estimation"], "Method"], [3936, 3939, ["schedule", "sampling", "methods"], "Method"], [3947, 3948, ["SeqGAN"], "Method"], [3970, 3973, ["adversarial", "training", "strategies"], "Method"], [3974, 3978, ["discrete", "sequence", "generative", "models"], "Method"], [3983, 3984, ["MLE"], "Method"], [3987, 3988, ["SeqGAN"], "Method"], [3989, 3992, ["PG", "-", "BLEU"], "Method"], [3999, 4000, ["GAN"], "Method"], [4011, 4012, ["BLEU"], "Metric"], [4016, 4018, ["generative", "policy"], "Method"], [4041, 4042, ["stability"], "Metric"], [4043, 4044, ["SeqGAN"], "Method"], [4047, 4049, ["training", "strategy"], "Method"], [4074, 4075, ["convergence"], "Metric"], [4078, 4079, ["SeqGAN"], "Method"], [4126, 4127, ["discriminator"], "Method"], [4134, 4135, ["convergence"], "Metric"], [4138, 4139, ["generator"], "Method"], [4143, 4144, ["discriminator"], "Method"], [4195, 4196, ["discriminator"], "Method"], [4216, 4217, ["SeqGAN"], "Method"], [4281, 4282, ["discriminator"], "Method"], [4292, 4293, ["bootstrapping"], "Task"], [4329, 4330, ["discriminator"], "Method"], [4370, 4373, ["generative", "adversarial", "nets"], "Method"], [4380, 4381, ["discriminator"], "Method"], [4392, 4393, ["discriminator"], "Method"], [4415, 4418, ["adversarial", "training", "process"], "Method"], [4426, 4430, ["Real", "-", "world", "Scenarios"], "Task"], [4439, 4440, ["SeqGAN"], "Method"], [4442, 4446, ["real", "-", "world", "tasks"], "Task"], [4447, 4450, ["i.e.", "poem", "composition"], "Task"], [4455, 4457, ["music", "generation"], "Task"], [4460, 4462, ["Text", "Generation"], "Task"], [4463, 4466, ["text", "generation", "scenarios"], "Task"], [4471, 4472, ["SeqGAN"], "Method"], [4474, 4476, ["Chinese", "poems"], "Material"], [4477, 4481, ["Barack", "Obama", "political", "speeches"], "Material"], [4484, 4487, ["poem", "composition", "task"], "Task"], [4496, 4498, ["Chinese", "quatrains"], "Material"], [4514, 4516, ["automatic", "solution"], "Task"], [4532, 4534, ["Chinese", "poems"], "Material"], [4542, 4547, ["Obama", "political", "speech", "generation", "task"], "Task"], [4563, 4567, ["Obama", "’s", "political", "speeches"], "Material"], [4570, 4572, ["BLEU", "score"], "Metric"], [4574, 4576, ["evaluation", "metric"], "Metric"], [4579, 4581, ["similarity", "degree"], "Metric"], [4592, 4593, ["BLEU"], "Metric"], [4600, 4603, ["machine", "translation", "quality"], "Metric"], [4628, 4630, ["poem", "evaluation"], "Task"], [4640, 4643, ["BLEU", "-", "2"], "Metric"], [4652, 4654, ["Chinese", "poems"], "Material"], [4668, 4671, ["BLEU", "-", "3"], "Metric"], [4672, 4675, ["BLEU", "-", "4"], "Metric"], [4677, 4680, ["Obama", "speech", "generation"], "Task"], [4715, 4717, ["generation", "tasks"], "Task"], [4741, 4742, ["BLEU"], "Metric"], [4746, 4748, ["poem", "generation"], "Task"], [4752, 4754, ["human", "judgement"], "Task"], [4756, 4757, ["poem"], "Material"], [4759, 4762, ["creative", "text", "construction"], "Task"], [4763, 4765, ["human", "evaluation"], "Task"], [4781, 4782, ["SeqGAN"], "Method"], [4783, 4784, ["MLE"], "Method"], [4789, 4791, ["Chinese", "poems"], "Material"], [4858, 4859, ["SeqGAN"], "Method"], [4861, 4862, ["MLE"], "Method"], [4863, 4865, ["text", "generation"], "Task"], [4869, 4871, ["poem", "composition"], "Task"], [4872, 4873, ["SeqGAN"], "Method"], [4882, 4884, ["Music", "Generation"], "Task"], [4885, 4887, ["music", "composition"], "Task"], [4890, 4892, ["Nottingham", "dataset"], "Material"], [4946, 4948, ["pitch", "sampling"], "Method"], [4981, 4982, ["BLEU"], "Metric"], [4986, 4988, ["evaluation", "metric"], "Metric"], [5001, 5004, ["mean", "squared", "error"], "Metric"], [5005, 5006, ["MSE"], "Metric"], [5021, 5022, ["SeqGAN"], "Method"], [5024, 5025, ["MLE"], "Method"], [5031, 5034, ["music", "generation", "task"], "Task"], [5045, 5048, ["sequence", "generation", "method"], "Method"], [5049, 5050, ["SeqGAN"], "Method"], [5054, 5057, ["generative", "adversarial", "nets"], "Method"], [5058, 5061, ["structured", "sequences", "generation"], "Task"], [5062, 5064, ["policy", "gradient"], "Method"], [5076, 5077, ["GANs"], "Method"], [5093, 5096, ["oracle", "evaluation", "mechanism"], "Method"], [5102, 5103, ["SeqGAN"], "Method"], [5109, 5113, ["real", "-", "world", "scenarios"], "Task"], [5116, 5117, ["poems"], "Task"], [5118, 5120, ["speech", "language"], "Task"], [5121, 5123, ["music", "generation"], "Task"], [5124, 5125, ["SeqGAN"], "Method"], [5144, 5145, ["robustness"], "Metric"], [5146, 5147, ["stability"], "Metric"], [5149, 5150, ["SeqGAN"], "Method"], [5159, 5163, ["Monte", "Carlo", "tree", "search"], "Method"], [5164, 5166, ["value", "network"], "Method"], [5168, 5171, ["action", "decision", "making"], "Task"], [5180, 5184, ["longer", "-", "term", "planning"], "Task"], [5241, 5243, ["generative", "model"], "Method"], [5245, 5247, ["discriminative", "model"], "Method"], [5262, 5264, ["ablation", "study"], "Task"], [5310, 5314, ["Sequence", "Generative", "Adversarial", "Nets"], "Method"], [5452, 5454, ["generative", "model"], "Method"], [5455, 5457, ["discriminative", "model"], "Method"], [5465, 5467, ["Generative", "Model"], "Method"], [5468, 5469, ["Sequences"], "Task"], [5471, 5474, ["recurrent", "neural", "networks"], "Method"], [5475, 5476, ["RNNs"], "Method"], [5479, 5481, ["generative", "model"], "Method"], [5483, 5484, ["RNN"], "Method"], [5508, 5511, ["softmax", "output", "layer"], "Method"], [5533, 5538, ["vanishing", "and", "exploding", "gradient", "problem"], "Task"], [5539, 5542, ["backpropagation", "through", "time"], "Task"], [5543, 5545, ["BPTT", ")"], "Method"], [5549, 5554, ["learning", "long", "-", "term", "dependencies"], "Task"], [5555, 5558, ["recurrent", "neural", "network"], "Method"], [5564, 5566, ["gated", "RNNs"], "Method"], [5590, 5592, ["gated", "RNNs"], "Method"], [5607, 5609, ["generative", "networks"], "Method"], [5632, 5633, ["LSTM"], "Method"], [5635, 5636, ["generator"], "Method"], [5646, 5648, ["RNN", "variants"], "Method"], [5652, 5655, ["gated", "recurrent", "unit"], "Method"], [5656, 5657, ["GRU"], "Method"], [5659, 5662, ["soft", "attention", "mechanism"], "Method"], [5668, 5669, ["generator"], "Method"], [5670, 5671, ["SeqGAN"], "Method"], [5678, 5679, ["RNN"], "Method"], [5681, 5684, ["maximum", "likelihood", "estimation"], "Method"], [5685, 5687, ["MLE", ")"], "Method"], [5707, 5708, ["MLE"], "Method"], [5709, 5711, ["generative", "models"], "Method"], [5719, 5720, ["generating"], "Task"], [5726, 5727, ["subsubsection"], "Method"], [5729, 5731, ["Discriminative", "Model"], "Method"], [5733, 5736, ["Deep", "discriminative", "models"], "Method"], [5738, 5741, ["deep", "neural", "network"], "Method"], [5742, 5743, ["DNN"], "Method"], [5745, 5748, ["convolutional", "neural", "network"], "Method"], [5749, 5750, ["CNN"], "Method"], [5757, 5758, ["RCNN"], "Method"], [5765, 5769, ["complicated", "sequence", "classification", "tasks"], "Task"], [5777, 5778, ["CNN"], "Method"], [5780, 5781, ["discriminator"], "Method"], [5782, 5783, ["CNN"], "Method"], [5791, 5797, ["text", "(", "token", "sequence", ")", "classification"], "Task"], [5811, 5813, ["discriminative", "models"], "Method"], [5816, 5817, ["classification"], "Task"], [5839, 5840, ["classifier"], "Method"], [5866, 5867, ["discriminator"], "Method"], [5895, 5899, ["-", "dimensional", "token", "embedding"], "Method"], [5912, 5913, ["kernel"], "Method"], [5915, 5917, ["convolutional", "operation"], "Method"], [5934, 5938, ["summation", "of", "elementwise", "production"], "Method"], [5946, 5950, ["non", "-", "linear", "function"], "Method"], [5970, 5971, ["kernel"], "Method"], [5985, 5987, ["feature", "map"], "Method"], [5991, 5998, ["max", "-", "over", "-", "time", "pooling", "operation"], "Method"], [6012, 6016, ["fully", "connected", "softmax", "layer"], "Method"], [6059, 6060, ["discriminator"], "Method"], [6070, 6072, ["highway", "architecture"], "Method"], [6075, 6078, ["fully", "connected", "layer"], "Method"], [6102, 6105, ["rectified", "linear", "unit"], "Method"], [6106, 6107, ["ReLU"], "Method"], [6127, 6129, ["sigmoid", "transformation"], "Method"], [6152, 6154, ["discriminative", "models"], "Method"], [6155, 6157, ["supervised", "training"], "Method"], [6162, 6164, ["cross", "entropy"], "Metric"], [6171, 6173, ["objective", "function"], "Metric"], [6174, 6178, ["classification", "and", "prediction", "tasks"], "Task"], [6196, 6198, ["discriminative", "models"], "Method"], [6202, 6204, ["Ablation", "Study"], "Task"], [6220, 6222, ["ablation", "study"], "Task"], [6226, 6227, ["SeqGAN"], "Method"], [6253, 6254, ["training"], "Task"], [6255, 6256, ["SeqGAN"], "Method"], [6266, 6269, ["adversarial", "training", "process"], "Method"], [6273, 6278, ["MLE", "supervised", "pre", "-", "training"], "Method"], [6289, 6290, ["SeqGAN"], "Method"], [6292, 6296, ["supervised", "pre", "-", "training"], "Task"], [6313, 6315, ["generative", "model"], "Method"], [6317, 6319, ["MLE", "methods"], "Method"], [6332, 6335, ["adversarial", "training", "process"], "Method"], [6348, 6349, ["SeqGAN"], "Method"], [6351, 6353, ["discriminative", "model"], "Method"], [6369, 6370, ["discriminator"], "Method"], [6412, 6414, ["training", "procedure"], "Method"], [6422, 6425, ["adversarial", "training", "strategies"], "Method"], [6426, 6429, ["sequence", "generative", "models"], "Method"], [6432, 6435, ["pre", "-", "training"], "Method"], [185, 186, ["RL"], "Method"], [191, 192, ["GAN"], "Method"], [838, 840, ["reinforcement", "learning"], "Method"], [1042, 1044, ["speech", "language"], "Task"], [1408, 1410, ["scheduled", "sampling"], "Method"], [1422, 1423, ["SS"], "Method"], [1485, 1487, ["reinforcement", "learning"], "Method"], [1600, 1601, ["RL"], "Method"], [2969, 2972, ["convolutional", "neural", "network"], "Method"], [3288, 3289, ["LSTM"], "Method"], [3555, 3556, ["LSTM"], "Method"], [3593, 3594, ["SeqGAN"], "Method"], [3697, 3698, ["LSTM"], "Method"], [3797, 3800, ["PG", "-", "BLEU"], "Method"], [4451, 4453, ["speech", "language"], "Task"], [4775, 4776, ["poems"], "Task"], [5596, 5601, ["Long", "Short", "-", "Term", "Memory"], "Method"], [5602, 5603, ["LSTM"], "Method"], [5753, 5756, ["convolutional", "neural", "network"], "Method"]]} {"doc_id": "24730424236724d3f798dec02901e7a1f1c4710e", "text": ["Joint", "Maximum", "Purity", "Forest", "with", "Application", "to", "Image", "Super", "-", "Resolution", "section", ":", "Abstract", "-", "In", "this", "paper", ",", "we", "propose", "a", "novel", "random", "-", "forest", "scheme", ",", "namely", "Joint", "Maximum", "Purity", "Forest", "(", "JMPF", ")", ",", "for", "classification", ",", "clustering", ",", "and", "regression", "tasks", ".", "In", "the", "JMPF", "scheme", ",", "the", "original", "feature", "space", "is", "transformed", "into", "a", "compactly", "pre", "-", "clustered", "feature", "space", ",", "via", "a", "trained", "rotation", "matrix", ".", "The", "rotation", "matrix", "is", "obtained", "through", "an", "iterative", "quantization", "process", ",", "where", "the", "input", "data", "belonging", "to", "different", "classes", "are", "clustered", "to", "the", "respective", "vertices", "of", "the", "new", "feature", "space", "with", "maximum", "purity", ".", "In", "the", "new", "feature", "space", ",", "orthogonal", "hyperplanes", ",", "which", "are", "employed", "at", "the", "split", "-", "nodes", "of", "decision", "trees", "in", "random", "forests", ",", "can", "tackle", "the", "clustering", "problems", "effectively", ".", "We", "evaluated", "our", "proposed", "method", "on", "public", "benchmark", "datasets", "for", "regression", "and", "classification", "tasks", ",", "and", "experiments", "showed", "that", "JMPF", "remarkably", "outperforms", "other", "state", "-", "of", "-", "the", "-", "art", "random", "-", "forest", "-", "based", "approaches", ".", "Furthermore", ",", "we", "applied", "JMPF", "to", "image", "super", "-", "resolution", ",", "because", "the", "transformed", ",", "compact", "features", "are", "more", "discriminative", "to", "the", "clustering", "-", "regression", "scheme", ".", "Experiment", "results", "on", "several", "public", "benchmark", "datasets", "also", "showed", "that", "the", "JMPF", "-", "based", "image", "super", "-", "resolution", "scheme", "is", "consistently", "superior", "to", "recent", "state", "-", "of", "-", "the", "-", "art", "image", "super", "-", "resolution", "algorithms", ".", "section", ":", "from", "achieving", "the", "optimal", "hyperplanes", "as", "SVM", "does", "(", "i.e.", ",", "there", "is", "no", "orthogonal", "constraint", "in", "SVM", ")", "in", "some", "original", "feature", "space", ",", "as", "shown", "in", "Fig", ".", "1", "(", "a", ")", ".", "In", "this", "paper", ",", "we", "aim", "to", "solve", "this", "orthogonalconstraint", "limitation", ".", "With", "the", "fixed", "orthogonal", "hyperplanes", ",", "we", "propose", "to", "rotate", "the", "feature", "space", ",", "this", "is", "equivalent", "to", "rotating", "the", "hyperplanes", ",", "in", "such", "a", "way", "that", "global", "maximum", "purity", "on", "the", "clustered", "data", "can", "be", "achieved", ",", "as", "illustrated", "in", "Fig", ".", "2", ".", "This", "strategy", "can", "achieve", "a", "joint", "maximum", "purity", "for", "all", "the", "split", "-", "nodes", "when", "training", "a", "random", "forest", ".", "Image", "super", "-", "resolution", "can", "be", "performed", "based", "on", "clustering", "/", "classification", ",", "according", "to", "the", "recent", "emerging", "clustering", "-", "regression", "stream", "[", "reference", "][", "reference", "][", "reference", "]", ",", "and", "the", "JMPF", "scheme", "can", "achieve", "remarkable", "performance", "on", "both", "the", "classification", "and", "regression", "tasks", ".", "Therefore", ",", "JMPF", "is", "applied", "to", "single", "-", "image", "super", "-", "resolution", "in", "this", "paper", ".", "In", "our", "algorithm", ",", "principal", "component", "analysis", "(", "PCA", ")", "is", "applied", "to", "the", "features", "for", "dimensionality", "reduction", ".", "The", "projected", "feature", "space", "is", "then", "rotated", "to", "a", "compact", ",", "preclustered", "feature", "space", "via", "a", "learned", "rotation", "matrix", ".", "Finally", ",", "for", "all", "the", "split", "-", "nodes", "trained", "for", "a", "random", "forest", ",", "their", "thresholds", "are", "directly", "set", "to", "the", "inherent", "zero", "-", "center", "orthogonal", "hyperplanes", "in", "the", "rotated", "feature", "space", "to", "meet", "the", "maximum", "-", "purity", "criterion", ".", "Experiment", "results", "show", "that", "JMPF", "can", "achieve", "more", "accurate", "clustering", "/", "classification", "performance", "on", "random", "forests", ",", "and", "applying", "JMPF", "to", "image", "super", "-", "resolution", "can", "achieve", "superior", "quality", ",", "compared", "to", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "Having", "introduced", "the", "main", "idea", "of", "our", "proposed", "algorithm", ",", "the", "remainder", "of", "this", "paper", "is", "organized", "as", "follows", ".", "In", "Section", "II", ",", "we", "will", "describe", "our", "proposed", "scheme", ",", "the", "joint", "maximum", "purity", "forest", "scheme", ",", "and", "present", "in", "detail", "how", "to", "compute", "the", "rotation", "matrix", "via", "clustering", "data", "into", "the", "feature", "-", "space", "vertices", ".", "Section", "III", "will", "evaluate", "our", "proposed", "method", "and", "compare", "its", "performance", "with", "recent", "state", "-", "of", "-", "the", "-", "art", "random", "-", "forest", "-", "based", "approaches", "on", "regression", "and", "classification", "tasks", ".", "In", "Section", "IV", ",", "we", "will", "validate", "the", "performance", "of", "JMPF", "scheme", "on", "single", "-", "image", "super", "-", "resolution", ".", "Conclusions", "are", "given", "in", "Section", "V.", "section", ":", "II", ".", "JOINT", "MAXIMUM", "PURITY", "FOREST", "SCHEME", "section", ":", "II.1", "Random", "Forest", "and", "Our", "Insights", "A", "random", "forest", "is", "an", "ensemble", "of", "binary", "decision", "trees", "(", ")", ":", "→", "ℝ", ",", "where", "(", "=", "1", ",", "2", ",", "…", ",", ")", "is", "the", "index", "of", "the", "trees", ",", "∈", "ℝ", "is", "the", "m", "-", "dimension", "feature", "space", ",", "and", "ℝ", "=", "[", "0", ",", "1", "]", "represents", "the", "space", "of", "class", "probability", "distributions", "over", "the", "label", "space", "=", "{", "1", ",", ".", ".", ".", ",", "}", ".", "As", "shown", "in", "Fig", ".", "1", "(", "b", ")", ",", "the", "vertical", "dotted", "line", "forms", "a", "hyperplane", ",", "=", "0", ",", "chosen", "in", "the", "first", "split", "-", "node", "for", "separating", "training", "samples", ",", "and", "the", "horizontal", "dotted", "line", "is", "the", "hyperplane", ",", "=", "0", ",", "for", "the", "second", "split", "-", "node", "to", "cluster", "all", "the", "feature", "data", "assigned", "to", "this", "node", ".", "This", "results", "in", "separating", "the", "three", "data", "samples", "(", "Red", ",", "Green", "and", "Blue", ")", "into", "three", "leaf", "-", "nodes", ".", "It", "can", "be", "seen", "from", "Fig", ".", "1", "(", "b", ")", "that", ",", "for", "each", "split", "-", "node", ",", "the", "optimal", "hyperplane", "with", "more", "generalization", "capability", "is", "the", "one", "which", "can", "achieve", "maximum", "purity", "in", "clustering", "samples", "into", "two", "groups", ".", "For", "example", ",", "the", "vertical", "dotted", "line", "is", "the", "first", "optimal", "hyperplane", "because", "it", "clusters", "all", "the", "red", "training", "samples", "into", "the", "right", "node", ",", "while", "all", "the", "blue", "and", "green", "samples", "are", "clustered", "into", "the", "left", "node", ".", "Furthermore", ",", "the", "left", "margin", "and", "the", "right", "margin", "are", "equal", ".", "Although", "there", "is", "no", "guarantee", "that", "optimal", "hyperplanes", "can", "be", "determined", "for", "all", "the", "split", "-", "nodes", "in", "a", "random", "forest", ",", "approximated", "optimal", "hyperplanes", "can", "be", "obtained", "through", "a", "random", "bagging", "strategy", ".", "The", "training", "of", "a", "whole", "random", "forest", "is", "to", "train", "all", "of", "its", "decision", "trees", ",", "by", "choosing", "the", "candidate", "features", "and", "thresholds", "for", "each", "of", "the", "split", "-", "nodes", ",", "where", "the", "feature", "dimensions", "and", "thresholds", "are", "determined", "using", "a", "random", "bagging", "strategy", ".", "In", "the", "prediction", "stage", ",", "each", "decision", "tree", "returns", "a", "class", "probability", "(", "|", ")", "for", "a", "given", "query", "sample", "∈", "ℝ", ",", "and", "the", "final", "class", "label", "y", "*", "is", "then", "obtained", "via", "averaging", ",", "as", "follows", ":", "*", "=", "arg", "max", "∑", "(", "|", ")", ".", "The", "splitting", "function", "for", "a", "split", "-", "node", "is", "denoted", "as", "(", ";", "Θ", ")", ",", "where", "is", "a", "sample", "and", "Θ", "is", "typically", "parameterized", "by", "two", "values", ":", "(", "i", ")", "a", "feature", "dimension", "Θ", "Î{1", ",", ".", ".", ".", ",", "}", ",", "and", "(", "ii", ")", "a", "threshold", "Θ", "Îℝ.", "The", "splitting", "function", "is", "defined", "as", "follows", ":", "where", "the", "outcome", "defines", "to", "which", "child", "node", "the", "sample", "is", "routed", ",", "and", "0", "and", "1", "are", "the", "two", "labels", "for", "the", "left", "and", "right", "child", "nodes", ",", "respectively", ".", "Each", "node", "chooses", "the", "best", "splitting", "function", "Θ", "*", "out", "of", "a", "randomly", "sampled", "set", "{", "Θ", "}", "by", "optimizing", "the", "following", "function", ":", "where", "and", "are", "the", "sets", "of", "samples", "that", "are", "routed", "to", "the", "left", "and", "the", "right", "child", "nodes", ",", "and", "|", "|", "represents", "the", "number", "of", "samples", "in", "the", "set", ".", "During", "the", "training", "of", "a", "random", "forest", ",", "the", "decision", "trees", "are", "provided", "with", "a", "random", "subset", "of", "the", "training", "data", "(", "i.e.", "bagging", ")", ",", "and", "are", "trained", "independently", "of", "each", "other", ".", "Therefore", ",", "the", "decision", "trees", "are", "working", "as", "independent", "experts", ".", "Taking", "random", "-", "forest", "-", "based", "classification", "as", "an", "example", ",", "training", "a", "single", "decision", "tree", "involves", "recursively", "splitting", "each", "node", ",", "such", "that", "the", "training", "data", "in", "each", "newly", "created", "child", "node", "is", "clustered", "according", "to", "their", "corresponding", "class", "labels", ",", "so", "the", "purity", "at", "each", "node", "is", "increasing", "along", "a", "tree", ".", "Each", "tree", "is", "grown", "until", "a", "stopping", "criterion", "is", "reached", "(", "e.g.", "the", "number", "of", "samples", "in", "a", "node", "is", "less", "than", "a", "threshold", "or", "the", "tree", "depth", "reaches", "a", "maximum", "value", ")", "and", "the", "class", "probability", "distributions", "are", "estimated", "in", "the", "leaf", "-", "nodes", ".", "After", "fulfilling", "one", "of", "these", "criteria", ",", "a", "density", "model", "(", ")", "in", "the", "leaf", "-", "node", "is", "estimated", "by", "all", "samples", "falling", "into", "this", "leaf", "-", "node", "for", "predicting", "the", "target", "value", "in", "the", "testing", "stage", ".", "A", "simple", "way", "to", "estimate", "the", "probability", "distribution", "(", ")", "is", "averaging", "all", "the", "samples", "in", "the", "leaf", "-", "node", ",", "while", "there", "are", "also", "variant", "methods", ",", "such", "as", "fitting", "a", "Gaussian", "distribution", "or", "kernel", "density", "estimation", ",", "ridge", "regression", "[", "reference", "][", "reference", "][", "reference", "]", ",", "and", "so", "on", ".", "(", ")", "is", "the", "local", "score", "for", "a", "set", "of", "samples", "(", "is", "either", "or", ")", ",", "which", "normally", "is", "calculated", "using", "entropy", "as", "in", "Eqn", ".", "(", "4", ")", ",", "but", "it", "can", "be", "replaced", "by", "variance", "[", "reference", "][", "reference", "][", "reference", "]", "or", "the", "Gini", "index", "[", "reference", "]", ".", "where", "K", "is", "the", "number", "of", "classes", ",", "and", "(", "|", ")", "is", "the", "probability", "for", "class", ",", "given", "the", "set", ".", "For", "the", "regression", "problem", ",", "the", "differential", "entropy", ":", "over", "continuous", "outputs", "can", "be", "employed", ",", "where", "(", "|", ")", "denotes", "the", "conditional", "probability", "of", "a", "target", "variable", "given", "the", "input", "sample", ".", "Assuming", "(", ".", ",", ".", ")", "to", "be", "a", "Gaussian", "distribution", "and", "having", "only", "a", "finite", "set", "of", "samples", ",", "the", "differential", "entropy", "can", "be", "written", "in", "closed", "form", "as", "where", "det", "(", "Σ", ")", "is", "the", "determinant", "of", "the", "estimated", "covariance", "matrix", "of", "the", "target", "variables", "in", ".", "For", "training", "each", "decision", "tree", "in", "a", "random", "forest", ",", "the", "goal", "on", "each", "split", "-", "node", "is", "to", "maximize", "the", "information", "gain", "(", "IG", ")", "by", "reducing", "the", "entropy", "after", "splitting", ".", "IG", "is", "defined", "as", "follows", ":", "Since", "each", "decision", "tree", "is", "a", "binary", "tree", "and", "each", "step", "is", "to", "split", "a", "current", "node", "(", "a", "parent", "set", ")", "into", "two", "children", "nodes", "(", "and", "sets", ")", ",", "IG", "can", "be", "described", "as", "follows", ":", "where", "ℋ", "is", "the", "optimal", "hyperplane", "of", "the", "split", "-", "node", ",", "and", "Eqn", ".", "(", "8", ")", "is", "the", "target", "function", "of", "each", "splitnode", "when", "training", "each", "decision", "tree", "of", "a", "random", "forest", ".", "As", "we", "can", "see", "from", "Fig", ".", "1", "(", "b", ")", ",", "all", "the", "optimal", "hyperplanes", "from", "split", "-", "nodes", "are", "achieved", "independently", "and", "locally", ".", "Since", "each", "optimal", "hyperplane", "is", "obtained", "from", "a", "subset", "of", "feature", "-", "dimension", "candidates", "with", "the", "randomly", "bagging", "strategy", ",", "there", "is", "no", "guarantee", "of", "obtaining", "a", "global", "optimum", "with", "respect", "to", "all", "the", "hyperplanes", "in", "all", "the", "split", "-", "nodes", ".", "An", "intuitive", "thinking", ",", "which", "was", "inspired", "by", "the", "data", "distribution", "in", "Fig", ".", "1", "(", "b", ")", ",", "is", "to", "achieve", "a", "global", "optimum", "by", "jointly", "considering", "all", "the", "hyperplanes", "of", "all", "the", "split", "-", "nodes", ",", "in", "the", "form", "as", "follows", ":", "where", "is", "the", "total", "number", "of", "split", "-", "nodes", "that", "a", "training", "sample", "has", "routed", "through", "a", "decision", "tree", ".", "As", "there", "is", "no", "mathematical", "solution", "to", "the", "problem", "described", "in", "Eqn", ".", "[", "reference", "]", ",", "an", "alternative", "way", "(", "i.e.", ",", "an", "approximate", "method", ")", "to", "numerically", "solving", "Eqn", ".", "(", "9", ")", "is", "to", "jointly", "maximize", "the", "purity", "of", "the", "clustered", "data", "groups", "at", "each", "of", "the", "split", "-", "nodes", ".", "This", "also", "means", "that", "all", "the", "data", "is", "clustered", "into", "the", "corners", "(", "feature", "-", "space", "vertices", ")", "of", "the", "feature", "space", ",", "as", "shown", "in", "Fig", ".", "2", ".", "section", ":", "II.2", "The", "Joint", "Maximum", "Purity", "Forest", "Scheme", "To", "calculate", "the", "threshold", "for", "each", "split", "-", "node", "in", "each", "decision", "tree", "when", "training", "a", "random", "forest", ",", "we", "are", "attempting", "to", "determine", "an", "orthogonal", "hyperplane", "for", "a", "three", "-", "category", "classification", "problem", ",", "as", "shown", "in", "Fig", ".", "1", ".", "Since", "the", "hyperplanes", "for", "the", "split", "-", "nodes", "of", "a", "decision", "tree", "are", "required", "to", "be", "orthogonal", "to", "each", "other", ",", "seeking", "an", "optimal", "orthogonal", "hyperplane", "locally", "can", "not", "guarantee", "obtaining", "maximum", "purity", "for", "the", "whole", "tree", "globally", ".", "As", "shown", "in", "Fig", ".", "2", ",", "it", "is", "easy", "to", "determine", "the", "vertical", "hyperplane", "for", "maximum", "purity", ",", "but", "it", "is", "hard", "to", "obtain", "the", "horizontal", "hyperplane", "for", "maximum", "purity", "in", "the", "original", "feature", "space", ".", "To", "achieve", "an", "optimal", "classification", "performance", "for", "the", "whole", "decision", "tree", ",", "all", "the", "split", "-", "nodes", "should", "be", "considered", "globally", "or", "simultaneously", ".", "As", "shown", "in", "Fig", ".", "2", ",", "a", "number", "of", "split", "-", "nodes", ",", "which", "have", "their", "hyperplanes", "orthogonal", "to", "each", "other", ",", "are", "required", "to", "separate", "the", "samples", "into", "different", "nodes", ".", "However", ",", "if", "we", "can", "transform", "the", "samples", "(", "zerocentered", "feature", "data", ")", "to", "locate", "them", "at", "the", "respective", "corners", "of", "the", "feature", "space", ",", "i.e.", "{", "−1", ",", "1", "}", "for", "mdimensional", "features", ",", "the", "feature", "data", "can", "be", "easily", "and", "accurately", "separated", "by", "the", "orthogonal", "(", "either", "vertical", "or", "horizontal", ")", "hyperplanes", ",", "which", "contain", "the", "space", "center", "{", "0", "}", ",", "as", "illustrated", "in", "Fig", ".", "1", "(", "b", ")", ".", "The", "insight", "behind", "this", "is", "that", "the", "data", "is", "clustered", "into", "the", "feature", "-", "space", "vertices", "(", "the", "corners", "in", "a", "2", "-", "D", "feature", "space", "means", "that", "the", "data", "points", "belong", "to", "{", "−1", ",", "1", "}", "as", "the", "coordinate", "range", "is", "set", "to", "[", "−1", ",", "1", "]", ")", ".", "To", "tackle", "the", "original", "feature", "data", ",", "which", "is", "not", "ideally", "clustered", "in", "the", "vertices", "or", "corners", "of", "the", "feature", "space", "or", "close", "to", "them", ",", "as", "shown", "in", "Fig", ".", "1", "(", "a", ")", ",", "an", "intuitive", "idea", "is", "to", "rotate", "the", "feature", "space", "(", "this", "is", "equivalent", "to", "rotating", "the", "hyperplanes", ")", ".", "This", "transformation", "clusters", "the", "feature", "data", "compactly", "into", "feature", "-", "space", "vertices", "{", "−1", ",", "1", "}", "with", "a", "total", "of", "2", "vertices", ".", "Therefore", ",", "a", "possible", "solution", "to", "the", "problem", "described", "in", "Eqn", ".", "(", "10", ")", "is", "to", "rotate", "the", "data", "features", "by", "a", "rotation", "matrix", "ℛ", "×", ",", "as", "shown", "in", "Fig", ".", "2", ",", "through", "which", "the", "original", "feature", "space", "is", "transformed", "into", "a", "more", "compact", "clustered", "feature", "space", ",", "where", "all", "the", "feature", "data", "is", "clustered", "close", "to", "the", "feature", "-", "space", "vertices", ".", "This", "solution", "can", "be", "mathematically", "defined", "as", "follows", ":", "where", "∈", "ℝ", "×", "contains", "n", "samples", ",", "each", "of", "which", "is", "a", "-", "dimensional", "feature", "vector", "arranged", "in", "a", "row", ",", "and", "is", "zero", "-", "centered", ",", "i.e.", "all", "the", "feature", "vectors", "are", "demeaned", "by", "subtracting", "the", "mean", "vector", "from", "each", "feature", "vector", ".", "This", "idea", "of", "clustering", "data", "into", "the", "feature", "-", "space", "vertices", "can", "also", "be", "found", "in", "locality", "-", "sensitive", "hashing", "(", "LSH", ")", "[", "reference", "]", "and", "image", "representation", "[", "reference", "]", ".", "In", "[", "reference", "]", ",", "a", "simple", "and", "efficient", "alternating", "minimization", "scheme", "was", "proposed", "to", "find", "a", "rotation", "matrix", "for", "zero", "-", "centered", "feature", "data", ",", "which", "minimizes", "the", "quantization", "errors", "by", "mapping", "the", "feature", "data", "to", "the", "vertices", "of", "a", "zero", "-", "centered", "binary", "hypercube", ".", "The", "method", "is", "termed", "as", "iterative", "quantization", "(", "ITQ", ")", ",", "which", "can", "work", "on", "multi", "-", "class", "spectral", "clustering", "and", "orthogonal", "Procrustes", "problem", ".", "Yu", "et", "al", ".", "[", "reference", "]", "proposed", "using", "a", "circulant", "matrix", "to", "speed", "up", "the", "computation", ",", "because", "the", "circulant", "structure", "enables", "the", "use", "of", "Fast", "Fourier", "Transformation", "(", "FFT", ")", ".", "As", "the", "computation", "of", "the", "rotation", "matrix", "in", "the", "training", "and", "testing", "stage", "is", "ignorable", ",", "we", "choose", "a", "similar", "scheme", "to", "ITQ", "[", "reference", "]", "to", "determine", "the", "rotation", "matrix", "(", "we", "throw", "away", "the", "final", "quantization", "matrix", "described", "in", "Eqn", ".", "[", "reference", "]", ",", "which", "is", "used", "for", "hashing", "in", "[", "reference", "]", ")", ",", "through", "which", "the", "original", "feature", "space", "can", "be", "transformed", "into", "a", "new", "compact", "clustered", "feature", "space", ":", "=", "ℛ", ",", "where", "the", "data", "is", "located", "at", "the", "respective", "vertices", "in", "the", "new", "feature", "space", ".", "After", "this", "transformation", ",", "a", "random", "forest", "with", "globally", "joint", "maximum", "purity", "of", "all", "the", "clustered", "data", "can", "be", "trained", ",", "through", "all", "the", "hyperplanes", "in", "the", "split", "-", "nodes", "of", "each", "decision", "tree", ".", "Based", "on", "this", "idea", ",", "our", "proposed", "scheme", "is", "called", "joint", "maximum", "purity", "forest", "(", "JMPF", ")", ".", "section", ":", "II.3", "Learning", "the", "Rotation", "Matrix", "via", "Clustering", "Data", "into", "Feature", "-", "Space", "Vertices", "Assuming", "that", "∈", "ℝ", "is", "one", "point", "in", "the", "-", "dimensional", "feature", "space", "(", "zero", "-", "centered", "data", ")", ",", "the", "respective", "vertices", "in", "the", "zero", "-", "centered", "binary", "hypercube", "space", "can", "be", "denoted", "as", "(", ")", "∈", "{", "−1", ",", "1", "}", ",", "and", "there", "is", "a", "total", "of", "2", "vertices", "in", "the", "-", "dimensional", "feature", "space", ".", "It", "is", "easy", "to", "see", "from", "We", "denote", "a", "binary", "code", "matrix", "∈", "{", "−1", ",", "1", "}", "×", ",", "whose", "rows", "=", "(", ")", "∈", ".", "For", "a", "matrix", "or", "a", "vector", ",", "(", ".", ")", "applies", "the", "sign", "operation", "to", "it", "element", "-", "wise", ".", "Our", "objective", "is", "to", "minimize", "the", "error", "between", "the", "feature", "and", "the", "feature", "-", "space", "vertices", ",", "i.e.", ",", "min‖", "−", "‖", ".", "As", "we", "can", "see", "in", "Fig", ".", "2", ",", "when", "the", "feature", "space", "is", "rotated", ",", "the", "feature", "points", "will", "be", "more", "concentrated", "around", "their", "nearest", "vertices", ",", "which", "means", "that", "the", "quantization", "error", "will", "become", "smaller", ".", "Therefore", ",", "the", "minimization", "problem", "of", "min‖", "−", "‖", "is", "equivalent", "to", "minimizing", "the", "error", "of", "the", "zerocentered", "data", "with", "respect", "to", "the", "Frobenius", "norm", ",", "as", "in", "the", "following", "formulation", ":", "Therefore", ",", "the", "task", "of", "this", "minimization", "problem", "is", "to", "determine", "an", "optimal", "rotation", "matrix", "ℛ", "to", "satisfy", "Eqn", ".", "[", "reference", "]", ".", "Since", "there", "are", "two", "variables", "in", "Eqn", ".", "(", "11", ")", ",", "the", "expectation", "-", "maximization", "(", "E", "-", "M", ")", "algorithm", "is", "applied", "to", "cluster", "data", "into", "the", "feature", "-", "space", "vertices", ",", "such", "that", "a", "local", "minimum", "of", "the", "binary", "code", "matrix", "and", "the", "rotation", "matrix", "ℛ", "are", "computed", "simultaneously", ".", "The", "idea", "of", "rotating", "feature", "data", "to", "minimize", "the", "error", "between", "the", "transformed", "data", "and", "the", "featurespace", "vertices", "can", "also", "be", "found", "in", "[", "reference", "]", ",", "which", "showed", "that", "the", "rotation", "matrix", "ℛ", "can", "be", "initialized", "randomly", ",", "and", "then", "iterated", "to", "converge", "to", "the", "required", "rotation", "matrix", ".", "Two", "iteration", "steps", "will", "be", "performed", ":", "in", "every", "iteration", ",", "each", "feature", "vector", "in", "the", "feature", "space", "is", "firstly", "quantized", "to", "the", "nearest", "vertex", "of", "the", "binary", "hypercube", ",", "i.e.", "to", "a", "vertex", "in", ",", "and", "then", "the", "rotation", "matrix", "ℛ", "is", "updated", "to", "minimize", "the", "quantization", "error", "by", "fixing", ".", "These", "two", "alternating", "steps", "are", "described", "in", "detail", "below", ":", "(", "1", ")", "Fix", "ℛ", "and", "update", ":", "Because", "the", "zero", "-", "centered", "data", "matrix", "is", "fixed", ",", "minimizing", "Eqn", ".", "(", "12", ")", "is", "equivalent", "to", "maximizing", "the", "following", "term", ":", "where", "is", "an", "element", "of", "=", "ℛ.", "To", "maximize", "Eqn", ".", "(", "13", ")", "with", "respect", "to", ",", "=", "1", "whenever", "≥", "0", "and", "=", "−1", "otherwise", ",", "i.e.", "=", "(", "ℛ", ")", "∈", "{", "−1", ",", "1", "}", ".", "(", "2", ")", "Fix", "and", "update", "ℛ", ":", "The", "problem", "of", "fixing", "to", "obtain", "a", "rotation", "matrix", "based", "on", "the", "objective", "function", "Eqn", ".", "(", "11", ")", "is", "relative", "to", "the", "classic", "orthogonal", "Procrustes", "problem", "[", "reference", "][", "reference", "][", "reference", "]", ",", "in", "which", "a", "rotation", "matrix", "is", "determined", "to", "align", "one", "point", "set", "with", "another", ".", "In", "our", "algorithm", ",", "these", "two", "point", "sets", "are", "the", "zero", "-", "centered", "data", "set", "and", "the", "quantized", "matrix", ".", "Therefore", ",", "a", "closed", "-", "form", "solution", "for", "ℛ", "is", "available", ",", "by", "applying", "SVD", "on", "the", "×", "matrix", "to", "obtain", "Ω", "(", "Ω", "is", "a", "diagonal", "matrix", ")", ",", "then", "set", "ℛ", "=", "to", "update", "ℛ.", "section", ":", "II.4", "Proof", "of", "the", "Orthogonal", "Procrustes", "Problem", ":", "For", "completeness", ",", "we", "prove", "the", "orthogonal", "Procrustes", "problem", ",", "for", "which", "the", "solution", "can", "be", "found", "in", "[", "reference", "][", "reference", "][", "reference", "]", ":", "Proof", ":", "Thus", ",", "min", "ℛ", "‖", "−", "ℛ", "‖", "equals", "to", "maximizing", ":", "The", "last", "inequality", "holds", "because", "Z", "is", "also", "an", "orthonormal", "matrix", ",", "and", "∑", ",", "=", "1", ",", ",", "≤", "1", ".", "The", "objective", "function", "can", "be", "maximized", "if", "Z", "=", ",", "i.e.", "section", ":", "ℛ", "=", "∎", "section", ":", "III", ".", "JOINT", "MAXIMUM", "PURITY", "FOREST", "FOR", "REGRESSION", "AND", "CLASSIFICATION", "section", ":", "III.1", "The", "Workflow", "of", "Joint", "Maximum", "Purity", "Forest", "Random", "forest", "is", "a", "machine", "-", "learning", "method", "using", "an", "ensemble", "of", "randomized", "decision", "trees", "for", "classification", ".", "Each", "tree", "in", "a", "random", "forest", "consists", "of", "split", "-", "nodes", "and", "leaf", "-", "nodes", ",", "which", "can", "be", "trained", "recursively", ".", "A", "random", "forest", "is", "constructed", "recursively", ",", "where", "each", "node", "attempts", "to", "find", "a", "splitting", "function", "or", "a", "hyperplane", "to", "separate", "its", "samples", "into", "two", "leaf", "-", "nodes", ",", "such", "that", "the", "information", "gain", "is", "optimized", ".", "A", "tree", "stops", "growing", "if", "the", "maximum", "depth", "is", "reached", "or", "if", "a", "node", "has", "achieved", "maximum", "purity", ",", "i.e.", "it", "contains", "only", "samples", "from", "one", "class", ".", "Then", ",", "each", "leaf", "-", "node", "collects", "the", "statistics", "of", "the", "samples", "falling", "in", "it", ".", "In", "the", "evaluation", "phase", ",", "the", "probability", "of", "a", "query", "sample", "x", "belonging", "to", "class", "k", "is", "given", "by", "averaging", "all", "the", "trees", ",", "or", "by", "other", "methods", ".", "Most", "random", "-", "forest", "-", "based", "models", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "share", "a", "similar", "workflow", ",", "as", "shown", "in", "Fig", ".", "3", ",", "in", "which", "the", "main", "task", "on", "training", "a", "tree", "in", "a", "random", "forest", "is", "to", "decide", "thresholds", "in", "the", "split", "-", "nodes", "and", "learn", "the", "regressors", "or", "classes", "in", "the", "leaf", "-", "nodes", ".", "Rigid", "regression", "or", "linear", "regression", "is", "often", "employed", "in", "the", "leaf", "-", "nodes", "for", "the", "prediction", "task", ",", "because", "rigid", "regression", "has", "a", "closed", "-", "form", "solution", ",", "while", "linear", "regression", "is", "an", "efficient", "optimization", "tool", ",", "and", "the", "LibLinear", "package", "[", "reference", "]", "can", "be", "used", "to", "fine", "-", "tune", "its", "configurations", ".", "Compared", "to", "conventional", "random", "forests", ",", "our", "JMPF", "scheme", "has", "one", "more", "step", ",", "as", "shown", "in", "the", "left", "of", "Fig", ".", "3", ",", "the", "rotation", "matrix", ".", "The", "JMPF", "scheme", "transforms", "the", "original", "feature", "space", "by", "rotating", "it", "into", "a", "more", "compact", ",", "pre", "-", "clustered", "feature", "space", ",", "using", "a", "trained", "rotation", "matrix", "learned", "through", "clustering", "feature", "vectors", "iteratively", "into", "the", "vertices", "of", "a", "new", "feature", "space", ".", "The", "whole", "workflow", "of", "our", "proposed", "algorithm", ",", "the", "JMPF", "scheme", ",", "is", "outlined", "in", "Fig", ".", "3", ".", "The", "source", "code", "of", "our", "algorithm", "is", "available", "to", "download", "at", ":", "https:", "//", "github.com", "/", "HarleyHK", "/", "JMPF", ".", "a", "section", ":", "III.2", "The", "inherent", "zero", "-", "center", "hyperplanes", "as", "thresholds", "for", "split", "-", "nodes", "In", "training", "a", "random", "forest", ",", "the", "two", "main", "operations", "for", "training", "(", "splitting", ")", "each", "split", "-", "node", "are", "to", "choose", "splitting", "feature", "(", "s", ")", ",", "and", "to", "determine", "the", "threshold", ",", "using", "a", "random", "bagging", "strategy", ",", "which", "can", "avoid", "over", "-", "fitting", "in", "training", "classifiers", ".", "In", "the", "rotated", "compact", "pre", "-", "clustered", "feature", "space", ",", "the", "inherent", "zerocenter", "hyperplanes", "are", "inherently", "the", "optimal", "thresholds", "(", "to", "meet", "the", "max", "-", "purity", "criterion", "on", "two", "clustered", "data", "groups", ")", "after", "training", "the", "rotation", "matrix", ".", "Therefore", ",", "these", "inherent", "zero", "-", "center", "hyperplanes", "can", "directly", "be", "set", "as", "the", "thresholds", "to", "achieve", "optimal", "classification", "performance", "on", "training", "a", "random", "forest", ".", "Compared", "to", "conventional", "random", "forests", ",", "our", "proposed", "JMPF", "only", "needs", "to", "choose", "which", "feature", "(", "s", ")", "to", "split", "data", "at", "split", "-", "nodes", ".", "This", "can", "speed", "up", "the", "training", "process", "for", "a", "random", "forest", ".", "Experimental", "results", "in", "the", "next", "subsection", "will", "validate", "this", "performance", ".", "section", ":", "III.3", ":", "Experimental", "results", "on", "JMPF", "regression", "and", "classification", "To", "evaluate", "the", "performances", "of", "the", "proposed", "JMPF", ",", "we", "test", "it", "with", "15", "standard", "machine", "-", "learning", "tasks", ",", "7", "for", "classification", "and", "8", "for", "regression", ".", "The", "datasets", "used", "in", "the", "experiments", "are", "summarized", "in", "Table", "-", "1", ".", "We", "use", "standard", "performance", "evaluation", "metrics", ":", "error", "rate", "for", "classification", "and", "root", "mean", "squared", "error", "(", "RMSE", ")", "for", "regression", ",", "unless", "otherwise", "specified", ".", "We", "firstly", "evaluate", "the", "proposed", "approach", "on", "two", "real", "applications", ",", "one", "for", "classification", "(", "Table", "-", "2", ")", "and", "one", "for", "regression", "(", "Table", "-", "3", ")", ".", "Our", "proposed", "JMPF", "is", "compared", "with", "the", "original", "random", "forest", "before", "refinement", "(", "denoted", "as", "RF", ")", ",", "and", "two", "state", "-", "of", "-", "the", "-", "art", "variants", ":", "alternating", "decision", "forests", "(", "ADF", ")", "[", "reference", "]", "and", "alternating", "regression", "forests", "(", "ARF", ")", "[", "reference", "]", ",", "for", "classification", "and", "regression", ",", "respectively", ".", "Furthermore", ",", "we", "compare", "with", "JMPF", "+", "ADF", "/", "ARF", ",", "for", "demonstrating", "that", "our", "algorithm", "can", "be", "combined", "with", "other", "methods", ".", "We", "follow", "the", "experiment", "settings", "in", "[", "reference", "][", "reference", "]", ".", "We", "set", "the", "maximum", "tree", "depth", "D", "at", "15", ",", "and", "the", "minimum", "sample", "number", "in", "a", "splitting", "node", "is", "set", "at", "5", ".", "The", "experiments", "were", "repeated", "five", "times", ",", "and", "the", "average", "error", "and", "standard", "deviation", "were", "measured", ".", "The", "results", "are", "presented", "in", "Table", "-", "2", "and", "Table", "-", "3", ",", "for", "the", "classification", "and", "regression", "tasks", ",", "respectively", ".", "In", "terms", "of", "accuracy", ",", "our", "proposed", "JMPF", "significantly", "outperforms", "the", "standard", "random", "forest", "on", "all", "classification", "and", "regression", "tasks", ".", "Compared", "to", "RF", ",", "JMPF", "achieves", "an", "average", "of", "23.57", "%", "improvement", "on", "the", "classification", "tasks", ",", "and", "an", "average", "of", "23.13", "%", "improvement", "on", "the", "regression", "tasks", ".", "Our", "method", "also", "consistently", "outperforms", "the", "state", "-", "of", "-", "theart", "variants", ":", "ADF", "/", "ARF", ".", "Moreover", ",", "the", "performance", "of", "our", "JMPF", "algorithm", "can", "be", "further", "improved", "by", "integrating", "with", "ADF", "and", "ARF", ",", "denoted", "as", "JMPF", "+", "ADF", "/", "ARF", ".", "As", "shown", "in", "Table", "-", "2", "and", "Table", "-", "[", "reference", "]", ",", "JMPF", ":", "proposed", "algorithm", ",", "JMPF", "+", "ARF", ":", "our", "proposed", "algorithm", "embedded", "into", "ARF", ".", "is", "the", "error", "scale", ".", "The", "number", "of", "randomly", "chosen", "hyperplanes", "#", "ℋ", "is", "3", ".", "The", "percentages", "in", "brackets", "for", "JMPF", "and", "JMPF", "+", "ARF", "are", "the", "reduction", "rates", "in", "RMSE", "compared", "with", "the", "RF", "algorithm", ".", "section", ":", "III.4", ":", "Discussions", "on", "Experimental", "Results", "The", "computational", "complexity", "of", "JMPF", "is", "similar", "to", "that", "of", "the", "standard", "random", "forest", ".", "As", "illustrated", "in", "the", "workflow", "of", "JMPF", "in", "Fig", ".", "3", ",", "only", "one", "additional", "step", ",", "which", "computes", "the", "rotation", "matrix", ",", "is", "required", ",", "when", "compared", "to", "the", "standard", "random", "forest", ".", "For", "a", "small", "dataset", "(", "e.g.", ",", "feature", "dimension", "size", "less", "than", "500", "and", "data", "size", "less", "than", "10", ",", "000", ")", ",", "the", "computation", "required", "to", "compute", "the", "rotation", "matrix", "for", "clustering", "data", "into", "the", "feature", "-", "space", "vertices", "is", "acceptable", "in", "the", "training", "stage", "(", "about", "10", "seconds", "per", "level", ",", "using", "MatLab", ")", "and", "negligible", "in", "the", "testing", "stage", ".", "When", "the", "dimension", "size", "becomes", "larger", ",", "PCA", "dimensionality", "reduction", "can", "be", "employed", ".", "If", "the", "size", "of", "the", "dataset", "increases", ",", "such", "that", "using", "PCA", "still", "involves", "heavy", "computation", ",", "bagging", "can", "be", "used", "to", "achieve", "comparable", "accuracy", "and", "the", "whole", "extra", "computation", "will", "be", "insignificant", ".", ",", "the", "number", "hyperplane", "(", "s", ")", "#", "ℋ", "on", "training", "the", "random", "forest", "is", "3", ")", ".", "To", "study", "the", "stability", "of", "JMPF", ",", "we", "choose", "the", "letterorig", "dataset", "for", "classification", "and", "the", "kin8", "nm", "dataset", "for", "regression", ",", "and", "the", "respective", "results", "are", "shown", "in", "Fig", ".", "4", "(", "a", ")", "and", "Fig", ".", "4", "(", "b", ")", ",", "respectively", ".", "In", "the", "experiments", ",", "the", "number", "of", "trees", ",", "i.e.", ",", "the", "number", "of", "weak", "classifiers", "in", "the", "random", "forest", ",", "varies", "from", "10", "to", "200", ",", "and", "we", "have", "three", "observations", ".", "Firstly", ",", "as", "shown", "in", "Fig", ".", "4", ",", "when", "the", "number", "of", "trees", "increases", ",", "the", "performance", "of", "all", "the", "algorithms", "improves", ".", "For", "classification", ",", "as", "shown", "in", "Fig", ".", "4", "(", "a", ")", ",", "when", "the", "number", "of", "trees", "is", "larger", "than", "100", ",", "the", "errors", "are", "converged", "to", "become", "steady", ".", "On", "the", "contrary", ",", "for", "the", "regression", "task", "as", "shown", "in", "Fig", ".", "4", "(", "b", ")", ",", "the", "errors", "are", "almost", "stable", ",", "ranged", "from", "10", "to", "200", ".", "Secondly", ",", "the", "results", "show", "that", "JMPF", "consistently", "outperforms", "ADF", "and", "RF", ",", "irrespective", "of", "the", "number", "of", "trees", "used", ".", "Finally", ",", "Fig", ".", "4", "clearly", "shows", "that", "JMPF", "can", "integrate", "with", "ADF", "or", "ARF", "to", "further", "improve", "its", "performance", ".", "section", ":", "IV", ".", "IMAGE", "SUPER", "-", "RESOLUTION", "BASED", "ON", "JOINT", "MAXIMUM", "PURITY", "FOREST", "section", ":", "IV.1", "Overview", "of", "Image", "Super", "-", "resolution", "and", "Related", "Works", "Image", "super", "-", "resolution", "(", "SR", ")", ",", "which", "recovers", "a", "high", "-", "resolution", "(", "HR", ")", "image", "from", "one", "single", "image", "or", "a", "number", "of", "low", "-", "resolution", "(", "LR", ")", "images", ",", "has", "been", "a", "hot", "research", "topic", "in", "the", "field", "of", "image", "processing", "for", "decades", ".", "SR", "is", "a", "well", "-", "known", "ill", "-", "posed", "problem", ",", "which", "needs", "artistic", "skills", "from", "mathematics", "and", "machine", "learning", ".", "Prior", "methods", "on", "SR", "are", "mainly", "based", "on", "edge", "preserving", ",", "such", "as", "New", "Edge", "-", "directed", "Interpolation", "(", "NEDI", ")", "[", "reference", "]", ",", "Soft", "-", "decision", "Adaptive", "Interpolation", "(", "SAI", ")", "[", "reference", "]", ",", "Directional", "Filtering", "and", "Data", "-", "Fusion", "(", "DFDF", ")", "[", "reference", "]", ",", "Modified", "Edge", "-", "Directed", "Interpolation", "(", "MEDI", ")", "[", "reference", "]", ",", "etc", ".", "The", "neighbor", "-", "embedding", "(", "NE", ")", "methods", "[", "reference", "][", "reference", "]", "set", "the", "milestone", "on", "the", "patch", "-", "learning", "-", "based", "superresolution", "approach", ".", "In", "this", "approach", ",", "each", "LR", "patch", "is", "approximated", "as", "a", "linear", "regression", "of", "its", "nearest", "LR", "neighbors", "in", "a", "collected", "dataset", ",", "while", "its", "HR", "counterpart", "can", "be", "reconstructed", "with", "the", "same", "coefficients", "of", "corresponding", "HR", "neighbors", ",", "based", "on", "the", "non", "-", "linear", "manifold", "structure", ".", "Although", "the", "NE", "method", "is", "simple", "and", "practical", ",", "it", "requires", "a", "huge", "dataset", "(", "millions", "of", "patches", ")", "to", "achieve", "good", "reconstruction", "quality", "and", "it", "is", "computationally", "intensive", ",", "because", "k", "-", "NN", "is", "used", "in", "searching", "neighboring", "patches", "in", "the", "huge", "dataset", ".", "Instead", "of", "using", "the", "patches", "extracted", "directly", "from", "natural", "images", ",", "Yang", "et", "al", ".", "[", "reference", "]", "employed", "sparse", "coding", "[", "reference", "][", "reference", "]", "to", "represent", "patch", "images", ",", "of", "large", "size", ",", "efficiently", ",", "which", "opens", "the", "era", "for", "sparse", "coding", "in", "the", "image", "inverse", "problems", ".", "The", "sparse", "-", "coding", "super", "-", "resolution", "(", "ScSR", ")", "approach", "is", "a", "framework", "that", "the", "HR", "counterpart", "of", "an", "LR", "patch", "can", "be", "reconstructed", "aided", "by", "two", "learned", "dictionaries", ",", "with", "the", "sparse", "constraint", "on", "the", "coefficients", "via", "the", "following", "formulations", ":", "The", "compact", "LR", "and", "HR", "dictionaries", "can", "be", "jointly", "learned", "with", "a", "sparsity", "constraint", ",", "using", "the", "following", "sparse", "representation", ":", "where", "and", "are", "the", "LR", "patch", "and", "the", "corresponding", "HR", "patch", ",", "respectively", ";", "and", "D", "and", "D", "are", "the", "LR", "and", "HR", "dictionaries", "learned", "from", "the", "LR", "and", "the", "corresponding", "HR", "patch", "samples", ",", "respectively", ".", "The", "value", "of", "in", "‖", "‖", "is", "the", "sparsity", "factor", "of", "the", "coefficients", ".", "‖", "‖", "is", "-", "norm", ",", "which", "means", "the", "non", "-", "zero", "count", "of", "the", "coefficients", "in", ".", "For", "each", "LR", "patch", "of", "an", "input", "LR", "image", ",", "the", "problem", "of", "finding", "the", "sparse", "coefficients", "can", "be", "formulated", "as", "follows", ":", "where", "is", "a", "linear", "or", "non", "-", "linear", "feature", "-", "extraction", "operator", "on", "the", "LR", "patches", ",", "which", "makes", "the", "LR", "patches", "more", "discriminative", "from", "each", "other", ".", "Typically", ",", "can", "be", "chosen", "as", "a", "high", "-", "pass", "filter", ",", "and", "a", "simple", "high", "-", "pass", "filter", "can", "be", "obtained", "by", "subtracting", "the", "input", "from", "the", "output", "of", "a", "low", "-", "pass", "filter", ",", "as", "in", "an", "early", "work", "[", "reference", "]", ".", "In", "The", "ideal", "regularization", "term", "for", "the", "sparse", "constraint", "on", "the", "coefficients", "α", "is", "the", "-", "norm", "(", "nonconvex", ")", ",", "but", ",", "based", "on", "greedy", "matching", ",", "it", "leads", "to", "an", "NP", "-", "hard", "problem", ".", "Alternatively", ",", "Yang", "et", "al", ".", "[", "reference", "]", "relaxed", "it", "to", "-", "norm", ",", "as", "shown", "in", "the", "following", "formulation", ":", "The", "Lagrange", "multiplier", "provides", "an", "equivalent", "formulation", "as", "follows", ":", "where", "the", "parameter", "balances", "the", "sparsity", "of", "the", "solution", "and", "the", "fidelity", "of", "the", "approximation", "to", ".", "However", ",", "the", "effectiveness", "of", "sparsity", "was", "challenged", "in", "[", "reference", "][", "reference", "]", ",", "as", "to", "whether", "real", "sparsity", "can", "help", "image", "classification", "and", "restoration", ",", "or", "locality", "property", "can", "achieve", "the", "same", "effect", ".", "Timofte", "et", "al", ".", "[", "reference", "]", "proposed", "an", "anchored", "neighborhood", "regression", "(", "ANR", ")", "framework", ",", "which", "relaxes", "the", "sparse", "decomposition", "optimization", "(", "-", "norm", ")", "of", "[", "reference", "][", "reference", "]", "to", "a", "ridge", "regression", "(", "-", "norm", ")", "problem", ".", "An", "important", "step", "in", "the", "ANR", "model", "is", "the", "relaxation", "of", "the", "-", "norm", "in", "Eqn", ".", "(", "23", ")", "to", "the", "-", "norm", "least", "-", "squares", "minimization", "constraint", ",", "as", "follows", ":", "where", "D", "and", "D", "are", "the", "LR", "and", "HR", "patch", "-", "based", "dictionaries", ",", "respectively", ".", "This", "-", "norm", "constraint", "problem", "can", "be", "solved", "with", "a", "closed", "-", "form", "solution", "from", "the", "ridge", "regression", "[", "reference", "]", "theory", ".", "Based", "on", "the", "Tikhonov", "regularization", "/", "ridge", "-", "regression", "theory", ",", "the", "closed", "-", "form", "solution", "of", "the", "coefficients", "is", "given", ":", "We", "assume", "that", "the", "HR", "patches", "share", "the", "same", "coefficient", "α", "from", "their", "counterpart", "LR", "patches", ",", "i.e.", ",", "=", "D", ".", "From", "Eqn", ".", "(", "25", ")", ",", "we", "have", ":", "Therefore", ",", "the", "HR", "patches", "can", "be", "reconstructed", "by", ":", "=", "y", ",", "where", "can", "be", "considered", "a", "projection", "matrix", ",", "which", "can", "be", "calculated", "offline", ",", "as", "follows", ":", "Ridge", "regression", "allows", "the", "coefficients", "to", "be", "calculated", "by", "multiplying", "the", "constant", "projection", "matrix", "with", "the", "new", "extracted", "feature", ",", "as", "described", "in", "Eqn", ".", "(", "26", ")", "and", "Eqn", ".", "[", "reference", "]", ".", "More", "importantly", ",", "the", "projection", "matrix", "can", "be", "pre", "-", "computed", ",", "and", "this", "offline", "learning", "enables", "significant", "speed", "-", "up", "at", "the", "prediction", "stage", ".", "Timofte", "et", "al", ".", "[", "reference", "]", "further", "extended", "the", "ANR", "approach", "to", "the", "A", "+", "approach", ",", "which", "learns", "regressors", "from", "all", "the", "training", "samples", ",", "rather", "than", "from", "a", "small", "quantity", "of", "neighbors", "of", "the", "anchor", "atoms", "as", "ANR", "does", ".", "Later", ",", "there", "are", "numerous", "variants", "and", "extended", "approaches", ",", "based", "on", "ANR", "and", "A", "+", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "By", "investigating", "the", "ANR", "model", ",", "Li", "et", "al", ".", "[", "reference", "]", "found", "that", "the", "weights", "of", "the", "supporting", "atoms", "can", "be", "of", "different", "values", "to", "represent", "their", "similarities", "to", "the", "anchor", "atom", ".", "Based", "on", "this", "idea", ",", "the", "normal", "collaborative", "representation", "(", "CR", ")", "model", "in", "ANR", "is", "generalized", "to", "a", "weighted", "model", ",", "named", "as", "weighted", "collaborative", "representation", "(", "WCR", ")", "model", ",", "as", "follows", ":", "where", "is", "a", "diagonal", "matrix", ".", "The", "weights", "on", "the", "diagonal", "atoms", "are", "proportional", "to", "their", "similarities", "to", "the", "anchor", "atom", ".", "Similarly", ",", "the", "new", "closed", "-", "form", "solution", "for", "the", "coefficients", "can", "be", "calculated", "offline", ",", "as", "follows", ":", "and", "the", "new", "projection", "matrix", "is", "given", "as", "follows", ":", "The", "WCR", "model", "can", "further", "improve", "the", "ANR", "or", "A", "+", "model", "in", "terms", "of", "image", "quality", ",", "but", "it", "is", "still", "a", "time", "-", "consuming", "problem", "to", "find", "the", "most", "similar", "anchor", "atoms", "in", "a", "dictionary", ",", "and", "this", "always", "hinders", "its", "applications", "where", "fast", "speed", "is", "greatly", "required", ".", "Schulter", "et", "al", ".", "[", "reference", "]", "adopted", "the", "random", "forest", "as", "a", "classifier", ",", "and", "the", "regressors", "are", "learned", "from", "the", "patches", "in", "the", "leaf", "-", "nodes", ".", "With", "the", "same", "number", "of", "regressors", ",", "these", "random", "-", "forest", "-", "based", "methods", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "can", "perform", "on", "a", "par", "with", "the", "A", "+", "method", "in", "terms", "of", "accuracy", ".", "However", ",", "they", "achieve", "an", "increase", "in", "speed", ",", "because", "the", "sublinear", "search", "property", "of", "random", "forest", "can", "remarkably", "reduce", "the", "regressors", "'", "search", "complexity", ".", "Recently", ",", "deep", "learning", "has", "become", "a", "hot", "research", "topic", ",", "which", "has", "been", "successfully", "applied", "to", "image", "super", "-", "resolution", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "and", "achieved", "promising", "performance", ",", "particularly", "in", "terms", "of", "image", "quality", ".", "In", "[", "reference", "][", "reference", "]", ",", "a", "convolutional", "neural", "-", "network", "-", "based", "image", "super", "-", "resolution", "(", "SRCNN", ")", "was", "proposed", ",", "in", "which", "an", "end", "-", "to", "-", "end", "mapping", "between", "LR", "and", "HR", "images", "is", "learned", "through", "a", "deep", "convolutional", "neural", "network", "(", "CNN", ")", ".", "The", "recent", "emerging", "stream", "[", "reference", "][", "reference", "]", "on", "single", "-", "image", "SR", "is", "to", "formulate", "the", "problem", "as", "a", "clusteringregression", "problem", ",", "which", "can", "be", "solved", "with", "machine", "-", "learning", "tools", ".", "These", "approaches", "are", "learningbased", "methods", ",", "which", "attempt", "to", "reconstruct", "an", "HR", "image", "from", "patches", "with", "the", "help", "of", "an", "external", "database", ".", "These", "methods", "first", "decompose", "an", "image", "into", "patches", ",", "then", "classify", "them", "into", "clusters", ".", "Regressors", "are", "then", "trained", "for", "each", "of", "the", "clusters", ",", "which", "generate", "mappings", "from", "an", "input", "LR", "patch", "'s", "feature", "to", "its", "corresponding", "HR", "patch", "(", "see", "Fig", ".", "5", ")", ".", "In", "the", "testing", "stage", ",", "an", "LR", "query", "image", "follows", "the", "same", "procedures", "to", "cut", "into", "patches", "and", "to", "extract", "features", ",", "which", "are", "then", "assigned", "to", "their", "corresponding", "clusters", "using", "the", "k", "-", "NN", "algorithm", "[", "reference", "][", "reference", "]", "or", "random", "forest", "[", "reference", "][", "reference", "][", "reference", "]", ".", "The", "respective", "HR", "patches", "are", "constructed", "through", "regressors", "learned", "for", "the", "clusters", "(", "see", "Fig", ".", "6", ")", ".", "This", "kind", "of", "clustering", "-", "regression", "algorithms", ",", "based", "on", "random", "forest", "[", "reference", "][", "reference", "][", "reference", "]", ",", "has", "achieved", "state", "-", "of", "-", "the", "-", "art", "performance", "in", "single", "image", "super", "-", "resolution", ",", "both", "in", "terms", "of", "accuracy", "and", "efficiency", ",", "because", "of", "the", "use", "of", "ensemble", "learning", "and", "sublinear", "search", ".", "As", "JMPF", "achieves", "promising", "results", "on", "both", "classification", "and", "regression", "tasks", ",", "it", "can", "be", "employed", "for", "image", "super", "-", "resolution", "for", "better", "performances", ".", "An", "overview", "of", "the", "training", "and", "testing", "processes", "of", "the", "proposed", "JMPF", "-", "based", "image", "SR", "method", "is", "illustrated", "in", "Fig", ".", "5", "and", "Fig", ".", "6", ",", "respectively", ".", "In", "our", "method", ",", "the", "first", "and", "second", "-", "order", "gradients", "are", "extracted", "as", "features", "from", "each", "patch", ",", "followed", "by", "PCA", "for", "dimensionality", "reduction", ".", "These", "features", "are", "then", "rotated", "into", "a", "more", "compact", ",", "pre", "-", "clustered", "feature", "space", ".", "Finally", ",", "all", "the", "thresholds", "are", "directly", "set", "to", "the", "inherent", "zero", "-", "center", "hyperplanes", "when", "training", "the", "random", "forest", ",", "and", "similar", "to", "other", "algorithms", ",", "the", "regressors", "at", "the", "leaf", "-", "nodes", "are", "computed", "using", "the", "rigid", "regression", "algorithms", ".", "This", "approach", "is", "named", "as", "JMPF", "-", "based", "image", "super", "-", "resolution", "method", ".", "section", ":", "IV.3", "The", "Working", "Processes", "of", "JMPF", "-", "based", "Image", "Super", "-", "resolution", "JMPF", "has", "been", "shown", "to", "achieve", "a", "better", "performance", "for", "clustering", "and", "classification", "than", "other", "random", "forest", "methods", ".", "Since", "image", "super", "-", "resolution", "can", "be", "considered", "as", "a", "clustering", "/", "classification", "problem", ",", "using", "JMPF", "is", "likely", "to", "result", "in", "better", "performance", ".", "This", "is", "mainly", "due", "to", "the", "features", "transformed", "to", "the", "vertices", "in", "the", "new", "feature", "space", ",", "so", "the", "features", "become", "more", "discriminative", ".", "The", "image", "super", "-", "resolution", "training", "and", "testing", "processes", "of", "our", "proposed", "JMPF", "-", "based", "method", "are", "described", "in", "Algorithm", "1", "and", "Algorithm", "2", ",", "respectively", ".", "section", ":", "IV.4", "Experimental", "Results", "on", "JMPF", "-", "based", "Image", "Super", "-", "Resolution", "In", "this", "section", ",", "we", "evaluate", "our", "image", "SR", "algorithm", "on", "some", "standard", "super", "-", "resolution", "datasets", ",", "including", "Set", "5", ",", "Set14", ",", "and", "B100", "[", "reference", "]", ",", "and", "compare", "it", "with", "a", "number", "of", "classical", "or", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "These", "include", "bicubic", "interpolation", ",", "sparse", "representation", "SR", "(", "Zeyde", ")", "[", "reference", "]", ",", "anchored", "neighborhood", "regression", "(", "ANR", ")", "[", "reference", "]", ",", "A", "+", "[", "reference", "]", ",", "standard", "random", "forest", "(", "RF", ")", "[", "reference", "]", ",", "and", "alternating", "regression", "forests", "(", "ARF", ")", "[", "reference", "]", ".", "We", "set", "the", "same", "parameters", "for", "all", "the", "random", "-", "forest", "-", "based", "algorithms", ":", "the", "number", "of", "trees", "in", "the", "random", "forest", "is", "10", ",", "and", "the", "maximum", "depth", "of", "each", "tree", "is", "15", ".", "Experiment", "results", "are", "tabulated", "in", "Tables", "-", "4", "and", "Tables", "-", "5", ",", "where", "JMPF", "is", "our", "proposed", "JMPF", "-", "based", "image", "super", "-", "resolution", "method", ",", "and", "JMPF", "", "is", "a", "trimmed", "version", ",", "such", "that", "the", "thresholds", "for", "the", "splitnodes", "are", "not", "the", "inherent", "zero", "-", "center", "hyperplanes", ",", "but", "set", "by", "the", "standard", "random", "-", "forest", "bagging", "strategy", ".", "We", "use", "the", "same", "training", "images", "(", "91", "images", ")", "for", "all", "the", "algorithms", "as", "previous", "works", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "do", ".", "However", ",", "for", "JMPF", "+", ",", "100", "more", "images", "from", "the", "General", "-", "100", "dataset", "[", "reference", "]", "are", "used", ",", "so", "as", "to", "check", "whether", "or", "not", "more", "training", "samples", "can", "further", "improve", "our", "proposed", "algorithm", ".", "Table", "-", "5", ":", "Detailed", "results", "of", "the", "proposed", "method", ",", "compared", "with", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "the", "dataset", "Set5", ",", "in", "terms", "of", "PSNR", "(", "dB", ")", "using", "three", "different", "magnification", "factors", "(", "×2", ",", "×3", ",", "×4", ")", ".", "Table", "-", "4", "tabulates", "the", "performances", ",", "in", "terms", "of", "the", "average", "peak", "signal", "to", "noise", "ratio", "(", "PSNR", ")", "scores", ",", "of", "our", "proposed", "algorithm", "and", "other", "image", "SR", "methods", ",", "on", "the", "3", "datasets", "with", "different", "magnification", "factors", ".", "For", "the", "Set5", "and", "Set14", "datasets", ",", "with", "different", "magnification", "factors", ",", "our", "proposed", "JMPF", "-", "based", "algorithm", "can", "achieve", "a", "comparable", "performance", "to", "other", "recent", "state", "-", "of", "-", "the", "-", "art", "methods", ",", "such", "as", "A", "+", "and", "ARF", ".", "As", "those", "random", "-", "forest", "-", "based", "algorithms", "may", "not", "be", "stable", "on", "small", "datasets", ",", "when", "evaluation", "works", "on", "extensive", "datasets", ",", "such", "as", "B100", ",", "our", "proposed", "algorithm", "JMPF", "can", "stably", "outperform", "A", "+", "and", "ARF", "for", "all", "magnification", "factors", "(", "×2", ",", "×3", ",", "×4", ")", ".", "Moreover", ",", "the", "objective", "quality", "metrics", "on", "PSNR", "also", "show", "that", "the", "JMPF", "algorithm", "can", "achieve", "a", "better", "performance", "when", "more", "samples", "are", "used", "for", "training", ",", "as", "shown", "from", "JMPF", "+", "in", "Table", "-", "4", ".", "Table", "-", "5", "provides", "more", "details", "of", "the", "performances", "in", "datasets", "Set5", ".", "To", "compare", "the", "visual", "quality", "of", "our", "proposed", "JMPF", "-", "based", "SR", "algorithm", "to", "other", "methods", ",", "Fig", ".", "7", ",", "shows", "the", "reconstructed", "HR", "images", "using", "different", "methods", ".", "Some", "regions", "in", "the", "reconstructed", "images", "are", "also", "enlarged", ",", "so", "as", "to", "show", "the", "details", "in", "the", "images", ".", "In", "general", ",", "our", "proposed", "method", "can", "produce", "better", "quality", "images", ",", "particularly", "in", "areas", "with", "rich", "texture", ",", "which", "verifies", "the", "feature", "discrimination", "of", "the", "proposed", "JMPF", "scheme", ".", "section", ":", "V.", "CONCLUSIONS", "In", "this", "paper", ",", "we", "have", "proposed", "a", "novel", "random", "-", "forest", "scheme", ",", "namely", "the", "Joint", "Maximum", "Purity", "Forest", "(", "JMPF", ")", "scheme", ",", "which", "rotates", "the", "feature", "space", "into", "a", "compact", ",", "clustered", "feature", "space", ",", "by", "jointly", "maximizing", "the", "purity", "of", "all", "the", "feature", "-", "space", "vertices", ".", "In", "the", "new", "pre", "-", "clustered", "feature", "space", ",", "orthogonal", "hyperplanes", "can", "be", "effectively", "used", "in", "the", "split", "-", "nodes", "of", "a", "decision", "tree", ",", "which", "can", "improve", "the", "performance", "of", "the", "trained", "random", "forest", ".", "Compared", "to", "the", "standard", "random", "forests", "and", "the", "recent", "state", "-", "of", "-", "the", "-", "art", "variants", ",", "such", "as", "alternating", "decision", "forests", "(", "ADF", ")", "[", "reference", "]", "and", "alternating", "regression", "forests", "(", "ARF", ")", "[", "reference", "]", ",", "our", "proposed", "random", "-", "forest", "method", "inherits", "the", "merits", "of", "random", "forests", "(", "fast", "training", "and", "testing", ",", "multi", "-", "class", "capability", ",", "etc", ".", ")", ",", "and", "also", "yields", "promising", "results", "on", "both", "classification", "and", "regression", "tasks", ".", "Experiments", "have", "shown", "that", "our", "method", "achieves", "an", "average", "improvement", "of", "about", "20", "%", "for", "classification", "and", "regression", "on", "publicly", "benchmarked", "datasets", ".", "Furthermore", ",", "our", "proposed", "scheme", "can", "integrate", "with", "other", "methods", ",", "such", "as", "ADF", "and", "ARF", ",", "to", "further", "improve", "the", "performance", ".", "We", "have", "also", "applied", "JMPF", "to", "single", "-", "image", "super", "-", "resolution", ".", "We", "tackle", "image", "super", "-", "resolution", "as", "a", "clustering", "-", "regression", "problem", ",", "and", "focus", "on", "the", "clustering", "stage", ",", "which", "happens", "at", "the", "split", "-", "nodes", "of", "each", "decision", "tree", ".", "By", "employing", "the", "JMPF", "strategy", ",", "we", "rotate", "the", "feature", "space", "into", "a", "pre", "-", "clustered", "feature", "space", ",", "which", "can", "cluster", "samples", "into", "different", "sub", "-", "spaces", "more", "compactly", "in", "an", "unsupervised", "problem", ".", "The", "compact", "pre", "-", "clustered", "feature", "space", "can", "provide", "the", "optimal", "thresholds", "for", "split", "-", "nodes", "in", "decision", "trees", ",", "which", "are", "the", "zero", "-", "centered", "orthogonal", "hyperplanes", ".", "Our", "experiment", "results", "on", "intensive", "image", "benchmark", "datasets", ",", "such", "as", "B100", ",", "show", "that", "the", "proposed", "JMPF", "-", "based", "image", "super", "-", "resolution", "approach", "can", "consistently", "outperform", "recent", "state", "-", "of", "-", "the", "-", "art", "algorithms", ",", "in", "terms", "of", "PSNR", "and", "visual", "quality", ".", "Our", "method", "also", "inherits", "the", "advantages", "of", "random", "forests", ",", "which", "have", "fast", "speed", "on", "both", "the", "training", "and", "inference", "processes", ".", "section", ":"], "entity": [[0, 4, ["Joint", "Maximum", "Purity", "Forest"], "Method"], [7, 11, ["Image", "Super", "-", "Resolution"], "Task"], [23, 27, ["random", "-", "forest", "scheme"], "Method"], [29, 33, ["Joint", "Maximum", "Purity", "Forest"], "Method"], [34, 35, ["JMPF"], "Method"], [38, 45, ["classification", ",", "clustering", ",", "and", "regression", "tasks"], "Task"], [79, 82, ["iterative", "quantization", "process"], "Method"], [127, 129, ["random", "forests"], "Method"], [133, 135, ["clustering", "problems"], "Task"], [147, 151, ["regression", "and", "classification", "tasks"], "Task"], [156, 157, ["JMPF"], "Method"], [167, 173, ["random", "-", "forest", "-", "based", "approaches"], "Method"], [178, 179, ["JMPF"], "Method"], [180, 184, ["image", "super", "-", "resolution"], "Task"], [196, 200, ["clustering", "-", "regression", "scheme"], "Method"], [232, 237, ["image", "super", "-", "resolution", "algorithms"], "Method"], [246, 247, ["SVM"], "Method"], [257, 258, ["SVM"], "Method"], [314, 317, ["global", "maximum", "purity"], "Metric"], [338, 340, ["maximum", "purity"], "Metric"], [349, 351, ["random", "forest"], "Method"], [352, 356, ["Image", "super", "-", "resolution"], "Task"], [361, 364, ["clustering", "/", "classification"], "Method"], [370, 374, ["clustering", "-", "regression", "stream"], "Method"], [393, 397, ["classification", "and", "regression", "tasks"], "Task"], [400, 401, ["JMPF"], "Method"], [404, 410, ["single", "-", "image", "super", "-", "resolution"], "Task"], [418, 421, ["principal", "component", "analysis"], "Method"], [422, 423, ["PCA"], "Method"], [430, 432, ["dimensionality", "reduction"], "Task"], [464, 466, ["random", "forest"], "Method"], [488, 492, ["maximum", "-", "purity", "criterion"], "Metric"], [497, 498, ["JMPF"], "Method"], [502, 505, ["clustering", "/", "classification"], "Metric"], [507, 509, ["random", "forests"], "Method"], [512, 513, ["JMPF"], "Method"], [514, 518, ["image", "super", "-", "resolution"], "Task"], [612, 618, ["random", "-", "forest", "-", "based", "approaches"], "Method"], [619, 623, ["regression", "and", "classification", "tasks"], "Task"], [637, 643, ["single", "-", "image", "super", "-", "resolution"], "Task"], [654, 659, ["JOINT", "MAXIMUM", "PURITY", "FOREST", "SCHEME"], "Method"], [662, 664, ["Random", "Forest"], "Method"], [668, 670, ["random", "forest"], "Method"], [672, 677, ["ensemble", "of", "binary", "decision", "trees"], "Method"], [933, 935, ["random", "forest"], "Method"], [944, 947, ["random", "bagging", "strategy"], "Method"], [953, 955, ["random", "forest"], "Method"], [961, 963, ["decision", "trees"], "Method"], [989, 992, ["random", "bagging", "strategy"], "Method"], [995, 997, ["prediction", "stage"], "Task"], [999, 1001, ["decision", "tree"], "Method"], [1027, 1028, ["averaging"], "Method"], [1042, 1044, ["splitting", "function"], "Method"], [1192, 1194, ["random", "forest"], "Method"], [1196, 1198, ["decision", "trees"], "Method"], [1224, 1226, ["decision", "trees"], "Method"], [1233, 1239, ["random", "-", "forest", "-", "based", "classification"], "Method"], [1246, 1248, ["decision", "tree"], "Method"], [1292, 1294, ["stopping", "criterion"], "Metric"], [1340, 1342, ["density", "model"], "Method"], [1402, 1404, ["Gaussian", "distribution"], "Method"], [1405, 1408, ["kernel", "density", "estimation"], "Method"], [1409, 1411, ["ridge", "regression"], "Method"], [1500, 1502, ["regression", "problem"], "Task"], [1540, 1542, ["Gaussian", "distribution"], "Method"], [1587, 1589, ["random", "forest"], "Method"], [1601, 1603, ["information", "gain"], "Metric"], [1604, 1605, ["IG"], "Metric"], [1613, 1614, ["IG"], "Method"], [1685, 1687, ["decision", "tree"], "Method"], [1689, 1691, ["random", "forest"], "Method"], [1734, 1737, ["randomly", "bagging", "strategy"], "Method"], [1821, 1823, ["decision", "tree"], "Method"], [1848, 1850, ["approximate", "method"], "Method"], [1852, 1855, ["numerically", "solving", "Eqn"], "Task"], [1933, 1935, ["random", "forest"], "Method"], [1948, 1951, ["category", "classification", "problem"], "Task"], [2039, 2040, ["classification"], "Task"], [2420, 2422, ["clustering", "data"], "Task"], [2433, 2440, ["locality", "-", "sensitive", "hashing", "(", "LSH", ")"], "Method"], [2444, 2446, ["image", "representation"], "Method"], [2459, 2462, ["alternating", "minimization", "scheme"], "Method"], [2479, 2481, ["quantization", "errors"], "Metric"], [2502, 2507, ["iterative", "quantization", "(", "ITQ", ")"], "Method"], [2512, 2517, ["multi", "-", "class", "spectral", "clustering"], "Task"], [2518, 2521, ["orthogonal", "Procrustes", "problem"], "Task"], [2532, 2534, ["circulant", "matrix"], "Method"], [2538, 2539, ["computation"], "Task"], [2548, 2551, ["Fast", "Fourier", "Transformation"], "Method"], [2552, 2553, ["FFT"], "Method"], [2564, 2568, ["training", "and", "testing", "stage"], "Task"], [2577, 2578, ["ITQ"], "Method"], [2606, 2607, ["hashing"], "Task"], [2653, 2655, ["random", "forest"], "Method"], [2693, 2697, ["joint", "maximum", "purity", "forest"], "Method"], [2877, 2879, ["quantization", "error"], "Metric"], [2886, 2890, ["minimization", "problem", "of", "min‖"], "Task"], [2897, 2898, ["error"], "Metric"], [2921, 2923, ["minimization", "problem"], "Task"], [2952, 2961, ["expectation", "-", "maximization", "(", "E", "-", "M", ")", "algorithm"], "Method"], [3089, 3091, ["quantization", "error"], "Metric"], [3196, 3199, ["objective", "function", "Eqn"], "Metric"], [3208, 3211, ["orthogonal", "Procrustes", "problem"], "Task"], [3257, 3261, ["closed", "-", "form", "solution"], "Method"], [3268, 3269, ["SVD"], "Method"], [3297, 3300, ["Orthogonal", "Procrustes", "Problem"], "Task"], [3307, 3310, ["orthogonal", "Procrustes", "problem"], "Task"], [3383, 3387, ["JOINT", "MAXIMUM", "PURITY", "FOREST"], "Method"], [3388, 3391, ["REGRESSION", "AND", "CLASSIFICATION"], "Task"], [3397, 3401, ["Joint", "Maximum", "Purity", "Forest"], "Method"], [3401, 3403, ["Random", "forest"], "Method"], [3405, 3409, ["machine", "-", "learning", "method"], "Method"], [3411, 3416, ["ensemble", "of", "randomized", "decision", "trees"], "Method"], [3417, 3418, ["classification"], "Task"], [3423, 3425, ["random", "forest"], "Method"], [3442, 3444, ["random", "forest"], "Method"], [3473, 3475, ["information", "gain"], "Metric"], [3479, 3480, ["tree"], "Method"], [3524, 3526, ["evaluation", "phase"], "Task"], [3552, 3558, ["random", "-", "forest", "-", "based", "models"], "Method"], [3590, 3592, ["random", "forest"], "Task"], [3613, 3615, ["Rigid", "regression"], "Method"], [3616, 3618, ["linear", "regression"], "Method"], [3628, 3630, ["prediction", "task"], "Task"], [3632, 3634, ["rigid", "regression"], "Method"], [3636, 3640, ["closed", "-", "form", "solution"], "Method"], [3642, 3644, ["linear", "regression"], "Method"], [3647, 3649, ["optimization", "tool"], "Method"], [3652, 3654, ["LibLinear", "package"], "Method"], [3670, 3672, ["random", "forests"], "Method"], [3795, 3797, ["random", "forest"], "Method"], [3805, 3806, ["splitting"], "Task"], [3828, 3831, ["random", "bagging", "strategy"], "Method"], [3840, 3841, ["classifiers"], "Method"], [3865, 3869, ["max", "-", "purity", "criterion"], "Metric"], [3899, 3900, ["classification"], "Task"], [3904, 3906, ["random", "forest"], "Method"], [3910, 3912, ["random", "forests"], "Method"], [3915, 3916, ["JMPF"], "Method"], [3938, 3940, ["training", "process"], "Method"], [3942, 3944, ["random", "forest"], "Method"], [3966, 3967, ["classification"], "Task"], [3974, 3975, ["JMPF"], "Method"], [3982, 3986, ["machine", "-", "learning", "tasks"], "Task"], [3989, 3990, ["classification"], "Task"], [3993, 3994, ["regression"], "Task"], [4011, 4014, ["performance", "evaluation", "metrics"], "Metric"], [4015, 4017, ["error", "rate"], "Metric"], [4018, 4019, ["classification"], "Metric"], [4020, 4027, ["root", "mean", "squared", "error", "(", "RMSE", ")"], "Metric"], [4028, 4029, ["regression"], "Task"], [4047, 4048, ["classification"], "Task"], [4056, 4057, ["regression"], "Task"], [4065, 4066, ["JMPF"], "Method"], [4071, 4075, ["random", "forest", "before", "refinement"], "Method"], [4078, 4079, ["RF"], "Method"], [4092, 4095, ["alternating", "decision", "forests"], "Method"], [4096, 4097, ["ADF"], "Method"], [4102, 4105, ["alternating", "regression", "forests"], "Method"], [4106, 4107, ["ARF"], "Method"], [4113, 4116, ["classification", "and", "regression"], "Task"], [4187, 4189, ["average", "error"], "Metric"], [4210, 4214, ["classification", "and", "regression", "tasks"], "Task"], [4220, 4221, ["accuracy"], "Metric"], [4224, 4225, ["JMPF"], "Method"], [4229, 4231, ["random", "forest"], "Method"], [4233, 4237, ["classification", "and", "regression", "tasks"], "Task"], [4240, 4241, ["RF"], "Method"], [4242, 4243, ["JMPF"], "Method"], [4252, 4254, ["classification", "tasks"], "Task"], [4264, 4266, ["regression", "tasks"], "Task"], [4280, 4283, ["ADF", "/", "ARF"], "Method"], [4299, 4302, ["ADF", "and", "ARF"], "Method"], [4324, 4325, ["JMPF"], "Method"], [4338, 4339, ["ARF"], "Method"], [4361, 4362, ["JMPF"], "Method"], [4368, 4372, ["reduction", "rates", "in", "RMSE"], "Metric"], [4375, 4377, ["RF", "algorithm"], "Method"], [4387, 4389, ["computational", "complexity"], "Metric"], [4390, 4391, ["JMPF"], "Method"], [4398, 4400, ["random", "forest"], "Method"], [4407, 4408, ["JMPF"], "Method"], [4432, 4434, ["random", "forest"], "Method"], [4489, 4490, ["MatLab"], "Method"], [4500, 4502, ["dimension", "size"], "Metric"], [4505, 4508, ["PCA", "dimensionality", "reduction"], "Method"], [4523, 4524, ["PCA"], "Method"], [4529, 4530, ["bagging"], "Method"], [4536, 4537, ["accuracy"], "Metric"], [4558, 4560, ["random", "forest"], "Method"], [4569, 4570, ["JMPF"], "Method"], [4577, 4578, ["classification"], "Task"], [4584, 4585, ["regression"], "Task"], [4623, 4625, ["weak", "classifiers"], "Method"], [4627, 4629, ["random", "forest"], "Method"], [4667, 4668, ["classification"], "Task"], [4703, 4705, ["regression", "task"], "Task"], [4733, 4734, ["JMPF"], "Method"], [4736, 4737, ["ADF"], "Method"], [4738, 4739, ["RF"], "Method"], [4756, 4757, ["JMPF"], "Method"], [4760, 4761, ["ADF"], "Method"], [4762, 4763, ["ARF"], "Method"], [4773, 4777, ["IMAGE", "SUPER", "-", "RESOLUTION"], "Task"], [4779, 4783, ["JOINT", "MAXIMUM", "PURITY", "FOREST"], "Method"], [4788, 4792, ["Image", "Super", "-", "resolution"], "Task"], [4795, 4799, ["Image", "super", "-", "resolution"], "Task"], [4800, 4801, ["SR"], "Task"], [4806, 4809, ["high", "-", "resolution"], "Task"], [4810, 4811, ["HR"], "Task"], [4821, 4824, ["low", "-", "resolution"], "Task"], [4825, 4826, ["LR"], "Task"], [4839, 4841, ["image", "processing"], "Task"], [4844, 4845, ["SR"], "Task"], [4850, 4854, ["ill", "-", "posed", "problem"], "Task"], [4862, 4864, ["machine", "learning"], "Task"], [4868, 4869, ["SR"], "Task"], [4873, 4875, ["edge", "preserving"], "Method"], [4878, 4883, ["New", "Edge", "-", "directed", "Interpolation"], "Method"], [4884, 4885, ["NEDI"], "Method"], [4890, 4895, ["Soft", "-", "decision", "Adaptive", "Interpolation"], "Method"], [4896, 4897, ["SAI"], "Method"], [4902, 4904, ["Directional", "Filtering"], "Method"], [4905, 4908, ["Data", "-", "Fusion"], "Method"], [4915, 4920, ["Modified", "Edge", "-", "Directed", "Interpolation"], "Method"], [4921, 4922, ["MEDI"], "Method"], [4930, 4933, ["neighbor", "-", "embedding"], "Method"], [4934, 4935, ["NE"], "Method"], [4947, 4954, ["patch", "-", "learning", "-", "based", "superresolution", "approach"], "Method"], [4966, 4968, ["linear", "regression"], "Method"], [5025, 5027, ["reconstruction", "quality"], "Metric"], [5034, 5037, ["k", "-", "NN"], "Method"], [5067, 5069, ["sparse", "coding"], "Method"], [5090, 5092, ["sparse", "coding"], "Task"], [5094, 5097, ["image", "inverse", "problems"], "Task"], [5099, 5105, ["sparse", "-", "coding", "super", "-", "resolution"], "Method"], [5106, 5107, ["ScSR"], "Method"], [5126, 5128, ["learned", "dictionaries"], "Method"], [5159, 5161, ["sparse", "representation"], "Method"], [5257, 5266, ["linear", "or", "non", "-", "linear", "feature", "-", "extraction", "operator"], "Method"], [5289, 5293, ["high", "-", "pass", "filter"], "Method"], [5297, 5301, ["high", "-", "pass", "filter"], "Method"], [5313, 5317, ["low", "-", "pass", "filter"], "Method"], [5352, 5354, ["greedy", "matching"], "Method"], [5359, 5363, ["NP", "-", "hard", "problem"], "Task"], [5435, 5437, ["image", "classification"], "Task"], [5438, 5439, ["restoration"], "Task"], [5458, 5461, ["anchored", "neighborhood", "regression"], "Method"], [5462, 5463, ["ANR"], "Method"], [5469, 5472, ["sparse", "decomposition", "optimization"], "Method"], [5484, 5486, ["ridge", "regression"], "Method"], [5487, 5491, ["-", "norm", ")", "problem"], "Task"], [5514, 5521, ["-", "norm", "least", "-", "squares", "minimization", "constraint"], "Task"], [5542, 5546, ["-", "norm", "constraint", "problem"], "Task"], [5551, 5555, ["closed", "-", "form", "solution"], "Method"], [5557, 5559, ["ridge", "regression"], "Method"], [5567, 5574, ["Tikhonov", "regularization", "/", "ridge", "-", "regression", "theory"], "Method"], [5576, 5580, ["closed", "-", "form", "solution"], "Method"], [5648, 5650, ["Ridge", "regression"], "Method"], [5697, 5699, ["offline", "learning"], "Method"], [5706, 5708, ["prediction", "stage"], "Task"], [5723, 5726, ["A", "+", "approach"], "Method"], [5764, 5765, ["ANR"], "Method"], [5828, 5830, ["collaborative", "representation"], "Method"], [5831, 5832, ["CR"], "Method"], [5835, 5836, ["ANR"], "Method"], [5840, 5842, ["weighted", "model"], "Method"], [5849, 5852, ["WCR", ")", "model"], "Method"], [5882, 5886, ["closed", "-", "form", "solution"], "Method"], [5908, 5910, ["WCR", "model"], "Method"], [5922, 5924, ["image", "quality"], "Metric"], [5952, 5954, ["fast", "speed"], "Metric"], [5967, 5969, ["random", "forest"], "Method"], [5971, 5972, ["classifier"], "Method"], [5995, 6001, ["random", "-", "forest", "-", "based", "methods"], "Method"], [6023, 6024, ["accuracy"], "Metric"], [6032, 6033, ["speed"], "Metric"], [6040, 6042, ["random", "forest"], "Method"], [6046, 6050, ["regressors", "'", "search", "complexity"], "Metric"], [6053, 6055, ["deep", "learning"], "Method"], [6068, 6072, ["image", "super", "-", "resolution"], "Task"], [6090, 6092, ["image", "quality"], "Metric"], [6101, 6111, ["convolutional", "neural", "-", "network", "-", "based", "image", "super", "-", "resolution"], "Method"], [6112, 6113, ["SRCNN"], "Method"], [6124, 6126, ["end", "mapping"], "Task"], [6127, 6128, ["LR"], "Task"], [6129, 6131, ["HR", "images"], "Task"], [6135, 6139, ["deep", "convolutional", "neural", "network"], "Method"], [6140, 6141, ["CNN"], "Method"], [6164, 6166, ["clusteringregression", "problem"], "Task"], [6172, 6176, ["machine", "-", "learning", "tools"], "Method"], [6180, 6182, ["learningbased", "methods"], "Method"], [6188, 6190, ["HR", "image"], "Task"], [6215, 6216, ["Regressors"], "Method"], [6253, 6256, ["LR", "query", "image"], "Task"], [6279, 6283, ["k", "-", "NN", "algorithm"], "Method"], [6289, 6291, ["random", "forest"], "Method"], [6306, 6307, ["regressors"], "Method"], [6321, 6325, ["clustering", "-", "regression", "algorithms"], "Method"], [6328, 6330, ["random", "forest"], "Method"], [6349, 6354, ["single", "image", "super", "-", "resolution"], "Task"], [6359, 6360, ["accuracy"], "Metric"], [6361, 6362, ["efficiency"], "Metric"], [6368, 6370, ["ensemble", "learning"], "Method"], [6371, 6373, ["sublinear", "search"], "Method"], [6375, 6376, ["JMPF"], "Method"], [6381, 6385, ["classification", "and", "regression", "tasks"], "Task"], [6391, 6395, ["image", "super", "-", "resolution"], "Task"], [6450, 6451, ["PCA"], "Method"], [6452, 6454, ["dimensionality", "reduction"], "Task"], [6489, 6491, ["random", "forest"], "Method"], [6509, 6512, ["rigid", "regression", "algorithms"], "Method"], [6541, 6542, ["JMPF"], "Method"], [6551, 6552, ["clustering"], "Task"], [6553, 6554, ["classification"], "Task"], [6556, 6559, ["random", "forest", "methods"], "Method"], [6561, 6565, ["image", "super", "-", "resolution"], "Task"], [6570, 6574, ["clustering", "/", "classification", "problem"], "Task"], [6576, 6577, ["JMPF"], "Method"], [6610, 6615, ["image", "super", "-", "resolution", "training"], "Task"], [6699, 6701, ["bicubic", "interpolation"], "Method"], [6712, 6715, ["anchored", "neighborhood", "regression"], "Method"], [6716, 6717, ["ANR"], "Method"], [6729, 6734, ["random", "forest", "(", "RF", ")"], "Method"], [6739, 6742, ["alternating", "regression", "forests"], "Method"], [6743, 6744, ["ARF"], "Method"], [6757, 6763, ["random", "-", "forest", "-", "based", "algorithms"], "Method"], [6770, 6772, ["random", "forest"], "Method"], [6799, 6800, ["JMPF"], "Method"], [6817, 6819, ["trimmed", "version"], "Method"], [6841, 6846, ["random", "-", "forest", "bagging", "strategy"], "Method"], [6878, 6880, ["JMPF", "+"], "Method"], [6886, 6890, ["General", "-", "100", "dataset"], "Material"], [6942, 6943, ["PSNR"], "Metric"], [6971, 6976, ["peak", "signal", "to", "noise", "ratio"], "Metric"], [6977, 6978, ["PSNR"], "Metric"], [7002, 7003, ["Set5"], "Material"], [7004, 7006, ["Set14", "datasets"], "Material"], [7040, 7041, ["ARF"], "Method"], [7044, 7050, ["random", "-", "forest", "-", "based", "algorithms"], "Method"], [7072, 7073, ["JMPF"], "Method"], [7079, 7080, ["ARF"], "Method"], [7095, 7098, ["objective", "quality", "metrics"], "Metric"], [7099, 7100, ["PSNR"], "Metric"], [7122, 7124, ["JMPF", "+"], "Method"], [7145, 7147, ["visual", "quality"], "Metric"], [7165, 7168, ["reconstructed", "HR", "images"], "Task"], [7235, 7239, ["random", "-", "forest", "scheme"], "Method"], [7242, 7246, ["Joint", "Maximum", "Purity", "Forest"], "Method"], [7247, 7248, ["JMPF"], "Method"], [7310, 7312, ["random", "forest"], "Method"], [7317, 7319, ["random", "forests"], "Method"], [7333, 7336, ["alternating", "decision", "forests"], "Method"], [7337, 7338, ["ADF"], "Method"], [7343, 7346, ["alternating", "regression", "forests"], "Method"], [7347, 7348, ["ARF"], "Method"], [7355, 7359, ["random", "-", "forest", "method"], "Method"], [7363, 7365, ["random", "forests"], "Method"], [7369, 7370, ["testing"], "Task"], [7371, 7375, ["multi", "-", "class", "capability"], "Metric"], [7387, 7391, ["classification", "and", "regression", "tasks"], "Task"], [7407, 7410, ["classification", "and", "regression"], "Task"], [7428, 7429, ["ADF"], "Method"], [7430, 7431, ["ARF"], "Method"], [7442, 7443, ["JMPF"], "Method"], [7444, 7450, ["single", "-", "image", "super", "-", "resolution"], "Task"], [7453, 7457, ["image", "super", "-", "resolution"], "Task"], [7459, 7463, ["clustering", "-", "regression", "problem"], "Task"], [7468, 7470, ["clustering", "stage"], "Task"], [7515, 7517, ["unsupervised", "problem"], "Task"], [7558, 7559, ["B100"], "Material"], [7588, 7589, ["PSNR"], "Metric"], [7590, 7592, ["visual", "quality"], "Metric"], [7600, 7602, ["random", "forests"], "Method"], [7610, 7614, ["training", "and", "inference", "processes"], "Task"], [48, 49, ["JMPF"], "Method"], [212, 213, ["JMPF"], "Method"], [384, 385, ["JMPF"], "Method"], [566, 570, ["joint", "maximum", "purity", "forest"], "Method"], [634, 635, ["JMPF"], "Method"], [1912, 1916, ["Joint", "Maximum", "Purity", "Forest"], "Method"], [2698, 2699, ["JMPF"], "Method"], [3674, 3675, ["JMPF"], "Method"], [3696, 3697, ["JMPF"], "Method"], [3746, 3747, ["JMPF"], "Method"], [3774, 3775, ["JMPF"], "Method"], [3963, 3964, ["JMPF"], "Method"], [4124, 4125, ["JMPF"], "Method"], [4290, 4291, ["JMPF"], "Method"], [4305, 4306, ["JMPF"], "Method"], [4329, 4330, ["JMPF"], "Method"], [4363, 4364, ["JMPF"], "Method"], [4960, 4961, ["LR"], "Task"], [4971, 4972, ["LR"], "Task"], [4980, 4981, ["HR"], "Task"], [4991, 4992, ["HR"], "Task"], [5005, 5006, ["NE"], "Method"], [5114, 5115, ["HR"], "Task"], [5118, 5119, ["LR"], "Task"], [5143, 5144, ["LR"], "Task"], [5145, 5146, ["HR"], "Task"], [5166, 5167, ["LR"], "Task"], [5171, 5172, ["HR"], "Task"], [5182, 5183, ["LR"], "Task"], [5184, 5185, ["HR"], "Task"], [5189, 5190, ["LR"], "Task"], [5193, 5194, ["HR"], "Task"], [5233, 5234, ["LR"], "Task"], [5238, 5239, ["LR"], "Task"], [5268, 5269, ["LR"], "Task"], [5274, 5275, ["LR"], "Task"], [5497, 5498, ["ANR"], "Method"], [5531, 5532, ["LR"], "Task"], [5533, 5534, ["HR"], "Task"], [5590, 5591, ["HR"], "Task"], [5600, 5601, ["LR"], "Task"], [5621, 5622, ["HR"], "Task"], [5719, 5720, ["ANR"], "Method"], [5749, 5750, ["ANR"], "Method"], [5789, 5790, ["ANR"], "Method"], [5846, 5848, ["collaborative", "representation"], "Method"], [5914, 5915, ["ANR"], "Method"], [6156, 6157, ["SR"], "Task"], [6231, 6232, ["LR"], "Task"], [6238, 6239, ["HR"], "Task"], [6301, 6302, ["HR"], "Task"], [6410, 6411, ["JMPF"], "Method"], [6414, 6415, ["SR"], "Task"], [6518, 6519, ["JMPF"], "Method"], [6534, 6535, ["JMPF"], "Method"], [6621, 6622, ["JMPF"], "Method"], [6642, 6643, ["JMPF"], "Method"], [6645, 6649, ["Image", "Super", "-", "Resolution"], "Task"], [6657, 6658, ["SR"], "Task"], [6674, 6675, ["B100"], "Material"], [6704, 6705, ["SR"], "Task"], [6803, 6804, ["JMPF"], "Method"], [6813, 6814, ["JMPF"], "Method"], [6937, 6938, ["Set5"], "Material"], [6988, 6989, ["SR"], "Task"], [7014, 7015, ["JMPF"], "Method"], [7067, 7068, ["B100"], "Material"], [7104, 7105, ["JMPF"], "Method"], [7140, 7141, ["Set5"], "Material"], [7150, 7151, ["JMPF"], "Method"], [7153, 7154, ["SR"], "Task"], [7219, 7220, ["JMPF"], "Method"], [7486, 7487, ["JMPF"], "Method"], [7564, 7565, ["JMPF"], "Method"]]} {"doc_id": "249b3b7421d3cdb932eecfe4b67203e0e46806b2", "text": ["document", ":", "Cell", "-", "aware", "Stacked", "LSTMs", "for", "Modeling", "Sentences", "We", "propose", "a", "method", "of", "stacking", "multiple", "long", "short", "-", "term", "memory", "(", "LSTM", ")", "layers", "for", "modeling", "sentences", ".", "In", "contrast", "to", "the", "conventional", "stacked", "LSTMs", "where", "only", "hidden", "states", "are", "fed", "as", "input", "to", "the", "next", "layer", ",", "our", "architecture", "accepts", "both", "hidden", "and", "memory", "cell", "states", "of", "the", "preceding", "layer", "and", "fuses", "information", "from", "the", "left", "and", "the", "lower", "context", "using", "the", "soft", "gating", "mechanism", "of", "LSTMs", ".", "Thus", "the", "proposed", "stacked", "LSTM", "architecture", "modulates", "the", "amount", "of", "information", "to", "be", "delivered", "not", "only", "in", "horizontal", "recurrence", "but", "also", "in", "vertical", "connections", ",", "from", "which", "useful", "features", "extracted", "from", "lower", "layers", "are", "effectively", "conveyed", "to", "upper", "layers", ".", "We", "dub", "this", "architecture", "Cell", "-", "aware", "Stacked", "LSTM", "(", "CAS", "-", "LSTM", ")", "and", "show", "from", "experiments", "that", "our", "models", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", "on", "benchmark", "datasets", "for", "natural", "language", "inference", ",", "paraphrase", "detection", ",", "and", "sentiment", "classification", ".", "section", ":", "Introduction", "In", "the", "field", "of", "natural", "language", "processing", "(", "NLP", ")", ",", "the", "most", "prevalent", "neural", "approach", "to", "obtaining", "sentence", "representations", "is", "to", "use", "recurrent", "neural", "networks", "(", "RNNs", ")", ",", "where", "words", "in", "a", "sentence", "are", "processed", "in", "a", "sequential", "and", "recurrent", "manner", ".", "Along", "with", "their", "intuitive", "design", ",", "RNNs", "have", "shown", "outstanding", "performance", "across", "various", "NLP", "tasks", "e.g.", "language", "modeling", ",", "machine", "translation", ",", "text", "classification", ",", "and", "parsing", ".", "Among", "several", "variants", "of", "the", "original", "RNN", ",", "gated", "recurrent", "architectures", "such", "as", "long", "short", "-", "term", "memory", "(", "LSTM", ")", "and", "gated", "recurrent", "unit", "(", "GRU", ")", "have", "been", "accepted", "as", "de", "-", "facto", "standard", "choices", "for", "RNNs", "due", "to", "their", "capability", "of", "addressing", "the", "vanishing", "and", "exploding", "gradient", "problem", "and", "considering", "long", "-", "term", "dependencies", ".", "Gated", "RNNs", "achieve", "these", "properties", "by", "introducing", "additional", "gating", "units", "that", "learn", "to", "control", "the", "amount", "of", "information", "to", "be", "transferred", "or", "forgotten", ",", "and", "are", "proven", "to", "work", "well", "without", "relying", "on", "complex", "optimization", "algorithms", "or", "careful", "initialization", ".", "Meanwhile", ",", "the", "common", "practice", "for", "further", "enhancing", "the", "expressiveness", "of", "RNNs", "is", "to", "stack", "multiple", "RNN", "layers", ",", "each", "of", "which", "has", "distinct", "parameter", "sets", "(", "stacked", "RNN", ")", ".", "In", "stacked", "RNNs", ",", "the", "hidden", "states", "of", "a", "layer", "are", "fed", "as", "input", "to", "the", "subsequent", "layer", ",", "and", "they", "are", "shown", "to", "work", "well", "due", "to", "increased", "depth", "or", "their", "ability", "to", "capture", "hierarchical", "time", "series", "which", "are", "inherent", "to", "the", "nature", "of", "the", "problem", "being", "modeled", ".", "0.25", "0.25", "However", "this", "setting", "of", "stacking", "RNNs", "might", "hinder", "the", "possibility", "of", "more", "sophisticated", "recurrence", "-", "based", "structures", "since", "the", "information", "from", "lower", "layers", "is", "simply", "treated", "as", "input", "to", "the", "next", "layer", ",", "rather", "than", "as", "another", "class", "of", "state", "that", "participates", "in", "core", "RNN", "computations", ".", "Especially", "for", "gated", "RNNs", "such", "as", "LSTMs", "and", "GRUs", ",", "this", "means", "that", "layer", "-", "to", "-", "layer", "connections", "can", "not", "fully", "benefit", "from", "the", "carefully", "constructed", "gating", "mechanism", "used", "in", "temporal", "transitions", ".", "Some", "recent", "work", "on", "stacking", "RNNs", "suggests", "alternative", "methods", "that", "encourage", "direct", "and", "effective", "interaction", "between", "RNN", "layers", "by", "adding", "residual", "connections", ",", "by", "shortcut", "connections", ",", "or", "by", "using", "cell", "states", "of", "LSTMs", ".", "In", "this", "paper", ",", "we", "propose", "a", "method", "of", "constructing", "multi", "-", "layer", "LSTMs", "where", "cell", "states", "are", "used", "in", "controlling", "the", "vertical", "information", "flow", ".", "This", "system", "utilizes", "states", "from", "the", "left", "and", "the", "lower", "context", "equally", "in", "computation", "of", "the", "new", "state", ",", "thus", "the", "information", "from", "lower", "layers", "is", "elaborately", "filtered", "and", "reflected", "through", "a", "soft", "gating", "mechanism", ".", "Our", "method", "is", "easy", "-", "to", "-", "implement", ",", "effective", ",", "and", "can", "replace", "conventional", "stacked", "LSTMs", "without", "much", "modification", "of", "the", "overall", "architecture", ".", "We", "call", "the", "proposed", "architecture", "Cell", "-", "aware", "Stacked", "LSTM", ",", "or", "CAS", "-", "LSTM", ",", "and", "evaluate", "our", "method", "on", "multiple", "benchmark", "datasets", ":", "SNLI", ",", "MultiNLI", ",", "Quora", "Question", "Pairs", ",", "and", "SST", ".", "From", "experiments", "we", "show", "that", "the", "CAS", "-", "LSTMs", "consistently", "outperform", "typical", "stacked", "LSTMs", ",", "opening", "the", "possibility", "of", "performance", "improvement", "of", "architectures", "that", "use", "stacked", "LSTMs", ".", "Our", "contribution", "is", "summarized", "as", "follows", ".", "We", "bring", "the", "idea", "of", "utilizing", "states", "coming", "from", "multiple", "directions", "to", "construction", "of", "stacked", "LSTM", "and", "apply", "the", "idea", "to", "the", "research", "of", "sentence", "representation", "learning", ".", "There", "is", "some", "prior", "work", "addressing", "the", "idea", "of", "incorporating", "more", "than", "one", "type", "of", "state", ",", "however", "to", "the", "best", "of", "our", "knowledge", "there", "is", "little", "work", "on", "applying", "the", "idea", "to", "sentence", "encoding", "and", "text", "classification", ".", "We", "conduct", "extensive", "evaluation", "of", "the", "proposed", "method", "and", "empirically", "prove", "its", "effectiveness", "on", "encoding", "sentences", ".", "Our", "models", "achieve", "new", "state", "-", "of", "-", "the", "-", "art", "results", "on", "SNLI", "and", "Quora", "Question", "Pairs", "datasets", ",", "and", "are", "on", "par", "with", "the", "best", "performing", "models", "on", "MultiNLI", "and", "SST", "datasets", ".", "This", "paper", "is", "organized", "as", "follows", ".", "We", "give", "a", "detailed", "description", "about", "the", "proposed", "method", "in", "§", "[", "reference", "]", ".", "Experimental", "results", "are", "given", "in", "§", "[", "reference", "]", ".", "We", "study", "prior", "work", "related", "to", "our", "objective", "in", "§", "[", "reference", "]", "and", "conclude", "in", "§", "[", "reference", "]", ".", "section", ":", "Model", "Description", "In", "this", "section", ",", "we", "give", "a", "detailed", "formulation", "of", "the", "architectures", "used", "in", "experiments", ".", "subsection", ":", "Notation", "Throughout", "this", "paper", ",", "we", "denote", "matrices", "as", "boldface", "capital", "letters", "(", ")", ",", "vectors", "as", "boldface", "lowercase", "letters", "(", ")", ",", "and", "scalars", "as", "normal", "italic", "letters", "(", ")", ".", "For", "LSTM", "states", ",", "we", "denote", "a", "hidden", "state", "as", "and", "a", "cell", "state", "as", ".", "Also", ",", "a", "layer", "index", "of", "or", "is", "denoted", "by", "superscript", "and", "a", "time", "index", "is", "denoted", "by", "a", "subscript", ",", "i.e.", "indicates", "the", "hidden", "state", "at", "time", "and", "layer", ".", "means", "the", "element", "-", "wise", "multiplication", "between", "two", "vectors", ".", "We", "write", "-", "th", "component", "of", "vector", "as", ".", "All", "vectors", "are", "assumed", "to", "be", "column", "vectors", ".", "subsection", ":", "Stacked", "LSTMs", "While", "there", "exist", "various", "versions", "of", "LSTM", "formulation", ",", "in", "this", "work", "we", "use", "the", "following", ",", "one", "of", "the", "most", "common", "versions", ":", "where", ",", ",", "and", ",", ",", "are", "trainable", "parameters", ".", "and", "are", "the", "sigmoid", "activation", "and", "the", "hyperbolic", "tangent", "activation", "function", "respectively", ".", "Also", "we", "assume", "that", "where", "is", "the", "-", "th", "input", "to", "the", "network", ".", "The", "input", "gate", "and", "the", "forget", "gate", "control", "the", "amount", "of", "information", "transmitted", "from", "and", ",", "the", "candidate", "cell", "state", "and", "the", "previous", "cell", "state", ",", "to", "the", "new", "cell", "state", ".", "Similarly", "the", "output", "gate", "soft", "-", "selects", "which", "portion", "of", "the", "cell", "state", "is", "to", "be", "used", "in", "the", "final", "hidden", "state", ".", "We", "can", "clearly", "see", "that", "cell", "states", "(", ",", ",", ")", "play", "a", "crucial", "role", "in", "forming", "horizontal", "recurrence", ".", "However", "the", "current", "formulation", "does", "not", "consider", ",", "the", "cell", "state", "from", "-", "th", "layer", ",", "in", "computation", "and", "thus", "the", "lower", "context", "is", "reflected", "only", "through", "the", "rudimentary", "way", ",", "hindering", "the", "possibility", "of", "controlling", "vertical", "information", "flow", ".", "subsection", ":", "Cell", "-", "aware", "Stacked", "LSTMs", "Now", "we", "extend", "the", "stacked", "LSTM", "formulation", "defined", "above", "to", "address", "the", "problem", "noted", "in", "the", "previous", "subsection", ".", "To", "enhance", "the", "interaction", "between", "layers", "in", "a", "way", "similar", "to", "how", "LSTMs", "keep", "and", "forget", "the", "information", "from", "the", "previous", "time", "step", ",", "we", "introduce", "the", "additional", "forget", "gate", "that", "determines", "whether", "to", "accept", "or", "ignore", "the", "signals", "coming", "from", "the", "previous", "layer", ".", "Therefore", "the", "proposed", "Cell", "-", "aware", "Stacked", "LSTM", "is", "formulated", "as", "follows", ":", "where", "and", ".", "can", "either", "be", "a", "vector", "of", "constants", "or", "parameters", ".", "When", ",", "the", "equations", "defined", "in", "the", "previous", "subsection", "are", "used", ".", "Therefore", ",", "it", "can", "be", "said", "that", "each", "non", "-", "bottom", "layer", "of", "CAS", "-", "LSTM", "accepts", "two", "sets", "of", "hidden", "and", "cell", "states", "—", "one", "from", "the", "left", "context", "and", "the", "other", "from", "the", "below", "context", ".", "The", "left", "and", "the", "below", "context", "participate", "in", "computation", "with", "the", "equivalent", "procedure", "so", "that", "the", "information", "from", "lower", "layers", "can", "be", "efficiently", "propagated", ".", "Fig", ".", "[", "reference", "]", "compares", "CAS", "-", "LSTM", "to", "the", "conventional", "stacked", "LSTM", "architecture", ",", "and", "Fig", ".", "[", "reference", "]", "depicts", "the", "computation", "flow", "of", "the", "CAS", "-", "LSTM", ".", "We", "argue", "that", "considering", "in", "computation", "is", "beneficial", "for", "the", "following", "reasons", ".", "First", ",", "contains", "additional", "information", "compared", "to", "since", "it", "is", "not", "filtered", "by", ".", "Thus", "a", "model", "that", "directly", "uses", "does", "not", "rely", "solely", "on", "for", "extracting", "information", ",", "due", "to", "the", "fact", "that", "it", "has", "access", "to", "the", "raw", "information", ",", "as", "in", "temporal", "connections", ".", "In", "other", "words", ",", "no", "longer", "has", "to", "take", "all", "responsibility", "for", "selecting", "useful", "features", "for", "both", "horizontal", "and", "vertical", "transitions", ",", "and", "the", "burden", "of", "selecting", "information", "is", "shared", "with", ".", "Another", "advantage", "of", "using", "the", "lies", "in", "the", "fact", "that", "it", "directly", "connects", "and", ".", "This", "direct", "connection", "helps", "and", "stabilizes", "training", ",", "since", "the", "terminal", "error", "signals", "can", "be", "easily", "backpropagated", "to", "model", "parameters", ".", "Fig", ".", "[", "reference", "]", "illustrates", "paths", "between", "the", "two", "cell", "states", ".", "We", "find", "experimentally", "that", "there", "is", "little", "difference", "between", "letting", "be", "constant", "and", "letting", "it", "be", "trainable", "parameters", ",", "thus", "we", "set", "in", "all", "experiments", ".", "We", "also", "experimented", "with", "the", "architecture", "without", "i.e.", "two", "cell", "states", "are", "combined", "by", "unweighted", "summation", "similar", "to", "multidimensional", "RNNs", ",", "and", "found", "that", "it", "leads", "to", "performance", "degradation", "and", "unstable", "convergence", ",", "likely", "due", "to", "mismatch", "in", "the", "range", "of", "cell", "state", "values", "between", "layers", "(", "for", "the", "first", "layer", "and", "for", "the", "others", ")", ".", "Experimental", "results", "on", "various", "are", "presented", "in", "§", "[", "reference", "]", ".", "paragraph", ":", "Connection", "to", "tree", "-", "structured", "RNNs", ".", "The", "idea", "of", "having", "multiple", "states", "is", "also", "related", "to", "tree", "-", "structured", "RNNs", ".", "Among", "them", ",", "tree", "-", "structured", "LSTMs", "(", "Tree", "-", "LSTMs", ")", "are", "similar", "to", "ours", "in", "that", "they", "use", "both", "hidden", "and", "cell", "states", "from", "children", "nodes", ".", "In", "Tree", "-", "LSTMs", ",", "states", "for", "all", "children", "nodes", "are", "regarded", "as", "input", ",", "and", "they", "participate", "in", "the", "computation", "equally", "through", "weight", "-", "shared", "(", "in", "Child", "-", "Sum", "Tree", "-", "LSTMs", ")", "or", "weight", "-", "unshared", "(", "in", "-", "ary", "Tree", "-", "LSTMs", ")", "projection", ".", "From", "this", "perspective", ",", "each", "CAS", "-", "LSTM", "layer", "(", "where", ")", "can", "be", "seen", "as", "a", "binary", "Tree", "-", "LSTM", "where", "the", "structures", "it", "operates", "on", "are", "fixed", "to", "right", "-", "branching", "trees", ".", "The", "use", "of", "cell", "state", "in", "computation", "could", "be", "one", "reason", "that", "Tree", "-", "LSTMs", "perform", "better", "than", "sequential", "LSTMs", "even", "when", "trivial", "trees", "(", "strictly", "left", "-", "or", "right", "-", "branching", ")", "are", "given", ".", "paragraph", ":", "Connection", "to", "multidimensional", "RNNs", ".", "Multidimensional", "RNNs", "(", "MDRNN", ")", "are", "an", "extension", "of", "1D", "sequential", "RNNs", "that", "can", "accept", "multidimensional", "input", "e.g.", "images", ",", "and", "have", "been", "successfully", "applied", "to", "image", "segmentation", "and", "handwriting", "recognition", ".", "Notably", "multidimensional", "LSTMs", "(", "MDLSTM", ")", "have", "an", "analogous", "formulation", "to", "ours", "except", "the", "term", "and", "the", "fact", "that", "we", "use", "distinct", "weights", "per", "column", "(", "or", "‘", "layer", "’", "in", "our", "case", ")", ".", "From", "this", "view", ",", "CAS", "-", "LSTM", "can", "be", "seen", "as", "a", "certain", "kind", "of", "MDLSTM", "that", "accepts", "a", "2D", "input", ".", "Grid", "LSTMs", "also", "take", "inputs", "but", "emit", "outputs", ",", "which", "is", "different", "from", "our", "case", "where", "a", "single", "set", "of", "hidden", "and", "cell", "states", "is", "produced", ".", "subsection", ":", "Sentence", "Encoders", "The", "sentence", "encoder", "network", "we", "use", "in", "our", "experiments", "takes", "words", "(", "assumed", "to", "be", "one", "-", "hot", "vectors", ")", "as", "input", ".", "The", "words", "are", "projected", "to", "corresponding", "word", "representations", ":", "where", ".", "Then", "is", "fed", "to", "a", "-", "layer", "CAS", "-", "LSTM", "model", ",", "resulting", "in", "the", "representations", ".", "The", "sentence", "representation", ",", ",", "is", "computed", "by", "max", "-", "pooling", "over", "time", "as", "in", "the", "work", "of", "conneau2017infersent", "conneau2017infersent", ".", "Similar", "to", "their", "results", ",", "from", "preliminary", "experiments", "we", "found", "that", "the", "max", "-", "pooling", "performs", "consistently", "better", "than", "mean", "-", "and", "last", "-", "pooling", ".", "To", "make", "models", "more", "expressive", ",", "a", "bidirectional", "CAS", "-", "LSTM", "network", "may", "also", "be", "used", ".", "In", "the", "bidirectional", "case", ",", "the", "forward", "representations", "and", "the", "backward", "representations", "are", "concatenated", "and", "max", "-", "pooled", "to", "yield", "the", "sentence", "representation", ".", "We", "call", "this", "bidirectional", "architecture", "Bi", "-", "CAS", "-", "LSTM", "in", "experiments", ".", "subsection", ":", "Top", "-", "layer", "Classifiers", "For", "the", "natural", "language", "inference", "experiments", ",", "we", "use", "the", "following", "heuristic", "function", "proposed", "by", "mou2016snli", "mou2016snli", "in", "feature", "extraction", ":", "where", "means", "vector", "concatenation", ",", "and", "and", "are", "applied", "element", "-", "wise", ".", "And", "we", "use", "the", "following", "function", "in", "paraphrase", "identification", "experiments", ":", "as", "in", "the", "work", "of", "ji2013discriminative", "ji2013discriminative", ".", "For", "sentiment", "classification", ",", "we", "use", "the", "sentence", "representation", "itself", ".", "We", "feed", "the", "feature", "extracted", "from", "as", "input", "to", "the", "MLP", "classifier", "with", "ReLU", "activation", "followed", "by", "the", "fully", "-", "connected", "softmax", "layer", "to", "predict", "the", "label", "distribution", ":", "where", ",", "is", "the", "number", "of", "label", "classes", ",", "and", "the", "dimension", "of", "the", "MLP", "output", ",", "section", ":", "Experiments", "We", "evaluate", "our", "method", "on", "natural", "language", "inference", "(", "NLI", ")", ",", "paraphrase", "identification", "(", "PI", ")", ",", "and", "sentiment", "classification", ".", "We", "also", "conduct", "analysis", "on", "gate", "values", "and", "experiments", "on", "model", "variants", ".", "For", "detailed", "experimental", "settings", ",", "we", "refer", "readers", "to", "the", "supplemental", "material", ".", "For", "the", "NLI", "and", "PI", "tasks", ",", "there", "exists", "recent", "work", "specializing", "in", "sentence", "pair", "classification", ".", "However", "in", "this", "work", "we", "confine", "our", "model", "to", "the", "architecture", "that", "encodes", "each", "sentence", "using", "a", "shared", "encoder", "without", "any", "inter", "-", "sentence", "interaction", ",", "in", "order", "to", "focus", "on", "the", "effectiveness", "of", "the", "models", "in", "extracting", "semantics", ".", "But", "note", "that", "the", "applicability", "of", "CAS", "-", "LSTM", "is", "not", "limited", "to", "sentence", "encoding", "based", "approaches", ".", "subsection", ":", "Natural", "Language", "Inference", "For", "the", "evaluation", "of", "performance", "of", "the", "proposed", "method", "on", "the", "NLI", "task", ",", "SNLI", "and", "MultiNLI", "datasets", "are", "used", ".", "The", "objective", "of", "both", "datasets", "is", "to", "predict", "the", "relationship", "between", "a", "premise", "and", "a", "hypothesis", "sentence", ":", "entailment", ",", "contradiction", ",", "and", "neutral", ".", "SNLI", "and", "MultiNLI", "datasets", "are", "composed", "of", "about", "570k", "and", "430k", "premise", "-", "hypothesis", "pairs", "respectively", ".", "GloVe", "pretrained", "word", "embeddings", "are", "used", "and", "remain", "fixed", "during", "training", ".", "The", "dimension", "of", "encoder", "states", "(", ")", "is", "set", "to", "300", "and", "a", "1024D", "MLP", "with", "one", "or", "two", "hidden", "layers", "is", "used", ".", "We", "apply", "dropout", "to", "the", "word", "embeddings", "and", "the", "MLP", "layers", ".", "The", "features", "used", "as", "input", "to", "the", "MLP", "classifier", "are", "extracted", "following", "Eq", ".", "[", "reference", "]", ".", "Table", "[", "reference", "]", "and", "[", "reference", "]", "contain", "results", "of", "the", "models", "on", "SNLI", "and", "MultiNLI", "datasets", ".", "In", "SNLI", ",", "our", "best", "model", "achieves", "the", "new", "state", "-", "of", "-", "the", "-", "art", "accuracy", "of", "87.0", "%", "with", "relatively", "fewer", "parameters", ".", "Similarly", "in", "MultiNLI", ",", "our", "models", "match", "the", "accuracy", "of", "state", "-", "of", "-", "the", "-", "art", "models", "in", "both", "in", "-", "domain", "(", "matched", ")", "and", "cross", "-", "domain", "(", "mismatched", ")", "test", "sets", ".", "Note", "that", "only", "the", "GloVe", "word", "vectors", "are", "used", "as", "word", "representations", ",", "as", "opposed", "to", "some", "models", "that", "introduce", "character", "-", "level", "features", ".", "It", "is", "also", "notable", "that", "our", "proposed", "architecture", "does", "not", "restrict", "the", "selection", "of", "pooling", "method", ";", "the", "performance", "could", "further", "be", "improved", "by", "replacing", "max", "-", "pooling", "with", "other", "advanced", "algorithms", "e.g.", "intra", "-", "sentence", "attention", "and", "generalized", "pooling", ".", "subsection", ":", "Paraphrase", "Identification", "We", "use", "Quora", "Question", "Pairs", "dataset", "in", "evaluating", "the", "performance", "of", "our", "method", "on", "the", "PI", "task", ".", "The", "dataset", "consists", "of", "over", "400k", "question", "pairs", ",", "and", "each", "pair", "is", "annotated", "with", "whether", "the", "two", "sentences", "are", "paraphrase", "of", "each", "other", "or", "not", ".", "Similar", "to", "the", "NLI", "experiments", ",", "GloVe", "pretrained", "vectors", ",", "300D", "encoders", ",", "and", "1024D", "MLP", "are", "used", ".", "The", "number", "of", "CAS", "-", "LSTM", "layers", "is", "fixed", "to", "2", "in", "PI", "experiments", ".", "Two", "sentence", "vectors", "are", "aggregated", "using", "Eq", ".", "[", "reference", "]", "and", "fed", "as", "input", "to", "the", "MLP", ".", "The", "results", "on", "the", "Quora", "Question", "Pairs", "dataset", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "Again", "we", "can", "see", "that", "our", "models", "outperform", "other", "models", "by", "large", "margin", ",", "achieving", "the", "new", "state", "of", "the", "art", ".", "0.22", "0.22", "0.22", "0.22", "0.22", "0.22", "subsection", ":", "Sentiment", "Classification", "In", "evaluating", "sentiment", "classification", "performance", ",", "the", "Stanford", "Sentiment", "Treebank", "(", "SST", ")", "is", "used", ".", "It", "consists", "of", "about", "12", ",", "000", "binary", "-", "parsed", "sentences", "where", "constituents", "(", "phrases", ")", "of", "each", "parse", "tree", "are", "annotated", "with", "a", "sentiment", "label", "(", "very", "positive", ",", "positive", ",", "neutral", ",", "negative", ",", "very", "negative", ")", ".", "Following", "the", "convention", "of", "prior", "work", ",", "all", "phrases", "and", "their", "labels", "are", "used", "in", "training", "but", "only", "the", "sentence", "-", "level", "data", "are", "used", "in", "evaluation", ".", "In", "evaluation", "we", "consider", "two", "settings", ",", "namely", "SST", "-", "2", "and", "SST", "-", "5", ",", "the", "two", "differing", "only", "in", "their", "level", "of", "granularity", "with", "regard", "to", "labels", ".", "In", "SST", "-", "2", ",", "data", "samples", "annotated", "with", "‘", "neutral", "’", "are", "ignored", "from", "training", "and", "evaluation", ".", "The", "two", "positive", "labels", "(", "very", "positive", ",", "positive", ")", "are", "considered", "as", "the", "same", "label", ",", "and", "similarly", "for", "the", "two", "negative", "labels", ".", "As", "a", "result", "98", ",", "794", "/", "872", "/", "1", ",", "821", "data", "samples", "are", "used", "in", "training", "/", "validation", "/", "test", ",", "and", "the", "task", "is", "considered", "as", "a", "binary", "classification", "problem", ".", "In", "SST", "-", "5", ",", "data", "are", "used", "as", "-", "is", "and", "thus", "the", "task", "is", "a", "5", "-", "class", "classification", "problem", ".", "All", "318", ",", "582", "/", "1", ",", "101", "/", "2", ",", "210", "data", "samples", "for", "training", "/", "validation", "/", "test", "are", "used", "in", "the", "SST", "-", "5", "setting", ".", "We", "use", "300D", "GloVe", "vectors", ",", "2", "-", "layer", "150D", "or", "300D", "encoders", ",", "and", "a", "300D", "MLP", "classifier", "for", "the", "models", ",", "however", "unlike", "previous", "experiments", "we", "tune", "the", "word", "embeddings", "during", "training", ".", "The", "results", "on", "SST", "are", "listed", "in", "Table", "[", "reference", "]", ".", "Our", "models", "achieve", "the", "new", "state", "-", "of", "-", "the", "-", "art", "accuracy", "on", "SST", "-", "2", "and", "competitive", "accuracy", "on", "SST", "-", "5", ",", "without", "utilizing", "parse", "tree", "information", ".", "subsection", ":", "Forget", "Gate", "Analysis", "To", "inspect", "the", "effect", "of", "the", "additional", "forget", "gate", ",", "we", "investigate", "how", "the", "values", "of", "vertical", "forget", "gates", "are", "distributed", ".", "We", "sample", "1", ",", "000", "random", "sentences", "from", "the", "development", "set", "of", "the", "SNLI", "dataset", ",", "and", "use", "the", "3", "-", "layer", "CAS", "-", "LSTM", "model", "trained", "on", "the", "SNLI", "dataset", "to", "compute", "gate", "values", ".", "If", "all", "values", "from", "a", "vertical", "forget", "gate", "were", "to", "be", "0", ",", "this", "would", "mean", "that", "the", "introduction", "of", "the", "additional", "forget", "gate", "is", "meaningless", "and", "the", "model", "would", "reduce", "to", "a", "plain", "stacked", "LSTM", ".", "On", "the", "contrary", "if", "all", "values", "were", "1", ",", "meaning", "that", "the", "vertical", "forget", "gates", "were", "always", "open", ",", "it", "would", "be", "impossible", "to", "say", "that", "the", "information", "is", "modulated", "effectively", ".", "Fig", ".", "[", "reference", "]", "and", "[", "reference", "]", "represent", "histograms", "of", "the", "vertical", "forget", "gate", "values", "from", "the", "second", "and", "the", "third", "layer", ".", "From", "the", "figures", "we", "can", "validate", "that", "the", "trained", "model", "does", "not", "fall", "into", "the", "degenerate", "case", "where", "vertical", "forget", "gates", "are", "ignored", ".", "Also", "the", "figures", "show", "that", "the", "values", "are", "right", "-", "skewed", ",", "which", "we", "conjecture", "to", "be", "a", "result", "of", "focusing", "more", "on", "a", "strong", "interaction", "between", "adjacent", "layers", ".", "To", "further", "verify", "that", "the", "gate", "values", "are", "diverse", "enough", "within", "each", "time", "step", ",", "we", "compute", "the", "distribution", "of", "the", "range", "of", "values", "per", "time", "step", ",", ",", "where", ".", "We", "plot", "the", "histograms", "in", "Fig", ".", "[", "reference", "]", "and", "[", "reference", "]", ".", "From", "the", "figure", "we", "see", "that", "a", "vertical", "forget", "gate", "controls", "the", "amount", "of", "information", "flow", "effectively", ",", "making", "the", "decision", "of", "retaining", "or", "discarding", "signals", ".", "Finally", ",", "to", "investigate", "the", "argument", "presented", "in", "§", "[", "reference", "]", "that", "the", "additional", "forget", "gate", "helps", "the", "previous", "output", "gate", "with", "reducing", "the", "burden", "of", "extracting", "all", "needed", "information", ",", "we", "inspect", "the", "distribution", "of", "the", "values", "from", ".", "This", "distribution", "indicates", "how", "differently", "the", "vertical", "forget", "gate", "and", "the", "previous", "output", "gate", "select", "information", "from", ".", "From", "Fig", ".", "[", "reference", "]", "and", "[", "reference", "]", "we", "can", "see", "that", "the", "two", "gates", "make", "fairly", "different", "decisions", ",", "from", "which", "we", "demonstrate", "that", "the", "direct", "path", "between", "and", "enables", "a", "model", "to", "utilize", "signals", "overlooked", "by", ".", "subsection", ":", "Model", "Variations", "In", "this", "subsection", ",", "we", "see", "the", "influence", "of", "each", "component", "of", "a", "model", "on", "performance", "by", "removing", "or", "replacing", "its", "components", ".", "the", "SNLI", "dataset", "is", "used", "for", "experiments", ",", "and", "the", "best", "performing", "configuration", "is", "used", "as", "a", "baseline", "for", "modifications", ".", "We", "consider", "the", "following", "variants", ":", "(", "i", ")", "models", "that", "use", "plain", "stacked", "LSTMs", ",", "(", "ii", ")", "models", "with", "different", ",", "(", "iii", ")", "models", "without", ",", "and", "(", "iv", ")", "models", "that", "integrate", "lower", "contexts", "via", "peephole", "connections", ".", "Variant", "(", "iv", ")", "integrates", "lower", "contexts", "via", "the", "following", "equations", ":", "where", "represent", "peephole", "weights", "that", "take", "cell", "states", "into", "account", ".", "Among", "the", "above", "equations", ",", "those", "that", "use", "the", "lower", "cell", "state", "are", "Eq", ".", "[", "reference", "]", "and", "[", "reference", "]", ".", "We", "can", "see", "that", "affects", "the", "value", "of", "only", "via", "peephole", "connections", ",", "which", "makes", "independent", "of", ".", "Table", "[", "reference", "]", "summarizes", "the", "results", "of", "model", "variants", ".", "We", "can", "again", "see", "that", "the", "use", "of", "cell", "states", "clearly", "improves", "sentence", "modeling", "performance", "(", "baseline", "vs.", "(", "i", ")", "and", "(", "iv", ")", "vs.", "(", "i", ")", ")", ".", "Also", "from", "the", "results", "of", "baseline", "and", "(", "ii", ")", ",", "we", "validate", "that", "the", "selection", "of", "does", "not", "significantly", "affect", "performance", "but", "introducing", "is", "beneficial", "(", "baseline", "vs.", "(", "iii", ")", ")", "possibly", "due", "to", "its", "effect", "on", "normalizing", "information", "from", "multiple", "sources", ",", "as", "mentioned", "in", "§", "[", "reference", "]", ".", "Finally", ",", "from", "the", "comparison", "between", "baseline", "and", "(", "iv", ")", ",", "we", "show", "that", "the", "proposed", "way", "of", "combining", "the", "left", "and", "the", "lower", "contexts", "leads", "to", "better", "modeling", "of", "sentence", "representations", "than", "that", "of", "zhang2016highway", "zhang2016highway", "in", "encoding", "sentences", ".", "section", ":", "Related", "Work", "paragraph", ":", "Stacked", "recurrent", "neural", "networks", ".", "There", "is", "some", "prior", "work", "on", "methods", "of", "stacking", "RNNs", "beyond", "the", "plain", "stacked", "RNNs", ".", "Residual", "LSTMs", "add", "residual", "connections", "between", "the", "hidden", "states", "computed", "at", "each", "LSTM", "layer", ",", "and", "shortcut", "-", "stacked", "LSTMs", "concatenate", "hidden", "states", "from", "all", "previous", "layers", "to", "make", "the", "backpropagation", "path", "short", ".", "In", "our", "method", ",", "the", "lower", "context", "is", "aggregated", "via", "a", "gating", "mechanism", ",", "and", "we", "believe", "it", "modulates", "the", "amount", "of", "information", "to", "be", "transmitted", "in", "a", "more", "efficient", "and", "effective", "way", "than", "vector", "addition", "or", "concatenation", ".", "Also", ",", "compared", "to", "concatenation", ",", "our", "method", "does", "not", "significantly", "increase", "the", "number", "of", "parameters", ".", "Highway", "LSTMs", "and", "depth", "-", "gated", "LSTMs", "are", "similar", "to", "our", "proposed", "models", "in", "that", "they", "use", "cell", "states", "from", "the", "previous", "layer", ",", "and", "they", "are", "successfully", "applied", "to", "the", "field", "of", "automatic", "speech", "recognition", "and", "language", "modeling", ".", "However", "in", "contrast", "to", "CAS", "-", "LSTM", ",", "where", "the", "additional", "forget", "gate", "aggregates", "the", "previous", "layer", "states", ",", "and", "thus", "contexts", "from", "the", "left", "and", "below", "participate", "in", "computation", "equitably", ",", "in", "Highway", "LSTMs", "and", "depth", "-", "gated", "LSTMs", "the", "previous", "layer", "states", "are", "considered", "only", "through", "peephole", "connections", ".", "The", "comparison", "of", "our", "models", "and", "this", "architecture", "is", "presented", "in", "§", "[", "reference", "]", ".", "paragraph", ":", "Multidimensional", "recurrent", "neural", "networks", ".", "There", "is", "another", "line", "of", "research", "that", "aims", "to", "extend", "RNNs", "to", "operate", "on", "multidimensional", "inputs", ".", "Grid", "LSTMs", "are", "a", "general", "-", "dimensional", "LSTM", "architecture", "that", "accepts", "sets", "of", "hidden", "and", "cell", "states", "as", "input", "and", "yields", "sets", "of", "states", "as", "output", ",", "in", "contrast", "to", "our", "architecture", ",", "which", "emits", "a", "single", "set", "of", "states", ".", "2D", "and", "3D", "Grid", "LSTMs", "bring", "a", "performance", "gain", "on", "character", "-", "level", "language", "modeling", "and", "machine", "translation", "respectively", ".", "Multidimensional", "RNNs", "have", "a", "similar", "formulation", "as", "ours", ",", "except", "that", "they", "do", "not", "normalize", "cell", "states", "and", "weights", "for", "all", "columns", "(", "layers", ")", "are", "tied", ".", "However", "they", "are", "often", "employed", "to", "model", "multidimensional", "data", "such", "as", "images", "of", "handwritten", "text", "with", "RNNs", ",", "rather", "than", "stacking", "RNN", "layers", "for", "modeling", "sequential", "data", ".", "paragraph", ":", "Deep", "recurrent", "transitions", ".", "Rather", "than", "stacking", "recurrent", "layers", ",", "some", "work", "focuses", "on", "increasing", "the", "depth", "of", "horizontal", "recurrence", ".", "pascanu2014construct", "pascanu2014construct", "have", "investigated", "various", "architectures", "to", "increase", "the", "depth", "of", "RNNs", ",", "inter", "alia", "Deep", "Transition", "RNNs", "address", "the", "problem", "of", "deep", "hidden", "-", "to", "-", "hidden", "transitions", ".", "graves2016adaptive", "graves2016adaptive", "proposed", "an", "adaptive", "computation", "time", "algorithm", "that", "learns", "how", "many", "micro", "time", "steps", "to", "take", "between", "receiving", "an", "input", "and", "emitting", "an", "output", ".", "Fast", "-", "Slow", "RNNs", "process", "data", "on", "different", "timescales", "by", "letting", "a", "fast", "cell", "iterate", "for", "a", "fixed", "number", "of", "time", "steps", "before", "a", "slow", "cell", "receives", "the", "next", "input", ".", "Multiscale", "RNNs", "e.g.", "Clockwork", "RNNs", "and", "Hierarchical", "Multiscale", "RNNs", "can", "be", "also", "regarded", "as", "architectures", "with", "increased", "recurrence", "depth", ".", "However", "as", "noted", "by", "zilly2017rhn", "zilly2017rhn", ",", "increase", "in", "recurrent", "depth", "results", "in", "a", "longer", "maximum", "path", "than", "stacking", "recurrent", "layers", "and", "makes", "training", "difficult", "without", "careful", "initialization", "or", "architectural", "choice", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "proposed", "a", "method", "of", "stacking", "multiple", "LSTM", "layers", "for", "modeling", "sentences", ",", "dubbed", "CAS", "-", "LSTM", ".", "It", "uses", "not", "only", "hidden", "states", "but", "also", "cell", "states", "from", "the", "previous", "layer", ",", "for", "the", "purpose", "of", "controlling", "the", "vertical", "information", "flow", "in", "a", "more", "elaborate", "way", ".", "We", "evaluated", "the", "proposed", "method", "on", "various", "benchmark", "tasks", ":", "natural", "language", "inference", ",", "paraphrase", "identification", ",", "and", "sentiment", "classification", ".", "Our", "models", "achieve", "the", "new", "state", "-", "of", "-", "the", "-", "art", "accuracy", "on", "SNLI", "and", "Quora", "Question", "Pairs", "datasets", "and", "obtain", "comparable", "results", "on", "MultiNLI", "and", "SST", "datasets", ".", "The", "proposed", "architecture", "can", "replace", "any", "stacked", "LSTM", "under", "one", "weak", "restriction", "—", "the", "size", "of", "states", "should", "be", "identical", "across", "all", "layers", ".", "For", "future", "work", "we", "plan", "to", "apply", "the", "CAS", "-", "LSTM", "architecture", "beyond", "sentence", "modeling", "tasks", ".", "Various", "problems", "e.g.", "sequence", "labeling", ",", "sequence", "generation", ",", "and", "language", "modeling", "might", "benefit", "from", "sophisticated", "modulation", "on", "context", "integration", ".", "Aggregating", "diverse", "contexts", "from", "sequential", "data", ",", "e.g.", "those", "from", "forward", "and", "backward", "reading", "of", "text", ",", "could", "also", "be", "an", "intriguing", "research", "direction", ".", "section", ":", "Acknowledgments", "We", "thank", "Dan", "Edmiston", "for", "the", "review", "of", "the", "manuscript", ".", "bibliography", ":", "References"], "entity": [[8, 10, ["Modeling", "Sentences"], "Task"], [16, 22, ["multiple", "long", "short", "-", "term", "memory"], "Method"], [23, 24, ["LSTM"], "Method"], [27, 29, ["modeling", "sentences"], "Task"], [75, 78, ["soft", "gating", "mechanism"], "Method"], [79, 80, ["LSTMs"], "Method"], [125, 130, ["Cell", "-", "aware", "Stacked", "LSTM"], "Method"], [131, 134, ["CAS", "-", "LSTM"], "Method"], [155, 158, ["natural", "language", "inference"], "Task"], [159, 161, ["paraphrase", "detection"], "Task"], [163, 165, ["sentiment", "classification"], "Task"], [173, 176, ["natural", "language", "processing"], "Task"], [177, 178, ["NLP"], "Task"], [183, 185, ["neural", "approach"], "Method"], [187, 189, ["sentence", "representations"], "Task"], [192, 195, ["recurrent", "neural", "networks"], "Method"], [196, 197, ["RNNs"], "Method"], [219, 220, ["RNNs"], "Method"], [229, 231, ["language", "modeling"], "Task"], [232, 234, ["machine", "translation"], "Task"], [235, 237, ["text", "classification"], "Task"], [239, 240, ["parsing"], "Task"], [247, 248, ["RNN"], "Method"], [249, 252, ["gated", "recurrent", "architectures"], "Method"], [254, 259, ["long", "short", "-", "term", "memory"], "Method"], [260, 261, ["LSTM"], "Method"], [263, 266, ["gated", "recurrent", "unit"], "Method"], [267, 268, ["GRU"], "Method"], [279, 280, ["RNNs"], "Method"], [287, 292, ["vanishing", "and", "exploding", "gradient", "problem"], "Task"], [299, 301, ["Gated", "RNNs"], "Method"], [307, 309, ["gating", "units"], "Method"], [333, 335, ["optimization", "algorithms"], "Method"], [348, 349, ["expressiveness"], "Metric"], [350, 351, ["RNNs"], "Task"], [355, 357, ["RNN", "layers"], "Method"], [366, 368, ["stacked", "RNN"], "Method"], [371, 373, ["stacked", "RNNs"], "Method"], [426, 428, ["stacking", "RNNs"], "Task"], [435, 439, ["recurrence", "-", "based", "structures"], "Method"], [465, 468, ["core", "RNN", "computations"], "Method"], [471, 473, ["gated", "RNNs"], "Method"], [475, 476, ["LSTMs"], "Method"], [477, 478, ["GRUs"], "Method"], [496, 498, ["gating", "mechanism"], "Method"], [507, 509, ["stacking", "RNNs"], "Task"], [548, 552, ["multi", "-", "layer", "LSTMs"], "Method"], [560, 563, ["vertical", "information", "flow"], "Task"], [596, 599, ["soft", "gating", "mechanism"], "Method"], [629, 635, ["architecture", "Cell", "-", "aware", "Stacked", "LSTM"], "Method"], [637, 640, ["CAS", "-", "LSTM"], "Method"], [650, 651, ["SNLI"], "Material"], [652, 653, ["MultiNLI"], "Material"], [654, 657, ["Quora", "Question", "Pairs"], "Material"], [659, 660, ["SST"], "Material"], [667, 670, ["CAS", "-", "LSTMs"], "Method"], [720, 723, ["sentence", "representation", "learning"], "Task"], [757, 759, ["sentence", "encoding"], "Task"], [760, 762, ["text", "classification"], "Task"], [777, 779, ["encoding", "sentences"], "Task"], [793, 794, ["SNLI"], "Material"], [795, 799, ["Quora", "Question", "Pairs", "datasets"], "Material"], [810, 811, ["MultiNLI"], "Material"], [812, 814, ["SST", "datasets"], "Material"], [1154, 1155, ["computation"], "Task"], [1173, 1176, ["vertical", "information", "flow"], "Task"], [1215, 1216, ["LSTMs"], "Method"], [1251, 1256, ["Cell", "-", "aware", "Stacked", "LSTM"], "Method"], [1294, 1302, ["non", "-", "bottom", "layer", "of", "CAS", "-", "LSTM"], "Method"], [1332, 1333, ["computation"], "Task"], [1355, 1358, ["CAS", "-", "LSTM"], "Method"], [1377, 1380, ["CAS", "-", "LSTM"], "Method"], [1385, 1387, ["in", "computation"], "Task"], [1420, 1422, ["extracting", "information"], "Task"], [1494, 1495, ["training"], "Task"], [1562, 1564, ["unweighted", "summation"], "Method"], [1566, 1568, ["multidimensional", "RNNs"], "Method"], [1578, 1580, ["unstable", "convergence"], "Metric"], [1621, 1625, ["tree", "-", "structured", "RNNs"], "Method"], [1636, 1640, ["tree", "-", "structured", "RNNs"], "Method"], [1724, 1728, ["CAS", "-", "LSTM", "layer"], "Method"], [1794, 1796, ["multidimensional", "RNNs"], "Task"], [1797, 1799, ["Multidimensional", "RNNs"], "Method"], [1800, 1802, ["MDRNN", ")"], "Method"], [1806, 1809, ["1D", "sequential", "RNNs"], "Method"], [1823, 1825, ["image", "segmentation"], "Task"], [1826, 1828, ["handwriting", "recognition"], "Task"], [1830, 1832, ["multidimensional", "LSTMs"], "Method"], [1833, 1834, ["MDLSTM"], "Method"], [1868, 1871, ["CAS", "-", "LSTM"], "Method"], [1879, 1880, ["MDLSTM"], "Method"], [1915, 1917, ["Sentence", "Encoders"], "Method"], [1918, 1921, ["sentence", "encoder", "network"], "Method"], [1946, 1948, ["word", "representations"], "Method"], [1958, 1962, ["CAS", "-", "LSTM", "model"], "Method"], [1969, 1971, ["sentence", "representation"], "Method"], [1976, 1979, ["max", "-", "pooling"], "Method"], [2001, 2004, ["max", "-", "pooling"], "Method"], [2008, 2014, ["mean", "-", "and", "last", "-", "pooling"], "Method"], [2022, 2027, ["bidirectional", "CAS", "-", "LSTM", "network"], "Method"], [2038, 2040, ["forward", "representations"], "Method"], [2042, 2044, ["backward", "representations"], "Method"], [2047, 2050, ["max", "-", "pooled"], "Method"], [2053, 2055, ["sentence", "representation"], "Method"], [2071, 2075, ["Top", "-", "layer", "Classifiers"], "Method"], [2077, 2080, ["natural", "language", "inference"], "Task"], [2086, 2088, ["heuristic", "function"], "Method"], [2093, 2095, ["feature", "extraction"], "Task"], [2098, 2100, ["vector", "concatenation"], "Method"], [2116, 2118, ["paraphrase", "identification"], "Task"], [2129, 2131, ["sentiment", "classification"], "Task"], [2135, 2137, ["sentence", "representation"], "Method"], [2149, 2151, ["MLP", "classifier"], "Method"], [2152, 2154, ["ReLU", "activation"], "Method"], [2157, 2162, ["fully", "-", "connected", "softmax", "layer"], "Method"], [2165, 2167, ["label", "distribution"], "Task"], [2193, 2196, ["natural", "language", "inference"], "Task"], [2197, 2198, ["NLI"], "Task"], [2200, 2202, ["paraphrase", "identification"], "Task"], [2203, 2204, ["PI"], "Task"], [2207, 2209, ["sentiment", "classification"], "Task"], [2220, 2222, ["model", "variants"], "Method"], [2238, 2239, ["NLI"], "Task"], [2240, 2242, ["PI", "tasks"], "Task"], [2249, 2252, ["sentence", "pair", "classification"], "Task"], [2270, 2272, ["shared", "encoder"], "Method"], [2290, 2292, ["extracting", "semantics"], "Task"], [2299, 2302, ["CAS", "-", "LSTM"], "Method"], [2306, 2310, ["sentence", "encoding", "based", "approaches"], "Method"], [2313, 2316, ["Natural", "Language", "Inference"], "Task"], [2327, 2329, ["NLI", "task"], "Task"], [2330, 2331, ["SNLI"], "Material"], [2332, 2334, ["MultiNLI", "datasets"], "Material"], [2362, 2363, ["SNLI"], "Material"], [2364, 2365, ["MultiNLI"], "Material"], [2404, 2406, ["1024D", "MLP"], "Method"], [2417, 2418, ["dropout"], "Method"], [2424, 2426, ["MLP", "layers"], "Method"], [2434, 2436, ["MLP", "classifier"], "Method"], [2459, 2460, ["SNLI"], "Material"], [2461, 2463, ["MultiNLI", "datasets"], "Material"], [2465, 2466, ["SNLI"], "Material"], [2480, 2481, ["accuracy"], "Metric"], [2491, 2492, ["MultiNLI"], "Material"], [2497, 2498, ["accuracy"], "Metric"], [2509, 2512, ["in", "-", "domain"], "Task"], [2535, 2537, ["word", "representations"], "Method"], [2564, 2566, ["pooling", "method"], "Method"], [2575, 2578, ["max", "-", "pooling"], "Method"], [2583, 2587, ["intra", "-", "sentence", "attention"], "Method"], [2588, 2590, ["generalized", "pooling"], "Method"], [2593, 2595, ["Paraphrase", "Identification"], "Task"], [2610, 2612, ["PI", "task"], "Task"], [2650, 2652, ["300D", "encoders"], "Method"], [2654, 2656, ["1024D", "MLP"], "Method"], [2662, 2666, ["CAS", "-", "LSTM", "layers"], "Method"], [2691, 2692, ["MLP"], "Method"], [2739, 2741, ["Sentiment", "Classification"], "Task"], [2743, 2745, ["sentiment", "classification"], "Task"], [2748, 2751, ["Stanford", "Sentiment", "Treebank"], "Material"], [2752, 2753, ["SST"], "Material"], [2823, 2824, ["evaluation"], "Task"], [2833, 2836, ["SST", "-", "2"], "Material"], [2837, 2840, ["SST", "-", "5"], "Material"], [2856, 2859, ["SST", "-", "2"], "Material"], [2929, 2932, ["binary", "classification", "problem"], "Task"], [2934, 2937, ["SST", "-", "5"], "Material"], [2950, 2955, ["5", "-", "class", "classification", "problem"], "Task"], [2994, 2998, ["150D", "or", "300D", "encoders"], "Method"], [3001, 3004, ["300D", "MLP", "classifier"], "Method"], [3023, 3024, ["SST"], "Material"], [3044, 3045, ["accuracy"], "Metric"], [3046, 3049, ["SST", "-", "2"], "Material"], [3053, 3056, ["SST", "-", "5"], "Material"], [3065, 3068, ["Forget", "Gate", "Analysis"], "Method"], [3103, 3105, ["SNLI", "dataset"], "Material"], [3109, 3116, ["3", "-", "layer", "CAS", "-", "LSTM", "model"], "Method"], [3119, 3121, ["SNLI", "dataset"], "Material"], [3342, 3346, ["retaining", "or", "discarding", "signals"], "Task"], [3475, 3477, ["SNLI", "dataset"], "Material"], [3493, 3494, ["modifications"], "Task"], [3624, 3626, ["sentence", "modeling"], "Task"], [3702, 3703, ["baseline"], "Metric"], [3727, 3729, ["sentence", "representations"], "Task"], [3732, 3734, ["zhang2016highway", "zhang2016highway"], "Method"], [3744, 3748, ["Stacked", "recurrent", "neural", "networks"], "Method"], [3757, 3759, ["stacking", "RNNs"], "Task"], [3761, 3764, ["plain", "stacked", "RNNs"], "Method"], [3777, 3779, ["LSTM", "layer"], "Method"], [3810, 3812, ["gating", "mechanism"], "Method"], [3833, 3835, ["vector", "addition"], "Method"], [3836, 3837, ["concatenation"], "Method"], [3842, 3843, ["concatenation"], "Method"], [3855, 3857, ["Highway", "LSTMs"], "Method"], [3858, 3862, ["depth", "-", "gated", "LSTMs"], "Method"], [3888, 3891, ["automatic", "speech", "recognition"], "Task"], [3892, 3894, ["language", "modeling"], "Task"], [3899, 3902, ["CAS", "-", "LSTM"], "Method"], [3928, 3930, ["Highway", "LSTMs"], "Method"], [3931, 3935, ["depth", "-", "gated", "LSTMs"], "Method"], [3964, 3968, ["Multidimensional", "recurrent", "neural", "networks"], "Method"], [3979, 3980, ["RNNs"], "Method"], [4037, 4042, ["character", "-", "level", "language", "modeling"], "Task"], [4043, 4045, ["machine", "translation"], "Task"], [4047, 4049, ["Multidimensional", "RNNs"], "Method"], [4091, 4092, ["RNNs"], "Method"], [4095, 4098, ["stacking", "RNN", "layers"], "Method"], [4141, 4144, ["Deep", "Transition", "RNNs"], "Method"], [4148, 4155, ["deep", "hidden", "-", "to", "-", "hidden", "transitions"], "Task"], [4160, 4164, ["adaptive", "computation", "time", "algorithm"], "Method"], [4182, 4186, ["Fast", "-", "Slow", "RNNs"], "Method"], [4213, 4215, ["Multiscale", "RNNs"], "Method"], [4216, 4218, ["Clockwork", "RNNs"], "Method"], [4219, 4222, ["Hierarchical", "Multiscale", "RNNs"], "Method"], [4251, 4254, ["stacking", "recurrent", "layers"], "Method"], [4256, 4257, ["training"], "Task"], [4282, 4284, ["modeling", "sentences"], "Task"], [4286, 4289, ["CAS", "-", "LSTM"], "Method"], [4330, 4333, ["natural", "language", "inference"], "Task"], [4334, 4336, ["paraphrase", "identification"], "Task"], [4338, 4340, ["sentiment", "classification"], "Task"], [4353, 4354, ["accuracy"], "Metric"], [4355, 4356, ["SNLI"], "Material"], [4357, 4361, ["Quora", "Question", "Pairs", "datasets"], "Material"], [4366, 4367, ["MultiNLI"], "Material"], [4368, 4370, ["SST", "datasets"], "Material"], [4408, 4411, ["sentence", "modeling", "tasks"], "Task"], [4415, 4417, ["sequence", "labeling"], "Task"], [4418, 4420, ["sequence", "generation"], "Task"], [4422, 4424, ["language", "modeling"], "Task"], [4430, 4432, ["context", "integration"], "Task"], [6, 7, ["LSTMs"], "Method"], [36, 37, ["LSTMs"], "Method"], [85, 86, ["LSTM"], "Method"], [226, 227, ["NLP"], "Task"], [536, 537, ["LSTMs"], "Method"], [616, 617, ["LSTMs"], "Method"], [674, 675, ["LSTMs"], "Method"], [687, 688, ["LSTMs"], "Method"], [711, 712, ["LSTM"], "Method"], [923, 924, ["LSTM"], "Method"], [1000, 1001, ["LSTMs"], "Method"], [1007, 1008, ["LSTM"], "Method"], [1183, 1184, ["LSTMs"], "Method"], [1189, 1190, ["LSTM"], "Method"], [1362, 1363, ["LSTM"], "Method"], [1647, 1648, ["LSTMs"], "Method"], [1651, 1652, ["LSTMs"], "Method"], [1673, 1674, ["LSTMs"], "Method"], [1703, 1704, ["LSTMs"], "Method"], [1715, 1716, ["LSTMs"], "Method"], [1739, 1740, ["LSTM"], "Method"], [1768, 1769, ["LSTMs"], "Method"], [1773, 1774, ["LSTMs"], "Method"], [1887, 1888, ["LSTMs"], "Method"], [2065, 2066, ["LSTM"], "Method"], [2597, 2600, ["Quora", "Question", "Pairs"], "Material"], [2643, 2644, ["NLI"], "Task"], [2697, 2700, ["Quora", "Question", "Pairs"], "Material"], [2980, 2981, ["SST"], "Material"], [3051, 3052, ["accuracy"], "Metric"], [3161, 3162, ["LSTM"], "Method"], [3509, 3510, ["LSTMs"], "Method"], [3766, 3767, ["LSTMs"], "Method"], [3784, 3785, ["LSTMs"], "Method"], [3987, 3988, ["LSTMs"], "Method"], [3993, 3994, ["LSTM"], "Method"], [4031, 4032, ["LSTMs"], "Method"], [4279, 4280, ["LSTM"], "Method"], [4378, 4379, ["LSTM"], "Method"], [4405, 4406, ["LSTM"], "Method"]]} {"doc_id": "25a784f7f8c94c42821ee078587fc38dffcd00a4", "text": ["document", ":", "Robust", "Face", "Detection", "via", "Learning", "Small", "Faces", "on", "Hard", "Images", "Recent", "anchor", "-", "based", "deep", "face", "detectors", "have", "achieved", "promising", "performance", ",", "but", "they", "are", "still", "struggling", "to", "detect", "hard", "faces", ",", "such", "as", "small", ",", "blurred", "and", "partially", "occluded", "faces", ".", "A", "reason", "is", "that", "they", "treat", "all", "images", "and", "faces", "equally", ",", "without", "putting", "more", "effort", "on", "hard", "ones", ";", "however", ",", "many", "training", "images", "only", "contain", "easy", "faces", ",", "which", "are", "less", "helpful", "to", "achieve", "better", "performance", "on", "hard", "images", ".", "In", "this", "paper", ",", "we", "propose", "that", "the", "robustness", "of", "a", "face", "detector", "against", "hard", "faces", "can", "be", "improved", "by", "learning", "small", "faces", "on", "hard", "images", ".", "Our", "intuitions", "are", "(", "1", ")", "hard", "images", "are", "the", "images", "which", "contain", "at", "least", "one", "hard", "face", ",", "thus", "they", "facilitate", "training", "robust", "face", "detectors", ";", "(", "2", ")", "most", "hard", "faces", "are", "small", "faces", "and", "other", "types", "of", "hard", "faces", "can", "be", "easily", "converted", "to", "small", "faces", "by", "shrinking", ".", "We", "build", "an", "anchor", "-", "based", "deep", "face", "detector", ",", "which", "only", "output", "a", "single", "feature", "map", "with", "small", "anchors", ",", "to", "specifically", "learn", "small", "faces", "and", "train", "it", "by", "a", "novel", "hard", "image", "mining", "strategy", ".", "Extensive", "experiments", "have", "been", "conducted", "on", "WIDER", "FACE", ",", "FDDB", ",", "Pascal", "Faces", ",", "and", "AFW", "datasets", "to", "show", "the", "effectiveness", "of", "our", "method", ".", "Our", "method", "achieves", "APs", "of", "95.7", ",", "94.9", "and", "89.7", "on", "easy", ",", "medium", "and", "hard", "WIDER", "FACE", "val", "dataset", "respectively", ",", "which", "surpass", "the", "previous", "state", "-", "of", "-", "the", "-", "arts", ",", "especially", "on", "the", "hard", "subset", ".", "Code", "and", "model", "are", "available", "at", ".", "section", ":", "Introduction", "Face", "detection", "is", "a", "fundamental", "and", "important", "computer", "vision", "problem", ",", "which", "is", "critical", "for", "many", "face", "-", "related", "tasks", ",", "such", "as", "face", "alignment", ",", "tracking", "and", "recognition", ".", "Stem", "from", "the", "recent", "successful", "development", "of", "deep", "neural", "networks", ",", "massive", "CNN", "-", "based", "face", "detection", "approaches", "have", "been", "proposed", "and", "achieved", "the", "state", "-", "of", "-", "the", "-", "art", "performance", ".", "However", ",", "face", "detection", "remains", "a", "challenging", "task", "due", "to", "occlusion", ",", "illumination", ",", "makeup", ",", "as", "well", "as", "pose", "and", "scale", "variance", ",", "as", "shown", "in", "the", "benchmark", "dataset", "WIDER", "FACE", ".", "Current", "state", "-", "of", "-", "the", "-", "art", "CNN", "-", "based", "face", "detectors", "attempt", "to", "address", "these", "challenges", "by", "employing", "more", "powerful", "backbone", "models", ",", "exploiting", "feature", "pyramid", "-", "style", "architectures", "to", "combine", "features", "from", "multiple", "detection", "feature", "maps", ",", "designing", "denser", "anchors", "and", "utilizing", "larger", "contextual", "information", ".", "These", "methods", "and", "techniques", "have", "been", "shown", "to", "be", "successful", "to", "build", "a", "robust", "face", "detector", ",", "and", "improve", "the", "performance", "towards", "human", "-", "level", "for", "most", "images", ".", "In", "spite", "of", "their", "success", "for", "most", "images", ",", "an", "evident", "performance", "gap", "still", "exists", "especially", "for", "those", "hard", "images", "which", "contain", "small", ",", "blurred", "and", "partially", "occluded", "faces", ".", "We", "realize", "that", "these", "hard", "images", "have", "become", "the", "main", "barriers", "for", "face", "detectors", "to", "achieve", "human", "-", "level", "detection", "performance", ".", "In", "Figure", "[", "reference", "]", ",", "we", "show", "that", ",", "even", "on", "the", "train", "set", "of", "WIDER", "FACE", ",", "the", "official", "pre", "-", "trained", "SSH", "still", "fails", "on", "some", "of", "the", "images", "with", "extremely", "hard", "faces", ".", "We", "show", "two", "such", "hard", "training", "images", "in", "the", "upper", "right", "corner", "in", "Figure", "[", "reference", "]", ".", "On", "the", "other", "hand", ",", "most", "training", "images", "with", "easy", "faces", "can", "be", "almost", "perfectly", "detected", "(", "see", "the", "illustration", "in", "the", "right", "lower", "corner", "of", "Figure", "[", "reference", "]", ")", ".", "As", "shown", "in", "left", "part", "of", "Figure", "[", "reference", "]", ",", "over", "two", "thirds", "of", "the", "training", "images", "already", "obtained", "perfect", "detection", "accuracy", ",", "which", "indicates", "that", "those", "easy", "images", "are", "less", "useful", "towards", "training", "a", "robust", "face", "detector", ".", "To", "address", "this", "issue", ",", "in", "this", "paper", ",", "we", "propose", "a", "robust", "face", "detector", "by", "putting", "more", "training", "focus", "on", "those", "hard", "images", ".", "This", "issue", "is", "most", "related", "to", "anchor", "-", "level", "hard", "example", "mining", "discussed", "in", "OHEM", ".", "However", ",", "due", "to", "the", "sparsity", "of", "ground", "-", "truth", "faces", "and", "positive", "anchors", ",", "traditional", "anchor", "-", "level", "hard", "example", "mining", "mainly", "focuses", "on", "mining", "hard", "negative", "anchors", ",", "and", "mining", "hard", "anchors", "on", "well", "-", "detected", "images", "exhibits", "less", "effectiveness", "since", "there", "is", "no", "useful", "information", "that", "can", "be", "further", "exploited", "in", "these", "easy", "images", ".", "To", "address", "this", "issue", ",", "we", "propose", "to", "mine", "hard", "examples", "at", "image", "level", "in", "parallel", "with", "anchor", "level", ".", "More", "specifically", ",", "we", "propose", "to", "dynamically", "assign", "difficulty", "scores", "to", "training", "images", "during", "the", "learning", "process", ",", "which", "can", "determine", "whether", "an", "image", "is", "already", "well", "-", "detected", "or", "still", "useful", "for", "further", "training", ".", "This", "allows", "us", "to", "fully", "utilize", "the", "images", "which", "were", "not", "perfectly", "detected", "to", "better", "facilitate", "the", "following", "learning", "process", ".", "We", "show", "this", "strategy", "can", "make", "our", "detector", "more", "robust", "towards", "hard", "faces", ",", "without", "involving", "more", "complex", "network", "architecture", "and", "computation", "overhead", ".", "Apart", "from", "mining", "the", "hard", "images", ",", "we", "also", "propose", "to", "improve", "the", "detection", "quality", "by", "exclusively", "exploiting", "small", "faces", ".", "Small", "faces", "are", "typically", "hard", "and", "have", "attracted", "extensive", "research", "attention", ".", "Existing", "methods", "aim", "at", "building", "a", "scale", "-", "invariant", "face", "detector", "to", "learn", "and", "infer", "on", "both", "small", "and", "big", "faces", ",", "with", "multiple", "levels", "of", "detection", "features", "and", "anchors", "of", "different", "sizes", ".", "Compared", "with", "these", "methods", ",", "our", "detector", "is", "more", "efficient", "since", "it", "is", "specially", "designed", "to", "aggressively", "leveraging", "the", "small", "faces", "during", "training", ".", "More", "specifically", ",", "large", "faces", "are", "automatically", "ignored", "during", "training", "due", "to", "our", "anchor", "design", ",", "so", "that", "the", "model", "can", "fully", "focus", "on", "the", "small", "hard", "faces", ".", "Additionally", ",", "experiments", "demonstrate", "that", "this", "design", "effectively", "achieves", "improvements", "on", "detecting", "all", "faces", "in", "spite", "of", "its", "simple", "and", "shallow", "architecture", ".", "To", "conclude", ",", "in", "this", "paper", ",", "we", "propose", "a", "novel", "face", "detector", "with", "the", "following", "contributions", ":", "We", "propose", "a", "hard", "image", "mining", "strategy", ",", "to", "improve", "the", "robustness", "of", "our", "detector", "to", "those", "extremely", "hard", "faces", ".", "This", "is", "done", "without", "any", "extra", "modules", ",", "parameters", "or", "computation", "overhead", "added", "on", "the", "existing", "detector", ".", "We", "design", "a", "single", "shot", "detector", "with", "only", "one", "detection", "feature", "map", ",", "which", "focuses", "on", "small", "faces", "with", "a", "specific", "range", "of", "sizes", ".", "This", "allows", "our", "model", "to", "be", "simple", "and", "focus", "on", "difficult", "small", "faces", "without", "struggling", "with", "scale", "variance", ".", "Our", "face", "detector", "establishes", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "all", "popular", "face", "detection", "datasets", ",", "including", "WIDER", "FACE", ",", "FDDB", ",", "Pascal", "Faces", ",", "and", "AFW", ".", "We", "achieve", "95.7", ",", "94.9", "and", "89.7", "on", "easy", ",", "medium", "and", "hard", "WIDER", "FACE", "val", "dataset", ".", "Our", "method", "also", "achieves", "APs", "of", "99.00", "and", "99.60", "on", "Pascal", "Faces", "and", "AFW", "respectively", ",", "as", "well", "as", "a", "TPR", "of", "98.7", "on", "FDDB", ".", "The", "remainder", "of", "this", "paper", "is", "organized", "as", "follows", ".", "In", "Section", "[", "reference", "]", ",", "we", "discuss", "some", "studies", "have", "been", "done", "which", "are", "related", "to", "our", "paper", ".", "In", "Section", "[", "reference", "]", ",", "we", "dive", "into", "details", "of", "our", "proposed", "method", ",", "and", "we", "discuss", "experiment", "results", "and", "ablation", "experiments", "in", "Section", "[", "reference", "]", ".", "Finally", ",", "conclusions", "are", "drawn", "in", "Section", "[", "reference", "]", ".", "section", ":", "Related", "work", "Face", "detection", "has", "received", "extensive", "research", "attention", ".", "With", "the", "emergence", "of", "modern", "CNN", "and", "object", "detector", ",", "there", "are", "many", "face", "detectors", "proposed", "to", "achieve", "promising", "performances", ",", "by", "adapting", "general", "object", "detection", "framework", "into", "face", "detection", "domain", ".", "We", "briefly", "review", "hard", "example", "mining", ",", "face", "detection", "architecture", ",", "and", "anchor", "design", "&", "matching", ".", "subsection", ":", "Hard", "example", "mining", "Hard", "example", "mining", "is", "an", "important", "strategy", "to", "improve", "model", "quality", ",", "and", "has", "been", "studied", "extensively", "in", "image", "classification", "and", "general", "object", "detection", ".", "The", "main", "idea", "is", "to", "find", "some", "hard", "positive", "and", "hard", "negative", "examples", "at", "each", "step", ",", "and", "put", "more", "effort", "into", "training", "on", "those", "hard", "examples", ".", "Recently", ",", "with", "modern", "detection", "frameworks", "proposed", "to", "boost", "the", "performance", ",", "OHEM", "and", "Focal", "loss", "have", "been", "proposed", "to", "select", "hard", "examples", ".", "OHEM", "computed", "the", "gradients", "of", "the", "networks", "by", "selecting", "the", "proposals", "with", "highest", "losses", "in", "every", "minibatch", ";", "while", "Focal", "loss", "aimed", "at", "naturally", "putting", "more", "focus", "on", "hard", "and", "misclassified", "examples", "by", "adding", "a", "factor", "to", "the", "standard", "cross", "entropy", "criterion", ".", "However", ",", "these", "algorithms", "mainly", "focused", "on", "anchor", "-", "level", "or", "proposal", "-", "level", "mining", ".", "It", "can", "not", "handle", "the", "imbalance", "of", "easy", "and", "hard", "images", "in", "the", "dataset", ".", "In", "our", "paper", ",", "we", "propose", "to", "exploit", "hard", "example", "mining", "on", "image", "level", ",", "hard", "image", "mining", ",", "to", "improve", "the", "quality", "of", "face", "detector", "on", "extremely", "hard", "faces", ".", "More", "specifically", ",", "we", "assign", "difficulty", "scores", "to", "training", "images", "while", "training", "with", "an", "SGD", "mechanism", ",", "and", "re", "-", "sample", "the", "training", "images", "to", "build", "a", "new", "training", "subset", "at", "the", "next", "epoch", ".", "subsection", ":", "Face", "Detection", "Architecture", "Recent", "state", "-", "of", "-", "the", "-", "art", "face", "detectors", "are", "generally", "built", "based", "on", "Faster", "-", "RCNN", ",", "R", "-", "FCN", "or", "SSD", ".", "SSH", "exploited", "the", "RPN", "(", "Region", "Proposal", "Network", ")", "from", "Faster", "-", "RCNN", "to", "detect", "faces", ",", "by", "building", "three", "detection", "feature", "maps", "and", "designing", "six", "anchors", "with", "different", "sizes", "attached", "to", "the", "detection", "feature", "maps", ".", "S", "FD", "and", "PyramidBox", ",", "on", "the", "other", "hand", ",", "adopted", "SSD", "as", "their", "detection", "architecture", "with", "six", "different", "detection", "feature", "maps", ".", "Different", "from", "S", "FD", ",", "PyramidBox", "exploited", "a", "feature", "pyramid", "-", "style", "structure", "to", "combine", "features", "from", "different", "detection", "feature", "maps", ".", "Our", "proposed", "method", ",", "on", "the", "other", "hand", ",", "only", "builds", "single", "level", "detection", "feature", "map", ",", "based", "on", "VGG16", ",", "for", "classification", "and", "bounding", "-", "box", "regression", ",", "which", "is", "both", "simple", "and", "effective", ".", "subsection", ":", "Anchor", "design", "and", "matching", "Usually", ",", "anchors", "are", "designed", "to", "have", "different", "sizes", "to", "detect", "objects", "with", "different", "scales", ",", "in", "order", "to", "build", "a", "scale", "-", "invariant", "detector", ".", "SSD", "as", "well", "as", "its", "follow", "-", "up", "detectors", "S", "FD", "and", "PyramidBox", ",", "had", "six", "sets", "of", "anchors", "with", "different", "sizes", ",", "ranging", "from", "(", ")", "to", "(", ")", ",", "and", "their", "network", "architectures", "had", "six", "levels", "of", "detection", "feature", "maps", ",", "with", "resolutions", "ranging", "from", "to", ",", "respectively", ".", "Similarly", ",", "SSH", "had", "the", "same", "anchor", "setting", ",", "and", "those", "anchors", "were", "attached", "to", "three", "levels", "of", "detection", "feature", "maps", "with", "resolutions", "ranging", "from", "to", ".", "The", "difference", "between", "SSH", "and", "S", "DF", "is", "that", "in", "SSH", ",", "anchors", "with", "two", "neighboring", "sizes", "shared", "the", "same", "detection", "feature", "map", ",", "while", "in", "S", "DF", ",", "anchors", "with", "different", "sizes", "are", "attached", "to", "different", "detection", "feature", "maps", ".", "SNIP", "discussed", "an", "alternative", "approach", "to", "handle", "scales", ".", "It", "showed", "that", "CNNs", "are", "not", "robust", "to", "changes", "in", "scale", ",", "so", "training", "and", "testing", "on", "the", "same", "scales", "of", "an", "image", "pyramid", "can", "be", "a", "more", "optimal", "strategy", ".", "In", "our", "paper", ",", "we", "exploit", "this", "idea", "by", "limiting", "the", "anchor", "sizes", "to", "be", "(", ")", ",", "(", ")", "and", "(", ")", ".", "Then", "those", "faces", "with", "either", "too", "small", "or", "too", "big", "sizes", "will", "not", "be", "matched", "to", "any", "of", "the", "anchors", ",", "thus", "will", "be", "ignored", "during", "the", "training", "and", "testing", ".", "By", "removing", "those", "large", "anchors", "with", "sizes", "larger", "than", "(", ")", ",", "our", "network", "focuses", "more", "on", "small", "faces", "which", "are", "potentially", "more", "difficult", ".", "To", "deal", "with", "large", "faces", ",", "we", "use", "multiscale", "training", "and", "testing", "to", "resize", "them", "to", "match", "our", "anchors", ".", "Experiments", "show", "this", "design", "performs", "well", "on", "both", "small", "and", "big", "faces", ",", "although", "it", "has", "fewer", "detection", "feature", "maps", "and", "anchor", "sizes", ".", "section", ":", "Proposed", "method", "In", "this", "section", ",", "we", "introduce", "our", "proposed", "method", "for", "effective", "face", "detection", ".", "We", "first", "discuss", "the", "architecture", "of", "our", "detector", "in", "Section", "[", "reference", "]", ",", "then", "we", "elaborate", "our", "hard", "image", "mining", "strategy", "in", "Section", "[", "reference", "]", ",", "as", "well", "as", "some", "other", "useful", "training", "techniques", "in", "Section", "[", "reference", "]", ".", "subsection", ":", "Single", "-", "level", "small", "face", "detection", "framework", "The", "framework", "of", "our", "face", "detector", "is", "illustrated", "in", "Figure", "[", "reference", "]", ".", "We", "use", "VGG16", "network", "as", "our", "backbone", "CNN", ",", "and", "combine", "conv4_3", "and", "conv5_3", "features", ",", "to", "build", "the", "detection", "feature", "map", "with", "both", "low", "-", "level", "and", "high", "-", "level", "semantic", "information", ".", "Similar", "to", "SSH", ",", "we", "apply", "1", "1", "convolution", "layers", "after", "conv4_3", "and", "conv5_3", "to", "reduce", "dimension", ",", "and", "then", "apply", "a", "3", "3", "convolution", "layer", "on", "the", "concatenation", "of", "these", "two", "dimension", "reduced", "features", ".", "The", "output", "feature", "of", "the", "3", "3", "convolution", "layer", "is", "the", "final", "detection", "feature", "map", ",", "which", "will", "be", "fed", "into", "the", "detection", "head", "for", "classification", "and", "bounding", "-", "box", "regression", ".", "The", "detection", "feature", "map", "has", "a", "resolution", "of", "of", "the", "original", "image", "(", "of", "size", ")", ".", "We", "attach", "three", "anchors", "at", "each", "point", "in", "the", "grid", "as", "default", "face", "detection", "boxes", ".", "Then", "we", "do", "classification", "and", "bounding", "-", "box", "regression", "on", "those", "anchors", ".", "Unlike", "many", "other", "face", "detectors", "which", "build", "multiple", "feature", "maps", "to", "detect", "face", "with", "a", "variant", "range", "of", "scales", ",", "inspired", "by", "SNIP", ",", "faces", "are", "trained", "and", "inferred", "with", "roughly", "the", "same", "scales", ".", "We", "only", "have", "one", "detection", "feature", "map", ",", "with", "three", "sets", "of", "anchors", "attached", "to", "it", ".", "The", "anchors", "have", "sizes", "of", "(", ")", ",", "(", ")", "and", "(", ")", ",", "and", "the", "aspect", "ratio", "is", "set", "to", "be", "1", ".", "By", "making", "this", "configuration", ",", "our", "network", "only", "trains", "and", "infers", "on", "small", "and", "medium", "size", "of", "faces", ";", "and", "we", "propose", "to", "handle", "large", "faces", "by", "shrinking", "the", "images", "in", "the", "test", "phase", ".", "We", "argue", "that", "there", "is", "no", "speed", "or", "accuracy", "degradation", "for", "large", "faces", ",", "since", "inferring", "on", "a", "tiny", "image", "(", "with", "short", "side", "containing", "100", "or", "300", "pixels", ")", "is", "very", "fast", ",", "and", "the", "shrinked", "large", "face", "will", "still", "have", "enough", "information", "to", "be", "recognized", ".", "To", "handle", "the", "difference", "of", "anchor", "sizes", "attached", "to", "the", "same", "detection", "feature", "map", ",", "we", "propose", "a", "detection", "head", "which", "uses", "different", "dilation", "rates", "for", "anchors", "with", "different", "sizes", ",", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "The", "intuition", "is", "that", "in", "order", "to", "detect", "faces", "with", "different", "sizes", ",", "different", "effective", "receptive", "fields", "are", "required", ".", "This", "naturally", "requires", "the", "backbone", "feature", "map", "to", "be", "invariant", "to", "scales", ".", "To", "this", "end", ",", "we", "adopt", "different", "dilation", "rates", "for", "anchors", "with", "different", "sizes", ".", "For", "anchors", "with", "size", "(", ")", ",", "(", ")", "and", "(", ")", ",", "we", "use", "a", "convolution", "with", "kernel", "size", "of", "and", "dilation", "rate", "of", ",", "and", "to", "gather", "context", "features", "at", "different", "scales", ".", "These", "three", "convolution", "layers", "share", "weights", "to", "reduce", "the", "model", "size", ".", "With", "this", "design", ",", "the", "input", "of", "the", "convolution", ",", "will", "be", "aligned", "to", "the", "same", "location", "of", "faces", ",", "regardless", "of", "the", "size", "of", "faces", "and", "anchors", ".", "Ablation", "experiments", "show", "the", "effectiveness", "of", "this", "multi", "-", "dilation", "design", ".", "subsection", ":", "Hard", "image", "mining", "Different", "from", "OHEM", "discussed", "in", "Section", "[", "reference", "]", ",", "which", "selects", "proposals", "or", "anchors", "with", "the", "highest", "losses", ",", "we", "propose", "a", "novel", "hard", "image", "mining", "strategy", "at", "image", "level", ".", "The", "intuition", "is", "that", "most", "images", "in", "the", "dataset", "are", "very", "easy", ",", "and", "we", "can", "achieve", "a", "very", "high", "AP", "even", "on", "the", "hard", "subset", "of", "the", "WIDER", "FACE", "val", "dataset", "with", "our", "baseline", "model", ".", "We", "believe", "not", "all", "training", "images", "should", "be", "treated", "equally", ",", "and", "well", "-", "recognized", "images", "will", "not", "help", "towards", "training", "a", "more", "robust", "face", "detector", ".", "To", "put", "more", "attention", "on", "training", "hard", "images", "instead", "of", "easy", "ones", ",", "we", "use", "a", "subset", "of", "all", "training", "images", ",", "to", "contain", "hard", "ones", "for", "training", ".", "At", "the", "beginning", "of", "each", "epoch", ",", "we", "build", "based", "on", "the", "difficulty", "scores", "obtained", "in", "the", "previous", "epoch", ".", "We", "initially", "use", "all", "training", "images", "to", "train", "our", "model", "(", ")", ".", "This", "is", "due", "to", "the", "fact", "that", "our", "initial", "ImageNet", "pre", "-", "trained", "model", "will", "only", "give", "random", "guess", "towards", "face", "detection", ".", "In", "this", "case", ",", "there", "is", "no", "easy", "image", ".", "In", "other", "words", ",", "every", "image", "is", "considered", "as", "hard", "image", "and", "fed", "to", "the", "network", "for", "training", "at", "the", "first", "epoch", ".", "During", "the", "training", "procedure", ",", "we", "dynamically", "assign", "different", "difficulty", "scores", "to", "training", "images", ",", "which", "is", "defined", "by", "the", "metric", "Worst", "Positive", "Anchor", "Score", "(", "WPAS", ")", ":", "where", "is", "the", "set", "of", "positive", "anchors", "for", "image", ",", "with", "IoU", "over", "0.5", "against", "ground", "-", "truth", "boxes", ",", "is", "the", "classification", "logit", "and", ",", "are", "the", "logits", "of", "anchor", "for", "image", "to", "be", "foreground", "face", "and", "background", ".", "All", "images", "are", "initially", "marked", "as", "hard", ",", "and", "any", "image", "with", "WPAS", "greater", "than", "a", "threshold", "will", "be", "marked", "as", "easy", "image", ".", "At", "the", "beginning", "of", "each", "epoch", ",", "we", "first", "randomly", "shuffle", "the", "training", "dataset", "to", "generate", "the", "complete", "training", "list", "for", "the", "following", "epoch", "of", "training", ".", "Then", "given", "an", "image", "marked", "as", "easy", ",", "we", "remove", "it", "from", "with", "a", "probability", "of", ".", "The", "remaining", "training", "list", ",", "which", "focuses", "more", "on", "hard", "images", ",", "will", "be", "used", "for", "training", "at", "this", "epoch", ".", "Note", "that", "for", "multi", "-", "GPU", "training", ",", "each", "GPU", "will", "maintain", "its", "training", "list", "independently", ".", "In", "our", "experiments", ",", "we", "set", "the", "probability", "to", "be", "0.7", ",", "and", "the", "threshold", "to", "be", "0.85", ".", "subsection", ":", "Training", "strategy", "subsubsection", ":", "Multi", "-", "scale", "training", "and", "anchor", "matching", "Since", "we", "only", "have", "anchors", "covering", "a", "limited", "range", "of", "face", "scales", ",", "we", "train", "our", "model", "by", "varying", "the", "sizes", "of", "training", "images", ".", "During", "the", "training", "phase", ",", "we", "resize", "the", "training", "images", "so", "that", "the", "short", "side", "of", "the", "image", "contains", "pixels", ",", "where", "is", "randomly", "selected", "from", ".", "We", "also", "set", "an", "upper", "bound", "of", "2000", "pixels", "to", "the", "long", "side", "of", "the", "image", "considering", "the", "GPU", "memory", "limitation", ".", "For", "each", "anchor", ",", "we", "assign", "a", "label", "based", "on", "how", "well", "it", "matches", "with", "any", "ground", "-", "truth", "face", "bounding", "box", ".", "If", "an", "anchor", "has", "an", "IoU", "(", "Intersection", "over", "Union", ")", "over", "0.5", "against", "a", "ground", "-", "truth", "face", "bounding", "box", ",", "we", "assign", "to", "that", "anchor", ".", "On", "the", "other", "hand", ",", "if", "the", "IoU", "against", "any", "ground", "-", "truth", "face", "bounding", "box", "is", "lower", "than", "0.3", ",", "we", "assign", "to", "that", "anchor", ".", "All", "other", "anchors", "will", "be", "given", "as", "the", "label", ",", "and", "thus", "will", "be", "ignored", "in", "the", "classification", "loss", ".", "By", "doing", "so", ",", "we", "only", "train", "on", "faces", "with", "designated", "scales", ".", "Those", "faces", "with", "no", "anchor", "matching", "will", "be", "simply", "ignored", ",", "since", "we", "do", "not", "assign", "the", "anchor", "with", "largest", "IoU", "to", "it", "(", "thus", "assign", "the", "corresponding", "anchor", "label", ")", "as", "Faster", "-", "RCNN", "does", ".", "This", "anchor", "matching", "strategy", "will", "ignore", "the", "large", "faces", ",", "and", "our", "model", "can", "put", "more", "capacity", "on", "learning", "different", "face", "patterns", "on", "hard", "small", "faces", "instead", "of", "memorizing", "the", "change", "in", "scales", ".", "For", "the", "regression", "loss", ",", "all", "anchors", "with", "IoU", "greater", "than", "0.3", "against", "ground", "-", "truth", "faces", "will", "be", "taken", "into", "account", "and", "contribute", "to", "the", "smooth", "loss", ".", "We", "use", "a", "smaller", "threshold", "(", "0.3", ")", "because", "(", "1", ")", "this", "will", "allow", "imperfectly", "matched", "anchors", "to", "be", "able", "to", "localize", "the", "face", ",", "which", "may", "be", "useful", "during", "the", "testing", "and", "(", "2", ")", "the", "regression", "task", "has", "less", "supervision", "since", "unlike", "classification", ",", "there", "are", "no", "negative", "anchors", "for", "computing", "loss", "and", "the", "positive", "anchors", "are", "usually", "sparse", ".", "subsubsection", ":", "Anchor", "-", "level", "hard", "example", "mining", "OHEM", "has", "been", "proven", "to", "be", "useful", "for", "object", "detection", "and", "face", "detection", "in", ".", "During", "our", "training", ",", "in", "parallel", "with", "our", "newly", "proposed", "hard", "image", "mining", ",", "we", "also", "exploit", "the", "traditional", "hard", "anchor", "mining", "method", "to", "focus", "more", "on", "the", "hard", "and", "misclassificed", "anchors", ".", "Given", "a", "training", "image", "with", "size", ",", "there", "are", "anchors", "at", "the", "detection", "head", ",", "and", "we", "only", "select", "256", "of", "them", "to", "be", "involved", "in", "computing", "the", "classification", "loss", ".", "For", "all", "positive", "anchors", "with", "IoU", "greater", "than", "against", "ground", "-", "truth", "boxes", ",", "we", "select", "the", "top", "64", "of", "them", "with", "lowest", "confidences", "to", "be", "recognized", "as", "face", ".", "After", "selecting", "positive", "anchors", ",", "(", ")", "negative", "anchors", "with", "highest", "face", "confidence", "are", "selected", "to", "compute", "the", "classification", "loss", "as", "the", "hard", "negative", "anchors", ".", "Note", "that", "we", "only", "perform", "OHEM", "for", "classification", "loss", ",", "and", "we", "keep", "all", "anchors", "with", "IoU", "greater", "than", "0.3", "for", "computing", "regression", "loss", ",", "without", "selecting", "a", "subset", "based", "on", "either", "classification", "loss", "or", "bounding", "-", "box", "regression", "loss", ".", "subsubsection", ":", "Data", "augmentation", "Data", "augmentation", "is", "extremely", "useful", "to", "make", "the", "model", "robust", "to", "light", ",", "scale", "changes", "and", "small", "shifts", ".", "In", "our", "proposed", "method", ",", "we", "exploit", "cropping", "and", "photometric", "distortion", "as", "data", "augmentation", ".", "Given", "a", "training", "image", "after", "resizing", ",", "we", "crop", "a", "patch", "of", "it", "with", "a", "probability", "of", ".", "The", "patch", "has", "a", "height", "of", "and", "a", "width", "of", "which", "are", "independently", "drawn", "from", "and", ",", "where", "is", "the", "uniform", "distribution", "and", ",", "are", "the", "height", "and", "width", "of", "the", "resized", "training", "image", ".", "All", "ground", "-", "truth", "boxes", "whose", "centers", "are", "located", "inside", "the", "patch", "are", "kept", ".", "After", "the", "random", "cropping", ",", "we", "apply", "photometric", "distortion", "following", "SSD", "by", "randomly", "modifying", "the", "brightness", ",", "contrast", ",", "saturation", "and", "hue", "of", "the", "cropped", "image", "randomly", ".", "[", "b", "]", "0.33", "easy", "subset", "[", "b", "]", "0.33", "medium", "subset", "[", "b", "]", "0.33", "hard", "subset", "[", "b", "]", "0.33", "[", "b", "]", "0.33", "[", "b", "]", "0.33", "section", ":", "Experiments", "To", "verify", "the", "effectiveness", "of", "our", "model", "and", "proposed", "method", ",", "we", "conduct", "extensive", "experiments", "on", "popular", "face", "detection", "datasets", ",", "including", "WIDER", "FACE", ",", "FDDB", ",", "Pascal", "Faces", "and", "AFW", ".", "It", "is", "worth", "noting", "that", "the", "training", "is", "only", "performed", "on", "the", "train", "set", "of", "WIDER", "FACE", ",", "and", "we", "use", "the", "same", "model", "for", "evaluation", "on", "all", "these", "datasets", "without", "further", "fine", "-", "tuning", ".", "subsection", ":", "Experimental", "settings", "We", "train", "our", "model", "on", "the", "train", "set", "of", "WIDER", "FACE", ",", "which", "has", "12880", "images", "with", "159k", "faces", "annotated", ".", "We", "flip", "all", "images", "horizontally", ",", "to", "double", "the", "size", "of", "our", "training", "dataset", "to", "25760", ".", "For", "each", "training", "image", ",", "we", "first", "randomly", "resize", "it", ",", "and", "then", "we", "use", "the", "cropping", "and", "photometric", "distortion", "data", "augmentation", "methods", "discussed", "in", "Section", "[", "reference", "]", "to", "pre", "-", "process", "the", "resized", "image", ".", "We", "use", "an", "ImageNet", "pre", "-", "trained", "VGG16", "model", "to", "initialize", "our", "network", "backbone", ",", "and", "our", "newly", "introduced", "layers", "are", "randomly", "initialized", "with", "Gaussian", "initialization", ".", "We", "train", "the", "model", "with", "the", "itersize", "to", "be", "2", ",", "for", "46k", "iterations", ",", "with", "a", "learning", "rate", "of", ",", "and", "then", "for", "another", "14k", "iterations", "with", "a", "smaller", "learning", "rate", "of", ".", "During", "training", ",", "we", "use", "4", "GPUs", "to", "simultaneously", "to", "compute", "the", "gradient", "and", "update", "the", "weight", "by", "synchronized", "SGD", "with", "Momentum", ".", "The", "first", "two", "blocks", "of", "VGG16", "are", "frozen", "during", "the", "training", ",", "and", "the", "rest", "layers", "of", "VGG16", "are", "set", "to", "have", "a", "double", "learning", "rate", ".", "Since", "our", "model", "is", "designed", "and", "trained", "on", "only", "small", "faces", ",", "we", "use", "a", "multiscale", "image", "pyramid", "for", "testing", "to", "deal", "with", "faces", "larger", "than", "our", "anchors", ".", "Specifically", ",", "we", "resize", "the", "testing", "image", "so", "that", "the", "short", "side", "contains", "100", ",", "300", ",", "600", ",", "1000", "and", "1400", "pixels", "for", "evaluation", "on", "WIDER", "FACE", "dataset", ".", "We", "also", "follow", "the", "testing", "strategies", "used", "in", "PyramidBox", "such", "as", "horizontal", "flip", "and", "bounding", "-", "box", "voting", ".", "subsection", ":", "Experiment", "results", "WIDER", "FACE", "dataset", "includes", "3226", "images", "and", "39708", "faces", "labelled", "in", "the", "val", "dataset", ",", "with", "three", "subsets", "–", "easy", ",", "medium", "and", "hard", ".", "In", "Figure", "[", "reference", "]", ",", "we", "show", "the", "precision", "-", "recall", "(", "PR", ")", "curve", "and", "average", "precision", "(", "AP", ")", "for", "our", "model", "compared", "with", "many", "other", "state", "-", "of", "-", "the", "-", "arts", "on", "these", "three", "subsets", ".", "As", "we", "can", "see", ",", "our", "method", "achieves", "the", "best", "performance", "on", "the", "hard", "subset", ",", "and", "outperforms", "the", "current", "state", "-", "of", "-", "the", "-", "art", "by", "a", "large", "margin", ".", "Since", "the", "hard", "set", "is", "a", "super", "set", "of", "small", "and", "medium", ",", "which", "contains", "all", "faces", "taller", "than", "10", "pixels", ",", "the", "performance", "on", "hard", "set", "can", "represent", "the", "performance", "on", "the", "full", "testing", "dataset", "more", "accurately", ".", "Our", "performance", "on", "the", "medium", "subset", "is", "comparable", "to", "the", "most", "recent", "state", "-", "of", "-", "the", "-", "art", "and", "the", "performance", "on", "the", "easy", "subset", "is", "a", "bit", "worse", "since", "our", "method", "focuses", "on", "learning", "hard", "faces", ",", "and", "the", "architecture", "of", "our", "model", "is", "simpler", "compared", "with", "other", "state", "-", "of", "-", "the", "-", "arts", ".", "There", "is", "also", "a", "WIDER", "FACE", "test", "dataset", "with", "no", "annotations", "provided", "publicly", ".", "It", "contains", "16097", "images", ",", "and", "is", "evaluated", "by", "WIDER", "FACE", "author", "team", ".", "We", "report", "the", "performance", "of", "our", "method", "at", "Figure", "[", "reference", "]", "for", "the", "hard", "subset", ".", "FDDB", "dataset", "includes", "5171", "faces", "on", "a", "set", "of", "2845", "images", ",", "and", "we", "use", "our", "model", "trained", "on", "WIDER", "FACE", "train", "set", "to", "infer", "on", "the", "FDDB", "dataset", ".", "We", "use", "the", "raw", "bounding", "-", "box", "result", "without", "fitting", "it", "into", "ellipse", "to", "compute", "ROC", ".", "We", "show", "the", "discontinuous", "ROC", "curve", "at", "Figure", "[", "reference", "]", "compared", "with", ",", "and", "our", "method", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "of", "TPR=98.7", "%", "given", "1000", "false", "positives", ".", "Pascal", "Faces", "dataset", "includes", "1335", "labeled", "faces", "on", "a", "set", "of", "851", "images", "extracted", "for", "the", "Pascal", "VOC", "dataset", ".", "We", "show", "the", "PR", "curve", "at", "Figure", "[", "reference", "]", "compared", "with", ",", "and", "our", "method", "achieves", "a", "new", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "of", "AP=99.0", ".", "AFW", "dataset", "includes", "473", "faces", "labelled", "in", "a", "set", "of", "205", "images", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "compared", "with", ",", "our", "method", "achieves", "state", "-", "of", "-", "the", "-", "art", "and", "almost", "perfect", "performance", ",", "with", "an", "AP", "of", "99.60", ".", "subsection", ":", "Ablation", "study", "and", "diagnosis", "subsubsection", ":", "Ablation", "experiments", "In", "order", "to", "verify", "the", "performance", "of", "our", "single", "level", "face", "detector", ",", "as", "well", "as", "the", "effectiveness", "of", "our", "proposed", "hard", "image", "mining", ",", "the", "dilated", "-", "head", "classification", "and", "regression", "structure", ",", "we", "conduct", "various", "ablation", "experiments", "on", "the", "WIDER", "FACE", "val", "dataset", ".", "All", "results", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "From", "Table", "[", "reference", "]", ",", "we", "can", "see", "that", "our", "single", "level", "baseline", "model", "can", "achieve", "performance", "comparable", "to", "the", "current", "state", "-", "of", "-", "the", "-", "art", "face", "detector", ",", "especially", "on", "the", "hard", "subset", ".", "Our", "model", "with", "single", "detection", "feature", "map", "performs", "better", "than", "the", "one", "with", "three", "detection", "feature", "maps", ",", "despite", "its", "shallower", "structure", ",", "fewer", "parameters", "and", "anchors", ".", "This", "confirms", "the", "effectiveness", "of", "our", "simple", "face", "detector", "with", "single", "detection", "feature", "map", "focusing", "on", "small", "faces", ".", "We", "also", "separately", "verify", "our", "newly", "proposed", "hard", "image", "mining", "(", "HIM", ")", "and", "dilated", "head", "architecture", "(", "DH", ")", "described", "in", "Subsection", "[", "reference", "]", "and", "Figure", "[", "reference", "]", "respectively", ".", "HIM", "can", "improve", "the", "performance", "on", "hard", "subset", "significantly", "without", "involving", "more", "complex", "network", "architecture", "nor", "computation", "overhead", ".", "DH", "itself", "can", "also", "boost", "the", "performance", ",", "which", "shows", "the", "effectiveness", "of", "designing", "larger", "convolution", "for", "larger", "anchors", ".", "Combining", "HIM", "and", "DH", "together", "can", "improve", "further", "towards", "the", "state", "-", "of", "-", "the", "-", "art", "performance", ".", "subsubsection", ":", "Diagnosis", "of", "hard", "image", "mining", "We", "investigate", "the", "effects", "of", "our", "hard", "image", "mining", "mechanism", ".", "We", "show", "the", "ratio", "of", "and", "(", "the", "ratio", "of", "the", "number", "of", "selected", "training", "images", "to", "the", "number", "of", "ignored", "training", "images", ")", "in", "Figure", "[", "reference", "]", "for", "each", "epoch", ".", "We", "can", "see", "that", "at", "the", "first", "epoch", ",", "all", "training", "images", "are", "used", "to", "train", "the", "model", ".", "Meanwhile", ",", "as", "the", "training", "process", "continues", ",", "more", "and", "more", "training", "images", "will", "be", "ignored", ".", "At", "the", "last", "epoch", ",", "over", "a", "half", "images", "will", "be", "ignored", "and", "thus", "will", "not", "be", "included", "in", ".", "subsubsection", ":", "Diagnosis", "of", "data", "augmentation", "We", "investigate", "the", "effectiveness", "of", "the", "photometric", "distortion", "as", "well", "as", "the", "cropping", "mechanisms", "as", "discussed", "in", "Subsection", "[", "reference", "]", ".", "The", "ablation", "results", "evaluated", "on", "WIDER", "FACE", "val", "dataset", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Both", "photometric", "distortion", "and", "cropping", "can", "contribute", "to", "a", "more", "robust", "face", "detector", ".", "subsubsection", ":", "Diagnosis", "of", "multi", "-", "scale", "testing", "Our", "face", "detector", "with", "one", "detection", "feature", "map", "is", "design", "for", "small", "face", "detection", ",", "and", "our", "anchors", "are", "only", "capable", "of", "capturing", "faces", "with", "sizes", "ranging", "from", "(", ")", "to", "(", ")", ".", "As", "a", "result", ",", "it", "is", "critical", "to", "adopt", "multi", "-", "scale", "testing", "to", "deal", "with", "large", "faces", ".", "Different", "from", "SSH", ",", "S", "FD", "and", "PyramidBox", ",", "our", "testing", "pyramid", "includes", "some", "extreme", "small", "scales", "(", "short", "side", "contains", "only", "100", "or", "300", "pixels", ")", ".", "In", "Table", "[", "reference", "]", ",", "we", "show", "the", "effectiveness", "of", "these", "extreme", "small", "scales", "to", "deal", "with", "easy", "and", "large", "images", ".", "Our", "full", "evaluation", "resizes", "the", "image", "so", "that", "the", "short", "side", "contains", "100", ",", "300", ",", "600", ",", "1000", "and", "1400", "pixels", "respectively", ",", "to", "build", "an", "image", "pyramid", ".", "We", "diagnose", "the", "impact", "of", "the", "extra", "small", "scales", "(", "100", "and", "300", ")", "by", "removing", "them", "from", "the", "image", "pyramid", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "the", "extra", "small", "scales", "are", "crucial", "to", "detect", "easy", "faces", ".", "Without", "resizing", "the", "short", "side", "to", "contain", "100", "and", "300", "pixels", ",", "the", "performance", "on", "easy", "subset", "is", "only", ",", "which", "is", "even", "lower", "than", "the", "performance", "on", "medium", "and", "hard", "which", "contain", "much", "harder", "faces", ".", "We", "will", "show", "in", "the", "next", "subsection", "that", "these", "extra", "small", "scales", "(", "and", ")", "lead", "to", "negligible", "computation", "overhead", ",", "due", "to", "the", "lower", "resolution", ".", "subsubsection", ":", "Diagnosis", "of", "accuracy", "/", "speed", "trade", "-", "off", "We", "evaluate", "the", "speed", "of", "our", "method", "as", "well", "as", "some", "other", "popular", "face", "detectors", "in", "Table", "[", "reference", "]", ".", "For", "fair", "comparison", ",", "we", "run", "all", "methods", "on", "the", "same", "machine", ",", "with", "one", "Titan", "X", "(", "Maxwell", ")", "GPU", ",", "and", "Intel", "Core", "i7", "-", "4770", "K", "3.50GHz", ".", "All", "methods", "except", "for", "PyramidBox", "are", "based", "on", "Caffe1", "implementation", ",", "which", "is", "compiled", "with", "CUDA", "9.0", "and", "CUDNN", "7", ".", "For", "PyramidBox", ",", "we", "follow", "the", "official", "fluid", "code", "and", "the", "default", "configurations", ".", "We", "use", "the", "officially", "built", "PaddlePaddle", "with", "CUDA", "9.0", "and", "CUDNN", "7", ".", "For", "SSH", ",", "S", "FD", "and", "Pyramid", ",", "we", "use", "the", "official", "inference", "code", "and", "configurations", ".", "For", "SSH", ",", "we", "use", "multi", "-", "scale", "testing", "with", "the", "short", "side", "containing", "500", ",", "800", ",", "1200", "and", "1600", "pixels", ",", "and", "for", "S", "FD", ",", "we", "execute", "the", "official", "evaluation", "code", "with", "both", "multi", "-", "scale", "testing", "and", "horizontal", "flip", ".", "PyramidBox", "takes", "a", "similar", "testing", "configuration", "as", "S", "FD", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "our", "detector", "can", "outperform", "SSH", ",", "S", "FD", "and", "PyramidBox", "significantly", "with", "a", "smaller", "inference", "time", ".", "Based", "on", "that", ",", "using", "horizontal", "flip", "can", "further", "improve", "the", "performance", "slightly", ".", "In", "terms", "of", "GPU", "memory", "usage", ",", "our", "method", "uses", "only", "a", "half", "of", "what", "PyramidBox", "occupies", ",", "while", "achieving", "better", "performance", ".", "Ours", "in", "Table", "[", "reference", "]", "indicates", "our", "method", "without", "extra", "small", "scales", "in", "inference", ",", ",", "evaluated", "with", "scales", "[", "600", ",", "1000", ",", "1400", "]", ".", "It", "is", "only", "faster", "than", "evaluation", "with", "[", "100", ",", "300", ",", "600", ",", "1000", ",", "1400", "]", "(", "1.59", "compared", "with", "1.70", ")", ".", "This", "proves", "that", "although", "our", "face", "detector", "is", "only", "trained", "on", "small", "faces", ",", "it", "can", "perform", "well", "on", "large", "faces", ",", "by", "simply", "shrinking", "the", "testing", "image", "with", "negligible", "computation", "overhead", ".", "section", ":", "Conclusion", "To", "conclude", ",", "we", "propose", "a", "novel", "face", "detector", "to", "focus", "on", "learning", "small", "faces", "on", "hard", "images", ",", "which", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "all", "popular", "face", "detection", "datasets", ".", "We", "propose", "a", "hard", "image", "mining", "strategy", "by", "dynamically", "assigning", "difficulty", "scores", "to", "training", "images", ",", "and", "re", "-", "sampling", "subsets", "with", "hard", "images", "for", "training", "before", "each", "epoch", ".", "We", "also", "design", "a", "single", "shot", "face", "detector", "with", "only", "one", "detection", "feature", "map", ",", "to", "train", "and", "test", "on", "small", "faces", ".", "With", "these", "designs", ",", "our", "model", "can", "put", "more", "attention", "on", "learning", "small", "hard", "faces", "instead", "of", "memorizing", "change", "of", "scales", ".", "Extensive", "experiments", "and", "ablations", "have", "been", "done", "to", "show", "the", "effectiveness", "of", "our", "method", ",", "and", "our", "face", "detector", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "all", "popular", "face", "detection", "datasets", ",", "including", "WIDER", "FACE", ",", "FDDB", ",", "Pascal", "Faces", "and", "AFW", ".", "Our", "face", "detector", "also", "enjoys", "faster", "multi", "-", "scale", "inference", "speed", "and", "less", "GPU", "memory", "usage", ".", "Our", "proposed", "method", "are", "flexible", "and", "can", "be", "applied", "to", "other", "backbones", "and", "tasks", ",", "which", "we", "remain", "as", "future", "work", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["Robust", "Face", "Detection"], "Task"], [6, 9, ["Learning", "Small", "Faces"], "Task"], [10, 12, ["Hard", "Images"], "Material"], [13, 19, ["anchor", "-", "based", "deep", "face", "detectors"], "Method"], [83, 85, ["hard", "images"], "Material"], [94, 95, ["robustness"], "Metric"], [97, 99, ["face", "detector"], "Method"], [110, 112, ["hard", "images"], "Material"], [119, 121, ["hard", "images"], "Material"], [137, 139, ["face", "detectors"], "Method"], [168, 174, ["anchor", "-", "based", "deep", "face", "detector"], "Method"], [197, 201, ["hard", "image", "mining", "strategy"], "Method"], [208, 210, ["WIDER", "FACE"], "Material"], [211, 212, ["FDDB"], "Material"], [213, 215, ["Pascal", "Faces"], "Material"], [217, 219, ["AFW", "datasets"], "Material"], [230, 231, ["APs"], "Metric"], [244, 247, ["FACE", "val", "dataset"], "Material"], [277, 279, ["Face", "detection"], "Task"], [284, 287, ["computer", "vision", "problem"], "Task"], [293, 297, ["face", "-", "related", "tasks"], "Task"], [300, 302, ["face", "alignment"], "Task"], [303, 306, ["tracking", "and", "recognition"], "Task"], [314, 317, ["deep", "neural", "networks"], "Method"], [318, 325, ["massive", "CNN", "-", "based", "face", "detection", "approaches"], "Method"], [342, 344, ["face", "detection"], "Task"], [368, 372, ["benchmark", "dataset", "WIDER", "FACE"], "Material"], [381, 386, ["CNN", "-", "based", "face", "detectors"], "Method"], [395, 397, ["backbone", "models"], "Method"], [399, 404, ["feature", "pyramid", "-", "style", "architectures"], "Method"], [435, 438, ["robust", "face", "detector"], "Method"], [469, 471, ["hard", "images"], "Material"], [485, 487, ["hard", "images"], "Material"], [493, 495, ["face", "detectors"], "Task"], [497, 501, ["human", "-", "level", "detection"], "Task"], [519, 521, ["WIDER", "FACE"], "Material"], [527, 528, ["SSH"], "Method"], [611, 613, ["detection", "accuracy"], "Metric"], [626, 629, ["robust", "face", "detector"], "Method"], [642, 645, ["robust", "face", "detector"], "Method"], [661, 667, ["anchor", "-", "level", "hard", "example", "mining"], "Task"], [669, 670, ["OHEM"], "Method"], [687, 693, ["anchor", "-", "level", "hard", "example", "mining"], "Method"], [764, 766, ["learning", "process"], "Task"], [803, 805, ["learning", "process"], "Task"], [834, 836, ["hard", "images"], "Material"], [843, 845, ["detection", "quality"], "Metric"], [869, 874, ["scale", "-", "invariant", "face", "detector"], "Method"], [934, 936, ["anchor", "design"], "Method"], [970, 972, ["shallow", "architecture"], "Method"], [984, 986, ["face", "detector"], "Method"], [994, 998, ["hard", "image", "mining", "strategy"], "Method"], [1002, 1003, ["robustness"], "Metric"], [1033, 1036, ["single", "shot", "detector"], "Method"], [1039, 1042, ["detection", "feature", "map"], "Method"], [1075, 1077, ["face", "detector"], "Method"], [1094, 1096, ["WIDER", "FACE"], "Material"], [1097, 1098, ["FDDB"], "Material"], [1099, 1101, ["Pascal", "Faces"], "Material"], [1103, 1104, ["AFW"], "Material"], [1119, 1122, ["FACE", "val", "dataset"], "Material"], [1127, 1128, ["APs"], "Metric"], [1133, 1135, ["Pascal", "Faces"], "Material"], [1143, 1144, ["TPR"], "Metric"], [1147, 1148, ["FDDB"], "Material"], [1223, 1225, ["Face", "detection"], "Task"], [1236, 1240, ["CNN", "and", "object", "detector"], "Method"], [1244, 1246, ["face", "detectors"], "Method"], [1259, 1262, ["face", "detection", "domain"], "Task"], [1266, 1269, ["hard", "example", "mining"], "Task"], [1270, 1273, ["face", "detection", "architecture"], "Method"], [1275, 1279, ["anchor", "design", "&", "matching"], "Task"], [1282, 1285, ["Hard", "example", "mining"], "Task"], [1285, 1288, ["Hard", "example", "mining"], "Method"], [1294, 1296, ["model", "quality"], "Metric"], [1303, 1305, ["image", "classification"], "Task"], [1342, 1344, ["detection", "frameworks"], "Method"], [1350, 1351, ["OHEM"], "Method"], [1362, 1363, ["OHEM"], "Method"], [1401, 1404, ["cross", "entropy", "criterion"], "Metric"], [1412, 1420, ["anchor", "-", "level", "or", "proposal", "-", "level", "mining"], "Task"], [1430, 1432, ["hard", "images"], "Material"], [1444, 1447, ["hard", "example", "mining"], "Method"], [1448, 1450, ["image", "level"], "Task"], [1451, 1454, ["hard", "image", "mining"], "Task"], [1460, 1462, ["face", "detector"], "Task"], [1481, 1483, ["SGD", "mechanism"], "Method"], [1504, 1507, ["Face", "Detection", "Architecture"], "Method"], [1515, 1517, ["face", "detectors"], "Method"], [1522, 1525, ["Faster", "-", "RCNN"], "Method"], [1526, 1529, ["R", "-", "FCN"], "Method"], [1530, 1531, ["SSD"], "Method"], [1532, 1533, ["SSH"], "Method"], [1535, 1540, ["RPN", "(", "Region", "Proposal", "Network"], "Method"], [1542, 1545, ["Faster", "-", "RCNN"], "Method"], [1552, 1555, ["detection", "feature", "maps"], "Method"], [1569, 1571, ["S", "FD"], "Method"], [1572, 1573, ["PyramidBox"], "Method"], [1580, 1581, ["SSD"], "Method"], [1583, 1585, ["detection", "architecture"], "Method"], [1594, 1596, ["S", "FD"], "Method"], [1597, 1598, ["PyramidBox"], "Method"], [1633, 1634, ["VGG16"], "Method"], [1636, 1637, ["classification"], "Task"], [1638, 1642, ["bounding", "-", "box", "regression"], "Task"], [1652, 1654, ["Anchor", "design"], "Method"], [1655, 1656, ["matching"], "Task"], [1677, 1681, ["scale", "-", "invariant", "detector"], "Method"], [1682, 1683, ["SSD"], "Method"], [1690, 1693, ["detectors", "S", "FD"], "Method"], [1694, 1695, ["PyramidBox"], "Method"], [1715, 1717, ["network", "architectures"], "Method"], [1721, 1724, ["detection", "feature", "maps"], "Method"], [1735, 1736, ["SSH"], "Method"], [1763, 1764, ["SSH"], "Method"], [1765, 1767, ["S", "DF"], "Method"], [1770, 1771, ["SSH"], "Method"], [1786, 1788, ["S", "DF"], "Method"], [1801, 1802, ["SNIP"], "Method"], [1813, 1814, ["CNNs"], "Method"], [1894, 1895, ["testing"], "Task"], [1929, 1931, ["multiscale", "training"], "Method"], [1979, 1982, ["effective", "face", "detection"], "Task"], [2001, 2005, ["hard", "image", "mining", "strategy"], "Method"], [2017, 2019, ["training", "techniques"], "Method"], [2027, 2034, ["Single", "-", "level", "small", "face", "detection", "framework"], "Method"], [2038, 2040, ["face", "detector"], "Method"], [2050, 2052, ["VGG16", "network"], "Method"], [2054, 2056, ["backbone", "CNN"], "Method"], [2067, 2070, ["detection", "feature", "map"], "Method"], [2084, 2085, ["SSH"], "Method"], [2089, 2092, ["1", "convolution", "layers"], "Method"], [2093, 2094, ["conv4_3"], "Method"], [2095, 2096, ["conv5_3"], "Method"], [2105, 2108, ["3", "convolution", "layer"], "Method"], [2123, 2127, ["3", "3", "convolution", "layer"], "Method"], [2140, 2142, ["detection", "head"], "Method"], [2143, 2144, ["classification"], "Task"], [2145, 2149, ["bounding", "-", "box", "regression"], "Task"], [2151, 2154, ["detection", "feature", "map"], "Method"], [2186, 2187, ["classification"], "Method"], [2188, 2192, ["bounding", "-", "box", "regression"], "Method"], [2199, 2201, ["face", "detectors"], "Method"], [2218, 2219, ["SNIP"], "Method"], [2235, 2238, ["detection", "feature", "map"], "Method"], [2315, 2317, ["accuracy", "degradation"], "Metric"], [2318, 2320, ["large", "faces"], "Task"], [2373, 2375, ["detection", "head"], "Method"], [2458, 2459, ["convolution"], "Method"], [2479, 2481, ["convolution", "layers"], "Method"], [2497, 2498, ["convolution"], "Method"], [2525, 2529, ["multi", "-", "dilation", "design"], "Method"], [2532, 2535, ["Hard", "image", "mining"], "Task"], [2537, 2538, ["OHEM"], "Method"], [2559, 2563, ["hard", "image", "mining", "strategy"], "Method"], [2587, 2588, ["AP"], "Metric"], [2595, 2599, ["WIDER", "FACE", "val", "dataset"], "Material"], [2628, 2630, ["face", "detector"], "Method"], [2702, 2707, ["ImageNet", "pre", "-", "trained", "model"], "Method"], [2713, 2715, ["face", "detection"], "Task"], [2769, 2774, ["metric", "Worst", "Positive", "Anchor", "Score"], "Metric"], [2775, 2776, ["WPAS"], "Metric"], [2895, 2897, ["hard", "images"], "Material"], [2910, 2914, ["multi", "-", "GPU", "training"], "Task"], [2949, 2953, ["Multi", "-", "scale", "training"], "Task"], [2954, 2956, ["anchor", "matching"], "Task"], [3125, 3127, ["classification", "loss"], "Metric"], [3175, 3176, ["RCNN"], "Method"], [3179, 3182, ["anchor", "matching", "strategy"], "Method"], [3214, 3216, ["regression", "loss"], "Task"], [3238, 3240, ["smooth", "loss"], "Metric"], [3279, 3281, ["regression", "task"], "Task"], [3286, 3287, ["classification"], "Task"], [3310, 3312, ["example", "mining"], "Method"], [3312, 3313, ["OHEM"], "Method"], [3320, 3322, ["object", "detection"], "Task"], [3323, 3325, ["face", "detection"], "Task"], [3337, 3340, ["hard", "image", "mining"], "Task"], [3346, 3350, ["hard", "anchor", "mining", "method"], "Method"], [3388, 3390, ["classification", "loss"], "Task"], [3439, 3441, ["classification", "loss"], "Metric"], [3452, 3453, ["OHEM"], "Method"], [3454, 3456, ["classification", "loss"], "Task"], [3469, 3471, ["regression", "loss"], "Task"], [3479, 3481, ["classification", "loss"], "Method"], [3482, 3487, ["bounding", "-", "box", "regression", "loss"], "Method"], [3490, 3492, ["Data", "augmentation"], "Task"], [3492, 3494, ["Data", "augmentation"], "Task"], [3523, 3525, ["data", "augmentation"], "Task"], [3596, 3598, ["random", "cropping"], "Task"], [3601, 3603, ["photometric", "distortion"], "Method"], [3604, 3605, ["SSD"], "Method"], [3677, 3679, ["WIDER", "FACE"], "Material"], [3680, 3681, ["FDDB"], "Material"], [3682, 3684, ["Pascal", "Faces"], "Material"], [3685, 3686, ["AFW"], "Material"], [3702, 3704, ["WIDER", "FACE"], "Material"], [3719, 3722, ["fine", "-", "tuning"], "Task"], [3736, 3738, ["WIDER", "FACE"], "Material"], [3781, 3788, ["cropping", "and", "photometric", "distortion", "data", "augmentation", "methods"], "Method"], [3805, 3811, ["ImageNet", "pre", "-", "trained", "VGG16", "model"], "Method"], [3814, 3816, ["network", "backbone"], "Method"], [3826, 3828, ["Gaussian", "initialization"], "Method"], [3846, 3848, ["learning", "rate"], "Metric"], [3859, 3861, ["learning", "rate"], "Metric"], [3864, 3865, ["training"], "Task"], [3881, 3883, ["synchronized", "SGD"], "Method"], [3884, 3885, ["Momentum"], "Method"], [3891, 3892, ["VGG16"], "Method"], [3903, 3904, ["VGG16"], "Method"], [3910, 3912, ["learning", "rate"], "Metric"], [3968, 3971, ["WIDER", "FACE", "dataset"], "Material"], [3976, 3978, ["testing", "strategies"], "Method"], [3980, 3981, ["PyramidBox"], "Method"], [3986, 3990, ["bounding", "-", "box", "voting"], "Method"], [3995, 3998, ["WIDER", "FACE", "dataset"], "Material"], [4029, 4032, ["precision", "-", "recall"], "Metric"], [4033, 4034, ["PR"], "Metric"], [4037, 4039, ["average", "precision"], "Metric"], [4040, 4041, ["AP"], "Metric"], [4095, 4097, ["hard", "set"], "Material"], [4167, 4170, ["learning", "hard", "faces"], "Task"], [4194, 4198, ["WIDER", "FACE", "test", "dataset"], "Material"], [4213, 4217, ["WIDER", "FACE", "author", "team"], "Material"], [4254, 4258, ["WIDER", "FACE", "train", "set"], "Material"], [4310, 4311, ["TPR=98.7"], "Metric"], [4314, 4316, ["false", "positives"], "Metric"], [4333, 4336, ["Pascal", "VOC", "dataset"], "Material"], [4366, 4367, ["AP=99.0"], "Metric"], [4368, 4370, ["AFW", "dataset"], "Material"], [4408, 4409, ["AP"], "Metric"], [4414, 4416, ["Ablation", "study"], "Task"], [4417, 4418, ["diagnosis"], "Task"], [4420, 4421, ["Ablation"], "Task"], [4430, 4434, ["single", "level", "face", "detector"], "Method"], [4443, 4446, ["hard", "image", "mining"], "Method"], [4448, 4452, ["dilated", "-", "head", "classification"], "Method"], [4453, 4455, ["regression", "structure"], "Method"], [4463, 4467, ["WIDER", "FACE", "val", "dataset"], "Material"], [4489, 4493, ["single", "level", "baseline", "model"], "Method"], [4507, 4509, ["face", "detector"], "Method"], [4513, 4515, ["hard", "subset"], "Material"], [4519, 4523, ["single", "detection", "feature", "map"], "Method"], [4530, 4533, ["detection", "feature", "maps"], "Method"], [4551, 4553, ["face", "detector"], "Method"], [4554, 4558, ["single", "detection", "feature", "map"], "Method"], [4570, 4576, ["hard", "image", "mining", "(", "HIM", ")"], "Method"], [4577, 4580, ["dilated", "head", "architecture"], "Method"], [4581, 4582, ["DH"], "Method"], [4596, 4597, ["HIM"], "Method"], [4612, 4614, ["computation", "overhead"], "Metric"], [4615, 4616, ["DH"], "Method"], [4630, 4631, ["convolution"], "Method"], [4636, 4637, ["HIM"], "Method"], [4638, 4639, ["DH"], "Method"], [4658, 4661, ["hard", "image", "mining"], "Task"], [4667, 4671, ["hard", "image", "mining", "mechanism"], "Method"], [4765, 4767, ["data", "augmentation"], "Task"], [4779, 4781, ["cropping", "mechanisms"], "Method"], [4794, 4798, ["WIDER", "FACE", "val", "dataset"], "Material"], [4810, 4811, ["cropping"], "Method"], [4817, 4819, ["face", "detector"], "Task"], [4824, 4828, ["multi", "-", "scale", "testing"], "Task"], [4829, 4831, ["face", "detector"], "Method"], [4833, 4836, ["detection", "feature", "map"], "Method"], [4839, 4842, ["small", "face", "detection"], "Task"], [4845, 4846, ["anchors"], "Method"], [4871, 4875, ["multi", "-", "scale", "testing"], "Task"], [4883, 4884, ["SSH"], "Method"], [4885, 4887, ["S", "FD"], "Method"], [4888, 4889, ["PyramidBox"], "Method"], [4892, 4893, ["pyramid"], "Method"], [5058, 5060, ["computation", "overhead"], "Metric"], [5071, 5077, ["accuracy", "/", "speed", "trade", "-", "off"], "Metric"], [5090, 5092, ["face", "detectors"], "Method"], [5133, 5134, ["PyramidBox"], "Method"], [5137, 5139, ["Caffe1", "implementation"], "Method"], [5147, 5149, ["CUDNN", "7"], "Method"], [5151, 5152, ["PyramidBox"], "Task"], [5157, 5159, ["fluid", "code"], "Method"], [5169, 5170, ["PaddlePaddle"], "Method"], [5171, 5173, ["CUDA", "9.0"], "Method"], [5174, 5176, ["CUDNN", "7"], "Method"], [5178, 5179, ["SSH"], "Method"], [5180, 5182, ["S", "FD"], "Method"], [5183, 5184, ["Pyramid"], "Method"], [5188, 5191, ["official", "inference", "code"], "Method"], [5195, 5196, ["SSH"], "Method"], [5199, 5203, ["multi", "-", "scale", "testing"], "Method"], [5219, 5221, ["S", "FD"], "Method"], [5230, 5234, ["multi", "-", "scale", "testing"], "Method"], [5238, 5239, ["PyramidBox"], "Method"], [5260, 5261, ["SSH"], "Method"], [5262, 5264, ["S", "FD"], "Method"], [5265, 5266, ["PyramidBox"], "Method"], [5270, 5272, ["inference", "time"], "Metric"], [5290, 5293, ["GPU", "memory", "usage"], "Metric"], [5324, 5325, ["inference"], "Task"], [5368, 5370, ["face", "detector"], "Method"], [5393, 5395, ["computation", "overhead"], "Metric"], [5406, 5408, ["face", "detector"], "Method"], [5412, 5414, ["small", "faces"], "Task"], [5415, 5417, ["hard", "images"], "Material"], [5439, 5443, ["hard", "image", "mining", "strategy"], "Method"], [5470, 5474, ["single", "shot", "face", "detector"], "Method"], [5477, 5480, ["detection", "feature", "map"], "Method"], [5500, 5504, ["learning", "small", "hard", "faces"], "Task"], [5528, 5530, ["face", "detector"], "Method"], [5548, 5550, ["WIDER", "FACE"], "Material"], [5551, 5552, ["FDDB"], "Material"], [5553, 5555, ["Pascal", "Faces"], "Material"], [5556, 5557, ["AFW"], "Material"], [5559, 5561, ["face", "detector"], "Method"], [5564, 5569, ["multi", "-", "scale", "inference", "speed"], "Metric"], [5571, 5574, ["GPU", "memory", "usage"], "Metric"], [1255, 1257, ["object", "detection"], "Task"], [1307, 1309, ["object", "detection"], "Task"], [2830, 2831, ["WPAS"], "Metric"], [4235, 4236, ["FDDB"], "Material"], [4262, 4263, ["FDDB"], "Material"], [4317, 4319, ["Pascal", "Faces"], "Material"], [4340, 4341, ["PR"], "Metric"]]} {"doc_id": "25f5df29342a04936ba0d308b4d1b8245a7e8f5c", "text": ["Convolutional", "Pose", "Machines", "section", ":", "Abstract", "Pose", "Machines", "provide", "a", "sequential", "prediction", "framework", "for", "learning", "rich", "implicit", "spatial", "models", ".", "In", "this", "work", "we", "show", "a", "systematic", "design", "for", "how", "convolutional", "networks", "can", "be", "incorporated", "into", "the", "pose", "machine", "framework", "for", "learning", "image", "features", "and", "image", "-", "dependent", "spatial", "models", "for", "the", "task", "of", "pose", "estimation", ".", "The", "contribution", "of", "this", "paper", "is", "to", "implicitly", "model", "long", "-", "range", "dependencies", "between", "variables", "in", "structured", "prediction", "tasks", "such", "as", "articulated", "pose", "estimation", ".", "We", "achieve", "this", "by", "designing", "a", "sequential", "architecture", "composed", "of", "convolutional", "networks", "that", "directly", "operate", "on", "belief", "maps", "from", "previous", "stages", ",", "producing", "increasingly", "refined", "estimates", "for", "part", "locations", ",", "without", "the", "need", "for", "explicit", "graphical", "model", "-", "style", "inference", ".", "Our", "approach", "addresses", "the", "characteristic", "difficulty", "of", "vanishing", "gradients", "during", "training", "by", "providing", "a", "natural", "learning", "objective", "function", "that", "enforces", "intermediate", "supervision", ",", "thereby", "replenishing", "back", "-", "propagated", "gradients", "and", "conditioning", "the", "learning", "procedure", ".", "We", "demonstrate", "state", "-", "of", "-", "the", "-", "art", "performance", "and", "outperform", "competing", "methods", "on", "standard", "benchmarks", "including", "the", "MPII", ",", "LSP", ",", "and", "FLIC", "datasets", ".", "section", ":", "Introduction", "We", "introduce", "Convolutional", "Pose", "Machines", "(", "CPMs", ")", "for", "the", "task", "of", "articulated", "pose", "estimation", ".", "CPMs", "inherit", "the", "benefits", "of", "the", "pose", "machine", "[", "reference", "]", "architecture", "-", "the", "implicit", "learning", "of", "long", "-", "range", "dependencies", "between", "image", "and", "multi", "-", "part", "cues", ",", "tight", "integration", "between", "learning", "and", "inference", ",", "a", "modular", "sequential", "design", "-", "and", "combine", "them", "with", "the", "advantages", "afforded", "by", "convolutional", "architectures", ":", "the", "ability", "to", "learn", "feature", "representations", "for", "both", "image", "and", "spatial", "context", "directly", "from", "data", ";", "a", "differentiable", "architecture", "that", "allows", "for", "globally", "joint", "training", "with", "backpropagation", ";", "and", "the", "ability", "to", "efficiently", "handle", "large", "training", "datasets", ".", "CPMs", "consist", "of", "a", "sequence", "of", "convolutional", "networks", "that", "repeatedly", "produce", "2D", "belief", "maps", "[", "reference", "]", "for", "the", "location", "[", "reference", "]", "We", "use", "the", "term", "belief", "in", "a", "slightly", "loose", "sense", ",", "however", "the", "belief", "of", "each", "part", ".", "At", "each", "stage", "in", "a", "CPM", ",", "image", "features", "and", "the", "belief", "maps", "produced", "by", "the", "previous", "stage", "are", "used", "as", "input", ".", "The", "belief", "maps", "provide", "the", "subsequent", "stage", "an", "expressive", "non", "-", "parametric", "encoding", "of", "the", "spatial", "uncertainty", "of", "location", "for", "each", "part", ",", "allowing", "the", "CPM", "to", "learn", "rich", "image", "-", "dependent", "spatial", "models", "of", "the", "relationships", "between", "parts", ".", "Instead", "of", "explicitly", "parsing", "such", "belief", "maps", "either", "using", "graphical", "models", "[", "reference", "][", "reference", "][", "reference", "]", "or", "specialized", "post", "-", "processing", "steps", "[", "reference", "][", "reference", "]", ",", "we", "learn", "convolutional", "networks", "that", "directly", "operate", "on", "intermediate", "belief", "maps", "and", "learn", "implicit", "image", "-", "dependent", "spatial", "models", "of", "the", "relationships", "between", "parts", ".", "The", "overall", "proposed", "multistage", "architecture", "is", "fully", "differentiable", "and", "therefore", "can", "be", "trained", "in", "an", "end", "-", "to", "-", "end", "fashion", "using", "backpropagation", ".", "At", "a", "particular", "stage", "in", "the", "CPM", ",", "the", "spatial", "context", "of", "part", "beliefs", "provide", "strong", "disambiguating", "cues", "to", "a", "subsequent", "stage", ".", "As", "a", "result", ",", "each", "stage", "of", "a", "CPM", "produces", "belief", "maps", "with", "increasingly", "refined", "estimates", "for", "the", "locations", "of", "each", "part", "(", "see", "Figure", "1", ")", ".", "In", "order", "to", "capture", "longrange", "interactions", "between", "parts", ",", "the", "design", "of", "the", "network", "in", "each", "stage", "of", "our", "sequential", "prediction", "framework", "is", "motivated", "by", "the", "goal", "of", "achieving", "a", "large", "receptive", "field", "on", "both", "the", "image", "and", "the", "belief", "maps", ".", "We", "find", ",", "through", "experiments", ",", "that", "large", "receptive", "fields", "on", "the", "belief", "maps", "are", "crucial", "for", "learning", "long", "range", "spatial", "relationships", "and", "remaps", "described", "are", "closely", "related", "to", "beliefs", "produced", "in", "message", "passing", "inference", "in", "graphical", "models", ".", "The", "overall", "architecture", "can", "be", "viewed", "as", "an", "unrolled", "mean", "-", "field", "message", "passing", "inference", "algorithm", "[", "reference", "]", "that", "is", "learned", "end", "-", "to", "-", "end", "using", "backpropagation", ".", "sult", "in", "improved", "accuracy", ".", "Composing", "multiple", "convolutional", "networks", "in", "a", "CPM", "results", "in", "an", "overall", "network", "with", "many", "layers", "that", "is", "at", "risk", "of", "the", "problem", "of", "vanishing", "gradients", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "during", "learning", ".", "This", "problem", "can", "occur", "because", "backpropagated", "gradients", "diminish", "in", "strength", "as", "they", "are", "propagated", "through", "the", "many", "layers", "of", "the", "network", ".", "While", "there", "exists", "recent", "work", "2", "which", "shows", "that", "supervising", "very", "deep", "networks", "at", "intermediate", "layers", "aids", "in", "learning", "[", "reference", "][", "reference", "]", ",", "they", "have", "mostly", "been", "restricted", "to", "classification", "problems", ".", "In", "this", "work", ",", "we", "show", "how", "for", "a", "structured", "prediction", "problem", "such", "as", "pose", "estimation", ",", "CPMs", "naturally", "suggest", "a", "systematic", "framework", "that", "replenishes", "gradients", "and", "guides", "the", "network", "to", "produce", "increasingly", "accurate", "belief", "maps", "by", "enforcing", "intermediate", "supervision", "periodically", "through", "the", "network", ".", "We", "also", "discuss", "different", "training", "schemes", "of", "such", "a", "sequential", "prediction", "architecture", ".", "Our", "main", "contributions", "are", "(", "a", ")", "learning", "implicit", "spatial", "models", "via", "a", "sequential", "composition", "of", "convolutional", "architectures", "and", "(", "b", ")", "a", "systematic", "approach", "to", "designing", "and", "training", "such", "an", "architecture", "to", "learn", "both", "image", "features", "and", "image", "-", "dependent", "spatial", "models", "for", "structured", "prediction", "tasks", ",", "without", "the", "need", "for", "any", "graphical", "model", "style", "inference", ".", "We", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", "on", "standard", "benchmarks", "including", "the", "MPII", ",", "LSP", ",", "and", "FLIC", "datasets", ",", "and", "analyze", "the", "effects", "of", "jointly", "training", "a", "multi", "-", "staged", "architecture", "with", "repeated", "intermediate", "supervision", ".", "section", ":", "Related", "Work", "The", "classical", "approach", "to", "articulated", "pose", "estimation", "is", "the", "pictorial", "structures", "model", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "in", "which", "spatial", "correlations", "between", "parts", "of", "the", "body", "are", "expressed", "as", "a", "tree", "-", "structured", "graphical", "model", "with", "kinematic", "priors", "that", "couple", "connected", "limbs", ".", "These", "methods", "have", "been", "successful", "on", "images", "where", "all", "the", "limbs", "of", "the", "person", "are", "visible", ",", "but", "are", "prone", "to", "characteristic", "errors", "such", "as", "double", "-", "counting", "image", "evidence", ",", "which", "occur", "because", "of", "correlations", "between", "variables", "that", "are", "not", "captured", "by", "a", "tree", "-", "structured", "model", ".", "The", "work", "of", "Kiefel", "et", "al", ".", "[", "reference", "]", "is", "based", "on", "the", "pictorial", "structures", "model", "but", "differs", "in", "the", "underlying", "graph", "representation", ".", "Hierarchical", "models", "[", "reference", "][", "reference", "]", "represent", "the", "relationships", "between", "parts", "at", "different", "scales", "and", "sizes", "in", "a", "hierarchical", "tree", "structure", ".", "The", "underlying", "assumption", "of", "these", "models", "is", "that", "larger", "parts", "(", "that", "correspond", "to", "full", "limbs", "instead", "of", "joints", ")", "can", "often", "have", "discriminative", "image", "structure", "that", "can", "be", "easier", "to", "detect", "and", "consequently", "help", "reason", "about", "the", "location", "of", "smaller", ",", "harder", "-", "to", "-", "detect", "parts", ".", "Non", "-", "tree", "models", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "incorporate", "interactions", "that", "introduce", "loops", "to", "augment", "the", "tree", "structure", "with", "additional", "edges", "that", "capture", "symmetry", ",", "occlusion", "and", "long", "-", "range", "relation", "-", "ships", ".", "These", "methods", "usually", "have", "to", "rely", "on", "approximate", "inference", "during", "both", "learning", "and", "at", "test", "time", ",", "and", "therefore", "have", "to", "trade", "off", "accurate", "modeling", "of", "spatial", "relationships", "with", "models", "that", "allow", "efficient", "inference", ",", "often", "with", "a", "simple", "parametric", "form", "to", "allow", "for", "fast", "inference", ".", "In", "contrast", ",", "methods", "based", "on", "a", "sequential", "prediction", "framework", "[", "reference", "]", "learn", "an", "implicit", "spatial", "model", "with", "potentially", "complex", "interactions", "between", "variables", "by", "directly", "training", "an", "inference", "procedure", ",", "as", "in", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "There", "has", "been", "a", "recent", "surge", "of", "interest", "in", "models", "that", "employ", "convolutional", "architectures", "for", "the", "task", "of", "articulated", "pose", "estimation", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "Toshev", "et", "al", ".", "[", "reference", "]", "take", "the", "approach", "of", "directly", "regressing", "the", "Cartesian", "coordinates", "using", "a", "standard", "convolutional", "architecture", "[", "reference", "]", ".", "Recent", "work", "regresses", "image", "to", "confidence", "maps", ",", "and", "resort", "to", "graphical", "models", ",", "which", "require", "hand", "-", "designed", "energy", "functions", "or", "heuristic", "initialization", "of", "spatial", "probability", "priors", ",", "to", "remove", "outliers", "on", "the", "regressed", "confidence", "maps", ".", "Some", "of", "them", "also", "utilize", "a", "dedicated", "network", "module", "for", "precision", "refinement", "[", "reference", "][", "reference", "]", ".", "In", "this", "work", ",", "we", "show", "the", "regressed", "confidence", "maps", "are", "suitable", "to", "be", "inputted", "to", "further", "convolutional", "networks", "with", "large", "receptive", "fields", "to", "learn", "implicit", "spatial", "dependencies", "without", "the", "use", "of", "hand", "designed", "priors", ",", "and", "achieve", "state", "-", "of", "-", "the", "-", "art", "performance", "over", "all", "precision", "region", "without", "careful", "initialization", "and", "dedicated", "precision", "refinement", ".", "Pfister", "et", "al", ".", "[", "reference", "]", "also", "used", "a", "network", "module", "with", "large", "receptive", "field", "to", "capture", "implicit", "spatial", "models", ".", "Due", "to", "the", "differentiable", "nature", "of", "convolutions", ",", "our", "model", "can", "be", "globally", "trained", ",", "where", "Tompson", "et", "al", ".", "[", "reference", "]", "and", "Steward", "et", "al", ".", "[", "reference", "]", "also", "discussed", "the", "benefit", "of", "joint", "training", ".", "Carreira", "et", "al", ".", "[", "reference", "]", "train", "a", "deep", "network", "that", "iteratively", "improves", "part", "detections", "using", "error", "feedback", "but", "use", "a", "cartesian", "representation", "as", "in", "[", "reference", "]", "which", "does", "not", "preserve", "spatial", "uncertainty", "and", "results", "in", "lower", "accuracy", "in", "the", "highprecision", "regime", ".", "In", "this", "work", ",", "we", "show", "how", "the", "sequential", "prediction", "framework", "takes", "advantage", "of", "the", "preserved", "uncertainty", "in", "the", "confidence", "maps", "to", "encode", "the", "rich", "spatial", "context", ",", "with", "enforcing", "the", "intermediate", "local", "supervisions", "to", "address", "the", "problem", "of", "vanishing", "gradients", ".", "section", ":", "Method", "section", ":", "Pose", "Machines", "We", "denote", "the", "pixel", "location", "of", "the", "p", "-", "th", "anatomical", "landmark", "(", "which", "we", "refer", "to", "as", "a", "part", ")", ",", "Y", "p", "∈", "Z", "⊂", "R", "2", ",", "where", "Z", "is", "the", "set", "of", "all", "(", "u", ",", "v", ")", "locations", "in", "an", "image", ".", "Our", "goal", "is", "to", "predict", "the", "image", "locations", "Y", "=", "(", "Y", "1", ",", ".", ".", ".", ",", "Y", "P", ")", "for", "all", "P", "parts", ".", "A", "pose", "machine", "[", "reference", "]", "(", "see", "Figure", "2a", "and", "2b", ")", "consists", "of", "a", "sequence", "of", "multi", "-", "class", "predictors", ",", "g", "t", "(", "·", ")", ",", "that", "are", "trained", "to", "predict", "the", "location", "of", "each", "part", "in", "each", "level", "of", "the", "hierarchy", ".", "In", "each", "stage", "t", "∈", "{", "1", ".", ".", ".", "T", "}", ",", "the", "classifiers", "g", "t", "predict", "beliefs", "for", "assigning", "a", "location", "to", "each", "part", "Y", "p", "=", "z", ",", "∀z", "∈", "Z", ",", "based", "on", "features", "extracted", "from", "the", "image", "at", "the", "location", "z", "denoted", "by", "x", "z", "∈", "R", "d", "and", "contextual", "information", "from", "the", "preceding", "classifier", "in", "the", "neighbor", "-", "26", "⇥", "26", "60", "⇥", "60", "96", "⇥", "96", "160", "⇥", "160", "240", "⇥", "240", "320", "⇥", "320", "400", "⇥", "400", "hood", "around", "each", "Y", "p", "in", "stage", "t.", "A", "classifier", "in", "the", "first", "stage", "t", "=", "1", ",", "therefore", "produces", "the", "following", "belief", "values", ":", "where", "is", "the", "score", "predicted", "by", "the", "classifier", "g", "1", "for", "assigning", "the", "p", "th", "part", "in", "the", "first", "stage", "at", "image", "location", "z.", "We", "represent", "all", "the", "beliefs", "of", "part", "p", "evaluated", "at", "every", "location", "z", "=", "(", "u", ",", "v", ")", "T", "in", "the", "image", "as", "b", "p", "t", "∈", "R", "w×h", ",", "where", "w", "and", "h", "are", "the", "width", "and", "height", "of", "the", "image", ",", "respectively", ".", "That", "is", ",", "b", "For", "convenience", ",", "we", "denote", "the", "collection", "of", "belief", "maps", "for", "all", "the", "parts", "as", "b", "t", "∈", "R", "w×h×", "(", "P", "+", "1", ")", "(", "P", "parts", "plus", "one", "for", "background", ")", ".", "In", "subsequent", "stages", ",", "the", "classifier", "predicts", "a", "belief", "for", "assigning", "a", "location", "to", "each", "part", "Y", "p", "=", "z", ",", "∀z", "∈", "Z", ",", "based", "on", "(", "1", ")", "features", "of", "the", "image", "data", "x", "t", "z", "∈", "R", "d", "again", ",", "and", "(", "2", ")", "contextual", "information", "from", "the", "preceeding", "classifier", "in", "the", "neighborhood", "around", "each", "Y", "p", ":", "where", "ψ", "t>1", "(", "·", ")", "is", "a", "mapping", "from", "the", "beliefs", "b", "t−1", "to", "context", "features", ".", "In", "each", "stage", ",", "the", "computed", "beliefs", "provide", "an", "increasingly", "refined", "estimate", "for", "the", "location", "of", "each", "part", ".", "Note", "that", "we", "allow", "image", "features", "x", "z", "for", "subsequent", "stage", "to", "be", "different", "from", "the", "image", "feature", "used", "in", "the", "first", "stage", "x.", "The", "pose", "machine", "proposed", "in", "[", "reference", "]", "used", "boosted", "random", "forests", "for", "prediction", "(", "{", "g", "t", "}", ")", ",", "fixed", "hand", "-", "crafted", "image", "features", "across", "all", "stages", "(", "x", "=", "x", ")", ",", "and", "fixed", "hand", "-", "crafted", "context", "feature", "maps", "(", "ψ", "t", "(", "·", ")", ")", "to", "capture", "spatial", "context", "across", "all", "stages", ".", "section", ":", "Convolutional", "Pose", "Machines", "We", "show", "how", "the", "prediction", "and", "image", "feature", "computation", "modules", "of", "a", "pose", "machine", "can", "be", "replaced", "by", "a", "deep", "convolutional", "architecture", "allowing", "for", "both", "image", "and", "contextual", "feature", "representations", "to", "be", "learned", "directly", "from", "data", ".", "Convolutional", "architectures", "also", "have", "the", "advantage", "of", "being", "completely", "differentiable", ",", "thereby", "enabling", "endto", "-", "end", "joint", "training", "of", "all", "stages", "of", "a", "CPM", ".", "We", "describe", "our", "design", "for", "a", "CPM", "that", "combines", "the", "advantages", "of", "deep", "convolutional", "architectures", "with", "the", "implicit", "spatial", "modeling", "afforded", "by", "the", "pose", "machine", "framework", ".", "section", ":", "Keypoint", "Localization", "Using", "Local", "Image", "Evidence", "The", "first", "stage", "of", "a", "convolutional", "pose", "machine", "predicts", "part", "beliefs", "from", "only", "local", "image", "evidence", ".", "Figure", "2c", "shows", "the", "network", "structure", "used", "for", "part", "detection", "from", "local", "image", "evidence", "using", "a", "deep", "convolutional", "network", ".", "The", "evidence", "is", "local", "because", "the", "receptive", "field", "of", "the", "first", "stage", "of", "the", "network", "is", "constrained", "to", "a", "small", "patch", "around", "the", "output", "pixel", "location", ".", "We", "use", "a", "network", "structure", "composed", "of", "five", "convolutional", "layers", "followed", "by", "two", "1", "×", "1", "convolutional", "layers", "which", "results", "in", "a", "fully", "convolutional", "archi", "-", "Figure", "3", ":", "Spatial", "context", "from", "belief", "maps", "of", "easier", "-", "to", "-", "detect", "parts", "can", "provide", "strong", "cues", "for", "localizing", "difficult", "-", "to", "-", "detect", "parts", ".", "The", "spatial", "contexts", "from", "shoulder", ",", "neck", "and", "head", "can", "help", "eliminate", "wrong", "(", "red", ")", "and", "strengthen", "correct", "(", "green", ")", "estimations", "on", "the", "belief", "map", "of", "right", "elbow", "in", "the", "subsequent", "stages", ".", "tecture", "[", "reference", "]", ".", "In", "practice", ",", "to", "achieve", "certain", "precision", ",", "we", "normalize", "input", "cropped", "images", "to", "size", "368", "×", "368", "(", "see", "Section", "4.2", "for", "details", ")", ",", "and", "the", "receptive", "field", "of", "the", "network", "shown", "above", "is", "160", "×", "160", "pixels", ".", "The", "network", "can", "effectively", "be", "viewed", "as", "sliding", "a", "deep", "network", "across", "an", "image", "and", "regressing", "from", "the", "local", "image", "evidence", "in", "each", "160", "×", "160", "image", "patch", "to", "a", "P", "+", "1", "sized", "output", "vector", "that", "represents", "a", "score", "for", "each", "part", "at", "that", "image", "location", ".", "section", ":", "Sequential", "Prediction", "with", "Learned", "Spatial", "Context", "Features", "While", "the", "detection", "rate", "on", "landmarks", "with", "consistent", "appearance", ",", "such", "as", "the", "head", "and", "shoulders", ",", "can", "be", "favorable", ",", "the", "accuracies", "are", "often", "much", "lower", "for", "landmarks", "lower", "down", "the", "kinematic", "chain", "of", "the", "human", "skeleton", "due", "to", "their", "large", "variance", "in", "configuration", "and", "appearance", ".", "The", "landscape", "of", "the", "belief", "maps", "around", "a", "part", "location", ",", "albeit", "noisy", ",", "can", ",", "however", ",", "be", "very", "informative", ".", "Illustrated", "in", "Figure", "3", ",", "when", "detecting", "challenging", "parts", "such", "as", "right", "elbow", ",", "the", "belief", "map", "for", "right", "shoulder", "with", "a", "sharp", "peak", "can", "be", "used", "as", "a", "strong", "cue", ".", "A", "predictor", "in", "subsequent", "stages", "(", "g", "t>1", ")", "can", "use", "the", "spatial", "context", "(", "ψ", "t>1", "(", "·", ")", ")", "of", "the", "noisy", "belief", "maps", "in", "a", "region", "around", "the", "image", "location", "z", "and", "improve", "its", "predictions", "by", "leveraging", "the", "fact", "that", "parts", "occur", "in", "consistent", "geometric", "configurations", ".", "In", "the", "second", "stage", "of", "a", "pose", "machine", ",", "the", "classifier", "g", "2", "accepts", "as", "input", "the", "image", "features", "x", "2", "z", "and", "features", "computed", "on", "the", "beliefs", "via", "the", "feature", "function", "ψ", "for", "each", "of", "the", "parts", "in", "the", "previous", "stage", ".", "The", "feature", "function", "ψ", "serves", "to", "encode", "the", "landscape", "of", "the", "belief", "maps", "from", "the", "previous", "stage", "in", "a", "spatial", "region", "around", "the", "location", "z", "of", "the", "different", "parts", ".", "For", "a", "convolutional", "pose", "machine", ",", "we", "do", "not", "have", "an", "explicit", "function", "that", "computes", "context", "features", ".", "Instead", ",", "we", "define", "ψ", "as", "being", "the", "receptive", "field", "of", "the", "predictor", "on", "the", "beliefs", "from", "the", "previous", "stage", ".", "The", "design", "of", "the", "network", "is", "guided", "by", "achieving", "a", "receptive", "field", "at", "the", "output", "layer", "of", "the", "second", "stage", "network", "that", "is", "large", "enough", "to", "allow", "the", "learning", "of", "potentially", "complex", "and", "long", "-", "range", "correlations", "between", "parts", ".", "By", "simply", "supplying", "features", "on", "the", "outputs", "of", "the", "previous", "stage", "(", "as", "opposed", "to", "specifying", "potential", "functions", "in", "a", "graphical", "model", ")", ",", "the", "convolutional", "layers", "in", "the", "subsequent", "stage", "allow", "the", "classifier", "to", "freely", "combine", "contextual", "information", "by", "picking", "the", "most", "predictive", "features", ".", "The", "belief", "maps", "from", "the", "first", "stage", "are", "generated", "from", "a", "network", "that", "examined", "the", "image", "locally", "with", "a", "small", "receptive", "field", ".", "In", "the", "second", "stage", ",", "we", "design", "a", "network", "that", "drastically", "increases", "the", "equivalent", "receptive", "field", ".", "Large", "receptive", "fields", "can", "be", "achieved", "either", "by", "pooling", "at", "the", "expense", "of", "precision", ",", "increasing", "the", "kernel", "size", "of", "the", "convolutional", "filters", "at", "the", "expense", "of", "increasing", "the", "number", "of", "parameters", ",", "or", "by", "increasing", "the", "number", "of", "convolutional", "layers", "at", "the", "risk", "of", "encountering", "vanishing", "gradients", "during", "training", ".", "Our", "network", "design", "and", "corresponding", "receptive", "field", "for", "the", "subsequent", "stages", "(", "t", "≥", "2", ")", "is", "shown", "in", "Figure", "2d", ".", "We", "choose", "to", "use", "multiple", "convolutional", "layers", "to", "achieve", "large", "receptive", "field", "on", "the", "8×", "downscaled", "heatmaps", ",", "as", "it", "allows", "us", "to", "be", "parsimonious", "with", "respect", "to", "the", "number", "of", "parameters", "of", "the", "model", ".", "We", "found", "that", "our", "stride", "-", "8", "network", "performs", "as", "well", "as", "a", "stride", "-", "4", "one", "even", "at", "high", "precision", "region", ",", "while", "it", "makes", "us", "easier", "to", "achieve", "larger", "receptive", "fields", ".", "We", "also", "repeat", "similar", "structure", "for", "image", "feature", "maps", "to", "make", "the", "spatial", "context", "be", "image", "-", "dependent", "and", "allow", "error", "correction", ",", "following", "the", "structure", "of", "pose", "machine", ".", "We", "find", "that", "accuracy", "improves", "with", "the", "size", "of", "the", "receptive", "field", ".", "In", "Figure", "4", "we", "show", "the", "improvement", "in", "accuracy", "on", "the", "FLIC", "dataset", "[", "reference", "]", "as", "the", "size", "of", "the", "receptive", "field", "on", "the", "original", "image", "is", "varied", "by", "varying", "the", "architecture", "without", "significantly", "changing", "the", "number", "of", "parameters", ",", "through", "a", "series", "of", "experimental", "trials", "on", "input", "images", "normalized", "to", "a", "size", "of", "304", "×", "304", ".", "We", "see", "that", "the", "accuracy", "improves", "as", "the", "effective", "receptive", "field", "increases", ",", "and", "starts", "to", "saturate", "around", "250", "pixels", ",", "which", "also", "happens", "to", "be", "roughly", "the", "size", "of", "the", "normalized", "object", ".", "This", "improvement", "in", "accuracy", "with", "receptive", "field", "size", "suggests", "that", "the", "network", "does", "indeed", "encode", "long", "range", "interactions", "between", "parts", "and", "that", "doing", "so", "is", "beneficial", ".", "In", "our", "best", "performing", "setting", "in", "Figure", "2", ",", "we", "normalize", "cropped", "images", "into", "a", "larger", "size", "of", "368", "×", "368", "pixels", "for", "better", "precision", ",", "and", "the", "receptive", "field", "of", "the", "second", "stage", "output", "on", "the", "belief", "maps", "of", "the", "first", "stage", "is", "set", "to", "31", "×", "31", ",", "which", "is", "equivalently", "400", "×", "400", "pixels", "on", "the", "original", "image", ",", "where", "the", "radius", "can", "usually", "cover", "any", "pair", "of", "the", "parts", ".", "With", "more", "stages", ",", "the", "effective", "receptive", "field", "is", "even", "larger", ".", "In", "the", "following", "section", "we", "show", "our", "results", "from", "up", "to", "6", "stages", ".", "section", ":", "Learning", "in", "Convolutional", "Pose", "Machines", "The", "design", "described", "above", "for", "a", "pose", "machine", "results", "in", "a", "deep", "architecture", "that", "can", "have", "a", "large", "number", "of", "layers", ".", "Training", "such", "a", "network", "with", "many", "layers", "can", "be", "prone", "to", "the", "problem", "of", "vanishing", "gradients", "[", "reference", "][", "reference", "][", "reference", "]", "where", ",", "as", "observed", "by", "Bradley", "[", "reference", "]", "and", "Bengio", "et", "al", ".", "[", "reference", "]", ",", "the", "magnitude", "of", "back", "-", "propagated", "gradients", "decreases", "in", "strength", "with", "the", "number", "of", "intermediate", "layers", "between", "the", "output", "layer", "and", "the", "input", "layer", ".", "Fortunately", ",", "the", "sequential", "prediction", "framework", "of", "the", "pose", "machine", "provides", "a", "natural", "approach", "to", "training", "our", "deep", "architecture", "that", "addresses", "this", "problem", ".", "Each", "stage", "of", "the", "pose", "machine", "is", "trained", "to", "repeatedly", "produce", "the", "belief", "maps", "for", "the", "locations", "of", "each", "of", "the", "parts", ".", "We", "encourage", "the", "network", "to", "repeatedly", "arrive", "at", "such", "a", "representation", "by", "defining", "a", "loss", "function", "at", "the", "output", "of", "each", "stage", "t", "that", "minimizes", "the", "l", "2", "distance", "between", "the", "predicted", "and", "ideal", "belief", "maps", "for", "each", "part", ".", "The", "ideal", "belief", "map", "for", "a", "part", "p", "is", "written", "as", "b", "p", "*", "(", "Y", "p", "=", "z", ")", ",", "which", "are", "created", "by", "putting", "Gaussian", "peaks", "at", "ground", "truth", "locations", "of", "each", "body", "part", "p.", "The", "cost", "function", "we", "aim", "to", "minimize", "at", "the", "output", "of", "each", "stage", "at", "each", "level", "is", "therefore", "given", "by", ":", "The", "overall", "objective", "for", "the", "full", "architecture", "is", "obtained", "by", "adding", "the", "losses", "at", "each", "stage", "and", "is", "given", "by", ":", "We", "use", "standard", "stochastic", "gradient", "descend", "to", "jointly", "train", "all", "the", "T", "stages", "in", "the", "network", ".", "To", "share", "the", "image", "feature", "x", "across", "all", "subsequent", "stages", ",", "we", "share", "the", "weights", "of", "corresponding", "convolutional", "layers", "(", "see", "Figure", "2", ")", "across", "stages", "t", "≥", "2", ".", "section", ":", "Evaluation", "section", ":", "Analysis", "Addressing", "vanishing", "gradients", ".", "The", "objective", "in", "Equation", "5", "describes", "a", "decomposable", "loss", "function", "that", "operates", "on", "different", "parts", "of", "the", "network", "(", "see", "Figure", "2", ")", ".", "Specifically", ",", "each", "term", "in", "the", "summation", "is", "applied", "to", "the", "network", "after", "each", "stage", "t", "effectively", "enforcing", "supervision", "in", "intermediate", "stages", "through", "the", "network", ".", "Intermediate", "supervision", "has", "the", "advantage", "that", ",", "even", "though", "the", "full", "architecture", "can", "have", "many", "layers", ",", "it", "does", "not", "fall", "prey", "to", "the", "vanishing", "gradient", "problem", "as", "the", "intermediate", "loss", "functions", "replenish", "the", "gradients", "at", "each", "stage", ".", "We", "verify", "this", "claim", "by", "observing", "histograms", "of", "gradient", "magnitude", "(", "see", "Figure", "5", ")", "at", "different", "depths", "in", "the", "architecture", "across", "training", "epochs", "for", "models", "with", "and", "without", "intermediate", "supervision", ".", "In", "early", "epochs", ",", "as", "we", "move", "from", "the", "output", "layer", "to", "the", "input", "layer", ",", "we", "observe", "on", "the", "model", "larger", "variance", "across", "all", "layers", ",", "suggesting", "that", "learning", "is", "indeed", "occurring", "in", "all", "the", "layers", "thanks", "to", "intermediate", "supervision", ".", "We", "also", "notice", "that", "as", "training", "progresses", ",", "the", "variance", "in", "the", "gradient", "magnitude", "distributions", "decreases", "pointing", "to", "model", "convergence", ".", "Benefit", "of", "end", "-", "to", "-", "end", "learning", ".", "We", "see", "in", "Figure", "6a", "that", "replacing", "the", "modules", "of", "a", "pose", "machine", "with", "the", "appropriately", "designed", "convolutional", "architecture", "provides", "a", "large", "boost", "of", "42.4", "percentage", "points", "over", "the", "previous", "approach", "of", "[", "reference", "]", "in", "the", "high", "precision", "regime", "(", "PCK@0.1", ")", "and", "30.9", "percentage", "points", "in", "the", "low", "precision", "regime", "(", "PCK@0.2", ")", ".", "Comparison", "on", "training", "schemes", ".", "We", "compare", "different", "variants", "of", "training", "the", "network", "in", "Figure", "6b", "on", "the", "LSP", "dataset", "with", "person", "-", "centric", "(", "PC", ")", "annotations", ".", "To", "demonstrate", "the", "benefit", "of", "intermediate", "supervision", "with", "joint", "training", "across", "stages", ",", "we", "train", "the", "model", "in", "four", "ways", ":", "(", "i", ")", "training", "from", "scratch", "using", "a", "global", "loss", "function", "that", "enforces", "intermediate", "supervision", "(", "ii", ")", "stage", "-", "wise", ";", "where", "each", "stage", "is", "trained", "in", "a", "feed", "-", "forward", "fashion", "and", "stacked", "(", "iii", ")", "as", "same", "as", "(", "i", ")", "but", "initialized", "with", "weights", "from", "(", "ii", ")", ",", "and", "(", "iv", ")", "as", "same", "as", "(", "i", ")", "but", "with", "no", "intermediate", "supervision", ".", "We", "find", "that", "network", "(", "i", ")", "outperforms", "all", "other", "training", "methods", ",", "showing", "that", "intermediate", "supervision", "and", "joint", "training", "across", "stage", "is", "indeed", "crucial", "in", "achieving", "good", "performance", ".", "The", "stagewise", "training", "in", "(", "ii", ")", "saturate", "at", "suboptimal", ",", "and", "the", "jointly", "fine", "-", "tuning", "in", "(", "iii", ")", "improves", "from", "this", "sub", "-", "optimal", "to", "the", "accuracy", "level", "closed", "to", "(", "i", ")", ",", "however", "with", "effectively", "longer", "training", "iterations", ".", "Performance", "across", "stages", ".", "We", "show", "a", "comparison", "of", "performance", "across", "each", "stage", "on", "the", "LSP", "dataset", "(", "PC", ")", "in", "Figure", "6c", ".", "We", "show", "that", "the", "performance", "increases", "monotonically", "until", "5", "stages", ",", "as", "the", "predictors", "in", "subsequent", "stages", "make", "use", "of", "contextual", "information", "in", "a", "large", "receptive", "field", "on", "the", "previous", "stage", "beliefs", "maps", "to", "resolve", "confusions", "between", "parts", "and", "background", ".", "We", "see", "diminishing", "returns", "at", "the", "6th", "stage", ",", "which", "is", "the", "number", "we", "choose", "for", "reporting", "our", "best", "results", "in", "this", "paper", "for", "LSP", "and", "MPII", "datasets", ".", "section", ":", "Datasets", "and", "Quantitative", "Analysis", "In", "this", "section", "we", "present", "our", "numerical", "results", "in", "various", "standard", "benchmarks", "including", "the", "MPII", ",", "LSP", ",", "and", "FLIC", "datasets", ".", "To", "have", "normalized", "input", "samples", "of", "368", "×", "368", "for", "training", ",", "we", "first", "resize", "the", "images", "to", "roughly", "make", "the", "samples", "into", "the", "same", "scale", ",", "and", "then", "crop", "or", "pad", "the", "image", "according", "to", "the", "center", "positions", "and", "rough", "scale", "estimations", "provided", "in", "the", "datasets", "if", "available", ".", "In", "datasets", "such", "as", "LSP", "without", "these", "information", ",", "we", "estimate", "them", "according", "to", "joint", "positions", "or", "image", "sizes", ".", "For", "testing", ",", "we", "perform", "similar", "resizing", "and", "cropping", "(", "or", "padding", ")", ",", "but", "estimate", "center", "position", "and", "scale", "only", "from", "image", "sizes", "when", "necessary", ".", "In", "addition", ",", "we", "merge", "the", "belief", "maps", "from", "different", "scales", "(", "perturbed", "around", "the", "given", "one", ")", "for", "final", "predictions", ",", "to", "handle", "the", "inaccuracy", "of", "the", "given", "scale", "estimation", ".", "We", "define", "and", "implement", "our", "model", "using", "the", "Caffe", "[", "13", "]", "libraries", "for", "deep", "learning", ".", "We", "publicly", "release", "the", "source", "code", "and", "details", "on", "the", "architecture", ",", "learning", "parameters", ",", "design", "decisions", "and", "data", "augmentation", "to", "ensure", "full", "reproducibility", ".", "[", "reference", "]", "MPII", "Human", "Pose", "Dataset", ".", "We", "show", "in", "Figure", "8", "our", "results", "on", "the", "MPII", "Human", "Pose", "dataset", "[", "reference", "]", "which", "consists", "more", "than", "28000", "training", "samples", ".", "We", "choose", "to", "randomly", "augment", "the", "data", "with", "rotation", "degrees", "in", "[", "−40", "•", ",", "40", "•", "]", ",", "scaling", "with", "factors", "in", "[", "0.7", ",", "1.3", "]", ",", "and", "horizonal", "flipping", ".", "The", "evaluation", "is", "based", "on", "PCKh", "metric", "[", "reference", "]", "where", "the", "error", "tolerance", "is", "normalized", "with", "respect", "to", "head", "size", "of", "the", "target", ".", "Because", "there", "often", "are", "multiple", "people", "in", "the", "proximity", "of", "the", "interested", "person", "(", "rough", "center", "position", "is", "given", "in", "the", "dataset", ")", ",", "we", "made", "two", "sets", "of", "ideal", "belief", "maps", "for", "training", ":", "one", "includes", "all", "the", "peaks", "for", "every", "person", "appearing", "in", "the", "proximity", "of", "the", "primary", "subject", "and", "the", "second", "type", "where", "we", "only", "place", "peaks", "for", "the", "primary", "subject", ".", "We", "supply", "the", "first", "set", "of", "belief", "maps", "to", "the", "loss", "layers", "in", "the", "first", "stage", "as", "the", "initial", "stage", "only", "relies", "on", "local", "image", "evidence", "to", "make", "predictions", ".", "We", "supply", "the", "second", "type", "of", "belief", "maps", "to", "the", "Wrists", "Elbows", "(", "a", ")", "(", "b", ")", "loss", "layers", "of", "all", "subsequent", "stages", ".", "We", "also", "find", "that", "supplying", "to", "all", "subsequent", "stages", "an", "additional", "heat", "-", "map", "with", "a", "Gaussian", "peak", "indicating", "center", "of", "the", "primary", "subject", "is", "beneficial", ".", "Our", "total", "PCKh", "-", "0.5", "score", "achieves", "state", "of", "the", "art", "at", "87.95", "%", "(", "88.52", "%", "when", "adding", "LSP", "training", "data", ")", ",", "which", "is", "6.11", "%", "higher", "than", "the", "closest", "competitor", ",", "and", "it", "is", "noteworthy", "that", "on", "the", "ankle", "(", "the", "most", "challenging", "part", ")", ",", "our", "PCKh", "-", "0.5", "score", "is", "78.28", "%", "(", "79.41", "%", "when", "adding", "LSP", "training", "data", ")", ",", "which", "is", "10.76", "%", "higher", "than", "the", "closest", "competitor", ".", "This", "result", "shows", "the", "capability", "of", "our", "model", "to", "capture", "long", "distance", "context", "given", "ankles", "are", "the", "farthest", "parts", "from", "head", "and", "other", "more", "recognizable", "parts", ".", "Figure", "11", "shows", "our", "accuracy", "is", "also", "consistently", "significantly", "higher", "than", "other", "methods", "across", "various", "view", "angles", "defined", "in", "[", "reference", "]", ",", "especially", "in", "those", "challenging", "non", "-", "frontal", "views", ".", "In", "summary", ",", "our", "method", "improves", "the", "accuracy", "in", "all", "parts", ",", "over", "all", "precisions", ",", "across", "all", "view", "angles", ",", "and", "is", "the", "first", "one", "achieving", "such", "high", "accuracy", "without", "any", "pre", "-", "training", "from", "other", "data", ",", "or", "post", "-", "inference", "parsing", "with", "hand", "-", "design", "priors", "or", "initialization", "of", "such", "a", "structured", "prediction", "task", "as", "in", "[", "reference", "][", "reference", "]", ".", "Our", "methods", "also", "does", "not", "need", "another", "module", "dedicated", "to", "location", "refinement", "as", "in", "[", "reference", "]", "to", "achieve", "great", "high", "-", "precision", "accuracy", "with", "a", "stride", "-", "8", "network", ".", "Leeds", "Sports", "Pose", "(", "LSP", ")", "Dataset", ".", "We", "evaluate", "our", "method", "on", "the", "Extended", "Leeds", "Sports", "Dataset", "[", "reference", "]", "that", "consists", "of", "11000", "images", "for", "training", "and", "1000", "images", "for", "testing", ".", "We", "trained", "on", "person", "-", "centric", "(", "PC", ")", "annotations", "and", "evaluate", "our", "method", "using", "the", "Percentage", "Correct", "Keypoints", "(", "PCK", ")", "metric", "[", "reference", "]", ".", "Using", "the", "same", "augmentation", "scheme", "as", "for", "the", "MPI", "dataset", ",", "our", "model", "again", "achieves", "state", "of", "the", "art", "at", "84.32", "%", "(", "90.5", "%", "when", "adding", "MPII", "train", "-", "MPII", "FLIC", "LSP", "Figure", "10", ":", "Qualitative", "results", "of", "our", "method", "on", "the", "MPII", ",", "LSP", "and", "FLIC", "datasets", "respectively", ".", "We", "see", "that", "the", "method", "is", "able", "to", "handle", "non", "-", "standard", "poses", "and", "resolve", "ambiguities", "between", "symmetric", "parts", "for", "a", "variety", "of", "different", "relative", "camera", "views", ".", "ing", "data", ")", ".", "Note", "that", "adding", "MPII", "data", "here", "significantly", "boosts", "our", "performance", ",", "due", "to", "its", "labeling", "quality", "being", "much", "better", "than", "LSP", ".", "Because", "of", "the", "noisy", "label", "in", "the", "LSP", "dataset", ",", "Pishchulin", "et", "al", ".", "[", "reference", "]", "reproduced", "the", "dataset", "with", "original", "high", "resolution", "images", "and", "better", "labeling", "quality", ".", "FLIC", "Dataset", ".", "We", "evaluate", "our", "method", "on", "the", "FLIC", "Dataset", "[", "reference", "]", "which", "consists", "of", "3987", "images", "for", "training", "and", "1016", "images", "for", "testing", ".", "We", "report", "accuracy", "as", "per", "the", "metric", "introduced", "in", "Sapp", "et", "al", ".", "[", "reference", "]", "for", "the", "elbow", "and", "wrist", "joints", "in", "Figure", "12", ".", "Again", ",", "we", "outperform", "all", "prior", "art", "at", "PCK@0.2", "with", "97.59", "%", "on", "elbows", "and", "95.03", "%", "on", "wrists", ".", "In", "higher", "precision", "region", "our", "advantage", "is", "even", "more", "significant", ":", "14.8", "percentage", "points", "on", "wrists", "and", "12.7", "percentage", "points", "on", "elbows", "at", "PCK@0.05", ",", "and", "8.9", "percentage", "points", "on", "wrists", "and", "9.3", "percentage", "points", "on", "elbows", "at", "PCK@0.1", ".", "section", ":", "Discussion", "Convolutional", "pose", "machines", "provide", "an", "end", "-", "to", "-", "end", "architecture", "for", "tackling", "structured", "prediction", "problems", "in", "computer", "vision", "without", "the", "need", "for", "graphical", "-", "model", "style", "inference", ".", "We", "showed", "that", "a", "sequential", "architecture", "composed", "of", "convolutional", "networks", "is", "capable", "of", "implicitly", "learning", "a", "spatial", "models", "for", "pose", "by", "communicating", "increasingly", "refined", "uncertainty", "-", "preserving", "beliefs", "between", "stages", ".", "Problems", "with", "spatial", "dependencies", "between", "variables", "arise", "in", "multiple", "domains", "of", "computer", "vision", "such", "as", "semantic", "image", "labeling", ",", "single", "image", "depth", "prediction", "and", "object", "detection", "and", "future", "work", "will", "involve", "extending", "our", "architecture", "to", "these", "problems", ".", "Our", "approach", "achieves", "state", "of", "the", "art", "accuracy", "on", "all", "primary", "benchmarks", ",", "however", "we", "do", "observe", "failure", "cases", "mainly", "when", "multiple", "people", "are", "in", "close", "proximity", ".", "Handling", "multiple", "people", "in", "a", "single", "end", "-", "to", "-", "end", "architecture", "is", "also", "a", "challenging", "problem", "and", "an", "interesting", "avenue", "for", "future", "work", ".", "section", ":"], "entity": [[0, 3, ["Convolutional", "Pose", "Machines"], "Method"], [6, 8, ["Pose", "Machines"], "Method"], [10, 13, ["sequential", "prediction", "framework"], "Method"], [16, 19, ["implicit", "spatial", "models"], "Method"], [30, 32, ["convolutional", "networks"], "Method"], [41, 44, ["learning", "image", "features"], "Task"], [45, 50, ["image", "-", "dependent", "spatial", "models"], "Method"], [54, 56, ["pose", "estimation"], "Task"], [66, 72, ["long", "-", "range", "dependencies", "between", "variables"], "Task"], [73, 76, ["structured", "prediction", "tasks"], "Task"], [88, 90, ["sequential", "architecture"], "Method"], [92, 94, ["convolutional", "networks"], "Method"], [117, 122, ["graphical", "model", "-", "style", "inference"], "Method"], [133, 134, ["training"], "Task"], [155, 157, ["learning", "procedure"], "Method"], [177, 178, ["MPII"], "Material"], [179, 180, ["LSP"], "Material"], [182, 184, ["FLIC", "datasets"], "Material"], [190, 193, ["Convolutional", "Pose", "Machines"], "Method"], [194, 196, ["CPMs", ")"], "Method"], [204, 205, ["CPMs"], "Method"], [210, 212, ["pose", "machine"], "Method"], [218, 225, ["implicit", "learning", "of", "long", "-", "range", "dependencies"], "Method"], [236, 237, ["learning"], "Task"], [238, 239, ["inference"], "Task"], [241, 244, ["modular", "sequential", "design"], "Method"], [253, 255, ["convolutional", "architectures"], "Method"], [260, 262, ["feature", "representations"], "Method"], [273, 275, ["differentiable", "architecture"], "Method"], [278, 281, ["globally", "joint", "training"], "Task"], [282, 283, ["backpropagation"], "Method"], [294, 295, ["CPMs"], "Method"], [300, 302, ["convolutional", "networks"], "Method"], [359, 361, ["belief", "maps"], "Method"], [367, 371, ["non", "-", "parametric", "encoding"], "Method"], [383, 384, ["CPM"], "Method"], [387, 392, ["image", "-", "dependent", "spatial", "models"], "Method"], [407, 409, ["graphical", "models"], "Method"], [418, 422, ["post", "-", "processing", "steps"], "Method"], [430, 432, ["convolutional", "networks"], "Method"], [441, 447, ["implicit", "image", "-", "dependent", "spatial", "models"], "Method"], [456, 458, ["multistage", "architecture"], "Method"], [475, 476, ["backpropagation"], "Method"], [508, 509, ["CPM"], "Method"], [510, 512, ["belief", "maps"], "Method"], [547, 550, ["sequential", "prediction", "framework"], "Method"], [587, 592, ["learning", "long", "range", "spatial", "relationships"], "Task"], [602, 605, ["message", "passing", "inference"], "Task"], [606, 608, ["graphical", "models"], "Method"], [617, 625, ["unrolled", "mean", "-", "field", "message", "passing", "inference", "algorithm"], "Method"], [637, 638, ["backpropagation"], "Method"], [642, 643, ["accuracy"], "Metric"], [646, 648, ["convolutional", "networks"], "Method"], [650, 651, ["CPM"], "Method"], [654, 656, ["overall", "network"], "Method"], [679, 680, ["learning"], "Task"], [714, 716, ["deep", "networks"], "Method"], [721, 722, ["learning"], "Task"], [734, 736, ["classification", "problems"], "Task"], [746, 749, ["structured", "prediction", "problem"], "Task"], [751, 753, ["pose", "estimation"], "Task"], [754, 755, ["CPMs"], "Method"], [786, 788, ["training", "schemes"], "Method"], [791, 794, ["sequential", "prediction", "architecture"], "Method"], [803, 806, ["implicit", "spatial", "models"], "Method"], [833, 838, ["image", "-", "dependent", "spatial", "models"], "Method"], [839, 842, ["structured", "prediction", "tasks"], "Task"], [848, 852, ["graphical", "model", "style", "inference"], "Method"], [868, 869, ["MPII"], "Material"], [870, 871, ["LSP"], "Material"], [873, 875, ["FLIC", "datasets"], "Material"], [884, 888, ["multi", "-", "staged", "architecture"], "Method"], [906, 909, ["pictorial", "structures", "model"], "Method"], [939, 944, ["tree", "-", "structured", "graphical", "model"], "Method"], [996, 1000, ["tree", "-", "structured", "model"], "Method"], [1015, 1018, ["pictorial", "structures", "model"], "Method"], [1023, 1025, ["graph", "representation"], "Method"], [1026, 1028, ["Hierarchical", "models"], "Method"], [1098, 1102, ["Non", "-", "tree", "models"], "Method"], [1146, 1148, ["approximate", "inference"], "Method"], [1150, 1151, ["learning"], "Task"], [1163, 1167, ["modeling", "of", "spatial", "relationships"], "Task"], [1172, 1173, ["inference"], "Task"], [1183, 1185, ["fast", "inference"], "Task"], [1193, 1196, ["sequential", "prediction", "framework"], "Method"], [1201, 1204, ["implicit", "spatial", "model"], "Method"], [1214, 1216, ["inference", "procedure"], "Method"], [1241, 1243, ["convolutional", "architectures"], "Method"], [1285, 1287, ["convolutional", "architecture"], "Method"], [1302, 1304, ["graphical", "models"], "Method"], [1313, 1315, ["heuristic", "initialization"], "Method"], [1336, 1338, ["network", "module"], "Method"], [1364, 1366, ["convolutional", "networks"], "Method"], [1415, 1417, ["network", "module"], "Method"], [1423, 1426, ["implicit", "spatial", "models"], "Method"], [1433, 1434, ["convolutions"], "Method"], [1463, 1465, ["joint", "training"], "Method"], [1475, 1477, ["deep", "network"], "Method"], [1488, 1490, ["cartesian", "representation"], "Method"], [1505, 1506, ["accuracy"], "Metric"], [1519, 1522, ["sequential", "prediction", "framework"], "Method"], [1558, 1560, ["Pose", "Machines"], "Method"], [1634, 1636, ["pose", "machine"], "Method"], [1651, 1655, ["multi", "-", "class", "predictors"], "Method"], [1738, 1739, ["classifier"], "Method"], [1773, 1774, ["classifier"], "Method"], [1796, 1798, ["classifier", "g"], "Method"], [1902, 1903, ["classifier"], "Method"], [1948, 1950, ["preceeding", "classifier"], "Method"], [2020, 2022, ["pose", "machine"], "Method"], [2028, 2031, ["boosted", "random", "forests"], "Method"], [2032, 2033, ["prediction"], "Task"], [2080, 2083, ["Convolutional", "Pose", "Machines"], "Method"], [2087, 2093, ["prediction", "and", "image", "feature", "computation", "modules"], "Method"], [2095, 2097, ["pose", "machine"], "Method"], [2102, 2105, ["deep", "convolutional", "architecture"], "Method"], [2108, 2113, ["image", "and", "contextual", "feature", "representations"], "Method"], [2120, 2122, ["Convolutional", "architectures"], "Method"], [2136, 2138, ["joint", "training"], "Task"], [2143, 2144, ["CPM"], "Method"], [2151, 2152, ["CPM"], "Method"], [2157, 2160, ["deep", "convolutional", "architectures"], "Method"], [2162, 2165, ["implicit", "spatial", "modeling"], "Method"], [2174, 2176, ["Keypoint", "Localization"], "Task"], [2185, 2188, ["convolutional", "pose", "machine"], "Method"], [2201, 2203, ["network", "structure"], "Method"], [2205, 2207, ["part", "detection"], "Task"], [2213, 2216, ["deep", "convolutional", "network"], "Method"], [2247, 2249, ["network", "structure"], "Method"], [2252, 2254, ["convolutional", "layers"], "Method"], [2260, 2262, ["convolutional", "layers"], "Method"], [2266, 2270, ["fully", "convolutional", "archi", "-"], "Method"], [2290, 2297, ["localizing", "difficult", "-", "to", "-", "detect", "parts"], "Task"], [2388, 2390, ["deep", "network"], "Method"], [2429, 2431, ["Sequential", "Prediction"], "Task"], [2438, 2440, ["detection", "rate"], "Metric"], [2458, 2459, ["accuracies"], "Metric"], [2594, 2596, ["pose", "machine"], "Method"], [2598, 2600, ["classifier", "g"], "Method"], [2663, 2666, ["convolutional", "pose", "machine"], "Method"], [2718, 2721, ["second", "stage", "network"], "Method"], [2760, 2762, ["graphical", "model"], "Method"], [2765, 2767, ["convolutional", "layers"], "Method"], [2773, 2774, ["classifier"], "Method"], [2787, 2789, ["belief", "maps"], "Method"], [2834, 2835, ["pooling"], "Method"], [2839, 2840, ["precision"], "Metric"], [2847, 2849, ["convolutional", "filters"], "Method"], [2878, 2880, ["network", "design"], "Method"], [2904, 2906, ["convolutional", "layers"], "Method"], [2939, 2943, ["stride", "-", "8", "network"], "Method"], [2948, 2952, ["stride", "-", "4", "one"], "Method"], [2989, 2991, ["error", "correction"], "Task"], [2996, 2998, ["pose", "machine"], "Method"], [3002, 3003, ["accuracy"], "Metric"], [3020, 3021, ["accuracy"], "Metric"], [3075, 3076, ["accuracy"], "Metric"], [3108, 3109, ["accuracy"], "Metric"], [3169, 3171, ["belief", "maps"], "Method"], [3236, 3239, ["Convolutional", "Pose", "Machines"], "Method"], [3245, 3247, ["pose", "machine"], "Method"], [3250, 3252, ["deep", "architecture"], "Method"], [3330, 3333, ["sequential", "prediction", "framework"], "Method"], [3335, 3337, ["pose", "machine"], "Method"], [3344, 3346, ["deep", "architecture"], "Method"], [3355, 3357, ["pose", "machine"], "Method"], [3452, 3454, ["cost", "function"], "Metric"], [3496, 3499, ["stochastic", "gradient", "descend"], "Method"], [3527, 3529, ["convolutional", "layers"], "Method"], [3546, 3549, ["Addressing", "vanishing", "gradients"], "Task"], [3557, 3560, ["decomposable", "loss", "function"], "Method"], [3600, 3602, ["Intermediate", "supervision"], "Method"], [3610, 3612, ["full", "architecture"], "Method"], [3624, 3627, ["vanishing", "gradient", "problem"], "Task"], [3700, 3701, ["learning"], "Task"], [3731, 3733, ["model", "convergence"], "Task"], [3740, 3742, ["end", "learning"], "Task"], [3754, 3756, ["pose", "machine"], "Method"], [3760, 3762, ["convolutional", "architecture"], "Method"], [3780, 3783, ["high", "precision", "regime"], "Metric"], [3784, 3785, ["PCK@0.1"], "Metric"], [3792, 3795, ["low", "precision", "regime"], "Metric"], [3796, 3797, ["PCK@0.2"], "Metric"], [3801, 3803, ["training", "schemes"], "Method"], [3817, 3819, ["LSP", "dataset"], "Material"], [3824, 3825, ["PC"], "Material"], [3833, 3835, ["intermediate", "supervision"], "Task"], [3836, 3838, ["joint", "training"], "Task"], [3878, 3882, ["feed", "-", "forward", "fashion"], "Method"], [3928, 3930, ["training", "methods"], "Method"], [3936, 3938, ["joint", "training"], "Method"], [3949, 3951, ["stagewise", "training"], "Method"], [3961, 3965, ["jointly", "fine", "-", "tuning"], "Task"], [3977, 3978, ["accuracy"], "Metric"], [4007, 4009, ["LSP", "dataset"], "Material"], [4010, 4011, ["PC"], "Material"], [4081, 4082, ["LSP"], "Material"], [4083, 4085, ["MPII", "datasets"], "Material"], [4090, 4092, ["Quantitative", "Analysis"], "Task"], [4106, 4107, ["MPII"], "Material"], [4108, 4109, ["LSP"], "Material"], [4111, 4113, ["FLIC", "datasets"], "Material"], [4168, 4169, ["LSP"], "Material"], [4190, 4193, ["resizing", "and", "cropping"], "Task"], [4240, 4242, ["scale", "estimation"], "Method"], [4251, 4252, ["Caffe"], "Method"], [4257, 4259, ["deep", "learning"], "Task"], [4278, 4280, ["data", "augmentation"], "Task"], [4355, 4357, ["PCKh", "metric"], "Method"], [4362, 4364, ["error", "tolerance"], "Metric"], [4405, 4407, ["belief", "maps"], "Method"], [4524, 4528, ["PCKh", "-", "0.5", "score"], "Metric"], [4541, 4544, ["LSP", "training", "data"], "Material"], [4572, 4576, ["PCKh", "-", "0.5", "score"], "Metric"], [4584, 4587, ["LSP", "training", "data"], "Material"], [4630, 4631, ["accuracy"], "Metric"], [4665, 4666, ["accuracy"], "Metric"], [4672, 4673, ["precisions"], "Metric"], [4687, 4688, ["accuracy"], "Metric"], [4698, 4702, ["post", "-", "inference", "parsing"], "Method"], [4712, 4715, ["structured", "prediction", "task"], "Task"], [4733, 4735, ["location", "refinement"], "Task"], [4749, 4753, ["stride", "-", "8", "network"], "Method"], [4754, 4757, ["Leeds", "Sports", "Pose"], "Material"], [4758, 4759, ["LSP"], "Material"], [4768, 4772, ["Extended", "Leeds", "Sports", "Dataset"], "Material"], [4804, 4807, ["Percentage", "Correct", "Keypoints"], "Metric"], [4808, 4809, ["PCK"], "Metric"], [4818, 4820, ["augmentation", "scheme"], "Method"], [4823, 4825, ["MPI", "dataset"], "Material"], [4842, 4843, ["MPII"], "Material"], [4845, 4846, ["MPII"], "Material"], [4846, 4847, ["FLIC"], "Material"], [4847, 4848, ["LSP"], "Material"], [4858, 4859, ["MPII"], "Material"], [4860, 4861, ["LSP"], "Material"], [4862, 4863, ["FLIC"], "Material"], [4912, 4914, ["labeling", "quality"], "Metric"], [4918, 4919, ["LSP"], "Material"], [4927, 4929, ["LSP", "dataset"], "Material"], [4947, 4949, ["labeling", "quality"], "Metric"], [4979, 4980, ["accuracy"], "Metric"], [5011, 5012, ["PCK@0.2"], "Metric"], [5016, 5017, ["elbows"], "Metric"], [5046, 5047, ["PCK@0.05"], "Metric"], [5061, 5062, ["PCK@0.1"], "Metric"], [5066, 5069, ["Convolutional", "pose", "machines"], "Method"], [5075, 5077, ["end", "architecture"], "Method"], [5079, 5082, ["structured", "prediction", "problems"], "Task"], [5083, 5085, ["computer", "vision"], "Task"], [5089, 5094, ["graphical", "-", "model", "style", "inference"], "Method"], [5099, 5101, ["sequential", "architecture"], "Method"], [5103, 5105, ["convolutional", "networks"], "Method"], [5111, 5113, ["spatial", "models"], "Method"], [5137, 5139, ["computer", "vision"], "Task"], [5141, 5144, ["semantic", "image", "labeling"], "Task"], [5145, 5149, ["single", "image", "depth", "prediction"], "Task"], [5150, 5152, ["object", "detection"], "Task"], [5171, 5172, ["accuracy"], "Metric"], [5198, 5204, ["end", "-", "to", "-", "end", "architecture"], "Task"], [37, 39, ["pose", "machine"], "Method"], [79, 81, ["pose", "estimation"], "Task"], [201, 203, ["pose", "estimation"], "Task"], [811, 813, ["convolutional", "architectures"], "Method"], [902, 904, ["pose", "estimation"], "Task"], [1248, 1250, ["pose", "estimation"], "Task"], [1339, 1340, ["precision"], "Metric"], [1395, 1396, ["precision"], "Metric"], [1402, 1403, ["precision"], "Metric"], [2168, 2170, ["pose", "machine"], "Method"], [2344, 2345, ["precision"], "Metric"], [2865, 2867, ["convolutional", "layers"], "Method"], [2955, 2956, ["precision"], "Metric"], [3023, 3024, ["FLIC"], "Material"], [3156, 3157, ["precision"], "Metric"], [4288, 4289, ["MPII"], "Material"], [4302, 4303, ["MPII"], "Material"], [4745, 4746, ["precision"], "Metric"], [4795, 4796, ["PC"], "Material"], [4901, 4902, ["MPII"], "Material"], [4950, 4951, ["FLIC"], "Material"], [4959, 4960, ["FLIC"], "Material"], [5025, 5026, ["precision"], "Metric"]]} {"doc_id": "269730dbbabed8b8b5ba720e44a4c31b1f51e8f1", "text": ["Published", "as", "a", "conference", "paper", "at", "ICLR", "2017", "QUERY", "-", "REDUCTION", "NETWORKS", "FOR", "QUESTION", "ANSWERING", "section", ":", "ABSTRACT", "In", "this", "paper", ",", "we", "study", "the", "problem", "of", "question", "answering", "when", "reasoning", "over", "multiple", "facts", "is", "required", ".", "We", "propose", "Query", "-", "Reduction", "Network", "(", "QRN", ")", ",", "a", "variant", "of", "Recurrent", "Neural", "Network", "(", "RNN", ")", "that", "effectively", "handles", "both", "short", "-", "term", "(", "local", ")", "and", "long", "-", "term", "(", "global", ")", "sequential", "dependencies", "to", "reason", "over", "multiple", "facts", ".", "QRN", "considers", "the", "context", "sentences", "as", "a", "sequence", "of", "state", "-", "changing", "triggers", ",", "and", "reduces", "the", "original", "query", "to", "a", "more", "informed", "query", "as", "it", "observes", "each", "trigger", "(", "context", "sentence", ")", "through", "time", ".", "Our", "experiments", "show", "that", "QRN", "produces", "the", "state", "-", "of", "-", "the", "-", "art", "results", "in", "bAbI", "QA", "and", "dialog", "tasks", ",", "and", "in", "a", "real", "goal", "-", "oriented", "dialog", "dataset", ".", "In", "addition", ",", "QRN", "formulation", "allows", "parallelization", "on", "RNN", "'s", "time", "axis", ",", "saving", "an", "order", "of", "magnitude", "in", "time", "complexity", "for", "training", "and", "inference", ".", "section", ":", "INTRODUCTION", "In", "this", "paper", ",", "we", "address", "the", "problem", "of", "question", "answering", "(", "QA", ")", "when", "reasoning", "over", "multiple", "facts", "is", "required", ".", "For", "example", ",", "consider", "we", "know", "that", "Frogs", "eat", "insects", "and", "Flies", "are", "insects", ".", "Then", "answering", "Do", "frogs", "eat", "flies", "?", "requires", "reasoning", "over", "both", "of", "the", "above", "facts", ".", "Question", "answering", ",", "more", "specifically", "context", "-", "based", "QA", ",", "has", "been", "extensively", "studied", "in", "machine", "comprehension", "tasks", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "However", ",", "most", "of", "the", "datasets", "are", "primarily", "focused", "on", "lexical", "and", "syntactic", "understanding", ",", "and", "hardly", "concentrate", "on", "inference", "over", "multiple", "facts", ".", "Recently", ",", "several", "datasets", "aimed", "for", "testing", "multi", "-", "hop", "reasoning", "have", "emerged", ";", "among", "them", "are", "story", "-", "based", "QA", "and", "the", "dialog", "task", ".", "Recurrent", "Neural", "Network", "(", "RNN", ")", "and", "its", "variants", ",", "such", "as", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "[", "reference", "]", "and", "Gated", "Recurrent", "Unit", "(", "GRU", ")", "[", "reference", "]", ",", "are", "popular", "choices", "for", "modeling", "natural", "language", ".", "However", ",", "when", "used", "for", "multi", "-", "hop", "reasoning", "in", "question", "answering", ",", "purely", "RNN", "-", "based", "models", "have", "shown", "to", "perform", "poorly", ".", "This", "is", "largely", "due", "to", "the", "fact", "that", "RNN", "'s", "internal", "memory", "is", "inherently", "unstable", "over", "a", "long", "term", ".", "For", "this", "reason", ",", "most", "recent", "approaches", "in", "the", "literature", "have", "mainly", "relied", "on", "global", "attention", "mechanism", "and", "shared", "external", "memory", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "The", "attention", "mechanism", "allows", "these", "models", "to", "focus", "on", "a", "single", "sentence", "in", "each", "layer", ".", "They", "can", "sequentially", "read", "multiple", "relevant", "sentences", "from", "the", "memory", "with", "multiple", "layers", "to", "perform", "multi", "-", "hop", "reasoning", ".", "However", ",", "one", "major", "drawback", "of", "these", "standard", "attention", "mechanisms", "is", "that", "they", "are", "insensitive", "to", "the", "time", "step", "(", "memory", "address", ")", "of", "the", "sentences", "when", "accessing", "them", ".", "Our", "proposed", "model", ",", "Query", "-", "Reduction", "Network", "1", "(", "QRN", ")", ",", "is", "a", "single", "recurrent", "unit", "that", "addresses", "the", "long", "-", "term", "dependency", "problem", "of", "most", "RNN", "-", "based", "models", "by", "simplifying", "the", "recurrent", "update", ",", "while", "taking", "the", "advantage", "of", "RNN", "'s", "capability", "to", "model", "sequential", "data", "(", "Figure", "1", ")", ".", "QRN", "considers", "the", "context", "sentences", "as", "a", "sequence", "of", "state", "-", "changing", "triggers", ",", "and", "transforms", "(", "reduces", ")", "the", "original", "query", "to", "a", "more", "informed", "query", "as", "it", "observes", "each", "trigger", "through", "time", ".", "For", "instance", "in", "Figure", "1b", ",", "the", "original", "question", ",", "Where", "is", "the", "apple", "?", ",", "can", "not", "be", "directly", "answered", "by", "any", "single", "sentence", "from", "the", "story", ".", "After", "observing", "the", "first", "sentence", ",", "Sandra", "got", "the", "apple", "there", ",", "QRN", "transforms", "the", "original", "question", "to", "a", "reduced", "query", "Where", "is", "Sandra", "?", ",", "which", "is", "presumably", "1", "Code", "is", "publicly", "available", "at", ":", "seominjoon.github.io", "/", "qrn", "/", "Figure", "1", ":", "(", "1a", ")", "QRN", "unit", ",", "(", "1b", ")", "2", "-", "layer", "QRN", "on", "5", "-", "sentence", "story", ",", "and", "(", "1c", ")", "entire", "QA", "system", "(", "QRN", "and", "input", "/", "output", "modules", ")", ".", "x", ",", "q", ",", "ŷ", "are", "the", "story", ",", "question", "and", "predicted", "answer", "in", "natural", "language", ",", "respectively", ".", "x", "=", "x1", ",", ".", ".", ".", ",", "xT", ",", "q", ",", "ŷ", "are", "their", "corresponding", "vector", "representations", "(", "upright", "font", ")", ".", "α", "and", "ρ", "are", "update", "gate", "and", "reduce", "functions", ",", "respectively.ŷ", "is", "assigned", "to", "be", "h", "2", "5", ",", "the", "local", "query", "at", "the", "last", "time", "step", "in", "the", "last", "layer", ".", "Also", ",", "red", "-", "colored", "text", "is", "the", "inferred", "meanings", "of", "the", "vectors", "(", "see", "'", "Interpretations", "'", "of", "Section", "5.3", ")", ".", "easier", "to", "answer", "than", "the", "original", "question", "given", "the", "context", "provided", "by", "the", "first", "sentence", ".", "2", "Unlike", "RNN", "-", "based", "models", ",", "QRN", "'s", "candidate", "state", "(", "h", "t", "in", "Figure", "1a", ")", "does", "not", "depend", "on", "the", "previous", "hidden", "state", "(", "h", "t−1", ")", ".", "Compared", "to", "memory", "-", "based", "approaches", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ",", "QRN", "can", "better", "encodes", "locality", "information", "because", "it", "does", "not", "use", "a", "global", "memory", "access", "controller", "(", "circle", "nodes", "in", "Figure", "2", ")", ",", "and", "the", "query", "updates", "are", "performed", "locally", ".", "In", "short", ",", "the", "main", "contribution", "of", "QRN", "is", "threefold", ".", "First", ",", "QRN", "is", "a", "simple", "variant", "of", "RNN", "that", "reduces", "the", "query", "given", "the", "context", "sentences", "in", "a", "differentiable", "manner", ".", "Second", ",", "QRN", "is", "situated", "between", "the", "attention", "mechanism", "and", "RNN", ",", "effectively", "handling", "time", "dependency", "and", "long", "-", "term", "dependency", "problems", "of", "each", "technique", ",", "respectively", ".", "Hence", "it", "is", "well", "-", "suited", "for", "sequential", "data", "with", "both", "local", "and", "global", "interactions", "(", "note", "that", "QRN", "is", "not", "the", "replacement", "of", "RNN", ",", "which", "is", "arguably", "better", "for", "modeling", "complex", "local", "interactions", ")", ".", "Third", ",", "unlike", "most", "RNN", "-", "based", "models", ",", "QRN", "can", "be", "parallelized", "over", "time", "by", "computing", "candidate", "reduced", "queries", "(", "h", "t", ")", "directly", "from", "local", "input", "queries", "(", "q", "t", ")", "and", "context", "sentence", "vectors", "(", "x", "t", ")", ".", "In", "fact", ",", "the", "parallelizability", "of", "QRN", "implies", "that", "QRN", "does", "not", "suffer", "from", "the", "vanishing", "gradient", "problem", "of", "RNN", ",", "hence", "effectively", "addressing", "the", "long", "-", "term", "dependency", ".", "We", "experimentally", "demonstrate", "these", "contributions", "by", "achieving", "the", "state", "-", "of", "-", "the", "-", "art", "results", "on", "story", "-", "based", "QA", "and", "interactive", "dialog", "datasets", ".", "section", ":", "MODEL", "In", "story", "-", "based", "QA", "(", "or", "dialog", "dataset", ")", ",", "the", "input", "is", "the", "context", "as", "a", "sequence", "of", "sentences", "(", "story", "or", "past", "conversations", ")", "and", "a", "question", "in", "natural", "language", "(", "equivalent", "to", "the", "user", "'s", "last", "utterance", "in", "the", "dialog", ")", ".", "The", "output", "is", "the", "predicted", "answer", "to", "the", "question", "in", "natural", "language", "(", "the", "system", "'s", "next", "utterance", "in", "the", "dialog", ")", ".", "The", "only", "supervision", "provided", "during", "training", "is", "the", "answer", "to", "the", "question", ".", "In", "this", "paper", "we", "particularly", "focus", "on", "end", "-", "to", "-", "end", "solutions", ",", "i.e.", ",", "the", "only", "supervision", "comes", "from", "questions", "and", "answers", ",", "and", "we", "restrain", "from", "using", "manually", "defined", "rules", "or", "external", "language", "resources", ",", "such", "as", "lexicon", "or", "dependency", "parser", ".", "Let", "x", "1", ",", ".", ".", ".", ",", "x", "T", "denote", "the", "sequence", "of", "sentences", ",", "where", "T", "is", "the", "number", "of", "sentences", "in", "the", "story", ",", "and", "let", "q", "denote", "the", "question", ".", "Letŷ", "denote", "the", "predicted", "answer", ",", "and", "y", "denote", "the", "true", "answer", ".", "Our", "proposed", "system", "for", "end", "-", "to", "-", "end", "QA", "task", "is", "divided", "into", "three", "modules", "(", "Figure", "1c", ")", ":", "input", "module", ",", "QRN", "layers", ",", "and", "output", "module", ".", "Input", "module", ".", "Input", "module", "maps", "each", "sentence", "x", "t", "and", "the", "question", "q", "to", "d", "-", "dimensional", "vector", "space", ",", "x", "t", "∈", "R", "d", "and", "q", "t", "∈", "R", "d", ".", "We", "adopt", "a", "previous", "solution", "for", "the", "input", "module", "(", "details", "in", "Section", "5", ")", ".", "QRN", "layers", ".", "QRN", "layers", "use", "the", "sentence", "vectors", "and", "the", "question", "vector", "from", "the", "input", "module", "to", "obtain", "the", "predicted", "answer", "in", "vector", "space", ",", "ŷ", "∈", "R", "d", ".", "A", "QRN", "layer", "refers", "to", "the", "recurrent", "application", "of", "a", "QRN", "unit", ",", "which", "can", "be", "considered", "as", "a", "variant", "of", "RNN", "with", "two", "inputs", ",", "two", "outputs", ",", "and", "a", "hidden", "state", "(", "reduced", "query", ")", ",", "all", "of", "which", "operate", "in", "vector", "space", ".", "The", "details", "of", "the", "QRN", "module", "is", "explained", "throughout", "this", "section", "(", "2.1", ",", "2.2", ")", ".", "Output", "module", ".", "Output", "module", "mapsŷ", "obtained", "from", "QRN", "to", "a", "natural", "language", "answerŷ", ".", "Similar", "to", "the", "input", "module", ",", "we", "adopt", "a", "standard", "solution", "for", "the", "output", "module", "(", "details", "in", "Section", "5", ")", ".", "We", "first", "formally", "define", "the", "base", "model", "of", "a", "QRN", "unit", ",", "and", "then", "we", "explain", "how", "we", "connect", "the", "input", "and", "output", "modules", "to", "it", "(", "Section", "2.1", ")", ".", "We", "also", "present", "a", "few", "extensions", "to", "the", "network", "that", "can", "improve", "QRN", "'s", "performance", "(", "Section", "2.2", ")", ".", "Finally", ",", "we", "show", "that", "QRN", "can", "be", "parallelized", "over", "time", ",", "giving", "computational", "advantage", "over", "most", "RNN", "-", "based", "models", "by", "one", "order", "of", "magnitude", "(", "Section", "3", ")", ".", "section", ":", "QRN", "UNIT", "As", "an", "RNN", "-", "based", "model", ",", "QRN", "is", "a", "single", "recurrent", "unit", "that", "updates", "its", "hidden", "state", "(", "reduced", "query", ")", "through", "time", "and", "layers", ".", "Figure", "1a", "depicts", "the", "schematic", "structure", "of", "a", "QRN", "unit", ",", "and", "Figure", "1b", "demonstrates", "how", "layers", "are", "stacked", ".", "A", "QRN", "unit", "accepts", "two", "inputs", "(", "local", "query", "vector", "q", "t", "∈", "R", "d", "and", "sentence", "vector", "x", "t", "∈", "R", "d", ")", ",", "and", "two", "outputs", "(", "reduced", "query", "vector", "h", "t", "∈", "R", "d", ",", "which", "is", "similar", "to", "the", "hidden", "state", "in", "RNN", ",", "and", "the", "sentence", "vector", "x", "t", "from", "the", "input", "without", "modification", ")", ".", "The", "local", "query", "vector", "is", "not", "necessarily", "identical", "to", "the", "original", "query", "(", "question", ")", "vector", "q.", "In", "order", "to", "compute", "the", "outputs", ",", "we", "use", "update", "gate", "function", "α", ":", "Intuitively", ",", "the", "update", "gate", "function", "measures", "the", "relevance", "between", "the", "sentence", "and", "the", "local", "query", "and", "is", "used", "to", "update", "the", "hidden", "state", ".", "The", "reduce", "function", "transforms", "the", "local", "query", "input", "to", "a", "candidate", "state", "which", "is", "a", "new", "reduced", "(", "easier", ")", "query", "given", "the", "sentence", ".", "The", "outputs", "are", "calculated", "with", "the", "following", "equations", ":", "where", "z", "t", "is", "the", "scalar", "update", "gate", ",", "h", "t", "is", "the", "candidate", "reduced", "query", ",", "and", "h", "t", "is", "the", "final", "reduced", "query", "at", "time", "step", "t", ",", "σ", "(", "·", ")", "is", "sigmoid", "activation", ",", "tanh", "(", "·", ")", "is", "hyperboolic", "tangent", "activation", "(", "applied", "element", "-", "wise", ")", ",", "•", "is", "element", "-", "wise", "vector", "multiplication", ",", "and", "[", ";", "]", "is", "vector", "concatenation", "along", "the", "row", ".", "As", "a", "base", "case", ",", "h", "0", "=", "0", ".", "Here", "we", "have", "explicitly", "defined", "α", "and", "ρ", ",", "but", "they", "can", "be", "any", "reasonable", "differentiable", "functions", ".", "The", "update", "gate", "is", "similar", "to", "the", "global", "attention", "mechanism", "[", "reference", "][", "reference", "]", "in", "that", "it", "measures", "the", "similarity", "between", "the", "sentence", "(", "a", "memory", "slot", ")", "and", "the", "query", ".", "However", ",", "a", "significant", "difference", "is", "that", "the", "update", "gate", "is", "computed", "using", "sigmoid", "(", "σ", ")", "function", "on", "the", "current", "memory", "slot", "only", "(", "hence", "internally", "embedded", "within", "the", "unit", ")", ",", "whereas", "the", "global", "attention", "is", "computed", "using", "softmax", "function", "over", "the", "entire", "memory", "(", "hence", "globally", "defined", ")", ".", "The", "update", "gate", "can", "be", "rather", "considered", "as", "local", "sigmoid", "attention", ".", "section", ":", "Stacking", "layers", "We", "just", "showed", "the", "single", "-", "layer", "case", "of", "QRN", ",", "but", "QRN", "with", "multiple", "layers", "is", "able", "to", "perform", "reasoning", "over", "multiple", "facts", "more", "effectively", ",", "as", "shown", "in", "the", "example", "of", "Figure", "1b", ".", "In", "order", "to", "stack", "several", "layers", "of", "QRN", ",", "the", "outputs", "of", "the", "current", "layer", "are", "used", "as", "the", "inputs", "to", "the", "next", "layer", ".", "That", "is", ",", "using", "superscript", "k", "to", "denote", "the", "current", "layer", "'s", "index", "(", "assuming", "1", "-", "based", "indexing", ")", ",", "we", "let", "q", "k", "+", "1", "t", "=", "h", "k", "t", ".", "Note", "that", "x", "t", "is", "passed", "to", "the", "next", "layer", "without", "any", "modification", ",", "so", "we", "do", "not", "put", "a", "layer", "index", "on", "it", ".", "Bi", "-", "direction", ".", "So", "far", "we", "have", "assumed", "that", "QRN", "only", "needs", "to", "look", "at", "past", "sentences", ",", "whereas", "often", "times", ",", "query", "answers", "can", "depend", "on", "future", "sentences", ".", "For", "instance", ",", "consider", "a", "sentence", "\"", "John", "dropped", "the", "football", ".", "\"", "at", "time", "t.", "Then", ",", "even", "if", "there", "is", "no", "mention", "about", "the", "\"", "football", "\"", "in", "the", "past", "(", "at", "time", "i", "<", "t", ")", ",", "it", "can", "be", "implied", "that", "\"", "John", "\"", "has", "the", "\"", "football", "\"", "at", "the", "current", "time", "t.", "In", "order", "to", "incorporate", "the", "future", "dependency", ",", "we", "obtain", "−", "→", "h", "t", "and", "←", "−", "h", "t", "in", "both", "forward", "and", "backward", "directions", ",", "respectively", ",", "using", "Equation", "3", ".", "We", "then", "add", "them", "together", "to", "get", "q", "t", "for", "the", "next", "layer", ".", "That", "is", ",", "[", "reference", "]", "are", "shared", "between", "the", "two", "directions", ".", "Connecting", "input", "and", "output", "modules", ".", "Figure", "1c", "depicts", "how", "QRN", "is", "connected", "with", "the", "input", "and", "output", "modules", ".", "In", "the", "first", "layer", "of", "QRN", ",", "q", "1", "t", "=", "q", "for", "all", "t", ",", "where", "q", "is", "obtained", "from", "the", "input", "module", "by", "processing", "the", "natural", "language", "question", "input", "q.", "x", "t", "is", "also", "obtained", "from", "x", "t", "by", "the", "same", "input", "module", ".", "The", "output", "at", "the", "last", "time", "step", "in", "the", "last", "layer", "is", "passed", "to", "the", "output", "module", ".", "That", "is", ",", "y", "=", "h", "K", "t", "where", "K", "represent", "the", "number", "of", "layers", "in", "the", "network", ".", "Then", "the", "output", "module", "gives", "the", "predicted", "answerŷ", "in", "natural", "language", ".", "section", ":", "EXTENSIONS", "Here", "we", "introduce", "a", "few", "extensions", "of", "QRN", ",", "and", "later", "in", "our", "experiments", ",", "we", "test", "QRN", "'s", "performance", "with", "and", "without", "each", "of", "these", "extensions", ".", "Reset", "gate", ".", "Inspired", "by", "GRU", "[", "reference", "]", ",", "we", "found", "that", "it", "is", "useful", "to", "allow", "the", "QRN", "unit", "to", "reset", "(", "nullify", ")", "the", "candidate", "reduced", "query", "(", "i.e.", ",", "h", "t", ")", "when", "necessary", ".", "For", "this", "we", "use", "a", "reset", "gate", "function", "β", ":", ",", "which", "can", "be", "defined", "similarly", "to", "the", "update", "gate", "function", ":", "where", "W", "(", "r", ")", "∈", "R", "1×d", "is", "a", "weight", "matrix", ",", "and", "b", "(", "r", ")", "∈", "R", "is", "a", "bias", "term", ".", "Equation", "3", "is", "rewritten", "as", "Note", "that", "we", "do", "not", "use", "the", "reset", "gate", "in", "the", "last", "layer", ".", "Vector", "gates", ".", "As", "in", "LSTM", "and", "GRU", ",", "update", "and", "reset", "gates", "can", "be", "vectors", "instead", "of", "scalar", "values", "for", "fine", "-", "controlled", "gating", ".", "For", "vector", "gates", ",", "we", "modify", "the", "row", "dimension", "of", "weights", "and", "biases", "in", "Equation", "1", "and", "5", "from", "1", "to", "d.", "Then", "we", "obtain", "z", "t", ",", "r", "t", "∈", "R", "d", "(", "instead", "of", "z", "t", ",", "r", "t", "∈", "R", ")", ",", "and", "these", "can", "be", "element", "-", "wise", "multiplied", "(", "•", ")", "instead", "of", "being", "broadcasted", "in", "Equation", "3", "and", "6", ".", "section", ":", "PARALLELIZATION", "An", "important", "advantage", "of", "QRN", "is", "that", "the", "recurrent", "updates", "in", "Equation", "3", "and", "5", "can", "be", "computed", "in", "parallel", "across", "time", ".", "This", "is", "in", "contrast", "with", "most", "RNN", "-", "based", "models", "that", "can", "not", "be", "parallelized", ",", "where", "computing", "the", "candidate", "hidden", "state", "at", "time", "t", "explicitly", "requires", "the", "previous", "hidden", "state", ".", "In", "QRN", ",", "the", "final", "reduced", "queries", "(", "h", "t", ")", "can", "be", "decomposed", "into", "computing", "over", "candidate", "reduced", "queries", "(", "h", "t", ")", ",", "without", "looking", "at", "the", "previous", "reduced", "query", ".", "Here", "we", "primarily", "show", "that", "the", "query", "update", "in", "Equation", "3", "can", "be", "parallelized", "by", "rewriting", "the", "equation", "with", "matrix", "operations", ".", "The", "extension", "to", "Equation", "5", "is", "straightforward", ".", "The", "proof", "for", "QRN", "with", "vector", "gates", "is", "shown", "in", "Appendix", "B.", "The", "recursive", "definition", "of", "Equation", "3", "can", "be", "explicitly", "written", "as", "Let", "b", "i", "=", "log", "(", "1", "−", "z", "i", ")", "for", "brevity", ".", "Then", "we", "can", "rewrite", "Equation", "7", "as", "the", "following", "equation", ":", "[", "reference", "]", "Figure", "2", ":", "The", "schematics", "of", "QRN", "and", "the", "two", "state", "-", "of", "-", "the", "-", "art", "models", ",", "End", "-", "to", "-", "End", "Memory", "Networks", "(", "N2N", ")", "and", "Improved", "Dynamic", "Memory", "Networks", "(", "DMN", "+", ")", ",", "simplified", "to", "emphasize", "the", "differences", "among", "the", "models", ".", "AGRU", "is", "a", "variant", "of", "GRU", "where", "the", "update", "gate", "is", "replaced", "with", "soft", "attention", ",", "proposed", "by", "[", "reference", "]", ".", "where", "L", ",", "L", "∈", "R", "T", "×T", "are", "lower", "and", "strictly", "lower", "triangular", "matrices", "of", "1", "'s", ",", "respectively", ",", "•", "is", "elementwise", "multiplication", ",", "and", "B", "is", "a", "matrix", "where", "T", "b", "'s", "are", "tiled", "across", "the", "column", ",", "i.", "section", ":", "RELATED", "WORK", "QRN", "is", "inspired", "by", "RNN", "-", "based", "models", "with", "gating", "mechanism", ",", "such", "as", "LSTM", "[", "reference", "]", "and", "GRU", "[", "reference", "]", ".", "While", "GRU", "and", "LSTM", "use", "the", "previous", "hidden", "state", "and", "the", "current", "input", "to", "obtain", "the", "candidate", "hidden", "state", ",", "QRN", "only", "uses", "the", "current", "two", "inputs", "to", "obtain", "the", "candidate", "reduced", "query", "(", "equivalent", "to", "candidate", "hidden", "state", ")", ".", "We", "conjecture", "that", "this", "not", "only", "gives", "computational", "advantage", "via", "parallelization", ",", "but", "also", "makes", "training", "easier", ",", "i.e.", ",", "avoiding", "vanishing", "gradient", "(", "which", "is", "critical", "for", "long", "-", "term", "dependency", ")", ",", "overfitting", "(", "by", "simplifying", "the", "model", ")", ",", "and", "converging", "to", "local", "minima", ".", "The", "idea", "of", "structurally", "simplifying", "(", "constraining", ")", "RNNs", "for", "learning", "longer", "-", "term", "patterns", "has", "been", "explored", "in", "recent", "previous", "work", ",", "such", "as", "Structurally", "Constrained", "Recurrent", "Network", "[", "reference", "]", "and", "Strongly", "-", "Typed", "Recurrent", "Neural", "Network", "(", "STRNN", ")", "[", "reference", "]", ".", "QRN", "is", "similar", "to", "STRNN", "in", "that", "both", "architectures", "use", "gating", "mechanism", ",", "and", "the", "gates", "and", "the", "candidate", "hidden", "states", "do", "not", "depend", "on", "the", "previous", "hidden", "states", ",", "which", "simplifies", "the", "recurrent", "relation", ".", "However", ",", "QRN", "can", "be", "distinguished", "from", "STRNN", "in", "three", "ways", ".", "First", ",", "QRN", "'s", "update", "gate", "simulates", "attention", "mechanism", ",", "measuring", "the", "relevance", "between", "the", "input", "sentence", "and", "query", ".", "On", "the", "other", "hand", ",", "the", "gates", "in", "STRNN", "can", "be", "considered", "as", "the", "simplification", "of", "LSTM", "/", "GRU", "by", "removing", "their", "dependency", "on", "previous", "hidden", "state", ".", "Second", ",", "QRN", "is", "an", "RNN", "that", "is", "natively", "compatible", "with", "context", "-", "based", "QA", "tasks", ",", "where", "the", "QRN", "unit", "accepts", "two", "inputs", ",", "i.e.", "each", "context", "sentence", "and", "query", ".", "This", "is", "distinct", "from", "STRNN", "which", "has", "only", "one", "input", ".", "Third", ",", "we", "show", "that", "QRN", "is", "timewise", "-", "parallelizable", "on", "GPUs", ".", "Our", "parallelization", "algorithm", "is", "also", "applicable", "to", "STRNN", ".", "End", "-", "to", "-", "end", "Memory", "Network", "(", "N2N", ")", "[", "reference", "]", "uses", "external", "memory", "with", "multi", "-", "layer", "attention", "mechanism", "to", "focus", "on", "sentences", "that", "are", "relevant", "to", "the", "question", ".", "There", "are", "two", "key", "differences", "between", "N2N", "and", "our", "QRN", ".", "First", ",", "N2N", "summarizes", "the", "entire", "memory", "in", "each", "layer", "to", "control", "the", "attention", "in", "the", "next", "layer", "(", "circle", "nodes", "in", "Figure", "2b", ")", ".", "Instead", ",", "QRN", "does", "not", "have", "any", "controller", "node", "(", "Figure", "2a", ")", "and", "is", "able", "to", "focus", "on", "relevant", "sentences", "through", "the", "update", "gate", "that", "is", "internally", "embodied", "within", "its", "unit", ".", "Second", ",", "N2N", "adds", "time", "-", "dependent", "trainable", "weights", "to", "the", "sentence", "representations", "to", "model", "the", "time", "dependency", "of", "the", "sentences", "(", "as", "discussed", "in", "Section", "1", ")", ".", "QRN", "does", "not", "need", "such", "additional", "weights", "as", "its", "inherent", "RNN", "architecture", "allows", "QRN", "to", "effectively", "model", "the", "time", "dependency", ".", "Neural", "Reasoner", "[", "reference", "]", "and", "Gated", "End", "-", "toend", "Memory", "Network", "[", "reference", "]", ")", "are", "variants", "of", "MemN2N", "that", "share", "its", "fundamental", "characteristics", ".", "Improved", "Dynamic", "Memory", "Network", "(", "DMN", "+", ")", "[", "reference", "]", "uses", "the", "hybrid", "of", "the", "attention", "mechanism", "and", "the", "RNN", "architecture", "to", "model", "the", "sequence", "of", "sentences", ".", "It", "consists", "of", "two", "distinct", "GRUs", ",", "one", "for", "the", "time", "axis", "(", "rectangle", "nodes", "in", "Figure", "2c", ")", "and", "one", "for", "the", "layer", "axis", "(", "circle", "nodes", "in", "Figure", "2c", ")", ".", "Note", "that", "the", "update", "gate", "of", "the", "GRU", "for", "the", "time", "axis", "is", "replaced", "with", "external", "softmax", "attention", "weights", ".", "DMN", "+", "uses", "the", "time", "-", "axis", "GRU", "to", "summarizes", "the", "entire", "memory", "in", "each", "layer", ",", "and", "then", "the", "layer", "-", "axis", "GRU", "controls", "the", "attention", "weights", "in", "each", "layer", ".", "In", "contrast", ",", "QRN", "is", "simply", "a", "single", "recurrent", "unit", "without", "any", "controller", "node", ".", "section", ":", "EXPERIMENTS", "5.1", "DATA", "bAbI", "story", "-", "based", "QA", "dataset", "bAbI", "story", "-", "based", "QA", "dataset", "is", "composed", "of", "20", "different", "tasks", "(", "Appendix", "A", ")", ",", "each", "of", "which", "has", "1", ",", "000", "(", "1k", ")", "synthetically", "-", "generated", "story", "-", "question", "pair", ".", "A", "story", "can", "be", "as", "short", "as", "two", "sentences", "and", "as", "long", "as", "200", "+", "sentences", ".", "A", "system", "is", "evaluated", "on", "the", "accuracy", "of", "getting", "the", "correct", "answers", "to", "the", "questions", ".", "The", "answers", "are", "single", "words", "or", "lists", "(", "e.g.", "\"", "football", ",", "apple", "\"", ")", ".", "Answering", "questions", "in", "each", "task", "requires", "selecting", "a", "set", "of", "relevant", "sentences", "and", "applying", "different", "kinds", "of", "logical", "reasoning", "over", "them", ".", "The", "dataset", "also", "includes", "10k", "training", "data", "(", "for", "each", "task", ")", ",", "which", "allows", "training", "more", "complex", "models", ".", "Note", "that", "DMN", "+", "[", "reference", "]", "only", "reports", "on", "the", "10k", "dataset", ".", "bAbI", "dialog", "dataset", "bAbI", "dialog", "dataset", "consists", "of", "5", "different", "tasks", "(", "Table", "3", ")", ",", "each", "of", "which", "has", "1k", "synthetically", "-", "generated", "goal", "-", "oriented", "dialogs", "between", "a", "user", "and", "the", "system", "in", "the", "domain", "of", "restaurant", "reservation", ".", "Each", "dialog", "is", "as", "long", "as", "96", "utterances", "and", "comes", "with", "external", "knowledge", "base", "(", "KB", ")", "providing", "information", "of", "each", "restaurant", ".", "The", "authors", "also", "provide", "Out", "-", "Of", "-", "Vocabulary", "(", "OOV", ")", "version", "of", "the", "dataset", ",", "where", "many", "of", "the", "words", "and", "KB", "keywords", "in", "test", "data", "are", "not", "seen", "during", "training", ".", "A", "system", "is", "evaluated", "on", "the", "accuracy", "of", "its", "response", "to", "each", "utterance", "of", "the", "user", ",", "choosing", "from", "up", "to", "2500", "possible", "candidate", "responses", ".", "A", "system", "is", "required", "not", "only", "to", "understand", "the", "user", "'s", "request", "but", "also", "refer", "to", "previous", "conversations", "in", "order", "to", "obtain", "the", "context", "information", "of", "the", "current", "conversation", ".", "transformed", "the", "Second", "Dialog", "State", "Tracking", "Challenge", "(", "DSTC2", ")", "dataset", "[", "reference", "]", "into", "the", "same", "format", "as", "the", "bAbI", "dialog", "dataset", ",", "for", "the", "measurement", "of", "performance", "on", "a", "real", "dataset", ".", "Each", "dialog", "can", "be", "as", "long", "as", "800", "+", "utterances", ",", "and", "a", "system", "needs", "to", "choose", "from", "2407", "possible", "candidate", "responses", "for", "each", "utterance", "of", "the", "user", ".", "Note", "that", "the", "evaluation", "metric", "of", "the", "original", "DSTC2", "is", "different", "from", "that", "of", "the", "transformed", "DSTC2", ",", "so", "previous", "work", "on", "the", "original", "DSTC2", "should", "not", "be", "directly", "compared", "to", "our", "work", ".", "We", "will", "refer", "to", "this", "transformed", "DSTC2", "dataset", "by", "\"", "Task", "6", "\"", "of", "dialog", "dataset", ".", "section", ":", "DSTC2", "(", "Task", "6", ")", "dialog", "dataset", "section", ":", "MODEL", "DETAILS", "Input", "Module", ".", "In", "the", "input", "module", ",", "we", "are", "given", "sentences", "(", "previous", "conversations", "in", "dialog", ")", "x", "t", "and", "a", "question", "(", "most", "recent", "user", "utterance", ")", "q", ",", "and", "we", "want", "to", "obtain", "their", "vector", "representations", ",", "x", "t", ",", "q", "∈", "R", "d", ".", "We", "use", "a", "trainable", "embedding", "matrix", "A", "∈", "R", "d×V", "to", "encode", "the", "one", "-", "hot", "vector", "of", "each", "word", "x", "tj", "in", "each", "sentence", "x", "t", "into", "a", "d", "-", "dimensional", "vector", "x", "tj", "∈", "R", "d", ".", "Then", "the", "sentence", "representation", "x", "t", "is", "obtained", "by", "Position", "Encoder", ".", "The", "same", "encoder", "with", "the", "same", "embedding", "matrix", "is", "also", "used", "to", "obtain", "the", "question", "vector", "q", "from", "q.", "Output", "Module", "for", "story", "-", "based", "QA", ".", "In", "the", "output", "module", ",", "we", "are", "given", "the", "vector", "representation", "of", "the", "predicted", "answerŷ", "and", "we", "want", "to", "obtain", "the", "natural", "language", "form", "of", "the", "answer", ",", "ŷ", ".", "We", "use", "a", "V", "-", "way", "single", "-", "layer", "softmax", "classifier", "to", "mapŷ", "to", "a", "V", "-", "dimensional", "sparse", "vector", ",", "v", "=", "softmax", "W", "(", "y", ")", "ŷ", "∈", "R", "V", ",", "where", "W", "(", "y", ")", "∈", "R", "V", "×d", "is", "a", "weight", "matrix", ".", "Then", "the", "final", "answerŷ", "is", "simply", "the", "argmax", "word", "inv", ".", "To", "handle", "questions", "with", "multiple", "-", "word", "answers", ",", "we", "consider", "each", "of", "them", "as", "a", "single", "word", "that", "contains", "punctuations", "such", "as", "space", "and", "comma", ",", "and", "put", "it", "in", "the", "vocabulary", ".", "Output", "Module", "for", "dialog", ".", "We", "use", "a", "fixed", "number", "single", "-", "layer", "softmax", "classifiers", ",", "each", "of", "which", "is", "similar", "to", "that", "of", "the", "sotry", "-", "based", "QA", "model", ",", "to", "sequentially", "output", "each", "word", "of", "the", "system", "'s", "response", ".", "While", "it", "is", "similar", "in", "spirit", "to", "the", "RNN", "decoder", "[", "reference", "]", ",", "our", "output", "module", "does", "not", "have", "a", "recurrent", "hidden", "state", "or", "gating", "mechanism", ".", "Instead", ",", "it", "solely", "uses", "the", "final", "ouptut", "of", "the", "QRN", ",", "ŷ", ",", "and", "the", "current", "word", "output", "to", "influence", "the", "prediction", "of", "the", "next", "word", "among", "possible", "candidates", ".", "Training", ".", "We", "withhold", "10", "%", "of", "the", "training", "for", "development", ".", "We", "use", "the", "hidden", "state", "size", "of", "50", "by", "deafult", ".", "Batch", "sizes", "of", "32", "for", "bAbI", "story", "-", "based", "QA", "1k", ",", "bAbI", "dialog", "and", "DSTC2", "dialog", ",", "and", "128", "for", "bAbI", "QA", "10k", "are", "used", ".", "The", "weights", "in", "the", "input", "and", "output", "modules", "are", "initialized", "with", "zero", "mean", "and", "the", "standard", "deviation", "of", "1", "/", "√", "d.", "Weights", "in", "the", "QRN", "unit", "are", "initialized", "using", "techniques", "by", "[", "reference", "]", ",", "and", "are", "tied", "across", "the", "layers", ".", "Forget", "bias", "of", "2.5", "is", "used", "for", "update", "gates", "(", "no", "bias", "for", "reset", "gates", ")", ".", "L2", "weight", "decay", "of", "0.001", "(", "0.0005", "for", "QA", "10k", ")", "is", "used", "for", "all", "weights", ".", "The", "loss", "function", "is", "the", "cross", "entropy", "betweenv", "and", "the", "one", "-", "hot", "vector", "of", "the", "true", "answer", ".", "The", "loss", "is", "minimized", "by", "stochastic", "gradient", "descent", "for", "maximally", "500", "epochs", ",", "but", "training", "is", "early", "stopped", "if", "the", "loss", "on", "the", "development", "data", "does", "not", "decrease", "for", "50", "epochs", ".", "The", "learning", "rate", "is", "controlled", "by", "AdaGrad", "[", "reference", "]", "with", "the", "initial", "learning", "rate", "of", "0.5", "(", "0.1", "for", "QA", "10k", ")", ".", "Since", "the", "model", "is", "sensitive", "to", "the", "weight", "initialization", ",", "we", "repeat", "each", "training", "procedure", "10", "times", "(", "50", "times", "for", "10k", ")", "with", "the", "new", "random", "initialization", "of", "the", "weights", "and", "report", "the", "result", "on", "the", "test", "data", "with", "the", "lowest", "loss", "on", "the", "development", "data", ".", "section", ":", "RESULTS", ".", "We", "compare", "our", "model", "with", "baselines", "and", "previous", "state", "-", "of", "-", "the", "-", "art", "models", "on", "story", "-", "based", "and", "dialog", "tasks", "(", "Table", "1", ")", ".", "These", "include", "LSTM", "[", "reference", "]", ",", "End", "-", "to", "-", "end", "Memory", "Networks", "(", "N2N", ")", "[", "reference", "]", ",", "Dynamic", "Memory", "Networks", "(", "DMN", "+", ")", "[", "reference", "]", ",", "Gated", "End", "-", "to", "-", "end", "Memory", "Networks", "(", "GMemN2N", ")", "[", "reference", "]", ",", "and", "Differentiable", "Neural", "Computer", "(", "DNC", ")", "[", "reference", "]", ".", "Story", "-", "based", "QA", ".", "Table", "1", "(", "top", ")", "reports", "the", "summary", "of", "results", "of", "our", "model", "(", "QRN", ")", "and", "previous", "work", "on", "bAbI", "QA", "(", "task", "-", "wise", "results", "are", "shown", "in", "Table", "2", "in", "Appendix", ")", ".", "In", "1k", "data", ",", "QRN", "'s", "'", "2r", "'", "(", "2", "layers", "+", "reset", "gate", "+", "d", "=", "50", ")", "outperforms", "all", "other", "models", "by", "a", "large", "margin", "(", "2.8", "+", "%", ")", ".", "In", "10k", "dataset", ",", "the", "average", "accuracy", "of", "QRN", "'s", "'", "6r200", "'", "(", "6", "layers", "+", "reset", "gate", "+", "d", "=", "200", ")", "model", "outperforms", "all", "previous", "models", "by", "a", "large", "margin", "(", "2.5", "+", "%", ")", ",", "achieving", "a", "nearly", "perfect", "score", "of", "99.7", "%", ".", "Dialog", ".", "Table", "1", "(", "bottom", ")", "reports", "the", "summary", "of", "the", "results", "of", "our", "model", "(", "QRN", ")", "and", "previous", "work", "on", "bAbI", "dialog", "and", "Task", "6", "dialog", "(", "task", "-", "wise", "results", "are", "shown", "in", "Table", "3", "in", "Appendix", ")", ".", "As", "done", "in", "previous", "work", "[", "reference", "]", ",", "we", "also", "report", "results", "when", "we", "use", "'", "Match", "'", "for", "dialogs", ".", "'", "Match", "'", "is", "the", "extension", "to", "the", "model", "which", "additionally", "takes", "as", "input", "whether", "each", "answer", "candidate", "matches", "with", "context", "(", "more", "details", "on", "Appendix", ")", ".", "QRN", "outperforms", "previous", "work", "by", "a", "large", "margin", "(", "2.0", "+", "%", ")", "in", "every", "comparison", ".", "Ablations", ".", "We", "test", "four", "types", "of", "ablations", "(", "also", "discussed", "in", "Section", "2.2", ")", ":", "number", "of", "layers", "(", "1", ",", "2", ",", "3", ",", "or", "6", ")", ",", "reset", "gate", "(", "r", ")", ",", "and", "gate", "vectorization", "(", "v", ")", "and", "the", "dimension", "of", "the", "hidden", "vector", "(", "50", ",", "100", ")", ".", "We", "show", "a", "subset", "of", "combinations", "of", "the", "ablations", "for", "bAbI", "QA", "in", "Table", "1", "and", "Table", "2", ";", "other", "combinations", "performed", "poorly", "and", "/", "or", "did", "not", "give", "interesting", "observations", ".", "According", "to", "the", "ablation", "results", ",", "we", "infer", "that", ":", "(", "a", ")", "When", "the", "number", "of", "layers", "is", "only", "one", ",", "the", "model", "lacks", "reasoning", "capability", ".", "In", "the", "case", "of", "1k", "dataset", ",", "when", "there", "are", "too", "many", "layers", "(", "6", ")", ",", "it", "seems", "correctly", "training", "the", "model", "becomes", "increasingly", "difficult", ".", "In", "the", "case", "of", "10k", "dataset", ",", "many", "layers", "(", "6", ")", "and", "hidden", "dimensions", "(", "200", ")", "Parallelization", ".", "We", "implement", "QRN", "with", "and", "without", "parallelization", "in", "TensorFlow", "[", "reference", "]", ")", "on", "a", "single", "Titan", "X", "GPU", "to", "qunaitify", "the", "computational", "gain", "of", "the", "parallelization", ".", "For", "QRN", "without", "parallelization", ",", "we", "use", "the", "RNN", "library", "provided", "by", "TensorFlow", ".", "QRN", "with", "parallelization", "gives", "6.2", "times", "faster", "training", "and", "inference", "than", "QRN", "without", "parallelization", "on", "average", ".", "We", "expect", "that", "the", "speedup", "can", "be", "even", "higher", "for", "datasets", "with", "larger", "context", ".", "Interpretations", ".", "An", "advantage", "of", "QRN", "is", "that", "the", "intermediate", "query", "updates", "are", "interpretable", ".", "Figure", "1", "shows", "intermediate", "local", "queries", "(", "q", "k", "t", ")", "interpreted", "in", "natural", "language", ",", "such", "as", "\"", "Where", "is", "Sandra", "?", "\"", ".", "In", "order", "to", "obtain", "these", ",", "we", "place", "a", "decoder", "on", "the", "input", "question", "embedding", "q", "and", "add", "its", "loss", "for", "recovering", "the", "question", "to", "the", "classification", "loss", "(", "similarly", "to", "[", "reference", "]", ")", ".", "We", "then", "use", "the", "same", "decoder", "to", "decode", "the", "intermediate", "queries", ".", "This", "helps", "us", "understand", "the", "flow", "of", "information", "in", "the", "networks", ".", "In", "Figure", "1", ",", "the", "question", "Where", "is", "apple", "?", "is", "transformed", "into", "Where", "is", "Sandra", "?", "at", "t", "=", "1", ".", "At", "t", "=", "2", ",", "as", "Sandra", "dropped", "the", "apple", ",", "the", "apple", "is", "no", "more", "relevant", "to", "Sandra", ".", "We", "obtain", "Where", "is", "Daniel", "?", "at", "time", "t", "=", "3", ",", "and", "it", "is", "propagated", "until", "t", "=", "5", ",", "where", "we", "observe", "a", "sentence", "(", "fact", ")", "that", "can", "be", "used", "to", "answer", "the", "query", ".", "Visualization", ".", "Figure", "3", "shows", "vizualization", "of", "the", "(", "scalar", ")", "magnitudes", "of", "update", "and", "reset", "gates", "on", "story", "sentences", "and", "dialog", "utterances", ".", "More", "visualizations", "are", "shown", "in", "Appendices", ":", "Figure", "4", "and", "Figure", "5", ".", "In", "Figure", "3", ",", "we", "observe", "high", "values", "on", "facts", "that", "provide", "information", "to", "answer", "question", "(", "the", "system", "'s", "next", "utterance", "for", "dialog", ")", ".", "In", "QA", "Task", "2", "example", "(", "top", "left", ")", ",", "we", "observe", "high", "update", "gate", "values", "in", "the", "first", "layer", "on", "facts", "that", "state", "who", "has", "the", "apple", ",", "and", "in", "the", "second", "layer", ",", "the", "high", "update", "gate", "values", "are", "on", "those", "that", "inform", "where", "that", "person", "went", "to", ".", "We", "also", "observe", "that", "the", "forward", "reset", "gate", "at", "t", "=", "2", "in", "the", "first", "layer", "(", "−", "→", "r", "1", "2", ")", "is", "low", ",", "which", "is", "signifying", "that", "apple", "no", "more", "belongs", "to", "Sandra", ".", "In", "dialog", "Task", "3", "(", "bottom", "left", ")", ",", "the", "model", "is", "able", "to", "infer", "that", "three", "restaurants", "are", "already", "recommended", "so", "that", "it", "can", "recommend", "another", "one", ".", "In", "dialog", "Task", "6", "(", "bottom", ")", ",", "the", "model", "focuses", "on", "the", "sentences", "containing", "Spanish", ",", "and", "does", "not", "concentrate", "much", "on", "other", "facts", "such", "as", "I", "do", "n't", "care", ".", "section", ":", "CONCLUSION", "In", "this", "paper", ",", "we", "introduce", "Query", "-", "Reduction", "Network", "(", "QRN", ")", "to", "answer", "context", "-", "based", "questions", "and", "carry", "out", "conversations", "with", "users", "that", "require", "multi", "-", "hop", "reasoning", ".", "We", "show", "the", "state", "-", "of", "-", "theart", "results", "in", "the", "three", "datasets", "of", "story", "-", "based", "QA", "and", "dialog", ".", "We", "model", "a", "story", "or", "a", "dialog", "as", "a", "sequence", "of", "state", "-", "changing", "triggers", "and", "compute", "the", "final", "answer", "to", "the", "question", "or", "the", "system", "'s", "next", "utterance", "by", "recurrently", "updating", "(", "or", "reducing", ")", "the", "query", ".", "QRN", "is", "situated", "between", "the", "attention", "mechanism", "and", "RNN", ",", "effectively", "handling", "time", "dependency", "and", "long", "-", "term", "dependency", "problems", "of", "each", "technique", ",", "respectively", ".", "It", "addresses", "the", "long", "-", "term", "dependency", "problem", "of", "most", "RNNs", "by", "simplifying", "the", "recurrent", "update", ",", "in", "which", "the", "candidate", "hidden", "state", "(", "reduced", "query", ")", "does", "not", "depend", "on", "the", "previous", "state", ".", "Moreover", ",", "QRN", "can", "be", "parallelized", "and", "can", "address", "the", "well", "-", "known", "problem", "of", "RNN", "'s", "vanishing", "gradients", ".", "section", ":", "A", "TASK", "-", "WISE", "RESULTS", "Here", "we", "provide", "detailed", "per", "-", "task", "breakdown", "of", "our", "results", "in", "QA", "(", "Table", "2", ")", "and", "dialog", "datasets", "(", "Table", "3", ")", ".", "Table", "2", ":", "bAbI", "QA", "dataset", "error", "rates", "(", "%", ")", "of", "QRN", "and", "previous", "work", ":", "LSTM", ",", "End", "-", "to", "-", "end", "Memory", "Networks", "(", "N2N", ")", "[", "reference", "]", ",", "Dynamic", "Memory", "Networks", "(", "DMN", "+", ")", "[", "reference", "]", ",", "Gated", "End", "-", "to", "-", "end", "Memory", "Networks", "(", "GMemN2N", ")", "[", "reference", "]", ".", "Results", "within", "each", "task", "of", "Differentiable", "Neural", "Computer", "(", "DNC", ")", "were", "not", "provided", "in", "its", "paper", "[", "reference", "]", "Table", "3", ":", "bAbI", "dialog", "and", "DSTC2", "dialog", "dataset", "average", "error", "rates", "(", "%", ")", "of", "QRN", "and", "previous", "work", ":", "End", "-", "to", "-", "end", "Memory", "Networks", "(", "N2N", ")", "and", "Gated", "End", "-", "to", "-", "end", "Memory", "Networks", "(", "GMemN2N", "[", "reference", "]", ")", ".", "For", "QRN", ",", "a", "number", "in", "the", "front", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "indicates", "the", "number", "of", "layers", "and", "a", "number", "in", "the", "back", "(", "100", ")", "indicates", "the", "dimension", "of", "hidden", "vector", ",", "while", "the", "default", "value", "is", "50", ".", "'", "r", "'", "indicates", "that", "the", "reset", "gate", "is", "used", ",", "'", "v", "'", "indicates", "that", "the", "gates", "were", "vectorized", ",", "and", "'", "+", "'", "indicates", "that", "'", "match", "'", "was", "used", ".", "section", ":", "B", "VECTOR", "GATE", "PARALLELIZATION", "For", "vector", "gates", ",", "we", "have", "z", "t", "∈", "R", "d", "instead", "of", "z", "t", "∈", "R.", "Therefore", "the", "following", "equation", "replaces", "Equation", "7", ":", "where", "z", "j", "k", "is", "the", "k", "-", "th", "column", "vector", "of", "z", "j", ".", "Let", "b", "ij", "=", "log", "(", "1", "−", "z", "i", "j", ")", "for", "brevity", ".", "Then", ",", "we", "can", "rewrite", "Equation", "8", "as", "following", ":", ".", ".", ".", "where", "L", ",", "L", "∈", "R", "T", "×T", "are", "lower", "and", "strictly", "lower", "triangular", "matrices", "of", "1", "'s", "are", "tiled", "across", "the", "column", ".", "section", ":", "C", "MODEL", "DETAILS", "Match", ".", "While", "similar", "in", "spirit", ",", "our", "'", "Match", "'", "model", "is", "slightly", "different", "from", "previous", "work", "(", "Bordes", "and", "[", "reference", "]", ".", "We", "use", "answer", "candidate", "embedding", "matrix", ",", "and", "add", "2", "dimension", "of", "0", "-", "1", "matrix", "which", "expresses", "whether", "the", "answer", "candidate", "matches", "with", "any", "word", "in", "the", "paragraph", "and", "the", "question", ".", "In", "other", "words", ",", "the", "softmax", "is", "computed", "bŷ", "are", "trainable", "weight", "matrices", ",", "and", "M", "(", "y", ")", "∈", "R", "V", "×2", "is", "the", "0", "-", "1", "match", "matrix", ".", "section", ":", "D", "VISUALIZATIONS", "Visualization", "of", "Story", "-", "based", "QA", ".", "Figure", "4", "shows", "visualization", "of", "models", "for", "story", "-", "based", "QA", "tasks", ".", "In", "the", "task", "3", "(", "left", ")", ",", "the", "model", "focuses", "on", "the", "facts", "that", "contain", "'", "football", "'", "in", "the", "first", "layer", ",", "and", "found", "out", "where", "Mary", "journeyed", "to", "before", "the", "bathroom", "in", "the", "second", "layer", ".", "In", "task", "7", "(", "right", ")", ",", "the", "model", "focuses", "on", "the", "facts", "that", "provide", "information", "about", "the", "location", "of", "Sandra", ".", "Can", "you", "make", "a", "restaurant", "reservation", "for", "eight", "in", "a", "cheap", "price", "range", "in", "madrid", "0.00", "1.00", "0.93", "1.00", "I", "'", "m", "on", "it", ".", "0.00", "1.00", "0.74", "0.00", "Any", "preference", "on", "a", "type", "of", "cuisine", ".", "0.00", "0.11", "1.00", "0.01", "I", "love", "british", "food", ".", "0.00", "0.99", "0.99", "0.57", "Okay", "let", "me", "look", "into", "some", "options", "for", "you", ".", "Here", "it", "is", ":", "resto", "-", "paris", "-", "expen", "-", "spanish", "-", "8stars", "-", "address", "Figure", "5", ":", "Visualization", "of", "update", "and", "reset", "gates", "in", "QRN", "'", "2r", "'", "model", "for", "on", "several", "tasks", "of", "bAbI", "dialog", "and", "DSTC2", "dialog", "(", "Table", "3", ")", ".", "We", "do", "not", "put", "reset", "gate", "in", "the", "last", "layer", ".", "Note", "that", "we", "only", "show", "some", "of", "recent", "sentences", "here", ",", "even", "the", "dialog", "has", "more", "sentences", ".", "Visualization", "of", "Dialog", ".", "Figure", "5", "shows", "visualization", "of", "models", "for", "dialog", "tasks", ".", "In", "the", "first", "dialog", "of", "task", "1", ",", "the", "model", "focuses", "on", "the", "user", "utterance", "that", "mentions", "the", "user", "'s", "desired", "cuisine", "and", "location", ",", "and", "the", "current", "query", "(", "user", "'s", "last", "utterance", ")", "informs", "the", "system", "of", "the", "number", "of", "people", ",", "so", "the", "system", "is", "able", "to", "learn", "that", "it", "now", "needs", "to", "ask", "the", "user", "about", "the", "desired", "price", "range", ".", "In", "the", "second", "dialog", "of", "task", "1", ",", "the", "model", "focuses", "on", "the", "facts", "that", "provide", "information", "about", "the", "requests", "of", "the", "user", ".", "In", "task", "4", "(", "third", ")", ",", "the", "model", "focuses", "on", "what", "restaurant", "a", "user", "is", "talking", "about", "and", "the", "information", "about", "the", "restaurant", ".", "section", ":", "section", ":", "ACKNOWLEDGMENTS", "This", "research", "was", "supported", "by", "the", "NSF", "(", "IIS", "1616112", ")", ",", "Allen", "Institute", "for", "AI", "(", "66", "-", "9175", ")", ",", "Allen", "Distinguished", "Investigator", "Award", ",", "Google", "Research", "Faculty", "Award", ",", "and", "Samsung", "GRO", "Award", ".", "We", "thank", "the", "anonymous", "reviewers", "for", "their", "helpful", "comments", ".", "section", ":"], "entity": [[8, 12, ["QUERY", "-", "REDUCTION", "NETWORKS"], "Method"], [13, 15, ["QUESTION", "ANSWERING"], "Task"], [27, 29, ["question", "answering"], "Task"], [39, 43, ["Query", "-", "Reduction", "Network"], "Method"], [44, 45, ["QRN"], "Method"], [50, 53, ["Recurrent", "Neural", "Network"], "Method"], [54, 55, ["RNN"], "Method"], [81, 82, ["QRN"], "Method"], [121, 122, ["QRN"], "Method"], [168, 170, ["time", "complexity"], "Metric"], [171, 172, ["training"], "Task"], [173, 174, ["inference"], "Task"], [187, 189, ["question", "answering"], "Task"], [190, 191, ["QA"], "Task"], [231, 233, ["Question", "answering"], "Task"], [236, 240, ["context", "-", "based", "QA"], "Task"], [246, 249, ["machine", "comprehension", "tasks"], "Task"], [269, 273, ["lexical", "and", "syntactic", "understanding"], "Task"], [278, 279, ["inference"], "Task"], [290, 294, ["multi", "-", "hop", "reasoning"], "Task"], [306, 308, ["dialog", "task"], "Task"], [309, 312, ["Recurrent", "Neural", "Network"], "Method"], [313, 314, ["RNN"], "Method"], [321, 326, ["Long", "Short", "-", "Term", "Memory"], "Method"], [327, 328, ["LSTM"], "Method"], [333, 336, ["Gated", "Recurrent", "Unit"], "Method"], [337, 338, ["GRU"], "Method"], [347, 350, ["modeling", "natural", "language"], "Task"], [356, 360, ["multi", "-", "hop", "reasoning"], "Task"], [361, 363, ["question", "answering"], "Task"], [409, 412, ["global", "attention", "mechanism"], "Method"], [427, 429, ["attention", "mechanism"], "Method"], [457, 461, ["multi", "-", "hop", "reasoning"], "Task"], [470, 472, ["attention", "mechanisms"], "Method"], [496, 501, ["Query", "-", "Reduction", "Network", "1"], "Method"], [502, 503, ["QRN"], "Method"], [508, 510, ["recurrent", "unit"], "Method"], [513, 518, ["long", "-", "term", "dependency", "problem"], "Task"], [527, 529, ["recurrent", "update"], "Method"], [535, 536, ["RNN"], "Method"], [547, 548, ["QRN"], "Method"], [623, 624, ["QRN"], "Method"], [666, 667, ["QRN"], "Method"], [668, 672, ["5", "-", "sentence", "story"], "Material"], [681, 682, ["QRN"], "Method"], [683, 687, ["input", "/", "output", "modules"], "Method"], [724, 726, ["vector", "representations"], "Method"], [835, 839, ["memory", "-", "based", "approaches"], "Method"], [840, 847, ["reference", "][", "reference", "][", "reference", "][", "reference"], "Method"], [849, 850, ["QRN"], "Method"], [888, 889, ["QRN"], "Method"], [894, 895, ["QRN"], "Method"], [900, 901, ["RNN"], "Method"], [916, 917, ["QRN"], "Method"], [921, 923, ["attention", "mechanism"], "Method"], [924, 925, ["RNN"], "Method"], [928, 930, ["time", "dependency"], "Task"], [931, 936, ["long", "-", "term", "dependency", "problems"], "Task"], [949, 951, ["sequential", "data"], "Material"], [960, 961, ["QRN"], "Method"], [966, 967, ["RNN"], "Method"], [988, 989, ["QRN"], "Method"], [1030, 1031, ["QRN"], "Method"], [1073, 1076, ["interactive", "dialog", "datasets"], "Material"], [1087, 1089, ["dialog", "dataset"], "Material"], [1111, 1113, ["natural", "language"], "Material"], [1136, 1138, ["natural", "language"], "Material"], [1169, 1175, ["end", "-", "to", "-", "end", "solutions"], "Task"], [1202, 1203, ["lexicon"], "Method"], [1204, 1206, ["dependency", "parser"], "Method"], [1275, 1277, ["input", "module"], "Method"], [1282, 1284, ["output", "module"], "Method"], [1285, 1287, ["Input", "module"], "Method"], [1371, 1373, ["recurrent", "application"], "Method"], [1386, 1387, ["RNN"], "Method"], [1428, 1430, ["Output", "module"], "Method"], [1431, 1433, ["Output", "module"], "Method"], [1436, 1437, ["QRN"], "Method"], [1470, 1472, ["base", "model"], "Method"], [1508, 1509, ["QRN"], "Method"], [1521, 1522, ["QRN"], "Method"], [1529, 1530, ["computational"], "Metric"], [1558, 1559, ["QRN"], "Method"], [1562, 1564, ["recurrent", "unit"], "Method"], [1644, 1645, ["RNN"], "Method"], [1716, 1718, ["reduce", "function"], "Method"], [1804, 1809, ["element", "-", "wise", "vector", "multiplication"], "Method"], [1850, 1852, ["update", "gate"], "Method"], [1856, 1859, ["global", "attention", "mechanism"], "Method"], [1895, 1900, ["sigmoid", "(", "σ", ")", "function"], "Method"], [1922, 1924, ["softmax", "function"], "Method"], [1935, 1937, ["update", "gate"], "Method"], [1948, 1950, ["Stacking", "layers"], "Method"], [1959, 1960, ["QRN"], "Method"], [1962, 1963, ["QRN"], "Method"], [1993, 1994, ["QRN"], "Method"], [2026, 2030, ["1", "-", "based", "indexing"], "Task"], [2079, 2080, ["QRN"], "Method"], [2227, 2228, ["QRN"], "Method"], [2242, 2243, ["QRN"], "Method"], [2298, 2300, ["output", "module"], "Method"], [2342, 2343, ["QRN"], "Method"], [2352, 2353, ["QRN"], "Method"], [2489, 2493, ["fine", "-", "controlled", "gating"], "Task"], [2495, 2497, ["vector", "gates"], "Task"], [2562, 2563, ["PARALLELIZATION"], "Task"], [2567, 2568, ["QRN"], "Method"], [2619, 2620, ["QRN"], "Method"], [2657, 2659, ["query", "update"], "Task"], [2670, 2672, ["matrix", "operations"], "Method"], [2684, 2685, ["QRN"], "Method"], [2738, 2739, ["QRN"], "Method"], [2751, 2758, ["End", "-", "to", "-", "End", "Memory", "Networks"], "Method"], [2759, 2760, ["N2N"], "Method"], [2762, 2766, ["Improved", "Dynamic", "Memory", "Networks"], "Method"], [2767, 2769, ["DMN", "+"], "Method"], [2780, 2781, ["AGRU"], "Method"], [2785, 2786, ["GRU"], "Method"], [2848, 2849, ["QRN"], "Method"], [2857, 2859, ["gating", "mechanism"], "Method"], [2862, 2863, ["LSTM"], "Method"], [2867, 2868, ["GRU"], "Method"], [2892, 2893, ["QRN"], "Method"], [2920, 2921, ["computational"], "Metric"], [2923, 2924, ["parallelization"], "Method"], [2928, 2929, ["training"], "Task"], [2964, 2970, ["structurally", "simplifying", "(", "constraining", ")", "RNNs"], "Method"], [2971, 2976, ["learning", "longer", "-", "term", "patterns"], "Task"], [2986, 2990, ["Structurally", "Constrained", "Recurrent", "Network"], "Method"], [3001, 3002, ["STRNN"], "Method"], [3007, 3008, ["QRN"], "Method"], [3011, 3012, ["STRNN"], "Method"], [3017, 3019, ["gating", "mechanism"], "Method"], [3045, 3046, ["QRN"], "Method"], [3050, 3051, ["STRNN"], "Method"], [3062, 3064, ["attention", "mechanism"], "Method"], [3083, 3084, ["STRNN"], "Method"], [3105, 3106, ["QRN"], "Method"], [3108, 3109, ["RNN"], "Method"], [3139, 3140, ["STRNN"], "Method"], [3151, 3152, ["QRN"], "Method"], [3157, 3158, ["GPUs"], "Method"], [3160, 3162, ["parallelization", "algorithm"], "Method"], [3166, 3167, ["STRNN"], "Method"], [3168, 3175, ["End", "-", "to", "-", "end", "Memory", "Network"], "Method"], [3176, 3177, ["N2N"], "Method"], [3185, 3190, ["multi", "-", "layer", "attention", "mechanism"], "Method"], [3207, 3208, ["N2N"], "Method"], [3210, 3211, ["QRN"], "Method"], [3240, 3241, ["QRN"], "Method"], [3273, 3274, ["N2N"], "Method"], [3282, 3284, ["sentence", "representations"], "Method"], [3300, 3301, ["QRN"], "Method"], [3313, 3314, ["QRN"], "Method"], [3321, 3323, ["Neural", "Reasoner"], "Method"], [3327, 3333, ["Gated", "End", "-", "toend", "Memory", "Network"], "Method"], [3340, 3341, ["MemN2N"], "Method"], [3347, 3351, ["Improved", "Dynamic", "Memory", "Network"], "Method"], [3363, 3365, ["attention", "mechanism"], "Method"], [3381, 3382, ["GRUs"], "Method"], [3416, 3417, ["GRU"], "Method"], [3429, 3431, ["DMN", "+"], "Method"], [3433, 3437, ["time", "-", "axis", "GRU"], "Method"], [3452, 3453, ["GRU"], "Method"], [3464, 3465, ["QRN"], "Method"], [3469, 3471, ["recurrent", "unit"], "Method"], [3545, 3546, ["accuracy"], "Metric"], [3588, 3590, ["logical", "reasoning"], "Method"], [3615, 3617, ["DMN", "+"], "Method"], [3627, 3630, ["bAbI", "dialog", "dataset"], "Material"], [3630, 3633, ["bAbI", "dialog", "dataset"], "Material"], [3665, 3667, ["restaurant", "reservation"], "Task"], [3695, 3700, ["Out", "-", "Of", "-", "Vocabulary"], "Method"], [3701, 3702, ["OOV"], "Method"], [3731, 3732, ["accuracy"], "Metric"], [3784, 3788, ["Dialog", "State", "Tracking", "Challenge"], "Material"], [3789, 3790, ["DSTC2"], "Material"], [3801, 3804, ["bAbI", "dialog", "dataset"], "Material"], [3847, 3849, ["evaluation", "metric"], "Metric"], [3852, 3853, ["DSTC2"], "Material"], [3868, 3869, ["DSTC2"], "Material"], [3897, 3898, ["DSTC2"], "Material"], [3908, 3910, ["Input", "Module"], "Method"], [3945, 3947, ["vector", "representations"], "Method"], [3959, 3962, ["trainable", "embedding", "matrix"], "Method"], [3997, 3999, ["sentence", "representation"], "Method"], [4004, 4006, ["Position", "Encoder"], "Method"], [4026, 4028, ["Output", "Module"], "Method"], [4043, 4045, ["vector", "representation"], "Method"], [4067, 4075, ["V", "-", "way", "single", "-", "layer", "softmax", "classifier"], "Method"], [4156, 4158, ["Output", "Module"], "Method"], [4159, 4160, ["dialog"], "Task"], [4166, 4171, ["single", "-", "layer", "softmax", "classifiers"], "Method"], [4213, 4215, ["output", "module"], "Method"], [4219, 4225, ["recurrent", "hidden", "state", "or", "gating", "mechanism"], "Method"], [4236, 4237, ["QRN"], "Method"], [4267, 4268, ["development"], "Task"], [4285, 4290, ["bAbI", "story", "-", "based", "QA"], "Material"], [4292, 4294, ["bAbI", "dialog"], "Material"], [4301, 4303, ["bAbI", "QA"], "Material"], [4367, 4370, ["L2", "weight", "decay"], "Metric"], [4385, 4387, ["loss", "function"], "Metric"], [4408, 4411, ["stochastic", "gradient", "descent"], "Method"], [4436, 4438, ["learning", "rate"], "Metric"], [4441, 4442, ["AdaGrad"], "Method"], [4448, 4450, ["learning", "rate"], "Metric"], [4466, 4468, ["weight", "initialization"], "Method"], [4501, 4502, ["loss"], "Metric"], [4523, 4527, ["the", "-", "art", "models"], "Method"], [4528, 4534, ["story", "-", "based", "and", "dialog", "tasks"], "Task"], [4541, 4542, ["LSTM"], "Method"], [4546, 4553, ["End", "-", "to", "-", "end", "Memory", "Networks"], "Method"], [4554, 4555, ["N2N"], "Method"], [4560, 4563, ["Dynamic", "Memory", "Networks"], "Method"], [4564, 4566, ["DMN", "+"], "Method"], [4571, 4579, ["Gated", "End", "-", "to", "-", "end", "Memory", "Networks"], "Method"], [4580, 4581, ["GMemN2N"], "Method"], [4587, 4590, ["Differentiable", "Neural", "Computer"], "Method"], [4591, 4592, ["DNC"], "Method"], [4616, 4617, ["QRN"], "Method"], [4622, 4624, ["bAbI", "QA"], "Material"], [4639, 4641, ["1k", "data"], "Material"], [4642, 4643, ["QRN"], "Method"], [4673, 4675, ["10k", "dataset"], "Material"], [4677, 4679, ["average", "accuracy"], "Metric"], [4680, 4681, ["QRN"], "Method"], [4737, 4738, ["QRN"], "Method"], [4743, 4745, ["bAbI", "dialog"], "Material"], [4786, 4788, ["Match", "'"], "Method"], [4813, 4814, ["QRN"], "Method"], [4830, 4831, ["Ablations"], "Method"], [4837, 4838, ["ablations"], "Method"], [4893, 4894, ["ablations"], "Method"], [4895, 4897, ["bAbI", "QA"], "Material"], [4949, 4951, ["1k", "dataset"], "Material"], [4994, 4995, ["QRN"], "Method"], [4998, 4999, ["parallelization"], "Method"], [5000, 5001, ["TensorFlow"], "Method"], [5014, 5015, ["computational"], "Metric"], [5018, 5019, ["parallelization"], "Method"], [5032, 5033, ["TensorFlow"], "Method"], [5034, 5035, ["QRN"], "Method"], [5036, 5037, ["parallelization"], "Method"], [5041, 5042, ["training"], "Task"], [5043, 5044, ["inference"], "Task"], [5045, 5046, ["QRN"], "Method"], [5047, 5048, ["parallelization"], "Method"], [5071, 5072, ["QRN"], "Method"], [5115, 5116, ["decoder"], "Method"], [5147, 5148, ["decoder"], "Method"], [5246, 5247, ["Visualization"], "Task"], [5398, 5400, ["dialog", "Task"], "Task"], [5427, 5429, ["dialog", "Task"], "Task"], [5441, 5442, ["Spanish"], "Material"], [5467, 5471, ["Query", "-", "Reduction", "Network"], "Method"], [5472, 5473, ["QRN"], "Method"], [5476, 5480, ["context", "-", "based", "questions"], "Task"], [5488, 5492, ["multi", "-", "hop", "reasoning"], "Method"], [5553, 5554, ["QRN"], "Method"], [5558, 5560, ["attention", "mechanism"], "Method"], [5561, 5562, ["RNN"], "Method"], [5565, 5567, ["time", "dependency"], "Task"], [5568, 5573, ["long", "-", "term", "dependency", "problems"], "Task"], [5582, 5587, ["long", "-", "term", "dependency", "problem"], "Task"], [5589, 5590, ["RNNs"], "Method"], [5593, 5595, ["recurrent", "update"], "Method"], [5616, 5617, ["QRN"], "Method"], [5653, 5654, ["QA"], "Task"], [5678, 5679, ["QRN"], "Method"], [5683, 5684, ["LSTM"], "Method"], [5685, 5692, ["End", "-", "to", "-", "end", "Memory", "Networks"], "Method"], [5693, 5694, ["N2N"], "Method"], [5699, 5702, ["Dynamic", "Memory", "Networks"], "Method"], [5703, 5705, ["DMN", "+"], "Method"], [5730, 5733, ["Differentiable", "Neural", "Computer"], "Method"], [5734, 5735, ["DNC"], "Method"], [5748, 5750, ["bAbI", "dialog"], "Material"], [5755, 5757, ["error", "rates"], "Metric"], [5761, 5762, ["QRN"], "Method"], [5766, 5773, ["End", "-", "to", "-", "end", "Memory", "Networks"], "Method"], [5774, 5775, ["N2N"], "Method"], [5777, 5785, ["Gated", "End", "-", "to", "-", "end", "Memory", "Networks"], "Method"], [5786, 5787, ["GMemN2N"], "Method"], [5793, 5794, ["QRN"], "Method"], [5873, 5876, ["VECTOR", "GATE", "PARALLELIZATION"], "Task"], [5877, 5879, ["vector", "gates"], "Task"], [5982, 5985, ["Match", "'", "model"], "Method"], [6000, 6004, ["answer", "candidate", "embedding", "matrix"], "Method"], [6076, 6079, ["visualization", "of", "models"], "Method"], [6242, 6244, ["bAbI", "dialog"], "Material"], [6281, 6284, ["Visualization", "of", "Dialog"], "Task"], [6292, 6294, ["dialog", "tasks"], "Task"], [134, 135, ["QA"], "Task"], [152, 153, ["QRN"], "Method"], [157, 158, ["RNN"], "Method"], [303, 304, ["QA"], "Task"], [365, 366, ["RNN"], "Method"], [383, 384, ["RNN"], "Method"], [520, 521, ["RNN"], "Method"], [657, 658, ["QRN"], "Method"], [678, 679, ["QA"], "Task"], [804, 805, ["RNN"], "Method"], [809, 810, ["QRN"], "Method"], [983, 984, ["RNN"], "Method"], [1027, 1028, ["QRN"], "Method"], [1040, 1041, ["RNN"], "Method"], [1071, 1072, ["QA"], "Task"], [1084, 1085, ["QA"], "Task"], [1263, 1264, ["QA"], "Task"], [1278, 1279, ["QRN"], "Method"], [1334, 1335, ["QRN"], "Method"], [1337, 1338, ["QRN"], "Method"], [1366, 1367, ["QRN"], "Method"], [1375, 1376, ["QRN"], "Method"], [1415, 1416, ["QRN"], "Method"], [1474, 1475, ["QRN"], "Method"], [1533, 1534, ["RNN"], "Method"], [1549, 1550, ["QRN"], "Method"], [1553, 1554, ["RNN"], "Method"], [1586, 1587, ["QRN"], "Method"], [1599, 1600, ["QRN"], "Method"], [2382, 2383, ["QRN"], "Method"], [2473, 2474, ["LSTM"], "Method"], [2592, 2593, ["RNN"], "Method"], [2852, 2853, ["RNN"], "Method"], [2875, 2876, ["LSTM"], "Method"], [2997, 3000, ["Recurrent", "Neural", "Network"], "Method"], [3057, 3058, ["QRN"], "Method"], [3091, 3092, ["LSTM"], "Method"], [3117, 3118, ["QA"], "Task"], [3122, 3123, ["QRN"], "Method"], [3214, 3215, ["N2N"], "Method"], [3310, 3311, ["RNN"], "Method"], [3367, 3368, ["RNN"], "Method"], [3485, 3486, ["QA"], "Task"], [3491, 3492, ["QA"], "Task"], [3860, 3861, ["DSTC2"], "Material"], [3884, 3885, ["DSTC2"], "Material"], [4032, 4033, ["QA"], "Task"], [4184, 4185, ["QA"], "Task"], [4206, 4207, ["RNN"], "Method"], [4295, 4296, ["DSTC2"], "Material"], [4332, 4333, ["QRN"], "Method"], [4375, 4376, ["QA"], "Task"], [4455, 4456, ["QA"], "Task"], [4600, 4601, ["QA"], "Task"], [5021, 5022, ["QRN"], "Method"], [5028, 5029, ["RNN"], "Method"], [5310, 5311, ["QA"], "Task"], [5510, 5511, ["QA"], "Task"], [5629, 5630, ["RNN"], "Method"], [5670, 5671, ["QA"], "Task"], [5672, 5674, ["error", "rates"], "Metric"], [5711, 5718, ["End", "-", "to", "-", "end", "Memory", "Networks"], "Method"], [5719, 5720, ["GMemN2N"], "Method"], [5751, 5752, ["DSTC2"], "Material"], [6071, 6072, ["QA"], "Task"], [6083, 6084, ["QA"], "Task"], [6232, 6233, ["QRN"], "Method"], [6245, 6246, ["DSTC2"], "Material"]]} {"doc_id": "270e65acc071b9e4e2a632720130c0e10cb6fa08", "text": ["document", ":", "Neural", "Tree", "Indexers", "for", "Text", "Understanding", "Recurrent", "neural", "networks", "(", "RNNs", ")", "process", "input", "text", "sequentially", "and", "model", "the", "conditional", "transition", "between", "word", "tokens", ".", "In", "contrast", ",", "the", "advantages", "of", "recursive", "networks", "include", "that", "they", "explicitly", "model", "the", "compositionality", "and", "the", "recursive", "structure", "of", "natural", "language", ".", "However", ",", "the", "current", "recursive", "architecture", "is", "limited", "by", "its", "dependence", "on", "syntactic", "tree", ".", "In", "this", "paper", ",", "we", "introduce", "a", "robust", "syntactic", "parsing", "-", "independent", "tree", "structured", "model", ",", "Neural", "Tree", "Indexers", "(", "NTI", ")", "that", "provides", "a", "middle", "ground", "between", "the", "sequential", "RNNs", "and", "the", "syntactic", "tree", "-", "based", "recursive", "models", ".", "NTI", "constructs", "a", "full", "n", "-", "ary", "tree", "by", "processing", "the", "input", "text", "with", "its", "node", "function", "in", "a", "bottom", "-", "up", "fashion", ".", "Attention", "mechanism", "can", "then", "be", "applied", "to", "both", "structure", "and", "node", "function", ".", "We", "implemented", "and", "evaluated", "a", "binary", "-", "tree", "model", "of", "NTI", ",", "showing", "the", "model", "achieved", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "three", "different", "NLP", "tasks", ":", "natural", "language", "inference", ",", "answer", "sentence", "selection", ",", "and", "sentence", "classification", ",", "outperforming", "state", "-", "of", "-", "the", "-", "art", "recurrent", "and", "recursive", "neural", "networks", ".", "section", ":", "Introduction", "Recurrent", "neural", "networks", "(", "RNNs", ")", "have", "been", "successful", "for", "modeling", "sequence", "data", ".", "RNNs", "equipped", "with", "gated", "hidden", "units", "and", "internal", "short", "-", "term", "memories", ",", "such", "as", "long", "short", "-", "term", "memories", "(", "LSTM", ")", "have", "achieved", "a", "notable", "success", "in", "several", "NLP", "tasks", "including", "named", "entity", "recognition", ",", "constituency", "parsing", ",", "textual", "entailment", "recognition", ",", "question", "answering", ",", "and", "machine", "translation", ".", "However", ",", "most", "LSTM", "models", "explored", "so", "far", "are", "sequential", ".", "It", "encodes", "text", "sequentially", "from", "left", "to", "right", "or", "vice", "versa", "and", "do", "not", "naturally", "support", "compositionality", "of", "language", ".", "Sequential", "LSTM", "models", "seem", "to", "learn", "syntactic", "structure", "from", "the", "natural", "language", "however", "their", "generalization", "on", "unseen", "text", "is", "relatively", "poor", "comparing", "with", "models", "that", "exploit", "syntactic", "tree", "structure", ".", "Unlike", "sequential", "models", ",", "recursive", "neural", "networks", "compose", "word", "phrases", "over", "syntactic", "tree", "structure", "and", "have", "shown", "improved", "performance", "in", "sentiment", "analysis", ".", "However", "its", "dependence", "on", "a", "syntactic", "tree", "architecture", "limits", "practical", "NLP", "applications", ".", "In", "this", "study", ",", "we", "introduce", "Neural", "Tree", "Indexers", "(", "NTI", ")", ",", "a", "class", "of", "tree", "structured", "models", "for", "NLP", "tasks", ".", "NTI", "takes", "a", "sequence", "of", "tokens", "and", "produces", "its", "representation", "by", "constructing", "a", "full", "n", "-", "ary", "tree", "in", "a", "bottom", "-", "up", "fashion", ".", "Each", "node", "in", "NTI", "is", "associated", "with", "one", "of", "the", "node", "transformation", "functions", ":", "leaf", "node", "mapping", "and", "non", "-", "leaf", "node", "composition", "functions", ".", "Unlike", "previous", "recursive", "models", ",", "the", "tree", "structure", "for", "NTI", "is", "relaxed", ",", "i.e.", ",", "NTI", "does", "not", "require", "the", "input", "sequences", "to", "be", "parsed", "syntactically", ";", "and", "therefore", "it", "is", "flexible", "and", "can", "be", "directly", "applied", "to", "a", "wide", "range", "of", "NLP", "tasks", "beyond", "sentence", "modeling", ".", "Furthermore", ",", "we", "propose", "different", "variants", "of", "node", "composition", "function", "and", "attention", "over", "tree", "for", "our", "NTI", "models", ".", "When", "a", "sequential", "leaf", "node", "transformer", "such", "as", "LSTM", "is", "chosen", ",", "the", "NTI", "network", "forms", "a", "sequence", "-", "tree", "hybrid", "model", "taking", "advantage", "of", "both", "conditional", "and", "compositional", "powers", "of", "sequential", "and", "recursive", "models", ".", "Figure", "shows", "a", "binary", "-", "tree", "model", "of", "NTI", ".", "Although", "the", "model", "does", "not", "follow", "the", "syntactic", "tree", "structure", ",", "we", "empirically", "show", "that", "it", "achieved", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "three", "different", "NLP", "applications", ":", "natural", "language", "inference", ",", "answer", "sentence", "selection", ",", "and", "sentence", "classification", ".", "section", ":", "Related", "Work", "subsection", ":", "Recurrent", "Neural", "Networks", "and", "Attention", "Mechanism", "RNNs", "model", "input", "text", "sequentially", "by", "taking", "a", "single", "token", "at", "each", "time", "step", "and", "producing", "a", "corresponding", "hidden", "state", ".", "The", "hidden", "state", "is", "then", "passed", "along", "through", "the", "next", "time", "step", "to", "provide", "historical", "sequence", "information", ".", "Although", "a", "great", "success", "in", "a", "variety", "of", "tasks", ",", "RNNs", "have", "limitations", ".", "Among", "them", ",", "it", "is", "not", "efficient", "at", "memorizing", "long", "or", "distant", "sequence", ".", "This", "is", "frequently", "called", "as", "information", "flow", "bottleneck", ".", "Approaches", "have", "therefore", "been", "developed", "to", "overcome", "the", "limitations", ".", "For", "example", ",", "to", "mitigate", "the", "information", "flow", "bottleneck", ",", "bahdanau:15", "extended", "RNNs", "with", "a", "soft", "attention", "mechanism", "in", "the", "context", "of", "neural", "machine", "translation", ",", "leading", "to", "improved", "the", "results", "in", "translating", "longer", "sentences", ".", "RNNs", "are", "linear", "chain", "-", "structured", ";", "this", "limits", "its", "potential", "for", "natural", "language", "which", "can", "be", "represented", "by", "complex", "structures", "including", "syntactic", "structure", ".", "In", "this", "study", ",", "we", "propose", "models", "to", "mitigate", "this", "limitation", ".", "subsection", ":", "Recursive", "Neural", "Networks", "Unlike", "RNNs", ",", "recursive", "neural", "networks", "explicitly", "model", "the", "compositionality", "and", "the", "recursive", "structure", "of", "natural", "language", "over", "tree", ".", "The", "tree", "structure", "can", "be", "predefined", "by", "a", "syntactic", "parser", ".", "Each", "non", "-", "leaf", "tree", "node", "is", "associated", "with", "a", "node", "composition", "function", "which", "combines", "its", "children", "nodes", "and", "produces", "its", "own", "representation", ".", "The", "model", "is", "then", "trained", "by", "back", "-", "propagating", "error", "through", "structures", ".", "The", "node", "composition", "function", "can", "be", "varied", ".", "A", "single", "layer", "network", "with", "non", "-", "linearity", "was", "adopted", "in", "recursive", "auto", "-", "associate", "memories", "and", "recursive", "autoencoders", ".", "socher2012semantic", "extended", "this", "network", "with", "an", "additional", "matrix", "representation", "for", "each", "node", "to", "augment", "the", "expressive", "power", "of", "the", "model", ".", "Tensor", "networks", "have", "also", "been", "used", "as", "composition", "function", "for", "sentence", "-", "level", "sentiment", "analysis", "task", ".", "Recently", ",", "zhu2015long", "introduced", "S", "-", "LSTM", "which", "extends", "LSTM", "units", "to", "compose", "tree", "nodes", "in", "a", "recursive", "fashion", ".", "In", "this", "paper", ",", "we", "introduce", "a", "novel", "attentive", "node", "composition", "function", "that", "is", "based", "on", "S", "-", "LSTM", ".", "Our", "NTI", "model", "does", "not", "rely", "on", "either", "a", "parser", "output", "or", "a", "fine", "-", "grained", "supervision", "of", "non", "-", "leaf", "nodes", ",", "both", "required", "in", "previous", "work", ".", "In", "NTI", ",", "the", "supervision", "from", "the", "target", "labels", "is", "provided", "at", "the", "root", "node", ".", "As", "such", ",", "our", "NTI", "model", "is", "robust", "and", "applicable", "to", "a", "wide", "range", "of", "NLP", "tasks", ".", "We", "introduce", "attention", "over", "tree", "in", "NTI", "to", "overcome", "the", "vanishing", "/", "explode", "gradients", "challenges", "as", "shown", "in", "RNNs", ".", "section", ":", "Methods", "Our", "training", "set", "consists", "of", "examples", ",", "where", "the", "input", "is", "a", "sequence", "of", "word", "tokens", "and", "the", "output", "can", "be", "either", "a", "single", "target", "or", "a", "sequence", ".", "Each", "input", "word", "token", "is", "represented", "by", "its", "word", "embedding", ".", "NTI", "is", "a", "full", "n", "-", "ary", "tree", "(", "and", "the", "sub", "-", "trees", "can", "be", "overlapped", ")", ".", "It", "has", "two", "types", "of", "transformation", "function", ":", "non", "-", "leaf", "node", "function", "and", "leaf", "node", "function", ".", "computes", "a", "(", "possibly", "non", "-", "linear", ")", "transformation", "of", "the", "input", "word", "embedding", ".", "is", "a", "function", "of", "its", "child", "nodes", "representation", ",", "where", "is", "the", "total", "number", "of", "child", "nodes", "of", "this", "non", "-", "leaf", "node", ".", "NTI", "can", "be", "implemented", "with", "different", "tree", "structures", ".", "In", "this", "study", "we", "implemented", "and", "evaluated", "a", "binary", "tree", "form", "of", "NTI", ":", "a", "non", "-", "leaf", "node", "can", "take", "in", "only", "two", "direct", "child", "nodes", "(", "i.e.", ",", ")", ".", "Therefore", ",", "the", "function", "composes", "its", "left", "child", "node", "and", "right", "child", "node", ".", "Figure", "illustrates", "our", "NTI", "model", "that", "is", "applied", "to", "question", "answering", "(", "a", ")", "and", "natural", "language", "inference", "tasks", "(", "b", ")", ".", "Note", "that", "the", "node", "and", "leaf", "node", "functions", "are", "neural", "networks", "and", "are", "the", "only", "training", "parameters", "in", "NTI", ".", "We", "explored", "two", "different", "approaches", "to", "compose", "node", "representations", ":", "an", "extended", "LSTM", "and", "attentive", "node", "composition", "functions", ",", "to", "be", "described", "below", ".", "subsection", ":", "Non", "-", "Leaf", "Node", "Composition", "Functions", "We", "define", "two", "different", "methods", "for", "non", "-", "leaf", "node", "function", ".", "LSTM", "-", "based", "Non", "-", "leaf", "Node", "Function", "(", "S", "-", "LSTM", ")", ":", "We", "initiate", "with", "LSTM", ".", "For", "non", "-", "leaf", "node", ",", "we", "adopt", "S", "-", "LSTM", "zhu2015long", ",", "an", "extension", "of", "LSTM", "to", "tree", "structures", ",", "to", "learn", "a", "node", "representation", "by", "its", "children", "nodes", ".", "Let", ",", ",", "and", "be", "vector", "representations", "and", "cell", "states", "for", "the", "left", "and", "right", "children", ".", "An", "S", "-", "LSTM", "computes", "a", "parent", "node", "representation", "and", "a", "node", "cell", "state", "as", "where", "and", "biases", "(", "for", "brevity", "we", "eliminated", "the", "bias", "terms", ")", "are", "the", "training", "parameters", ".", "and", "denote", "the", "element", "-", "wise", "function", "and", "the", "element", "-", "wise", "vector", "multiplication", ".", "Extension", "of", "S", "-", "LSTM", "non", "-", "leaf", "node", "function", "to", "compose", "more", "children", "is", "straightforward", ".", "However", ",", "the", "number", "of", "parameters", "increases", "quadratically", "in", "S", "-", "LSTM", "as", "we", "add", "more", "child", "nodes", ".", "Attentive", "Non", "-", "leaf", "Node", "Function", "(", "ANF", ")", ":", "Some", "NLP", "applications", "(", "e.g.", ",", "QA", "and", "machine", "translation", ")", "would", "benefit", "from", "a", "dynamic", "query", "dependent", "composition", "function", ".", "We", "introduce", "ANF", "as", "a", "new", "non", "-", "leaf", "node", "function", ".", "Unlike", "S", "-", "LSTM", ",", "ANF", "composes", "the", "child", "nodes", "attentively", "in", "respect", "to", "another", "relevant", "input", "vector", ".", "The", "input", "vector", "can", "be", "a", "learnable", "representation", "from", "a", "sequence", "representation", ".", "Given", "a", "matrix", "resulted", "by", "concatenating", "the", "child", "node", "representations", ",", "and", "the", "third", "input", "vector", ",", "ANF", "is", "defined", "as", "where", "is", "a", "learnable", "matrix", ",", "the", "attention", "score", "and", "the", "attention", "weight", "vector", "for", "each", "child", ".", "is", "an", "attention", "scoring", "function", ",", "which", "can", "be", "implemented", "as", "a", "multi", "-", "layer", "perceptron", "(", "MLP", ")", "or", "a", "matrix", "-", "vector", "product", ".", "The", "matrices", "and", "and", "the", "vector", "are", "training", "parameters", ".", "is", "a", "vector", "of", "ones", "and", "the", "outer", "product", ".", "We", "use", "function", "for", "non", "-", "linear", "transformation", ".", "subsection", ":", "Attention", "Over", "Tree", "Comparing", "with", "sequential", "LSTM", "models", ",", "NTI", "has", "less", "recurrence", ",", "which", "is", "defined", "by", "the", "tree", "depth", ",", "for", "binary", "tree", "where", "is", "the", "length", "of", "the", "input", "sequence", ".", "However", ",", "NTI", "still", "needs", "to", "compress", "all", "the", "input", "information", "into", "a", "single", "representation", "vector", "of", "the", "root", ".", "This", "imposes", "practical", "difficulties", "when", "processing", "long", "sequences", ".", "We", "address", "this", "issue", "with", "attention", "mechanism", "over", "tree", ".", "In", "addition", ",", "the", "attention", "mechanism", "can", "be", "used", "for", "matching", "trees", "(", "described", "in", "Section", "4", "as", "Tree", "matching", "NTI", ")", "that", "carry", "different", "sequence", "information", ".", "We", "first", "define", "a", "global", "attention", "and", "then", "introduce", "a", "tree", "attention", "which", "considers", "the", "parent", "-", "child", "dependency", "for", "calculation", "of", "the", "attention", "weights", ".", "Global", "Attention", ":", "An", "attention", "neural", "network", "for", "the", "global", "attention", "takes", "all", "node", "representations", "as", "input", "and", "produces", "an", "attentively", "blended", "vector", "for", "the", "whole", "tree", ".", "This", "neural", "net", "is", "similar", "to", "ANF", ".", "Particularly", ",", "given", "a", "matrix", "resulted", "by", "concatenating", "the", "node", "representations", ",", "…", ",", "and", "the", "relevant", "input", "representation", ",", "the", "global", "attention", "is", "defined", "as", "where", "and", "are", "training", "parameters", "and", "the", "attention", "weight", "vector", "for", "each", "node", ".", "This", "attention", "mechanism", "is", "robust", "as", "it", "globally", "normalizes", "the", "attention", "score", "with", "to", "obtain", "the", "weights", ".", "However", ",", "it", "does", "not", "consider", "the", "tree", "structure", "when", "producing", "the", "final", "representation", ".", "Tree", "Attention", ":", "We", "modify", "the", "global", "attention", "network", "to", "the", "tree", "attention", "mechanism", ".", "The", "resulting", "tree", "attention", "network", "performs", "almost", "the", "same", "computation", "as", "ANF", "for", "each", "node", ".", "It", "compares", "the", "parent", "and", "children", "nodes", "to", "produce", "a", "new", "representation", "assuming", "that", "all", "node", "representations", "are", "constructed", ".", "Given", "a", "matrix", "resulted", "by", "concatenating", "the", "parent", "node", "representation", ",", "the", "left", "child", "and", "the", "right", "child", "and", "the", "relevant", "input", "representation", ",", "every", "non", "-", "leaf", "node", "simply", "updates", "its", "own", "representation", "by", "using", "the", "following", "equation", "in", "a", "bottom", "-", "up", "manner", ".", "and", "this", "equation", "is", "similarity", "to", "the", "global", "attention", ".", "However", ",", "now", "each", "non", "-", "leaf", "node", "attentively", "collects", "its", "own", "and", "children", "representations", "and", "passes", "towards", "the", "root", "which", "finally", "constructs", "the", "attentively", "blended", "tree", "representation", ".", "Note", "that", "unlike", "the", "global", "attention", ",", "the", "tree", "attention", "locally", "normalizes", "the", "attention", "scores", "with", ".", "section", ":", "Experiments", "We", "describe", "in", "this", "section", "experiments", "on", "three", "different", "NLP", "tasks", ",", "natural", "language", "inference", ",", "question", "answering", "and", "sentence", "classification", "to", "demonstrate", "the", "flexibility", "and", "the", "effectiveness", "of", "NTI", "in", "the", "different", "settings", ".", "We", "trained", "NTI", "using", "Adam", "with", "hyperparameters", "selected", "on", "development", "set", ".", "The", "pre", "-", "trained", "300", "-", "D", "Glove", "840B", "vectors", "were", "obtained", "for", "the", "word", "embeddings", ".", "The", "word", "embeddings", "are", "fixed", "during", "training", ".", "The", "embeddings", "for", "out", "-", "of", "-", "vocabulary", "words", "were", "set", "to", "zero", "vector", ".", "We", "pad", "the", "input", "sequence", "to", "form", "a", "full", "binary", "tree", ".", "A", "padding", "vector", "was", "inserted", "when", "padding", ".", "We", "analyzed", "the", "effects", "of", "the", "padding", "size", "and", "found", "out", "that", "it", "has", "no", "influence", "on", "the", "performance", "(", "see", "Appendix", "[", "reference", "]", ")", ".", "The", "size", "of", "hidden", "units", "of", "the", "NTI", "modules", "were", "set", "to", "300", ".", "The", "models", "were", "regularized", "by", "using", "dropouts", "and", "an", "weight", "decay", ".", "subsection", ":", "Natural", "Language", "Inference", "We", "conducted", "experiments", "on", "the", "Stanford", "Natural", "Language", "Inference", "(", "SNLI", ")", "dataset", ",", "which", "consists", "of", "549", ",", "367", "/", "9", ",", "842", "/", "9", ",", "824", "premise", "-", "hypothesis", "pairs", "for", "train", "/", "dev", "/", "test", "sets", "and", "target", "label", "indicating", "their", "relation", ".", "Unless", "otherwise", "noted", ",", "we", "follow", "the", "setting", "in", "the", "previous", "work", "and", "use", "an", "MLP", "for", "classification", "which", "takes", "in", "NTI", "outputs", "and", "computes", "the", "concatenation", ",", "absolute", "difference", "and", "elementwise", "product", "of", "the", "two", "sentence", "representations", ".", "The", "MLP", "has", "also", "an", "input", "layer", "with", "1024", "units", "with", "activation", "and", "a", "output", "layer", ".", "We", "explored", "nine", "different", "task", "-", "oriented", "NTI", "models", "with", "varying", "complexity", ",", "to", "be", "described", "below", ".", "For", "each", "model", ",", "we", "set", "the", "batch", "size", "to", "32", ".", "The", "initial", "learning", ",", "the", "regularization", "strength", "and", "the", "number", "of", "epoch", "to", "be", "trained", "are", "varied", "for", "each", "model", ".", "NTI", "-", "SLSTM", ":", "this", "model", "does", "not", "rely", "on", "transformer", "but", "uses", "the", "S", "-", "LSTM", "units", "for", "the", "non", "-", "leaf", "node", "function", ".", "We", "set", "the", "initial", "learning", "rate", "to", "1e", "-", "3", "and", "regularizer", "strength", "to", "3e", "-", "5", ",", "and", "train", "the", "model", "for", "90", "epochs", ".", "The", "neural", "net", "was", "regularized", "by", "10", "%", "input", "dropouts", "and", "the", "20", "%", "output", "dropouts", ".", "NTI", "-", "SLSTM", "-", "LSTM", ":", "we", "use", "LSTM", "for", "the", "leaf", "node", "function", ".", "Concretely", ",", "the", "LSTM", "output", "vectors", "are", "given", "to", "NTI", "-", "SLSTM", "and", "the", "memory", "cells", "of", "the", "lowest", "level", "S", "-", "LSTM", "were", "initialized", "with", "the", "LSTM", "memory", "states", ".", "The", "hyper", "-", "parameters", "are", "the", "same", "as", "the", "previous", "model", ".", "NTI", "-", "SLSTM", "node", "-", "by", "-", "node", "global", "attention", ":", "This", "model", "learns", "inter", "-", "sentence", "relation", "with", "the", "global", "attention", "over", "premise", "-", "indexed", "tree", ",", "which", "is", "similar", "to", "word", "-", "by", "-", "word", "attention", "model", "of", "rocktaschel:16", "in", "that", "it", "attends", "over", "the", "premise", "tree", "nodes", "at", "every", "time", "step", "of", "hypothesis", "encoding", ".", "We", "tie", "the", "weight", "parameters", "of", "the", "two", "NTI", "-", "SLSTMs", "for", "premise", "and", "hypothesis", "and", "no", "transformer", "used", ".", "We", "set", "the", "initial", "learning", "rate", "to", "3e", "-", "4", "and", "regularizer", "strength", "to", "1e", "-", "5", ",", "and", "train", "the", "model", "for", "40", "epochs", ".", "The", "neural", "net", "was", "regularized", "by", "15", "%", "input", "dropouts", "and", "the", "15", "%", "output", "dropouts", ".", "NTI", "-", "SLSTM", "node", "-", "by", "-", "node", "tree", "attention", ":", "this", "is", "a", "variation", "of", "the", "previous", "model", "with", "the", "tree", "attention", ".", "The", "hyper", "-", "parameters", "are", "the", "same", "as", "the", "previous", "model", ".", "NTI", "-", "SLSTM", "-", "LSTM", "node", "-", "by", "-", "node", "global", "attention", ":", "in", "this", "model", "we", "include", "LSTM", "as", "the", "leaf", "node", "function", ".", "Here", "we", "initialize", "the", "memory", "cell", "of", "S", "-", "LSTM", "with", "LSTM", "memory", "and", "hidden", "/", "memory", "state", "of", "hypothesis", "LSTM", "with", "premise", "LSTM", "(", "the", "later", "follows", "the", "work", "of", ")", ".", "We", "set", "the", "initial", "learning", "rate", "to", "3e", "-", "4", "and", "regularizer", "strength", "to", "1e", "-", "5", ",", "and", "train", "the", "model", "for", "10", "epochs", ".", "The", "neural", "net", "was", "regularized", "by", "10", "%", "input", "dropouts", "and", "the", "15", "%", "output", "dropouts", ".", "NTI", "-", "SLSTM", "-", "LSTM", "node", "-", "by", "-", "node", "tree", "attention", ":", "this", "is", "a", "variation", "of", "the", "previous", "model", "with", "the", "tree", "attention", ".", "The", "hyper", "-", "parameters", "are", "the", "same", "as", "the", "previous", "model", ".", "Tree", "matching", "NTI", "-", "SLSTM", "-", "LSTM", "global", "attention", ":", "this", "model", "first", "constructs", "the", "premise", "and", "hypothesis", "trees", "simultaneously", "with", "the", "NTI", "-", "SLSTM", "-", "LSTM", "model", "and", "then", "computes", "their", "matching", "vector", "by", "using", "the", "global", "attention", "and", "an", "additional", "LSTM", ".", "The", "attention", "vectors", "are", "produced", "at", "each", "hypothesis", "tree", "node", "and", "then", "are", "given", "to", "the", "LSTM", "model", "sequentially", ".", "The", "LSTM", "model", "compress", "the", "attention", "vectors", "and", "outputs", "a", "single", "matching", "vector", ",", "which", "is", "passed", "to", "an", "MLP", "for", "classification", ".", "The", "MLP", "for", "this", "tree", "matching", "setting", "has", "an", "input", "layer", "with", "1024", "units", "with", "activation", "and", "a", "output", "layer", ".", "Unlike", "WangJ15b", "’s", "matching", "LSTM", "model", "which", "is", "specific", "to", "matching", "sequences", ",", "we", "use", "the", "standard", "LSTM", "units", "and", "match", "trees", ".", "We", "set", "the", "initial", "learning", "rate", "to", "3e", "-", "4", "and", "regularizer", "strength", "to", "3e", "-", "5", ",", "and", "train", "the", "model", "for", "20", "epochs", ".", "The", "neural", "net", "was", "regularized", "by", "20", "%", "input", "dropouts", "and", "the", "20", "%", "output", "dropouts", ".", "Tree", "matching", "NTI", "-", "SLSTM", "-", "LSTM", "tree", "attention", ":", "we", "replace", "the", "global", "attention", "with", "the", "tree", "attention", ".", "The", "hyper", "-", "parameters", "are", "the", "same", "as", "the", "previous", "model", ".", "Full", "tree", "matching", "NTI", "-", "SLSTM", "-", "LSTM", "global", "attention", ":", "this", "model", "produces", "two", "sets", "of", "the", "attention", "vectors", ",", "one", "by", "attending", "over", "the", "premise", "tree", "regarding", "each", "hypothesis", "tree", "node", "and", "another", "by", "attending", "over", "the", "hypothesis", "tree", "regarding", "each", "premise", "tree", "node", ".", "Each", "set", "of", "the", "attention", "vectors", "is", "given", "to", "a", "LSTM", "model", "to", "achieve", "full", "tree", "matching", ".", "The", "last", "hidden", "states", "of", "the", "two", "LSTM", "models", "(", "i.e.", "one", "for", "each", "attention", "vector", "set", ")", "are", "concatenated", "for", "classification", ".", "The", "training", "weights", "are", "shared", "among", "the", "LSTM", "models", "The", "hyper", "-", "parameters", "are", "the", "same", "as", "the", "previous", "model", ".", "Table", "[", "reference", "]", "shows", "the", "results", "of", "our", "models", ".", "For", "comparison", ",", "we", "include", "the", "results", "from", "the", "published", "state", "-", "of", "-", "the", "-", "art", "systems", ".", "While", "most", "of", "the", "sentence", "encoder", "models", "rely", "solely", "on", "word", "embeddings", ",", "the", "dependency", "tree", "CNN", "and", "the", "SPINN", "-", "PI", "models", "make", "use", "of", "sentence", "parser", "output", ";", "which", "present", "strong", "baseline", "systems", ".", "The", "last", "set", "of", "methods", "designs", "inter", "-", "sentence", "relation", "with", "soft", "attention", ".", "Our", "best", "score", "on", "this", "task", "is", "87.3", "%", "accuracy", "obtained", "with", "the", "full", "tree", "matching", "NTI", "model", ".", "The", "previous", "best", "performing", "model", "on", "the", "task", "performs", "phrase", "matching", "by", "using", "the", "attention", "mechanism", ".", "Our", "results", "show", "that", "NTI", "-", "SLSTM", "improved", "the", "performance", "of", "the", "sequential", "LSTM", "encoder", "by", "approximately", "2", "%", ".", "Not", "surprisingly", ",", "using", "LSTM", "as", "leaf", "node", "function", "helps", "in", "learning", "better", "representations", ".", "Our", "NTI", "-", "SLSTM", "-", "LSTM", "is", "a", "hybrid", "model", "which", "encodes", "a", "sequence", "sequentially", "through", "its", "leaf", "node", "function", "and", "then", "hierarchically", "composes", "the", "output", "representations", ".", "The", "node", "-", "by", "-", "node", "attention", "models", "improve", "the", "performance", ",", "indicating", "that", "modeling", "inter", "-", "sentence", "interaction", "is", "an", "important", "element", "in", "NLI", ".", "Aggregating", "matching", "vector", "between", "trees", "or", "sequences", "with", "a", "separate", "LSTM", "model", "is", "effective", ".", "The", "global", "attention", "seems", "to", "be", "robust", "on", "this", "task", ".", "The", "tree", "attention", "were", "not", "helpful", "as", "it", "normalizes", "the", "attention", "scores", "locally", "in", "parent", "-", "child", "relationship", ".", "subsection", ":", "Answer", "Sentence", "Selection", "For", "this", "task", ",", "a", "model", "is", "trained", "to", "identify", "the", "correct", "sentences", "that", "answer", "a", "factual", "question", ",", "from", "a", "set", "of", "candidate", "sentences", ".", "We", "experiment", "on", "WikiQA", "dataset", "constructed", "from", "Wikipedia", ".", "The", "dataset", "contains", "20", ",", "360", "/", "2", ",", "733", "/", "6", ",", "165", "QA", "pairs", "for", "train", "/", "dev", "/", "test", "sets", ".", "We", "used", "the", "same", "setup", "in", "the", "language", "inference", "task", "except", "that", "we", "replace", "the", "layer", "with", "a", "layer", "and", "model", "the", "following", "conditional", "probability", "distribution", ".", "where", "and", "are", "the", "question", "and", "the", "answer", "encoded", "vectors", "and", "denotes", "the", "output", "of", "the", "hidden", "layer", "of", "the", "MLP", ".", "For", "this", "task", ",", "we", "use", "NTI", "-", "SLSTM", "-", "LSTM", "to", "encode", "answer", "candidate", "sentences", "and", "NTI", "-", "ANF", "-", "LSTM", "to", "encode", "the", "question", "sentences", ".", "Note", "that", "NTI", "-", "ANF", "-", "LSTM", "is", "relied", "on", "ANF", "as", "the", "non", "-", "leaf", "node", "function", ".", "vector", "for", "NTI", "-", "ANF", "-", "LSTM", "is", "the", "answer", "representation", "produced", "by", "the", "answer", "encoding", "NTI", "-", "SLSTM", "-", "LSTM", "model", ".", "We", "set", "the", "batch", "size", "to", "4", "and", "the", "initial", "learning", "rate", "to", "1e", "-", "3", ",", "and", "train", "the", "model", "for", "10", "epochs", ".", "We", "used", "20", "%", "input", "dropouts", "and", "no", "weight", "decay", ".", "Following", "previous", "work", ",", "we", "adopt", "MAP", "and", "MRR", "as", "the", "evaluation", "metrics", "for", "this", "task", ".", "Table", "[", "reference", "]", "presents", "the", "results", "of", "our", "model", "and", "the", "previous", "models", "for", "the", "task", ".", "The", "classifier", "with", "handcrafted", "features", "is", "a", "SVM", "model", "trained", "with", "a", "set", "of", "features", ".", "The", "Bigram", "-", "CNN", "model", "is", "a", "simple", "convolutional", "neural", "net", ".", "The", "Deep", "LSTM", "and", "LSTM", "attention", "models", "outperform", "the", "previous", "best", "result", "by", "a", "large", "margin", ",", "nearly", "5", "-", "6", "%", ".", "NASM", "improves", "the", "result", "further", "and", "sets", "a", "strong", "baseline", "by", "combining", "variational", "auto", "-", "encoder", "with", "the", "soft", "attention", ".", "In", "NASM", ",", "they", "adopt", "a", "deep", "three", "-", "layer", "LSTM", "and", "introduced", "a", "latent", "stochastic", "attention", "mechanism", "over", "the", "answer", "sentence", ".", "Our", "NTI", "model", "exceeds", "NASM", "by", "approximately", "0.4", "%", "on", "MAP", "for", "this", "task", ".", "subsection", ":", "Sentence", "Classification", "Lastly", ",", "we", "evaluated", "NTI", "on", "the", "Stanford", "Sentiment", "Treebank", "(", "SST", ")", ".", "This", "dataset", "comes", "with", "standard", "train", "/", "dev", "/", "test", "sets", "and", "two", "subtasks", ":", "binary", "sentence", "classification", "or", "fine", "-", "grained", "classification", "of", "five", "classes", ".", "We", "trained", "our", "model", "on", "the", "text", "spans", "corresponding", "to", "labeled", "phrases", "in", "the", "training", "set", "and", "evaluated", "the", "model", "on", "the", "full", "sentences", ".", "We", "use", "NTI", "-", "SLSTM", "and", "NTI", "-", "SLSTM", "-", "LSTM", "models", "to", "learn", "sentence", "representations", "for", "the", "task", ".", "The", "sentence", "representations", "were", "passed", "to", "a", "two", "-", "layer", "MLP", "for", "classification", ".", "We", "set", "the", "batch", "size", "to", "64", ",", "the", "initial", "learning", "rate", "to", "1e", "-", "3", "and", "regularizer", "strength", "to", "3e", "-", "5", ",", "and", "train", "each", "model", "for", "10", "epochs", ".", "The", "NTI", "-", "SLSTM", "model", "was", "regularized", "by", "10%", "/", "20", "%", "of", "input", "/", "output", "and", "20%", "/", "30", "%", "of", "input", "/", "output", "dropouts", "and", "the", "NTI", "-", "SLSTM", "-", "LSTM", "model", "20", "%", "of", "input", "and", "20%", "/", "30", "%", "of", "input", "/", "output", "dropouts", "for", "binary", "and", "fine", "-", "grained", "settings", ".", "NTI", "-", "SLSTM", "-", "LSTM", "(", "as", "shown", "in", "Table", "[", "reference", "]", ")", "set", "the", "state", "-", "of", "-", "the", "-", "art", "results", "on", "both", "subtasks", ".", "Our", "NTI", "-", "SLSTM", "model", "performed", "slightly", "worse", "than", "its", "constituency", "tree", "-", "based", "counter", "part", ",", "CT", "-", "LSTM", "model", ".", "The", "CT", "-", "LSTM", "model", "composes", "phrases", "according", "to", "the", "output", "of", "a", "sentence", "parser", "and", "uses", "a", "node", "composition", "function", "similar", "to", "S", "-", "LSTM", ".", "After", "we", "transformed", "the", "input", "with", "the", "LSTM", "leaf", "node", "function", ",", "we", "achieved", "the", "best", "performance", "on", "this", "task", ".", "section", ":", "Qualitative", "Analysis", "subsection", ":", "Attention", "and", "Compositionality", "To", "help", "analyzing", "the", "results", ",", "we", "output", "attention", "weights", "by", "our", "NTI", "-", "SLSTM", "node", "-", "by", "-", "node", "global", "attention", "model", ".", "Figure", "[", "reference", "]", "shows", "the", "attention", "heatmaps", "for", "two", "sentences", "in", "the", "SNLI", "test", "set", ".", "It", "shows", "that", "our", "model", "semantically", "aligns", "single", "or", "multiword", "expressions", "(", "”", "little", "child", "”", "and", "”", "toddler", "”", ";", "”", "rock", "wall", "”", "and", "”", "stone", "”", ")", ".", "In", "addition", ",", "our", "model", "is", "able", "to", "re", "-", "orient", "its", "attention", "over", "different", "parts", "of", "the", "hypothesis", "when", "the", "expression", "is", "more", "complex", ".", "For", "example", ",", "for", "(", "c", ")", "”", "rock", "wall", "in", "autumn", "”", ",", "NTI", "mostly", "focuses", "on", "the", "nodes", "in", "depth", "1", ",", "2", "and", "3", "representing", "contexts", "related", "to", "”", "a", "stone", "”", ",", "”", "leaves", ".", "”", "and", "”", "a", "stone", "wall", "surrounded", "”", ".", "Surprisingly", ",", "attention", "degree", "for", "the", "single", "word", "expression", "like", "”", "stone", "”", ",", "”", "wall", "”", "and", "”", "leaves", "”", "is", "lower", "to", "compare", "with", "multiword", "phrases", ".", "Sequence", "models", "lack", "this", "property", "as", "they", "have", "no", "explicit", "composition", "module", "to", "produce", "such", "mutiword", "phrases", ".", "Finally", ",", "the", "most", "interesting", "pattern", "is", "that", "the", "model", "attends", "over", "higher", "level", "(", "low", "depth", ")", "tree", "nodes", "with", "rich", "semantics", "when", "considering", "a", "(", "c", ")", "longer", "phrase", "or", "(", "d", ")", "full", "sentence", ".", "As", "shown", "in", "(", "d", ")", ",", "the", "NTI", "model", "aligns", "the", "root", "node", "representing", "the", "whole", "hypothesis", "sentence", "to", "the", "higher", "level", "tree", "nodes", "covering", "larger", "sub", "-", "trees", "in", "the", "premise", ".", "It", "certainly", "ignores", "the", "lower", "level", "single", "word", "expressions", "and", "only", "starts", "to", "attend", "when", "the", "words", "are", "collectively", "to", "form", "rich", "semantics", ".", "subsection", ":", "Learned", "Representations", "of", "Phrases", "and", "Sentences", "Using", "cosine", "similarity", "between", "their", "representations", "produced", "by", "the", "NTI", "-", "SLSTM", "model", ",", "we", "show", "that", "NTI", "is", "able", "to", "capture", "paraphrases", "on", "SNLI", "test", "data", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "NTI", "seems", "to", "distinguish", "plural", "from", "singular", "forms", "(", "similar", "phrases", "to", "”", "a", "person", "”", ")", ".", "In", "addition", ",", "NTI", "captures", "non", "-", "surface", "knowledge", ".", "For", "example", ",", "the", "phrases", "similar", "to", "”", "park", "for", "fun", "”", "tend", "to", "align", "to", "the", "semantic", "content", "of", "fun", "and", "park", ",", "including", "”", "people", "play", "frisbee", "outdoors", "”", ".", "The", "NTI", "model", "was", "able", "to", "relate", "”", "Santa", "Claus", "”", "to", "christmas", "and", "snow", ".", "Interestingly", ",", "the", "learned", "representations", "were", "also", "able", "to", "connect", "implicit", "semantics", ".", "For", "example", ",", "NTI", "found", "that", "”", "sad", ",", "depressed", ",", "and", "hatred", "”", "is", "close", "to", "the", "phrases", "like", "”", "an", "Obama", "supporter", "is", "upset", "”", ".", "Overall", "the", "NTI", "model", "is", "robust", "to", "the", "length", "of", "the", "phrases", "being", "matched", ".", "Given", "a", "short", "phrase", ",", "NTI", "can", "retrieve", "longer", "yet", "semantically", "coherent", "sequences", "from", "the", "SNLI", "test", "set", ".", "In", "Table", "[", "reference", "]", ",", "we", "show", "nearest", "-", "neighbor", "sentences", "from", "SNLI", "test", "set", ".", "Note", "that", "the", "sentences", "listed", "in", "the", "first", "two", "columns", "sound", "semantically", "coherent", "but", "not", "the", "ones", "in", "the", "last", "column", ".", "The", "query", "sentence", "”", "A", "dog", "sells", "a", "women", "a", "hat", "”", "does", "not", "actually", "represent", "a", "common", "-", "sense", "knowledge", "and", "this", "sentence", "now", "seem", "to", "confuse", "the", "NTI", "model", ".", "As", "a", "result", ",", "the", "retrieved", "sentence", "are", "arbitrary", "and", "not", "coherent", ".", "subsection", ":", "Effects", "of", "Padding", "Size", "We", "introduced", "a", "special", "padding", "character", "in", "order", "to", "construct", "full", "binary", "tree", ".", "Does", "this", "padding", "character", "influence", "the", "performance", "of", "the", "NTI", "models", "?", "In", "Figure", "[", "reference", "]", ",", "we", "show", "relationship", "between", "the", "padding", "size", "and", "the", "accuracy", "on", "Stanford", "sentiment", "analysis", "data", ".", "Each", "sentence", "was", "padded", "to", "form", "a", "full", "binary", "tree", ".", "The", "x", "-", "axis", "represents", "the", "number", "of", "padding", "characters", "introduced", ".", "When", "the", "padding", "size", "is", "less", "(", "up", "to", "10", ")", ",", "the", "NTI", "-", "SLSTM", "-", "LSTM", "model", "performs", "better", ".", "However", ",", "this", "model", "tends", "to", "perform", "poorly", "or", "equally", "when", "the", "padding", "size", "is", "large", ".", "Overall", "we", "do", "not", "observe", "any", "significant", "performance", "drop", "for", "both", "models", "as", "the", "padding", "size", "increases", ".", "This", "suggests", "that", "NTI", "learns", "to", "ignore", "the", "special", "padding", "character", "while", "processing", "padded", "sentences", ".", "The", "same", "scenario", "was", "also", "observed", "while", "analyzing", "attention", "weights", ".", "The", "attention", "over", "the", "padded", "nodes", "was", "nearly", "zero", ".", "section", ":", "Discussion", "and", "Conclusion", "We", "introduced", "Neural", "Tree", "Indexers", ",", "a", "class", "of", "tree", "structured", "recursive", "neural", "network", ".", "The", "NTI", "models", "achieved", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "different", "NLP", "tasks", ".", "Most", "of", "the", "NTI", "models", "form", "deep", "neural", "networks", "and", "we", "think", "this", "is", "one", "reason", "that", "NTI", "works", "well", "even", "if", "it", "lacks", "direct", "linguistic", "motivations", "followed", "by", "other", "syntactic", "-", "tree", "-", "structured", "recursive", "models", ".", "CNN", "and", "NTI", "are", "topologically", "related", ".", "Both", "NTI", "and", "CNNs", "are", "hierarchical", ".", "However", ",", "current", "implementation", "of", "NTI", "only", "operates", "on", "non", "-", "overlapping", "sub", "-", "trees", "while", "CNNs", "can", "slide", "over", "the", "input", "to", "produce", "higher", "-", "level", "representations", ".", "NTI", "is", "flexible", "in", "selecting", "the", "node", "function", "and", "the", "attention", "mechanism", ".", "Like", "CNN", ",", "the", "computation", "in", "the", "same", "tree", "-", "depth", "can", "be", "parallelized", "effectively", ";", "and", "therefore", "NTI", "is", "scalable", "and", "suitable", "for", "large", "-", "scale", "sequence", "processing", ".", "Note", "that", "NTI", "can", "be", "seen", "as", "a", "generalization", "of", "LSTM", ".", "If", "we", "construct", "left", "-", "branching", "trees", "in", "a", "bottom", "-", "up", "fashion", ",", "the", "model", "acts", "just", "like", "sequential", "LSTM", ".", "Different", "branching", "factors", "for", "the", "underlying", "tree", "structure", "have", "yet", "to", "be", "explored", ".", "NTI", "can", "be", "extended", "so", "it", "learns", "to", "select", "and", "compose", "dynamic", "number", "of", "nodes", "for", "efficiency", ",", "essentially", "discovering", "intrinsic", "hierarchical", "structure", "in", "the", "input", ".", "section", ":", "Acknowledgments", "We", "would", "like", "to", "thank", "the", "anonymous", "reviewers", "for", "their", "insightful", "comments", "and", "suggestions", ".", "This", "work", "was", "supported", "in", "part", "by", "the", "grant", "HL125089", "from", "the", "National", "Institutes", "of", "Health", "(", "NIH", ")", ".", "Any", "opinions", ",", "findings", "and", "conclusions", "or", "recommendations", "expressed", "in", "this", "material", "are", "those", "of", "the", "authors", "and", "do", "not", "necessarily", "reflect", "those", "of", "the", "sponsor", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["Neural", "Tree", "Indexers"], "Method"], [6, 8, ["Text", "Understanding"], "Task"], [8, 11, ["Recurrent", "neural", "networks"], "Method"], [12, 13, ["RNNs"], "Method"], [33, 35, ["recursive", "networks"], "Method"], [54, 56, ["recursive", "architecture"], "Method"], [81, 84, ["Neural", "Tree", "Indexers"], "Method"], [85, 86, ["NTI"], "Method"], [105, 106, ["NTI"], "Method"], [129, 131, ["Attention", "mechanism"], "Method"], [170, 172, ["NLP", "tasks"], "Task"], [173, 176, ["natural", "language", "inference"], "Task"], [177, 180, ["answer", "sentence", "selection"], "Task"], [182, 184, ["sentence", "classification"], "Task"], [193, 198, ["recurrent", "and", "recursive", "neural", "networks"], "Method"], [202, 205, ["Recurrent", "neural", "networks"], "Method"], [206, 207, ["RNNs"], "Method"], [216, 217, ["RNNs"], "Method"], [223, 228, ["internal", "short", "-", "term", "memories"], "Method"], [231, 236, ["long", "short", "-", "term", "memories"], "Method"], [237, 238, ["LSTM"], "Method"], [246, 248, ["NLP", "tasks"], "Task"], [249, 252, ["named", "entity", "recognition"], "Task"], [253, 255, ["constituency", "parsing"], "Task"], [256, 259, ["textual", "entailment", "recognition"], "Task"], [260, 262, ["question", "answering"], "Task"], [264, 266, ["machine", "translation"], "Task"], [329, 331, ["sequential", "models"], "Method"], [332, 335, ["recursive", "neural", "networks"], "Method"], [348, 350, ["sentiment", "analysis"], "Task"], [361, 363, ["NLP", "applications"], "Task"], [370, 373, ["Neural", "Tree", "Indexers"], "Method"], [374, 375, ["NTI"], "Method"], [384, 386, ["NLP", "tasks"], "Task"], [387, 388, ["NTI"], "Method"], [415, 416, ["NTI"], "Method"], [422, 425, ["node", "transformation", "functions"], "Method"], [426, 429, ["leaf", "node", "mapping"], "Method"], [439, 441, ["recursive", "models"], "Method"], [446, 447, ["NTI"], "Method"], [452, 453, ["NTI"], "Method"], [479, 481, ["NLP", "tasks"], "Task"], [482, 484, ["sentence", "modeling"], "Task"], [492, 495, ["node", "composition", "function"], "Method"], [496, 499, ["attention", "over", "tree"], "Method"], [506, 510, ["sequential", "leaf", "node", "transformer"], "Method"], [512, 513, ["LSTM"], "Method"], [530, 539, ["conditional", "and", "compositional", "powers", "of", "sequential", "and", "recursive", "models"], "Method"], [548, 549, ["NTI"], "Method"], [579, 581, ["NLP", "applications"], "Task"], [582, 585, ["natural", "language", "inference"], "Task"], [586, 589, ["answer", "sentence", "selection"], "Task"], [591, 593, ["sentence", "classification"], "Task"], [600, 603, ["Recurrent", "Neural", "Networks"], "Method"], [604, 606, ["Attention", "Mechanism"], "Method"], [606, 607, ["RNNs"], "Method"], [655, 656, ["RNNs"], "Method"], [678, 681, ["information", "flow", "bottleneck"], "Task"], [698, 701, ["information", "flow", "bottleneck"], "Task"], [704, 705, ["RNNs"], "Method"], [707, 710, ["soft", "attention", "mechanism"], "Method"], [714, 717, ["neural", "machine", "translation"], "Task"], [728, 729, ["RNNs"], "Method"], [730, 734, ["linear", "chain", "-", "structured"], "Method"], [767, 770, ["Recursive", "Neural", "Networks"], "Method"], [771, 772, ["RNNs"], "Method"], [773, 776, ["recursive", "neural", "networks"], "Method"], [798, 800, ["syntactic", "parser"], "Method"], [811, 814, ["node", "composition", "function"], "Method"], [847, 850, ["single", "layer", "network"], "Method"], [857, 862, ["recursive", "auto", "-", "associate", "memories"], "Method"], [863, 865, ["recursive", "autoencoders"], "Method"], [873, 875, ["matrix", "representation"], "Method"], [887, 889, ["Tensor", "networks"], "Method"], [894, 896, ["composition", "function"], "Method"], [897, 903, ["sentence", "-", "level", "sentiment", "analysis", "task"], "Task"], [908, 911, ["S", "-", "LSTM"], "Method"], [932, 936, ["attentive", "node", "composition", "function"], "Method"], [940, 943, ["S", "-", "LSTM"], "Method"], [974, 975, ["NTI"], "Method"], [1004, 1006, ["NLP", "tasks"], "Task"], [1013, 1014, ["NTI"], "Method"], [1017, 1022, ["vanishing", "/", "explode", "gradients", "challenges"], "Task"], [1025, 1026, ["RNNs"], "Method"], [1067, 1069, ["word", "embedding"], "Method"], [1070, 1071, ["NTI"], "Method"], [1119, 1121, ["word", "embedding"], "Method"], [1146, 1147, ["NTI"], "Method"], [1210, 1212, ["question", "answering"], "Task"], [1233, 1235, ["neural", "networks"], "Method"], [1242, 1243, ["NTI"], "Method"], [1251, 1253, ["node", "representations"], "Method"], [1256, 1257, ["LSTM"], "Method"], [1258, 1262, ["attentive", "node", "composition", "functions"], "Method"], [1282, 1287, ["non", "-", "leaf", "node", "function"], "Task"], [1297, 1300, ["S", "-", "LSTM"], "Method"], [1305, 1306, ["LSTM"], "Method"], [1308, 1312, ["non", "-", "leaf", "node"], "Task"], [1323, 1324, ["LSTM"], "Method"], [1331, 1333, ["node", "representation"], "Method"], [1356, 1359, ["S", "-", "LSTM"], "Method"], [1361, 1364, ["parent", "node", "representation"], "Method"], [1396, 1401, ["element", "-", "wise", "vector", "multiplication"], "Method"], [1428, 1431, ["S", "-", "LSTM"], "Method"], [1438, 1444, ["Attentive", "Non", "-", "leaf", "Node", "Function"], "Method"], [1445, 1446, ["ANF"], "Method"], [1449, 1451, ["NLP", "applications"], "Task"], [1454, 1455, ["QA"], "Task"], [1456, 1458, ["machine", "translation"], "Task"], [1463, 1468, ["dynamic", "query", "dependent", "composition", "function"], "Method"], [1471, 1472, ["ANF"], "Method"], [1482, 1485, ["S", "-", "LSTM"], "Method"], [1486, 1487, ["ANF"], "Method"], [1506, 1508, ["learnable", "representation"], "Method"], [1510, 1512, ["sequence", "representation"], "Method"], [1554, 1557, ["attention", "scoring", "function"], "Method"], [1564, 1570, ["multi", "-", "layer", "perceptron", "(", "MLP"], "Method"], [1573, 1577, ["matrix", "-", "vector", "product"], "Method"], [1602, 1606, ["non", "-", "linear", "transformation"], "Task"], [1618, 1619, ["NTI"], "Method"], [1645, 1646, ["NTI"], "Method"], [1677, 1679, ["attention", "mechanism"], "Method"], [1680, 1681, ["tree"], "Method"], [1686, 1688, ["attention", "mechanism"], "Method"], [1720, 1722, ["tree", "attention"], "Method"], [1736, 1738, ["Global", "Attention"], "Method"], [1740, 1743, ["attention", "neural", "network"], "Method"], [1745, 1747, ["global", "attention"], "Method"], [1765, 1767, ["neural", "net"], "Method"], [1770, 1771, ["ANF"], "Method"], [1813, 1815, ["attention", "mechanism"], "Method"], [1851, 1854, ["global", "attention", "network"], "Method"], [1891, 1893, ["node", "representations"], "Method"], [1903, 1906, ["parent", "node", "representation"], "Method"], [1989, 1991, ["tree", "attention"], "Method"], [2010, 2012, ["NLP", "tasks"], "Task"], [2013, 2016, ["natural", "language", "inference"], "Task"], [2017, 2019, ["question", "answering"], "Task"], [2020, 2022, ["sentence", "classification"], "Task"], [2030, 2031, ["NTI"], "Method"], [2038, 2039, ["NTI"], "Method"], [2040, 2041, ["Adam"], "Method"], [2042, 2043, ["hyperparameters"], "Method"], [2062, 2064, ["word", "embeddings"], "Task"], [2155, 2156, ["dropouts"], "Method"], [2158, 2160, ["weight", "decay"], "Method"], [2163, 2166, ["Natural", "Language", "Inference"], "Task"], [2171, 2175, ["Stanford", "Natural", "Language", "Inference"], "Material"], [2176, 2177, ["SNLI"], "Material"], [2227, 2228, ["MLP"], "Method"], [2229, 2230, ["classification"], "Task"], [2248, 2250, ["sentence", "representations"], "Method"], [2252, 2253, ["MLP"], "Method"], [2256, 2258, ["input", "layer"], "Method"], [2259, 2261, ["1024", "units"], "Method"], [2262, 2263, ["activation"], "Method"], [2265, 2267, ["output", "layer"], "Method"], [2299, 2301, ["initial", "learning"], "Method"], [2319, 2322, ["NTI", "-", "SLSTM"], "Method"], [2329, 2330, ["transformer"], "Method"], [2349, 2351, ["learning", "rate"], "Metric"], [2372, 2374, ["neural", "net"], "Method"], [2379, 2381, ["input", "dropouts"], "Method"], [2388, 2393, ["NTI", "-", "SLSTM", "-", "LSTM"], "Method"], [2396, 2397, ["LSTM"], "Method"], [2412, 2415, ["NTI", "-", "SLSTM"], "Method"], [2446, 2456, ["NTI", "-", "SLSTM", "node", "-", "by", "-", "node", "global", "attention"], "Method"], [2478, 2485, ["word", "-", "by", "-", "word", "attention", "model"], "Method"], [2501, 2503, ["hypothesis", "encoding"], "Method"], [2512, 2515, ["NTI", "-", "SLSTMs"], "Method"], [2528, 2530, ["learning", "rate"], "Metric"], [2551, 2553, ["neural", "net"], "Method"], [2567, 2577, ["NTI", "-", "SLSTM", "node", "-", "by", "-", "node", "tree", "attention"], "Method"], [2603, 2615, ["NTI", "-", "SLSTM", "-", "LSTM", "node", "-", "by", "-", "node", "global", "attention"], "Method"], [2621, 2622, ["LSTM"], "Method"], [2642, 2646, ["hidden", "/", "memory", "state"], "Method"], [2665, 2667, ["learning", "rate"], "Metric"], [2688, 2690, ["neural", "net"], "Method"], [2704, 2716, ["NTI", "-", "SLSTM", "-", "LSTM", "node", "-", "by", "-", "node", "tree", "attention"], "Method"], [2742, 2751, ["Tree", "matching", "NTI", "-", "SLSTM", "-", "LSTM", "global", "attention"], "Method"], [2764, 2770, ["NTI", "-", "SLSTM", "-", "LSTM", "model"], "Method"], [2779, 2781, ["global", "attention"], "Method"], [2784, 2785, ["LSTM"], "Method"], [2825, 2826, ["MLP"], "Method"], [2827, 2828, ["classification"], "Task"], [2830, 2831, ["MLP"], "Method"], [2838, 2840, ["input", "layer"], "Method"], [2847, 2849, ["output", "layer"], "Method"], [2870, 2872, ["match", "trees"], "Method"], [2877, 2879, ["learning", "rate"], "Metric"], [2900, 2902, ["neural", "net"], "Method"], [2907, 2909, ["input", "dropouts"], "Method"], [2916, 2925, ["Tree", "matching", "NTI", "-", "SLSTM", "-", "LSTM", "tree", "attention"], "Method"], [2948, 2958, ["Full", "tree", "matching", "NTI", "-", "SLSTM", "-", "LSTM", "global", "attention"], "Method"], [3034, 3035, ["classification"], "Task"], [3091, 3094, ["sentence", "encoder", "models"], "Method"], [3097, 3099, ["word", "embeddings"], "Method"], [3106, 3110, ["SPINN", "-", "PI", "models"], "Method"], [3129, 3133, ["inter", "-", "sentence", "relation"], "Task"], [3146, 3147, ["accuracy"], "Metric"], [3150, 3155, ["full", "tree", "matching", "NTI", "model"], "Method"], [3165, 3167, ["phrase", "matching"], "Task"], [3170, 3172, ["attention", "mechanism"], "Method"], [3177, 3180, ["NTI", "-", "SLSTM"], "Method"], [3185, 3188, ["sequential", "LSTM", "encoder"], "Method"], [3197, 3198, ["LSTM"], "Method"], [3206, 3207, ["representations"], "Task"], [3209, 3214, ["NTI", "-", "SLSTM", "-", "LSTM"], "Method"], [3216, 3218, ["hybrid", "model"], "Method"], [3233, 3235, ["output", "representations"], "Method"], [3237, 3244, ["node", "-", "by", "-", "node", "attention", "models"], "Method"], [3251, 3255, ["inter", "-", "sentence", "interaction"], "Task"], [3260, 3261, ["NLI"], "Task"], [3262, 3265, ["Aggregating", "matching", "vector"], "Task"], [3309, 3312, ["Answer", "Sentence", "Selection"], "Task"], [3341, 3343, ["WikiQA", "dataset"], "Material"], [3378, 3381, ["language", "inference", "task"], "Task"], [3414, 3416, ["hidden", "layer"], "Method"], [3418, 3419, ["MLP"], "Method"], [3426, 3431, ["NTI", "-", "SLSTM", "-", "LSTM"], "Method"], [3458, 3459, ["ANF"], "Method"], [3467, 3468, ["vector"], "Method"], [3476, 3478, ["answer", "representation"], "Method"], [3500, 3502, ["learning", "rate"], "Metric"], [3523, 3525, ["weight", "decay"], "Method"], [3532, 3533, ["MAP"], "Method"], [3534, 3535, ["MRR"], "Method"], [3537, 3539, ["evaluation", "metrics"], "Metric"], [3562, 3563, ["classifier"], "Method"], [3568, 3570, ["SVM", "model"], "Method"], [3578, 3582, ["Bigram", "-", "CNN", "model"], "Method"], [3585, 3588, ["convolutional", "neural", "net"], "Method"], [3590, 3592, ["Deep", "LSTM"], "Method"], [3593, 3596, ["LSTM", "attention", "models"], "Method"], [3612, 3613, ["NASM"], "Method"], [3624, 3628, ["variational", "auto", "-", "encoder"], "Method"], [3634, 3635, ["NASM"], "Method"], [3647, 3651, ["latent", "stochastic", "attention", "mechanism"], "Method"], [3660, 3661, ["NASM"], "Method"], [3666, 3667, ["MAP"], "Task"], [3673, 3675, ["Sentence", "Classification"], "Task"], [3679, 3680, ["NTI"], "Method"], [3682, 3685, ["Stanford", "Sentiment", "Treebank"], "Material"], [3686, 3687, ["SST"], "Material"], [3704, 3707, ["binary", "sentence", "classification"], "Task"], [3708, 3712, ["fine", "-", "grained", "classification"], "Task"], [3743, 3746, ["NTI", "-", "SLSTM"], "Method"], [3747, 3752, ["NTI", "-", "SLSTM", "-", "LSTM"], "Method"], [3755, 3757, ["sentence", "representations"], "Method"], [3762, 3764, ["sentence", "representations"], "Method"], [3770, 3772, ["layer", "MLP"], "Method"], [3773, 3774, ["classification"], "Task"], [3785, 3787, ["learning", "rate"], "Metric"], [3808, 3812, ["NTI", "-", "SLSTM", "model"], "Method"], [3829, 3833, ["input", "/", "output", "dropouts"], "Method"], [3835, 3841, ["NTI", "-", "SLSTM", "-", "LSTM", "model"], "Method"], [3856, 3862, ["binary", "and", "fine", "-", "grained", "settings"], "Task"], [3863, 3868, ["NTI", "-", "SLSTM", "-", "LSTM"], "Method"], [3892, 3896, ["NTI", "-", "SLSTM", "model"], "Method"], [3926, 3928, ["sentence", "parser"], "Method"], [3931, 3934, ["node", "composition", "function"], "Method"], [3936, 3939, ["S", "-", "LSTM"], "Method"], [3963, 3965, ["Qualitative", "Analysis"], "Task"], [3967, 3968, ["Attention"], "Task"], [3982, 3993, ["NTI", "-", "SLSTM", "node", "-", "by", "-", "node", "global", "attention", "model"], "Method"], [4082, 4083, ["NTI"], "Method"], [4145, 4147, ["Sequence", "models"], "Method"], [4155, 4157, ["composition", "module"], "Method"], [4276, 4280, ["NTI", "-", "SLSTM", "model"], "Method"], [4284, 4285, ["NTI"], "Method"], [4303, 4304, ["NTI"], "Method"], [4324, 4325, ["NTI"], "Method"], [4440, 4441, ["NTI"], "Method"], [4542, 4544, ["Padding", "Size"], "Metric"], [4585, 4586, ["accuracy"], "Metric"], [4587, 4591, ["Stanford", "sentiment", "analysis", "data"], "Material"], [4628, 4634, ["NTI", "-", "SLSTM", "-", "LSTM", "model"], "Method"], [4675, 4676, ["NTI"], "Method"], [4716, 4719, ["Neural", "Tree", "Indexers"], "Method"], [4743, 4745, ["NLP", "tasks"], "Task"], [4752, 4755, ["deep", "neural", "networks"], "Method"], [4763, 4764, ["NTI"], "Method"], [4784, 4785, ["CNN"], "Method"], [4786, 4787, ["NTI"], "Method"], [4792, 4793, ["NTI"], "Method"], [4794, 4795, ["CNNs"], "Method"], [4803, 4804, ["NTI"], "Method"], [4814, 4815, ["CNNs"], "Method"], [4824, 4826, ["level", "representations"], "Method"], [4827, 4828, ["NTI"], "Method"], [4837, 4839, ["attention", "mechanism"], "Method"], [4841, 4842, ["CNN"], "Method"], [4858, 4859, ["NTI"], "Method"], [4864, 4869, ["large", "-", "scale", "sequence", "processing"], "Task"], [4872, 4873, ["NTI"], "Method"], [4918, 4919, ["NTI"], "Method"], [63, 64, ["tree"], "Method"], [77, 78, ["tree"], "Method"], [95, 96, ["RNNs"], "Method"], [99, 100, ["tree"], "Method"], [112, 113, ["tree"], "Method"], [149, 150, ["tree"], "Method"], [152, 153, ["NTI"], "Method"], [270, 271, ["LSTM"], "Method"], [299, 300, ["LSTM"], "Method"], [325, 326, ["tree"], "Method"], [340, 341, ["tree"], "Method"], [357, 358, ["tree"], "Method"], [380, 381, ["tree"], "Method"], [404, 405, ["tree"], "Method"], [443, 444, ["tree"], "Method"], [501, 502, ["NTI"], "Method"], [517, 518, ["NTI"], "Method"], [523, 524, ["tree"], "Method"], [545, 546, ["tree"], "Method"], [558, 559, ["tree"], "Method"], [788, 789, ["tree"], "Method"], [791, 792, ["tree"], "Method"], [805, 806, ["tree"], "Method"], [913, 914, ["LSTM"], "Method"], [917, 918, ["tree"], "Method"], [945, 946, ["NTI"], "Method"], [993, 994, ["NTI"], "Method"], [1009, 1012, ["attention", "over", "tree"], "Method"], [1077, 1078, ["tree"], "Method"], [1152, 1153, ["tree"], "Method"], [1164, 1165, ["tree"], "Method"], [1167, 1168, ["NTI"], "Method"], [1204, 1205, ["NTI"], "Method"], [1216, 1219, ["natural", "language", "inference"], "Task"], [1288, 1289, ["LSTM"], "Method"], [1317, 1318, ["LSTM"], "Method"], [1325, 1326, ["tree"], "Method"], [1406, 1407, ["LSTM"], "Method"], [1615, 1616, ["LSTM"], "Method"], [1628, 1629, ["tree"], "Method"], [1633, 1634, ["tree"], "Method"], [1702, 1703, ["NTI"], "Method"], [1714, 1716, ["global", "attention"], "Method"], [1762, 1763, ["tree"], "Method"], [1793, 1795, ["global", "attention"], "Method"], [1837, 1838, ["tree"], "Method"], [1856, 1857, ["tree"], "Method"], [1862, 1863, ["tree"], "Method"], [1949, 1951, ["global", "attention"], "Method"], [1978, 1979, ["tree"], "Method"], [1985, 1987, ["global", "attention"], "Method"], [2098, 2099, ["tree"], "Method"], [2142, 2143, ["NTI"], "Method"], [2233, 2234, ["NTI"], "Method"], [2275, 2276, ["NTI"], "Method"], [2335, 2336, ["LSTM"], "Method"], [2406, 2407, ["LSTM"], "Method"], [2425, 2426, ["LSTM"], "Method"], [2430, 2431, ["LSTM"], "Method"], [2466, 2468, ["global", "attention"], "Method"], [2472, 2473, ["tree"], "Method"], [2494, 2495, ["tree"], "Method"], [2588, 2589, ["tree"], "Method"], [2637, 2638, ["LSTM"], "Method"], [2639, 2640, ["LSTM"], "Method"], [2648, 2649, ["LSTM"], "Method"], [2651, 2652, ["LSTM"], "Method"], [2727, 2728, ["tree"], "Method"], [2794, 2795, ["tree"], "Method"], [2802, 2803, ["LSTM"], "Method"], [2807, 2808, ["LSTM"], "Method"], [2833, 2834, ["tree"], "Method"], [2854, 2855, ["LSTM"], "Method"], [2867, 2868, ["LSTM"], "Method"], [2929, 2931, ["global", "attention"], "Method"], [2933, 2934, ["tree"], "Method"], [2975, 2976, ["tree"], "Method"], [2979, 2980, ["tree"], "Method"], [2988, 2989, ["tree"], "Method"], [2992, 2993, ["tree"], "Method"], [3005, 3006, ["LSTM"], "Method"], [3010, 3011, ["tree"], "Method"], [3020, 3021, ["LSTM"], "Method"], [3043, 3044, ["LSTM"], "Method"], [3102, 3103, ["tree"], "Method"], [3272, 3273, ["LSTM"], "Method"], [3278, 3280, ["global", "attention"], "Method"], [3289, 3290, ["tree"], "Method"], [3437, 3438, ["NTI"], "Method"], [3441, 3442, ["LSTM"], "Method"], [3450, 3451, ["NTI"], "Method"], [3454, 3455, ["LSTM"], "Method"], [3469, 3470, ["NTI"], "Method"], [3473, 3474, ["LSTM"], "Method"], [3483, 3484, ["NTI"], "Method"], [3487, 3488, ["LSTM"], "Method"], [3643, 3644, ["LSTM"], "Method"], [3657, 3658, ["NTI"], "Method"], [3902, 3903, ["tree"], "Method"], [3910, 3911, ["LSTM"], "Method"], [3916, 3917, ["LSTM"], "Method"], [3947, 3948, ["LSTM"], "Method"], [4007, 4008, ["SNLI"], "Material"], [4181, 4182, ["tree"], "Method"], [4209, 4210, ["NTI"], "Method"], [4224, 4225, ["tree"], "Method"], [4291, 4292, ["SNLI"], "Material"], [4364, 4365, ["NTI"], "Method"], [4395, 4396, ["NTI"], "Method"], [4422, 4423, ["NTI"], "Method"], [4450, 4451, ["SNLI"], "Material"], [4467, 4468, ["SNLI"], "Material"], [4522, 4523, ["NTI"], "Method"], [4556, 4557, ["tree"], "Method"], [4567, 4568, ["NTI"], "Method"], [4601, 4602, ["tree"], "Method"], [4723, 4724, ["tree"], "Method"], [4730, 4731, ["NTI"], "Method"], [4749, 4750, ["NTI"], "Method"], [4778, 4779, ["tree"], "Method"], [4848, 4849, ["tree"], "Method"], [4880, 4881, ["LSTM"], "Method"], [4902, 4903, ["LSTM"], "Method"], [4910, 4911, ["tree"], "Method"]]} {"doc_id": "2777cd26b2c257843273fe41ad4c5b8cdf1b1b75", "text": ["document", ":", "Understanding", "Humans", "in", "Crowded", "Scenes", ":", "Deep", "Nested", "Adversarial", "Learning", "and", "A", "New", "Benchmark", "for", "Multi", "-", "Human", "Parsing", "Despite", "the", "noticeable", "progress", "in", "perceptual", "tasks", "like", "detection", ",", "instance", "segmentation", "and", "human", "parsing", ",", "computers", "still", "perform", "unsatisfactorily", "on", "visually", "understanding", "humans", "in", "crowded", "scenes", ",", "such", "as", "group", "behavior", "analysis", ",", "person", "re", "-", "identification", "and", "autonomous", "driving", ",", "etc", ".", "To", "this", "end", ",", "models", "need", "to", "comprehensively", "perceive", "the", "semantic", "information", "and", "the", "differences", "between", "instances", "in", "a", "multi", "-", "human", "image", ",", "which", "is", "recently", "defined", "as", "the", "multi", "-", "human", "parsing", "task", ".", "In", "this", "paper", ",", "we", "present", "a", "new", "large", "-", "scale", "database", "“", "M", "ulti", "-", "H", "uman", "P", "arsing", "(", "MHP", ")", "”", "for", "algorithm", "development", "and", "evaluation", ",", "and", "advances", "the", "state", "-", "of", "-", "the", "-", "art", "in", "understanding", "humans", "in", "crowded", "scenes", ".", "MHP", "contains", "25", ",", "403", "elaborately", "annotated", "images", "with", "58", "fine", "-", "grained", "semantic", "category", "labels", ",", "involving", "2", "-", "26", "persons", "per", "image", "and", "captured", "in", "real", "-", "world", "scenes", "from", "various", "viewpoints", ",", "poses", ",", "occlusion", ",", "interactions", "and", "background", ".", "We", "further", "propose", "a", "novel", "deep", "N", "ested", "A", "dversarial", "N", "etwork", "(", "NAN", ")", "model", "for", "multi", "-", "human", "parsing", ".", "NAN", "consists", "of", "three", "G", "enerative", "A", "dversarial", "N", "etwork", "(", "GAN", ")-", "like", "sub", "-", "nets", ",", "respectively", "performing", "semantic", "saliency", "prediction", ",", "instance", "-", "agnostic", "parsing", "and", "instance", "-", "aware", "clustering", ".", "These", "sub", "-", "nets", "form", "a", "nested", "structure", "and", "are", "carefully", "designed", "to", "learn", "jointly", "in", "an", "end", "-", "to", "-", "end", "way", ".", "NAN", "consistently", "outperforms", "existing", "state", "-", "of", "-", "the", "-", "art", "solutions", "on", "our", "MHP", "and", "several", "other", "datasets", ",", "and", "serves", "as", "a", "strong", "baseline", "to", "drive", "the", "future", "research", "for", "multi", "-", "human", "parsing", ".", "section", ":", "Introduction", "One", "of", "the", "primary", "goals", "of", "intelligent", "human", "-", "computer", "interaction", "is", "understanding", "the", "humans", "in", "visual", "scenes", ".", "It", "involves", "several", "perceptual", "tasks", "including", "detection", ",", "i.e.", "localizing", "different", "persons", "at", "a", "coarse", ",", "bounding", "box", "level", "(", "Fig", ".", "[", "reference", "]", "(", "a", ")", ")", ",", "instance", "segmentation", ",", "i.e.", "labelling", "each", "pixel", "of", "each", "person", "uniquely", "(", "Fig", ".", "[", "reference", "]", "(", "b", ")", ")", ",", "and", "human", "parsing", ",", "i.e.", "decomposing", "persons", "into", "their", "semantic", "categories", "(", "Fig", ".", "[", "reference", "]", "(", "c", ")", ")", ".", "Recently", ",", "deep", "learning", "based", "methods", "have", "achieved", "remarkable", "sucess", "in", "these", "perceptual", "tasks", "thanks", "to", "the", "availability", "of", "plentiful", "annotated", "images", "for", "training", "and", "evaluation", "purposes", ".", "Though", "exciting", ",", "current", "progress", "is", "still", "far", "from", "the", "utimate", "goal", "of", "visually", "understanding", "humans", ".", "As", "Fig", ".", "[", "reference", "]", "shows", ",", "previous", "efforts", "on", "understanding", "humans", "in", "visual", "scenes", "either", "only", "consider", "coarse", "information", "or", "are", "agnostic", "to", "different", "instances", ".", "In", "the", "real", "-", "world", "scenarios", ",", "it", "is", "more", "likely", "that", "there", "simutaneously", "exist", "multiple", "persons", ",", "with", "various", "human", "interactions", ",", "poses", "and", "occlusion", ".", "Thus", ",", "it", "is", "more", "practically", "demanded", "to", "parse", "human", "body", "parts", "and", "fashion", "items", "at", "the", "instance", "level", ",", "which", "is", "recently", "defined", "as", "the", "multi", "-", "human", "parsing", "task", ".", "Multi", "-", "human", "parsing", "enables", "more", "detailed", "understanding", "of", "humans", "in", "crowded", "scenes", "and", "aligns", "better", "with", "many", "real", "-", "world", "applications", ",", "such", "as", "group", "behavior", "analysis", ",", "person", "re", "-", "identification", ",", "e", "-", "commerce", ",", "image", "editing", ",", "video", "surveillance", ",", "autonomous", "driving", "and", "virtual", "reality", ".", "However", ",", "the", "existing", "benchmark", "datasets", "are", "not", "suitable", "for", "such", "a", "new", "task", ".", "Even", "though", "Li", "et", "al", ".", "proposed", "a", "preliminary", "M", "ulti", "-", "H", "uman", "P", "arsing", "(", "MHP", "v1.0", ")", "dataset", ",", "it", "only", "contains", "4", ",", "980", "images", "annotated", "with", "18", "semantic", "labels", ".", "In", "this", "work", ",", "we", "propose", "a", "new", "large", "-", "scale", "benchmark", "“", "M", "ulti", "-", "H", "uman", "P", "arsing", "(", "MHP", "v2.0", ")", "”", ",", "aiming", "to", "push", "the", "frontiers", "of", "multi", "-", "human", "parsing", "research", "towards", "holistically", "understanding", "humans", "in", "crowded", "scenes", ".", "The", "data", "in", "MHP", "v2.0", "cover", "wide", "variability", "and", "complexity", "w.r.t", ".", "viewpoints", ",", "poses", ",", "occlusion", ",", "human", "interactions", "and", "background", ".", "It", "in", "total", "includes", "25", ",", "403", "human", "images", "with", "pixel", "-", "wise", "annotations", "of", "58", "semantic", "categories", ".", "We", "further", "propose", "a", "novel", "deep", "N", "ested", "A", "dversarial", "N", "etwork", "(", "NAN", ")", "model", "for", "solving", "the", "challenging", "multi", "-", "human", "parsing", "problem", ".", "Unlike", "most", "existing", "methods", "which", "rely", "on", "separate", "stages", "of", "instance", "localization", ",", "human", "parsing", "and", "result", "refinement", ",", "the", "proposed", "NAN", "parses", "semantic", "categories", "and", "differentiates", "different", "person", "instances", "simultaneously", "in", "an", "effective", "and", "time", "-", "efficient", "manner", ".", "NAN", "consists", "of", "three", "G", "enerative", "A", "dversarial", "N", "etwork", "(", "GAN", ")-", "like", "sub", "-", "nets", ",", "respectively", "performing", "semantic", "saliency", "prediction", ",", "instance", "-", "agnostic", "parsing", "and", "instance", "-", "aware", "clustering", ".", "Each", "sub", "-", "task", "is", "simpler", "than", "the", "original", "multi", "-", "human", "parsing", "task", ",", "and", "is", "more", "easily", "addressed", "by", "the", "corresponding", "sub", "-", "net", ".", "Unlike", "many", "multi", "-", "task", "learning", "applications", ",", "in", "our", "method", "the", "sub", "-", "nets", "depend", "on", "each", "other", ",", "forming", "a", "causal", "nest", "by", "dynamically", "boosting", "each", "other", "through", "an", "adversarial", "strategy", "(", "See", "Fig", ".", "[", "reference", "]", ")", ",", "which", "is", "hence", "called", "a", "“", "nested", "adversarial", "learning", "”", "structure", ".", "Such", "a", "structure", "enables", "effortless", "gradient", "B", "ackpro", "P", "agation", "(", "BP", ")", "in", "NAN", "such", "that", "it", "can", "be", "trained", "in", "a", "holistic", ",", "end", "-", "to", "-", "end", "way", ",", "which", "is", "favorable", "to", "both", "accuracy", "and", "speed", ".", "We", "conduct", "qualitative", "and", "quantitative", "experiments", "on", "the", "MHP", "v2.0", "dataset", "proposed", "in", "this", "work", ",", "as", "well", "as", "the", "MHP", "v1.0", ",", "PASCAL", "-", "Person", "-", "Part", "and", "Buffy", "benchmark", "datasets", ".", "The", "results", "demonstrate", "the", "superiority", "of", "NAN", "on", "multi", "-", "human", "parsing", "over", "the", "state", "-", "of", "-", "the", "-", "arts", ".", "Our", "contributions", "are", "summarized", "as", "follows", ".", "We", "propose", "a", "new", "large", "-", "scale", "benchmark", "and", "evaluation", "server", "to", "advance", "understanding", "of", "humans", "in", "crowded", "scenes", ",", "which", "contains", "25", ",", "403", "images", "annotated", "pixel", "-", "wisely", "with", "58", "semantic", "category", "labels", ".", "We", "propose", "a", "novel", "deep", "N", "ested", "A", "dversarial", "N", "etwork", "(", "NAN", ")", "model", "for", "multi", "-", "human", "parsing", ",", "which", "serves", "as", "a", "strong", "baseline", "to", "inspire", "more", "future", "research", "efforts", "on", "this", "task", ".", "Comprehensive", "evaluations", "on", "the", "MHP", "v2.0", "dataset", "proposed", "in", "this", "work", ",", "as", "well", "as", "the", "MHP", "v1.0", ",", "PASCAL", "-", "Person", "-", "Part", "and", "Buffy", "benchmark", "datasets", "verify", "the", "superiority", "of", "NAN", "on", "understanding", "humans", "in", "crowded", "scenes", "over", "the", "state", "-", "of", "-", "the", "-", "arts", ".", "section", ":", "Related", "Work", "paragraph", ":", "Human", "Parsing", "Datasets", "The", "statistics", "of", "popular", "publicly", "available", "datasets", "for", "human", "parsing", "are", "summarized", "in", "Tab", ".", "[", "reference", "]", ".", "The", "Buffy", "dataset", "was", "released", "in", "2011", "for", "human", "parsing", "and", "instance", "segmentation", ".", "It", "contains", "only", "748", "images", "annotated", "with", "13", "semantic", "categories", ".", "The", "Fashionista", "dataset", "was", "released", "in", "2012", "for", "human", "parsing", ",", "containing", "limited", "images", "annotated", "with", "56", "fashion", "categories", ".", "The", "PASCAL", "-", "Person", "-", "Part", "dataset", "was", "initially", "annotated", "by", "Chen", "et", "al", ".", "from", "the", "PASCAL", "-", "VOC", "-", "2010", "dataset", ".", "Chen", "et", "al", ".", "extended", "it", "for", "human", "parsing", "with", "7", "coarse", "body", "part", "labels", ".", "The", "ATR", "dataset", "was", "released", "in", "2015", "for", "human", "parsing", "with", "a", "large", "number", "of", "images", "annotated", "with", "18", "semantic", "categories", ".", "The", "LIP", "dataset", "further", "extended", "ATR", "by", "cropping", "person", "instances", "from", "Microsoft", "COCO", ".", "It", "is", "a", "large", "-", "scale", "human", "parsing", "dataset", "with", "densely", "pixel", "-", "wise", "annotations", "of", "20", "semantic", "categories", ".", "But", "it", "has", "two", "limitations", ".", "1", ")", "Despite", "the", "large", "data", "size", ",", "it", "contains", "limited", "semantic", "category", "annotations", ",", "which", "restricts", "the", "fine", "-", "grained", "understanding", "of", "humans", "in", "visual", "scenes", ".", "2", ")", "In", "LIP", ",", "only", "a", "small", "proportion", "of", "images", "involve", "multiple", "persons", "with", "interactions", ".", "Such", "an", "instance", "-", "agnostic", "setting", "severely", "deviates", "from", "reality", ".", "Even", "in", "the", "MHP", "v1.0", "dataset", "proposed", "by", "Li", "et", "al", ".", "for", "multi", "-", "human", "parsing", ",", "only", "4", ",", "980", "images", "are", "included", "and", "annotated", "with", "18", "semantic", "labels", ".", "Comparatively", ",", "our", "MHP", "v2.0", "dataset", "contains", "25", ",", "403", "elaborately", "annotated", "images", "with", "58", "fine", "-", "grained", "semantic", "part", "labels", ".", "It", "is", "the", "largest", "and", "most", "comprehensive", "multi", "-", "human", "parsing", "dataset", "to", "date", ",", "to", "our", "best", "knowledge", ".", "Visual", "comparisons", "between", "LIP", ",", "MHP", "v1.0", "and", "our", "MHP", "v2.0", "are", "provided", "in", "Fig", ".", "[", "reference", "]", ".", "paragraph", ":", "Human", "Parsing", "Approaches", "Recently", ",", "many", "research", "efforts", "have", "been", "devoted", "to", "human", "parsing", "due", "to", "its", "wide", "range", "of", "potential", "applications", ".", "For", "example", ",", "Liang", "et", "al", ".", "proposed", "a", "proposal", "-", "free", "network", "for", "instance", "segmentation", "by", "directly", "predicting", "the", "instance", "numbers", "of", "different", "categories", "and", "the", "pixel", "-", "level", "information", ".", "Gong", "et", "al", ".", "proposed", "a", "self", "-", "supervised", "structure", "-", "sensitive", "learning", "approach", ",", "which", "imposes", "human", "pose", "structures", "to", "parsing", "results", "without", "resorting", "to", "extra", "supervision", ".", "Liu", "et", "al", ".", "proposed", "a", "single", "frame", "video", "parsing", "method", "which", "integrates", "frame", "parsing", ",", "optical", "flow", "estimation", "and", "temporal", "fusion", "into", "a", "unified", "network", ".", "Zhao", "et", "al", ".", "proposed", "a", "self", "-", "supervised", "neural", "aggregation", "network", ",", "which", "learns", "to", "aggregate", "the", "multi", "-", "scale", "features", "and", "incorporates", "a", "self", "-", "supervised", "joint", "loss", "to", "ensure", "the", "consistency", "between", "parsing", "and", "pose", ".", "He", "et", "al", ".", "proposed", "the", "Mask", "R", "-", "CNN", ",", "which", "is", "extended", "from", "Faster", "R", "-", "CNN", "by", "adding", "a", "branch", "for", "predicting", "an", "object", "mask", "in", "parallel", "with", "the", "existing", "branch", "for", "bounding", "box", "recognition", ".", "Brabandere", "et", "al", ".", "proposed", "to", "tackle", "instance", "segmentation", "with", "a", "discriminative", "loss", "function", ",", "operating", "at", "the", "pixel", "level", ",", "which", "encourages", "a", "convolutional", "network", "to", "produce", "a", "representation", "of", "the", "image", "that", "can", "be", "easily", "clustered", "into", "instances", "with", "a", "simple", "post", "-", "processing", "step", ".", "However", ",", "these", "methods", "either", "only", "consider", "coarse", "semantic", "information", "or", "are", "agnostic", "to", "different", "instances", ".", "To", "enable", "more", "detailed", "human", "-", "centric", "analysis", ",", "Li", "et", "al", ".", "initially", "proposed", "the", "multi", "-", "human", "parsing", "task", ",", "which", "aligns", "better", "with", "the", "realistic", "scenarios", ".", "They", "also", "proposed", "a", "novel", "MH", "-", "Parser", "model", "as", "a", "reference", "method", "which", "generates", "parsing", "maps", "and", "instance", "masks", "simutaneously", "in", "a", "bottom", "-", "up", "fashion", ".", "Jiang", "et", "al", ".", "proposed", "a", "new", "approach", "to", "segment", "human", "instances", "and", "label", "their", "body", "parts", "using", "region", "assembly", ".", "Li", "et", "al", ".", "proposed", "a", "framework", "with", "a", "human", "detector", "and", "a", "category", "-", "level", "segmentation", "module", "to", "segment", "the", "parts", "of", "objects", "at", "the", "instance", "level", ".", "These", "methods", "involve", "mutiple", "separate", "stages", "for", "instance", "localization", ",", "human", "parsing", "and", "result", "refinement", ".", "In", "comparison", ",", "the", "proposed", "NAN", "produces", "accurate", "multi", "-", "human", "parsing", "results", "through", "a", "single", "forward", "-", "pass", "in", "a", "time", "-", "efficient", "manner", "without", "tedious", "pre", "-", "or", "post", "-", "processing", ".", "section", ":", "Multi", "-", "Human", "Parsing", "Benchmark", "In", "this", "section", ",", "we", "introduce", "the", "“", "M", "ulti", "-", "H", "uman", "P", "arsing", "(", "MHP", "v2.0", ")", "”", ",", "a", "new", "large", "-", "scale", "dataset", "focusing", "on", "semantic", "understanding", "of", "humans", "in", "crowded", "scenes", "with", "several", "appealing", "properties", ".", "1", ")", "It", "contains", "25", ",", "403", "elaborately", "annotated", "images", "with", "58", "fine", "-", "grained", "labels", "on", "body", "parts", ",", "fashion", "items", "and", "one", "background", "label", ",", "which", "is", "larger", "and", "more", "comprehensive", "than", "previous", "similar", "attempts", ".", "2", ")", "The", "images", "within", "MHP", "v2.0", "are", "collected", "from", "real", "-", "world", "scenarios", ",", "involving", "humans", "with", "various", "viewpoints", ",", "poses", ",", "occlusion", ",", "interactions", "and", "resolution", ".", "3", ")", "The", "background", "of", "images", "in", "MHP", "v2.0", "is", "more", "complex", "and", "diverse", "than", "previous", "datasets", ".", "Some", "examples", "are", "showed", "in", "Fig", ".", "[", "reference", "]", ".", "The", "MHP", "v2.0", "dataset", "is", "expected", "to", "provide", "a", "new", "benchmark", "suitable", "for", "multi", "-", "human", "parsing", "together", "with", "a", "standard", "evaluation", "server", "where", "the", "test", "set", "will", "be", "kept", "secret", "to", "avoid", "overfitting", ".", "subsection", ":", "Image", "Collection", "and", "Annotation", "We", "manually", "specify", "some", "underlying", "relationships", "(", "such", "as", "family", ",", "couple", ",", "team", ",", "etc", ".", ")", "and", "possible", "scenes", "(", "such", "as", "sports", ",", "conferences", ",", "banquets", ",", "etc", ".", ")", "to", "ensure", "the", "diversity", "of", "returned", "results", ".", "Based", "on", "any", "one", "of", "these", "specifications", ",", "corresponding", "multi", "-", "human", "images", "are", "located", "by", "performing", "Internet", "searches", "over", "Creative", "Commons", "licensed", "imagery", ".", "For", "each", "identified", "image", ",", "the", "contained", "human", "number", "and", "the", "corresponding", "URL", "are", "stored", "in", "a", "spreadsheet", ".", "Automated", "scrapping", "software", "is", "used", "to", "download", "the", "multi", "-", "human", "imagery", "and", "stores", "all", "relevant", "information", "in", "a", "relational", "database", ".", "Moreover", ",", "a", "pool", "of", "images", "containing", "clearly", "visible", "persons", "with", "interactions", "and", "rich", "fashion", "items", "is", "also", "constructed", "from", "the", "existing", "human", "-", "centric", "datasets", "to", "augment", "and", "complement", "Internet", "scraping", "results", ".", "After", "curating", "the", "imagery", ",", "manual", "annotation", "is", "conducted", "by", "professional", "data", "annotators", ",", "which", "includes", "two", "distinct", "tasks", ".", "The", "first", "task", "is", "manually", "counting", "the", "number", "of", "foreground", "persons", "and", "duplicating", "each", "image", "to", "several", "copies", "according", "to", "the", "count", "number", ".", "Each", "duplicated", "image", "is", "marked", "with", "the", "image", "ID", ",", "the", "contained", "person", "number", "and", "a", "self", "-", "index", ".", "The", "second", "is", "assigning", "the", "fine", "-", "grained", "pixel", "-", "wise", "label", "to", "each", "semantic", "category", "for", "each", "person", "instance", ".", "We", "implement", "an", "annotation", "tool", "and", "generate", "multi", "-", "scale", "superpixels", "of", "images", "based", "on", "to", "speed", "up", "the", "annotation", ".", "See", "Fig", ".", "[", "reference", "]", "for", "an", "example", ".", "Each", "multi", "-", "human", "image", "contains", "at", "least", "two", "instances", ".", "The", "annotation", "for", "each", "instance", "is", "done", "in", "a", "left", "-", "to", "-", "right", "order", ",", "corresponding", "to", "the", "duplicated", "image", "with", "the", "self", "-", "index", "from", "beginning", "to", "end", ".", "For", "each", "instance", ",", "58", "semantic", "categories", "are", "defined", "and", "annotated", ",", "including", "cap", "/", "hat", ",", "helmet", ",", "face", ",", "hair", ",", "left", "-", "arm", ",", "right", "-", "arm", ",", "left", "-", "hand", ",", "right", "-", "hand", ",", "protector", ",", "bikini", "/", "bra", ",", "jacket", "/", "windbreaker", "/", "hoodie", ",", "t", "-", "shirt", ",", "polo", "-", "shirt", ",", "sweater", ",", "singlet", ",", "torso", "-", "skin", ",", "pants", ",", "shorts", "/", "swim", "-", "shorts", ",", "skirt", ",", "stockings", ",", "socks", ",", "left", "-", "boot", ",", "right", "-", "boot", ",", "left", "-", "shoe", ",", "right", "-", "shoe", ",", "left", "-", "highheel", ",", "right", "-", "highheel", ",", "left", "-", "sandal", ",", "right", "-", "sandal", ",", "left", "-", "leg", ",", "right", "-", "leg", ",", "left", "-", "foot", ",", "right", "-", "foot", ",", "coat", ",", "dress", ",", "robe", ",", "jumpsuits", ",", "other", "-", "full", "-", "body", "-", "clothes", ",", "headwear", ",", "backpack", ",", "ball", ",", "bats", ",", "belt", ",", "bottle", ",", "carrybag", ",", "cases", ",", "sunglasses", ",", "eyewear", ",", "gloves", ",", "scarf", ",", "umbrella", ",", "wallet", "/", "purse", ",", "watch", ",", "wristband", ",", "tie", ",", "other", "-", "accessaries", ",", "other", "-", "upper", "-", "body", "-", "clothes", "and", "other", "-", "lower", "-", "body", "-", "clothes", ".", "Each", "instance", "has", "a", "complete", "set", "of", "annotations", "whenever", "the", "corresponding", "category", "appears", "in", "the", "current", "image", ".", "When", "annotating", "one", "instance", ",", "others", "are", "regarded", "as", "background", ".", "Thus", ",", "the", "resulting", "annotation", "set", "for", "each", "image", "consists", "of", "instance", "-", "level", "parsing", "masks", ",", "where", "is", "the", "number", "of", "persons", "in", "the", "image", ".", "After", "annotation", ",", "manual", "inspection", "is", "performed", "on", "all", "images", "and", "corresponding", "annotations", "to", "verify", "the", "correctness", ".", "In", "cases", "where", "annotations", "are", "erroneous", ",", "the", "information", "is", "manually", "rectified", "by", "5", "well", "informed", "analysts", ".", "The", "whole", "work", "took", "around", "three", "months", "to", "accomplish", "by", "25", "professional", "data", "annotators", ".", "subsection", ":", "Dataset", "Splits", "and", "Statistics", "In", "total", ",", "there", "are", "25", ",", "403", "images", "in", "the", "MHP", "v2.0", "dataset", ".", "Each", "image", "contains", "2", "-", "26", "person", "instances", ",", "with", "3", "on", "average", ".", "The", "resolution", "of", "the", "images", "ranges", "from", "85", "100", "to", "4", ",", "511", "6", ",", "919", ",", "with", "644", "718", "on", "average", ".", "We", "spit", "the", "images", "into", "training", ",", "validation", "and", "testing", "sets", ".", "Following", "random", "selection", ",", "we", "arrive", "at", "a", "unique", "split", "consisting", "of", "15", ",", "403", "training", "and", "5", ",", "000", "validation", "images", "with", "publicly", "available", "annotations", ",", "as", "well", "as", "5", ",", "000", "testing", "images", "with", "annotations", "withheld", "for", "benchmarking", "purpose", ".", "The", "statistics", "w.r.t", ".", "data", "distribution", "on", "59", "semantic", "categories", ",", "the", "average", "semantic", "category", "number", "per", "image", "and", "the", "average", "instance", "number", "per", "image", "in", "the", "MHP", "v2.0", "dataset", "are", "illustrated", "in", "Fig", ".", "[", "reference", "]", "(", "a", ")", ",", "(", "b", ")", "and", "(", "c", ")", ",", "respectively", ".", "In", "general", ",", "face", ",", "arms", "and", "legs", "are", "the", "most", "remarkable", "parts", "of", "a", "human", "body", ".", "However", ",", "understanding", "humans", "in", "crowded", "scenes", "needs", "to", "analyze", "fine", "-", "grained", "details", "of", "each", "person", "of", "interest", ",", "including", "different", "body", "parts", ",", "clothes", "and", "accessaries", ".", "We", "therefore", "define", "11", "body", "parts", ",", "and", "47", "clothes", "and", "accessaries", ".", "Among", "these", "11", "body", "parts", ",", "we", "divide", "arms", ",", "hands", ",", "legs", "and", "feet", "into", "left", "and", "right", "side", "for", "more", "precise", "analysis", ",", "which", "also", "increases", "the", "difficulty", "of", "the", "task", ".", "We", "define", "hair", ",", "face", "and", "torso", "-", "skin", "as", "the", "remaining", "three", "body", "parts", ",", "which", "can", "be", "used", "as", "auxiliary", "guidance", "for", "more", "comprehensive", "instance", "-", "level", "analysis", ".", "As", "for", "clothing", "categories", ",", "we", "have", "common", "clothes", "like", "coat", ",", "jacket", "/", "windbreaker", "/", "hoodie", ",", "sweater", ",", "singlet", ",", "pants", ",", "shorts", "/", "swim", "-", "shorts", "and", "shoes", ",", "confusing", "categories", "such", "as", "t", "-", "shirt", "v.s", ".", "polo", "-", "shirt", ",", "stockings", "v.s", ".", "socks", ",", "skirt", "v.s", ".", "dress", "and", "robe", ",", "and", "boots", "v.s", ".", "sandals", "and", "highheels", ",", "and", "infrequent", "categories", "such", "as", "cap", "/", "hat", ",", "helmet", ",", "protector", ",", "bikini", "/", "bra", ",", "jumpsuits", ",", "gloves", "and", "scarf", ".", "Furthermore", ",", "accessaries", "like", "sunglasses", ",", "belt", ",", "tie", ",", "watch", "and", "bags", "are", "also", "taken", "into", "account", ",", "which", "are", "common", "but", "hard", "to", "predict", ",", "especially", "for", "the", "small", "-", "scale", "ones", ".", "To", "summarize", ",", "the", "pre", "-", "defined", "semantic", "categories", "of", "MHP", "v2.0", "involve", "most", "body", "parts", ",", "clothes", "and", "accessaries", "of", "different", "styles", "for", "men", ",", "women", "and", "children", "in", "all", "seasons", ".", "The", "images", "in", "the", "MHP", "v2.0", "dataset", "contain", "diverse", "instance", "numbers", ",", "viewpoints", ",", "poses", ",", "occlusion", ",", "interactions", "and", "background", "complexities", ".", "MHP", "v2.0", "aligns", "better", "with", "real", "-", "world", "scenarios", "and", "serves", "as", "a", "more", "realistic", "benchmark", "for", "human", "-", "centric", "analysis", ",", "which", "pushes", "the", "frontiers", "of", "fine", "-", "grained", "multi", "-", "human", "parsing", "research", ".", "section", ":", "Deep", "Nested", "Adversarial", "Networks", "As", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "the", "proposed", "deep", "N", "ested", "A", "dversarial", "N", "etwork", "(", "NAN", ")", "model", "consists", "of", "three", "GAN", "-", "like", "sub", "-", "nets", "that", "jointly", "perform", "semantic", "saliency", "prediction", ",", "instance", "-", "agnostic", "parsing", "and", "instance", "-", "aware", "clustering", "end", "-", "to", "-", "end", ".", "NAN", "produces", "accurate", "multi", "-", "human", "parsing", "results", "through", "a", "single", "forward", "-", "pass", "in", "a", "time", "-", "efficient", "manner", "without", "tedious", "pre", "-", "or", "post", "-", "processing", ".", "We", "now", "present", "each", "component", "in", "details", ".", "subsection", ":", "Semantic", "Saliency", "Prediction", "Large", "modality", "and", "interaction", "variations", "are", "the", "main", "challenge", "to", "multi", "-", "human", "parsing", "and", "also", "the", "key", "obstacle", "to", "learning", "a", "well", "-", "performing", "human", "-", "centric", "analysis", "model", ".", "To", "address", "this", "problem", ",", "we", "propose", "to", "decompose", "the", "original", "task", "into", "three", "granularities", "and", "adaptively", "impose", "a", "prior", "on", "the", "specific", "process", ",", "each", "with", "the", "aid", "of", "a", "GAN", "-", "based", "sub", "-", "net", ".", "This", "reduces", "the", "training", "complexity", "and", "leads", "to", "better", "empirical", "performance", "with", "limited", "data", ".", "The", "first", "sub", "-", "net", "estimates", "semantic", "saliency", "maps", "to", "locate", "the", "most", "noticeable", "and", "eye", "-", "attracting", "human", "regions", "in", "images", ",", "which", "serves", "as", "a", "basic", "prior", "to", "facilitate", "further", "processing", "on", "humans", ",", "as", "illustrated", "in", "Fig", ".", "[", "reference", "]", "left", ".", "We", "formulate", "semantic", "saliency", "prediction", "as", "a", "binary", "pixel", "-", "wise", "labelling", "problem", "to", "segment", "out", "foreground", "v.s", ".", "background", ".", "Inspired", "by", "the", "recent", "success", "of", "F", "ully", "C", "onvolutional", "N", "etwork", "s", "(", "FCNs", ")", "based", "methods", "on", "image", "-", "to", "-", "image", "applications", ",", "we", "leverage", "an", "FCN", "backbone", "(", "FCN", "-", "8s", ")", "as", "the", "generator", "of", "NAN", "for", "semantic", "saliency", "prediction", ",", "where", "denotes", "the", "network", "parameters", ",", "and", ",", ",", "and", "denote", "the", "image", "height", ",", "width", ",", "channel", "number", "and", "semantic", "category", "(", "i.e.", ",", "foreground", "plus", "background", ")", "number", ",", "repectively", ".", "Formally", ",", "let", "the", "input", "RGB", "image", "be", "denoted", "by", "and", "the", "semantic", "saliency", "map", "be", "denoted", "by", ",", "then", "The", "key", "requirements", "for", "are", "that", "the", "semantic", "saliency", "map", "should", "present", "indistinguishable", "properities", "compared", "with", "a", "real", "one", "(", "i.e.", ",", "ground", "truth", ")", "in", "appearance", "while", "preserving", "the", "intrinsic", "contextually", "remarkable", "information", ".", "To", "this", "end", ",", "we", "propose", "to", "learn", "by", "minimizing", "a", "combination", "of", "two", "losses", ":", "where", "is", "the", "adv", "ersarial", "loss", "for", "refining", "realism", "and", "alleviating", "artifacts", ",", "is", "the", "s", "emantic", "s", "aliency", "loss", "for", "pixel", "-", "wise", "image", "labelling", ",", "are", "weighting", "parameters", "among", "different", "losses", ".", "is", "a", "pixel", "-", "wise", "cross", "-", "entropy", "loss", "calculated", "based", "on", "the", "binary", "pixel", "-", "wise", "annotations", "to", "learn", ":", "is", "proposed", "to", "narrow", "the", "gap", "between", "the", "distributions", "of", "generated", "and", "real", "results", ".", "To", "facilitate", "this", "process", ",", "we", "leverage", "a", "C", "onvolutional", "N", "eural", "N", "etwork", "(", "CNN", ")", "backbone", "as", "the", "discriminator", "to", "be", "as", "simple", "as", "possible", "to", "avoid", "typical", "GAN", "tricks", ".", "We", "alternatively", "optimize", "and", "to", "learn", "and", ":", "where", "denotes", "the", "binary", "real", "v.s", ".", "fake", "indicator", ".", "subsection", ":", "Instance", "-", "Agnostic", "Parsing", "The", "second", "sub", "-", "net", "concatenates", "the", "information", "from", "the", "original", "RGB", "image", "with", "semantic", "saliency", "prior", "as", "input", "and", "estimates", "a", "fine", "-", "grained", "instance", "-", "agnostic", "parsing", "map", ",", "which", "further", "serves", "as", "stronger", "semantic", "guidance", "from", "the", "global", "perspective", "to", "facilitate", "instance", "-", "aware", "clustering", ",", "as", "illustrated", "in", "Fig", ".", "[", "reference", "]", "middle", ".", "We", "formulate", "instance", "-", "agnostic", "parsing", "as", "a", "multi", "-", "class", "dense", "classification", "problem", "to", "mask", "semantically", "consistent", "regions", "of", "body", "parts", "and", "fashion", "items", ".", "Inspired", "by", "the", "leading", "performance", "of", "the", "skip", "-", "net", "on", "recognition", "tasks", ",", "we", "modify", "a", "skip", "-", "net", "(", "WS", "-", "ResNet", ")", "into", "an", "FCN", "-", "based", "architecture", "as", "the", "generator", "of", "NAN", "to", "learn", "a", "highly", "non", "-", "linear", "transformation", "for", "instance", "-", "agnostic", "parsing", ",", "where", "denotes", "the", "network", "parameters", "for", "the", "generator", "and", "denotes", "the", "semantic", "category", "number", ".", "The", "prediction", "is", "downsampled", "by", "for", "accuracy", "v.s", ".", "speed", "trade", "-", "off", ".", "Contextual", "information", "from", "global", "and", "local", "regions", "compensates", "each", "other", "and", "naturally", "benefits", "human", "parsing", ".", "The", "hierarchical", "features", "within", "a", "skip", "-", "net", "are", "multi", "-", "scale", "in", "nature", "due", "to", "the", "increasing", "receptive", "field", "sizes", ",", "which", "are", "combined", "together", "via", "skip", "connections", ".", "Such", "a", "combined", "representation", "comprehensively", "maintains", "the", "contextual", "information", ",", "which", "is", "crucial", "for", "generating", "smooth", "and", "accurate", "parsing", "results", ".", "Formally", ",", "let", "the", "instance", "-", "agnostic", "parsing", "map", "be", "denoted", "by", ",", "then", "Similar", "to", "the", "first", "sub", "-", "net", ",", "we", "propose", "to", "learn", "by", "minimizing", ":", "where", "is", "the", "g", "lobal", "p", "arsing", "loss", "for", "semantic", "part", "labelling", ".", "is", "a", "standard", "pixel", "-", "wise", "cross", "-", "entropy", "loss", "calculated", "based", "on", "the", "multi", "-", "class", "pixel", "-", "wise", "annotations", "to", "learn", ".", "is", "also", "slightly", "finetuned", "due", "to", "the", "hinged", "gradient", "backpropagation", "route", "within", "the", "nested", "structure", ":", "is", "proposed", "to", "ensure", "the", "correctness", "and", "realism", "of", "the", "current", "phase", "and", "also", "the", "previous", "one", "for", "information", "flow", "consistency", ".", "To", "facilitate", "this", "process", ",", "we", "leverage", "a", "same", "CNN", "backbone", "with", "as", "the", "discriminator", ",", "which", "are", "learned", "separately", ".", "We", "alternatively", "optimize", "and", "to", "learn", ",", "and", "slightly", "finetune", ":", "subsection", ":", "Instance", "-", "Aware", "Clustering", "The", "third", "sub", "-", "net", "concatenates", "the", "information", "from", "the", "original", "RGB", "image", "with", "semantic", "saliency", "and", "instance", "-", "agnostic", "parsing", "priors", "as", "input", "and", "estimates", "an", "instance", "-", "aware", "clustering", "map", "by", "associating", "each", "semantic", "parsing", "mask", "to", "one", "of", "the", "person", "instances", "in", "the", "scene", ",", "as", "illustrated", "in", "Fig", ".", "[", "reference", "]", "right", ".", "Inspired", "by", "the", "observation", "that", "a", "human", "glances", "at", "an", "image", "and", "instantly", "knows", "how", "many", "and", "where", "the", "objects", "are", "in", "the", "image", ",", "we", "formulate", "instance", "-", "aware", "clustering", "by", "parallelly", "inferring", "the", "instance", "number", "and", "pixel", "-", "wise", "instance", "location", ",", "discarding", "the", "requirement", "of", "time", "-", "consuming", "region", "proposal", "generation", ".", "We", "modify", "a", "same", "backbone", "architecture", "to", "incorporate", "two", "sibling", "branches", "as", "the", "generator", "of", "NAN", "for", "location", "-", "sensitive", "learning", ",", "where", "denotes", "the", "network", "parameters", "for", "the", "generator", "and", "denotes", "the", "pre", "-", "defined", "instance", "location", "coordinate", "number", ".", "As", "multi", "-", "scale", "features", "integrating", "both", "global", "and", "local", "contextual", "information", "are", "crucial", "for", "increasing", "location", "prediction", "accuracy", ",", "we", "further", "augment", "the", "pixel", "-", "wise", "instance", "location", "prediction", "branch", "with", "a", "M", "ulti", "-", "S", "cale", "F", "usion", "U", "nit", "(", "MSFU", ")", "to", "fuse", "shallow", "-", ",", "middle", "-", "and", "deep", "-", "level", "features", ",", "while", "using", "the", "feature", "maps", "downsampled", "by", "concatenated", "with", "feature", "maps", "from", "the", "first", "branch", "for", "instance", "number", "regression", ".", "Formally", ",", "let", "the", "pixel", "-", "wise", "instance", "location", "map", "be", "denoted", "by", "and", "the", "instance", "number", "be", "denoted", "by", ",", "then", "We", "propose", "to", "learn", "by", "minimizing", ":", "where", "is", "the", "p", "ixel", "-", "wise", "i", "nstance", "l", "ocation", "loss", "for", "pixel", "-", "wise", "instance", "location", "regression", "and", "is", "the", "i", "nstance", "n", "umber", "loss", "for", "instance", "number", "regression", ".", "is", "a", "standard", "smooth", "-", "loss", "calculated", "based", "on", "the", "foreground", "pixel", "-", "wise", "instance", "location", "annotations", "to", "learn", ".", "Since", "a", "person", "instance", "can", "be", "identified", "by", "its", "top", "-", "left", "corner", "and", "bottom", "-", "right", "corner", "of", "the", "surrounding", "bounding", "box", ",", "for", "each", "pixel", "belonging", "to", "the", "person", "instance", ",", "the", "pixel", "-", "wise", "instance", "location", "vector", "is", "defined", "as", ",", "where", "and", "are", "the", "width", "and", "height", "of", "the", "person", "instance", "for", "normalization", ",", "respectively", ".", "is", "a", "standard", "loss", "calculated", "based", "on", "the", "instance", "number", "annotations", "to", "learn", ".", "and", "are", "also", "slightly", "finetuned", "due", "to", "the", "chained", "schema", "within", "the", "nest", ":", "Given", "these", "information", ",", "instance", "-", "aware", "clustering", "maps", "can", "be", "effortlessly", "generated", "with", "little", "computational", "overhead", ",", "which", "are", "denoted", "by", ".", "Similar", "to", ",", "is", "proposed", "to", "ensure", "the", "correctness", "and", "realism", "of", "all", "phases", "for", "the", "information", "flow", "consistency", ".", "To", "facilitate", "this", "process", ",", "we", "leverage", "a", "same", "CNN", "backbone", "with", "as", "the", "discriminator", ",", "which", "are", "learned", "separately", ".", "We", "alternatively", "optimize", "and", "to", "learn", ",", "and", "slightly", "finetune", "and", ":", "subsection", ":", "Training", "and", "Inference", "The", "goal", "of", "NAN", "is", "to", "use", "sets", "of", "real", "targets", "to", "learn", "three", "GAN", "-", "like", "sub", "-", "nets", "that", "mutually", "boost", "and", "jointly", "accomplish", "multi", "-", "human", "parsing", ".", "Each", "separate", "loss", "serves", "as", "a", "deep", "supervision", "within", "the", "nested", "structure", "benefitting", "network", "convergence", ".", "The", "overall", "objective", "function", "for", "NAN", "is", "Clearly", ",", "the", "NAN", "is", "end", "-", "to", "-", "end", "trainable", "and", "can", "be", "optimized", "with", "the", "proposed", "nested", "adversarial", "learning", "strategy", "and", "BP", "algorithm", ".", "During", "testing", ",", "we", "simply", "feed", "the", "input", "image", "into", "NAN", "to", "get", "the", "instance", "-", "agnostic", "parsing", "map", "from", ",", "pixel", "-", "wise", "instance", "location", "map", "and", "instance", "number", "from", ".", "Then", "we", "employ", "an", "off", "-", "the", "-", "shelf", "clustering", "method", "to", "obtain", "the", "instance", "-", "aware", "clustering", "map", ".", "Example", "results", "are", "visualized", "in", "Fig", ".", "[", "reference", "]", ".", "section", ":", "Experiments", "We", "evaluate", "NAN", "qualitatively", "and", "quantitatively", "under", "various", "settings", "and", "granularities", "for", "understanding", "humans", "in", "crowded", "scenes", ".", "In", "particular", ",", "we", "evaluate", "multi", "-", "human", "parsing", "performance", "on", "the", "MHP", "v2.0", "dataset", "proposed", "in", "this", "work", ",", "as", "well", "as", "the", "MHP", "v1.0", "and", "PASCAL", "-", "Person", "-", "Part", "benchmark", "datasets", ".", "We", "also", "evaluate", "instance", "-", "agnostic", "parsing", "and", "instance", "-", "aware", "clustering", "results", "on", "the", "Buffy", "benchmark", "dataset", ",", "which", "are", "byproducts", "of", "NAN", ".", "subsection", ":", "Experimental", "Settings", "subsubsection", ":", "Implementation", "Details", "Throughout", "the", "experiments", ",", "the", "sizes", "of", "the", "RGB", "image", ",", "the", "semantic", "saliency", "prediction", ",", "inputs", "to", "the", "discriminator", "and", "inputs", "to", "the", "generator", "are", "fixed", "as", ";", "the", "sizes", "of", "the", "instance", "-", "agnostic", "parsing", "prediction", ",", "instance", "-", "aware", "clustering", "prediction", ",", "inputs", "to", "the", "discriminator", ",", "inputs", "to", "the", "generator", ",", "inputs", "to", "the", "discriminator", "and", "instance", "location", "map", "are", "fixed", "as", ";", "the", "channel", "number", "of", "the", "pixel", "-", "wise", "instance", "location", "map", "is", "fixed", "as", ",", "incorporating", "two", "corner", "points", "of", "the", "associated", "bounding", "box", ";", "the", "constraint", "factors", "are", "empirically", "fixed", "as", "and", ",", "respectively", ";", "the", "generator", "is", "initialized", "with", "FCN", "-", "8s", "by", "replacing", "the", "last", "layer", "with", "a", "new", "convolutional", "layer", "with", "kernel", "size", ",", "pretrained", "on", "PASCAL", "-", "VOC", "-", "2011", "and", "finetuned", "on", "the", "target", "dataset", ";", "the", "generator", "is", "initialized", "with", "WS", "-", "ResNet", "by", "eliminating", "the", "spatial", "pooling", "layers", ",", "increasing", "the", "strides", "of", "the", "first", "convolutional", "layers", "up", "to", "2", "in", "B", ",", "eliminating", "the", "top", "-", "most", "global", "pooling", "layer", "and", "the", "linear", "classifier", ",", "and", "adding", "two", "new", "convolutional", "layers", "with", "kernel", "sizes", "and", ",", "pretrained", "on", "ImageNet", "and", "PASCAL", "-", "VOC", "-", "2012", ",", "and", "finetuned", "on", "the", "target", "dataset", ";", "the", "generator", "is", "initialized", "with", "the", "same", "backbone", "architecture", "and", "pre", "-", "trained", "weights", "with", "(", "which", "are", "learned", "separately", ")", ",", "by", "further", "augmenting", "it", "with", "two", "sibling", "branches", "for", "pixel", "-", "wise", "instance", "location", "map", "prediction", "and", "instance", "number", "prediction", ",", "where", "the", "first", "branch", "utilizes", "a", "MSFU", "(", "three", "convolutional", "layers", "with", "kernal", "sizes", "for", "specific", "scale", "adaption", ")", "ended", "with", "a", "convolutional", "layer", "with", "kernel", "size", "for", "multi", "-", "scale", "feature", "aggregation", "and", "a", "final", "convolutional", "layer", "with", "kernel", "size", "for", "location", "regression", "and", "the", "second", "branch", "utilizes", "the", "feature", "maps", "downsampled", "by", "8", "concatenated", "with", "the", "feature", "maps", "from", "the", "first", "branch", "ended", "with", "a", "global", "pooling", "layer", ",", "a", "hidden", "512", "-", "way", "fully", "-", "connected", "layer", "and", "a", "final", "1", "-", "way", "fully", "-", "connected", "layer", "for", "instance", "number", "regression", ";", "the", "three", "discriminators", "(", "which", "are", "learned", "separately", ")", "are", "all", "initialized", "with", "a", "VGG", "-", "16", "by", "adding", "a", "new", "convolutional", "layer", "at", "the", "very", "begining", "with", "kernel", "size", "for", "input", "adaption", ",", "and", "replacing", "the", "last", "layer", "with", "a", "new", "1", "-", "way", "fully", "-", "connected", "layer", "activated", "by", "sigmoid", ",", "pre", "-", "trained", "on", "ImageNet", "and", "finetuned", "on", "the", "target", "dataset", ";", "the", "newly", "added", "layers", "are", "randomly", "initialized", "by", "drawing", "weights", "from", "a", "zero", "-", "mean", "Gaussian", "distribution", "with", "standard", "deviation", ";", "we", "employ", "an", "off", "-", "the", "-", "shelf", "clustering", "method", "to", "obtain", "the", "instance", "-", "aware", "clustering", "map", ";", "the", "dropout", "ratio", "is", "empirically", "fixed", "as", ";", "the", "weight", "decay", "and", "batch", "size", "are", "fixed", "as", "and", ",", "respectively", ";", "We", "use", "an", "initial", "learning", "rate", "of", "for", "pre", "-", "trained", "layers", ",", "and", "for", "newly", "added", "layers", "in", "all", "our", "experiments", ";", "we", "decrease", "the", "learning", "rate", "to", "of", "the", "previous", "one", "after", "20", "epochs", "and", "train", "the", "network", "for", "roughly", "60", "epochs", "one", "after", "the", "other", ";", "the", "proposed", "network", "is", "implemented", "based", "on", "the", "publicly", "available", "TensorFlow", "platform", ",", "which", "is", "trained", "using", "Adam", "(", ")", "on", "four", "NVIDIA", "GeForce", "GTX", "TITAN", "X", "GPUs", "with", "12", "G", "memory", ";", "the", "same", "training", "setting", "is", "utilized", "for", "all", "our", "compared", "network", "variants", ";", "we", "evaluate", "the", "testing", "time", "by", "averaging", "the", "running", "time", "for", "images", "on", "the", "target", "set", "on", "NVIDIA", "GeForce", "GTX", "TITAN", "X", "GPU", "and", "Intel", "Core", "i7", "-", "4930", "K", "CPU@3.40GHZ", ";", "our", "NAN", "can", "rapidly", "process", "one", "image", "in", "about", "1", "second", ",", "which", "compares", "much", "favorably", "to", "other", "state", "-", "of", "-", "the", "-", "art", "approaches", ",", "as", "the", "current", "state", "-", "of", "-", "the", "-", "art", "methods", "rely", "on", "region", "proposal", "preprocessing", "and", "complex", "processing", "steps", ".", "subsubsection", ":", "Evaluation", "Metrics", "Following", ",", "we", "use", "the", "A", "verage", "P", "recision", "based", "on", "p", "art", "(", ")", "and", "P", "ercentage", "of", "C", "orrectly", "parsed", "semantic", "P", "arts", "(", "PCP", ")", "metrics", "for", "multi", "-", "human", "parsing", "evaluation", ".", "Different", "from", "the", "A", "verage", "P", "recision", "based", "on", "r", "egion", "(", ")", "used", "in", "instance", "segmentation", ",", "uses", "part", "-", "level", "pixel", "I", "ntersection", "o", "ver", "U", "nion", "(", "IoU", ")", "of", "different", "semantic", "part", "categories", "within", "a", "person", "instance", "to", "determine", "if", "one", "instance", "is", "a", "true", "positive", ".", "We", "prefer", "over", "as", "we", "focus", "on", "human", "-", "centric", "analysis", "and", "we", "aim", "to", "investigate", "to", "how", "well", "a", "person", "instance", "as", "a", "whole", "is", "parsed", ".", "Additionally", ",", "we", "also", "report", "the", ",", "which", "is", "the", "mean", "of", "the", "at", "IoU", "thresholds", "ranging", "from", "to", ",", "in", "increments", "of", "0.1", ".", "As", "averages", "the", "IoU", "of", "each", "semantic", "part", "category", ",", "it", "fails", "to", "reflect", "how", "many", "semantic", "parts", "are", "correctly", "parsed", ".", "We", "further", "incorporate", "the", "PCP", ",", "originally", "used", "in", "human", "pose", "estimation", ",", "to", "evaluate", "the", "parsing", "quality", "within", "person", "instances", ".", "For", "each", "true", "-", "positive", "person", "instance", ",", "we", "find", "all", "the", "semantic", "categories", "(", "excluding", "background", ")", "with", "pixel", "IoU", "larger", "than", "a", "threshold", ",", "which", "are", "regarded", "as", "correctly", "parsed", ".", "The", "PCP", "of", "one", "person", "instance", "is", "the", "ratio", "between", "the", "correctly", "parsed", "semantic", "category", "number", "and", "the", "total", "semantic", "category", "number", "of", "that", "person", ".", "Missed", "person", "instances", "are", "assigned", "with", "PCP", ".", "The", "overall", "PCP", "is", "the", "average", "PCP", "for", "all", "person", "instances", ".", "Note", "that", "PCP", "is", "also", "a", "human", "-", "centric", "evaluation", "metric", ".", "subsection", ":", "Evaluations", "on", "the", "MHP", "v2.0", "Benchmark", "The", "MHP", "v2.0", "dataset", "proposed", "in", "this", "paper", "is", "the", "largest", "and", "most", "comprehensive", "multi", "-", "human", "parsing", "benchmark", "to", "date", ",", "which", "extends", "MHP", "v1.0", "to", "push", "the", "frontiers", "of", "understanding", "humans", "in", "crowded", "scenes", "by", "containing", "25", ",", "403", "elaborately", "annotated", "images", "with", "58", "fine", "-", "grained", "semantic", "category", "labels", ".", "Annotation", "examples", "are", "visualized", "in", "Fig", ".", "[", "reference", "]", "(", "c", ")", ".", "The", "data", "are", "randomly", "organized", "into", "3", "splits", ",", "consisting", "of", "15", ",", "403", "training", "and", "5", ",", "000", "validation", "images", "with", "publicly", "available", "annotations", ",", "as", "well", "as", "5", ",", "000", "testing", "images", "with", "annotations", "withheld", "for", "benchmarking", "purpose", ".", "Evaluation", "systems", "report", "the", "and", "PCP", "over", "the", "validation", "and", "testing", "sets", ".", "subsubsection", ":", "Component", "Analysis", "We", "first", "investigate", "different", "architectures", "and", "loss", "function", "combinations", "of", "NAN", "to", "see", "their", "respective", "roles", "in", "multi", "-", "human", "parsing", ".", "We", "compare", "16", "variants", "from", "four", "aspects", ",", "i.e.", ",", "different", "baselines", "(", "Mask", "R", "-", "CNN", "and", "MH", "-", "Parser", ")", ",", "different", "network", "structures", "(", "w", "/", "o", ",", "w", "/", "o", "concatenated", "input", "(", "RGB", "only", ")", ",", "w", "/", "o", "concatenated", "input", "(", "RGB", "only", ")", ",", "w", "/", "o", ",", "w", "/", "o", ",", "w", "/", "o", "concatenated", "input", ",", "w", "/", "o", ",", "w", "/", "o", "concatenated", "input", ",", "w", "/", "o", "MSFU", ")", ",", "our", "proposed", "NAN", ",", "and", "upperbounds", "(", ":", "use", "the", "ground", "truth", "semantic", "saliency", "maps", "instead", "of", "prediction", "while", "keeping", "other", "settings", "the", "same", ";", ":", "use", "the", "ground", "truth", "instance", "-", "agnostic", "parsing", "maps", "instead", "of", "prediction", "while", "keeping", "other", "settings", "the", "same", ";", ":", "use", "the", "ground", "truth", "instance", "number", "instead", "of", "prediction", "while", "keeping", "other", "settings", "the", "same", ";", ":", "use", "the", "ground", "truth", "pixel", "-", "wise", "instance", "location", "maps", "instead", "of", "prediction", "while", "keeping", "other", "settings", "the", "same", ")", ".", "The", "performance", "comparison", "in", "terms", "of", "@IoU=0.5", ",", "and", "PCP@IoU=0.5", "on", "the", "MHP", "v2.0", "validation", "set", "is", "reported", "in", "Tab", ".", "[", "reference", "]", ".", "By", "comaring", "the", "results", "from", "the", "v.s", ".", "panels", ",", "we", "observe", "that", "our", "proposed", "NAN", "consistently", "outperforms", "the", "baselines", "Mask", "R", "-", "CNN", "and", "MH", "-", "Parser", "by", "a", "large", "margin", ",", "i.e.", ",", "and", "in", "terms", "of", ",", "and", "in", "terms", "of", ",", "and", "and", "in", "terms", "of", "PCP", ".", "Mask", "R", "-", "CNN", "suffers", "difficulties", "to", "differentiate", "entangled", "humans", ".", "MH", "-", "Parser", "involves", "multiple", "stages", "for", "instance", "localization", ",", "human", "parsing", "and", "result", "refinement", "with", "high", "complexity", ",", "yielding", "sub", "-", "optimal", "results", ",", "whereas", "NAN", "parses", "semantic", "categories", ",", "differentiates", "different", "person", "instances", "and", "refines", "results", "simultaneously", "through", "deep", "nested", "adversarial", "learning", "in", "an", "effective", "yet", "time", "-", "efficient", "manner", ".", "By", "comaring", "the", "results", "from", "the", "v.s", ".", "panels", ",", "we", "observe", "that", "NAN", "consistently", "outperforms", "the", "9", "variants", "in", "terms", "of", "network", "structure", ".", "In", "particular", ",", "w", "/", "o", "refers", "to", "truncating", "the", "semantic", "saliency", "prediction", "sub", "-", "net", "from", "NAN", ",", "leading", "to", ",", "and", "performance", "drop", "in", "terms", "of", "all", "metrics", ".", "This", "verifies", "the", "necessity", "of", "semantic", "saliency", "prediction", "that", "locates", "the", "most", "noticeable", "human", "regions", "in", "images", "to", "serve", "as", "a", "basic", "prior", "to", "facilitate", "further", "human", "-", "centic", "processing", ".", "The", "superiority", "of", "incorporating", "adaptive", "prior", "information", "to", "specific", "process", "can", "be", "verified", "by", "comparing", "w", "/", "o", "concatenated", "input", "with", "NAN", ",", "i.e.", ",", ",", "and", ";", ",", "and", "differences", "in", "terms", "of", "all", "metrics", ".", "The", "superiority", "of", "incorporating", "adversarial", "learning", "to", "specific", "process", "can", "be", "verified", "by", "comparing", "w", "/", "o", "with", "NAN", ",", "i.e.", ",", ",", "and", ";", ",", "and", ";", ",", "and", "decrease", "in", "terms", "of", "all", "metrics", ".", "Nested", "adversarial", "learning", "strategy", "ensures", "the", "correctness", "and", "realism", "of", "all", "phases", "for", "information", "flow", "consistency", ",", "the", "superiority", "of", "which", "is", "verified", "by", "comparing", "w", "/", "o", "concatenated", "input", "with", "NAN", ",", "i.e.", ",", ",", "and", ";", ",", "and", "decline", "in", "terms", "of", "all", "metrics", ".", "MSFU", "dynamically", "fuses", "multi", "-", "scale", "features", "for", "enhancing", "instance", "-", "aware", "clustering", "accuracy", ",", "the", "superiority", "of", "which", "is", "verified", "by", "comparing", "w", "/", "o", "MSFU", "with", "NAN", ",", "i.e.", ",", ",", "and", "drop", "in", "terms", "of", "all", "metrics", ".", "Finally", ",", "we", "also", "evaluate", "the", "limitations", "of", "our", "current", "algorithm", ".", "By", "comparing", "with", "NAN", ",", "only", ",", "and", "improvement", "in", "term", "of", "all", "metrics", "are", "obtained", ",", "which", "shows", "that", "the", "errors", "from", "semantic", "saliency", "prediction", "are", "already", "small", "and", "have", "only", "little", "effect", "on", "the", "final", "results", ".", "A", "large", "gap", "between", ",", "and", "of", "and", ",", "and", "of", "NAN", "shows", "that", "a", "better", "instance", "-", "agnostic", "parsing", "network", "architecture", "can", "definitely", "help", "improve", "the", "performance", "of", "multi", "-", "human", "parsing", "under", "our", "NAN", "framework", ".", "By", "comparing", "and", "with", "NAN", ",", ",", "and", ";", ",", "and", "improvement", "in", "term", "of", "all", "metrics", "are", "obtained", ",", "which", "shows", "that", "accurate", "instance", "-", "aware", "clustering", "results", "are", "critical", "for", "superior", "multi", "-", "human", "parsing", ".", "subsubsection", ":", "Quantitative", "Comparison", "The", "performance", "comparison", "of", "the", "proposed", "NAN", "with", "two", "state", "-", "of", "-", "the", "-", "art", "methods", "in", "terms", "of", "@IoU=0.5", ",", "and", "PCP@IoU=0.5", "on", "the", "MHP", "v2.0", "testing", "set", "is", "reported", "in", "Tab", ".", "[", "reference", "]", ".", "Following", ",", "we", "conduct", "experiments", "under", "three", "settings", ":", "All", "reports", "the", "evaluation", "over", "the", "whole", "testing", "set", ";", "Inter%20", "reports", "the", "evaluation", "over", "the", "sub", "-", "set", "containing", "the", "images", "with", "top", "20", "%", "interaction", "intensity", ";", "Inter%10", "reports", "the", "evaluation", "over", "the", "sub", "-", "set", "containing", "the", "images", "with", "top", "10", "%", "interaction", "intensity", ".", "Our", "NAN", "is", "significantly", "superior", "over", "other", "state", "-", "of", "-", "the", "-", "arts", "on", "setting", "-", "1", ".", "In", "particular", ",", "NAN", "improves", "the", "-", "best", "by", ",", "and", "in", "terms", "of", "all", "metrics", ".", "For", "the", "more", "challenging", "scenarios", "with", "intensive", "interactions", "(", "setting", "-", "2", ",", "3", ")", ",", "NAN", "also", "consistently", "achieves", "the", "best", "performance", ".", "In", "particular", ",", "for", "Inter%20", "and", "Inter%10", ",", "NAN", "improves", "the", "-", "best", "by", ",", "and", ";", ",", "and", "in", "terms", "of", "all", "metrics", ".", "This", "verifies", "the", "effectiveness", "of", "our", "NAN", "for", "multi", "-", "human", "parsing", "and", "understanding", "humans", "in", "crowded", "scenes", ".", "Moreover", ",", "NAN", "can", "rapidly", "process", "one", "512", "512", "image", "in", "about", "1", "second", "with", "acceptable", "resource", "consumption", ",", "which", "is", "attractive", "to", "real", "applications", ".", "This", "compares", "much", "favorably", "to", "MH", "-", "Parser", "(", "14.94", "img", "/", "s", ")", ",", "which", "relies", "on", "separate", "and", "complex", "post", "-", "processing", "(", "including", "CRF", ")", "steps", ".", "subsubsection", ":", "Qualitative", "Comparison", "Fig", ".", "[", "reference", "]", "visualizes", "the", "qualitative", "comparison", "of", "the", "proposed", "NAN", "with", "two", "state", "-", "of", "-", "the", "-", "art", "methods", "and", "corresponding", "ground", "truths", "on", "the", "MHP", "v2.0", "dataset", ".", "Note", "that", "Mask", "R", "-", "CNN", "only", "offers", "silhouettes", "of", "different", "person", "instances", ",", "we", "only", "compare", "our", "instance", "-", "aware", "clustering", "results", "with", "it", "while", "comparing", "our", "holistic", "results", "with", "MH", "-", "Parser", ".", "It", "can", "be", "observed", "that", "the", "proposed", "NAN", "performs", "well", "in", "multi", "-", "human", "parsing", "with", "a", "wide", "range", "of", "viewpoints", ",", "poses", ",", "occlusion", ",", "interactions", "and", "background", "complexity", ".", "The", "instance", "-", "agnostic", "parsing", "and", "instance", "-", "aware", "clustering", "predictions", "of", "NAN", "present", "high", "consistency", "with", "corresponding", "ground", "truths", ",", "thanks", "to", "the", "novel", "network", "structure", "and", "effective", "training", "strategy", ".", "In", "contrast", ",", "Mask", "R", "-", "CNN", "suffers", "difficulties", "to", "differentiate", "entangled", "humans", ",", "while", "MH", "-", "Parser", "struggles", "to", "generate", "fine", "-", "grained", "parsing", "results", "and", "clearly", "segmented", "instance", "masks", ".", "This", "further", "desmonstrates", "the", "effectiveness", "of", "the", "proposed", "NAN", ".", "We", "also", "show", "some", "failure", "cases", "of", "our", "NAN", "in", "Fig", ".", "[", "reference", "]", ".", "As", "can", "be", "observed", ",", "humans", "in", "crowded", "scenes", "with", "heavy", "occlusion", ",", "extreme", "poses", "and", "intensive", "interactions", "are", "difficult", "to", "identify", "and", "segment", ".", "Some", "small", "-", "scale", "semantic", "categories", "within", "person", "instances", "are", "difficult", "to", "parse", ".", "This", "confirms", "that", "MHP", "v2.0", "aligns", "with", "real", "-", "world", "situations", "and", "deserves", "more", "furture", "attention", "and", "research", "efforts", ".", "subsection", ":", "Evaluations", "on", "the", "MHP", "v1.0", "Benchmark", "The", "MHP", "v1.0", "dataset", "is", "the", "first", "multi", "-", "human", "parsing", "benchmark", ",", "originally", "proposed", "by", "Li", "et", "al", ".", ",", "which", "contains", "4", ",", "980", "images", "annotated", "with", "18", "semantic", "labels", ".", "Annotation", "examples", "are", "visualized", "in", "Fig", ".", "[", "reference", "]", "(", "b", ")", ".", "The", "data", "are", "randomly", "organized", "into", "3", "splits", ",", "consisting", "of", "3", ",", "000", "training", ",", "1", ",", "000", "validation", "and", "1", ",", "000", "testing", "images", "with", "publicly", "available", "annotations", ".", "Evaluation", "systems", "report", "the", "and", "PCP", "over", "the", "testing", "set", ".", "Refer", "to", "for", "more", "details", ".", "The", "performance", "comparison", "of", "the", "proposed", "NAN", "with", "three", "state", "-", "of", "-", "the", "-", "art", "methods", "in", "terms", "of", "@IoU=0.5", ",", "and", "PCP@IoU=0.5", "on", "the", "MHP", "v1.0", "testing", "set", "is", "reported", "in", "Tab", ".", "[", "reference", "]", ".", "With", "the", "nested", "adversarial", "learning", "of", "semantic", "saliency", "prediction", ",", "instance", "-", "agnostic", "parsing", "and", "instance", "-", "aware", "clustering", ",", "our", "method", "outperforms", "the", "-", "best", "by", "for", ",", "for", "and", "for", "PCP", ".", "Visual", "comparison", "of", "multi", "-", "human", "parsing", "results", "by", "NAN", "and", "three", "state", "-", "of", "-", "the", "-", "art", "methods", "is", "provided", "in", "Fig", ".", "[", "reference", "]", ",", "which", "further", "validates", "the", "advantages", "of", "our", "NAN", "over", "existing", "solutions", ".", "subsection", ":", "Evaluations", "on", "the", "PASCAL", "-", "Person", "-", "Part", "Benchmark", "The", "PASCAL", "-", "Person", "-", "Part", "dataset", "is", "a", "set", "of", "additional", "annotations", "for", "PASCAL", "-", "VOC", "-", "2010", ".", "It", "goes", "beyond", "the", "original", "PASCAL", "object", "detection", "task", "by", "providing", "pixel", "-", "wise", "labels", "for", "six", "human", "body", "parts", ",", "i.e.", ",", "head", ",", "torso", ",", "upper", "-/", "lower", "-", "arms", ",", "and", "upper", "-/", "lower", "-", "legs", ".", "The", "rest", "of", "each", "image", "is", "considered", "as", "background", ".", "There", "are", "3", ",", "535", "images", "in", "the", "PASCAL", "-", "Person", "-", "Part", "dataset", ",", "which", "is", "split", "into", "separate", "training", "set", "containing", "1", ",", "717", "images", "and", "testing", "set", "containing", "1", ",", "818", "images", ".", "For", "fair", "comparison", ",", "we", "report", "the", "over", "the", "testing", "set", "for", "multi", "-", "human", "parsing", ".", "Refer", "to", "for", "more", "details", ".", "The", "performance", "comparison", "of", "the", "proposed", "NAN", "with", "two", "state", "-", "of", "-", "the", "-", "art", "methods", "in", "terms", "of", "@IoU=", "and", "on", "the", "PASCAL", "-", "Person", "-", "Part", "testing", "set", "is", "reported", "in", "Tab", ".", "[", "reference", "]", ".", "Our", "method", "dramatically", "surpasses", "the", "-", "best", "by", "for", "and", "for", ".", "Qualitative", "multi", "-", "human", "parsing", "results", "by", "NAN", "are", "visualized", "in", "Fig", ".", "[", "reference", "]", ",", "which", "possess", "a", "high", "concordance", "with", "corresponding", "ground", "truths", ".", "This", "again", "verifies", "the", "effectiveness", "of", "our", "method", "for", "human", "-", "centric", "analysis", ".", "subsection", ":", "Evaluations", "on", "the", "Buffy", "Benchmark", "The", "Buffy", "dataset", "was", "released", "in", "2011", "for", "human", "parsing", "and", "instance", "segmentation", ",", "which", "contains", "748", "images", "annotated", "with", "12", "semantic", "labels", ".", "The", "data", "are", "randomly", "organized", "into", "2", "splits", ",", "consisting", "of", "452", "training", "and", "296", "testing", "images", "with", "publicly", "available", "annotations", ".", "For", "fair", "comparison", ",", "we", "report", "the", "F", "orward", "(", "F", ")", "and", "B", "ackward", "(", "B", ")", "scores", "over", "the", "episode", "4", ",", "5", "and", "6", "for", "instance", "segmentation", "evaluation", ".", "Refer", "to", "for", "more", "details", ".", "The", "performance", "comparison", "of", "the", "proposed", "NAN", "with", "three", "state", "-", "of", "-", "the", "-", "art", "methods", "in", "terms", "of", "F", "and", "B", "scores", "on", "the", "Buffy", "dataset", "episode", "4", ",", "5", "and", "6", "is", "reported", "in", "Tab", ".", "[", "reference", "]", ".", "Our", "NAN", "consistently", "achieves", "the", "best", "performance", "for", "all", "metrics", ".", "In", "particualr", ",", "NAN", "significantly", "improves", "the", "-", "best", "by", "for", "F", "score", "and", "for", "B", "score", ",", "with", "an", "average", "boost", "of", ".", "Qualitative", "instance", "-", "agnostic", "parsing", "and", "instance", "-", "aware", "clustering", "results", "by", "NAN", "are", "visualized", "in", "Fig", ".", "[", "reference", "]", ",", "which", "well", "shows", "the", "promising", "potential", "of", "our", "method", "for", "fine", "-", "grained", "understanding", "humans", "in", "crowded", "scenes", ".", "section", ":", "Conclusions", "In", "this", "work", ",", "we", "presented", "“", "M", "ulti", "-", "H", "uman", "P", "arsing", "(", "MHP", "v2.0", ")", "”", ",", "a", "large", "-", "scale", "multi", "-", "human", "parsing", "dataset", "and", "a", "carefully", "designed", "benchmark", "to", "spark", "progress", "in", "understanding", "humans", "in", "crowded", "scenes", ".", "MHP", "v2.0", "contains", "25", ",", "403", "images", ",", "which", "are", "richly", "labelled", "with", "59", "semantic", "categories", ".", "We", "also", "proposed", "a", "novel", "deep", "N", "ested", "A", "dversarial", "N", "etwork", "(", "NAN", ")", "model", "to", "address", "this", "challenging", "problem", "and", "performed", "detailed", "evaluations", "of", "the", "proposed", "method", "with", "current", "state", "-", "of", "-", "the", "-", "arts", "on", "MHP", "v2.0", "and", "several", "other", "datasets", ".", "We", "envision", "the", "proposed", "MHP", "v2.0", "dataset", "and", "the", "baseline", "method", "would", "drive", "the", "human", "parsing", "research", "towards", "real", "-", "world", "application", "scenario", "with", "simultaneous", "presence", "of", "multiple", "persons", "and", "complex", "interactions", "among", "them", ".", "In", "future", ",", "we", "will", "continue", "to", "take", "efforts", "to", "construct", "a", "more", "comprehensive", "multi", "-", "human", "parsing", "benchmark", "dataset", "with", "more", "images", "and", "more", "detailed", "semantic", "category", "annotations", "to", "further", "push", "the", "frontiers", "of", "multi", "-", "human", "parsing", "research", ".", "section", ":", "Acknowledgement", "The", "work", "of", "Jian", "Zhao", "was", "partially", "supported", "by", "C", "hina", "S", "cholarship", "C", "ouncil", "(", "CSC", ")", "grant", "201503170248", ".", "The", "work", "of", "Jiashi", "Feng", "was", "partially", "supported", "by", "NUS", "startup", "R", "-", "263", "-", "000", "-", "C08", "-", "133", ",", "MOE", "Tier", "-", "I", "R", "-", "263", "-", "000", "-", "C21", "-", "112", ",", "NUS", "IDS", "R", "-", "263", "-", "000", "-", "C67", "-", "646", "and", "ECRA", "R", "-", "263", "-", "000", "-", "C87", "-", "133", ".", "bibliography", ":", "References"], "entity": [[2, 7, ["Understanding", "Humans", "in", "Crowded", "Scenes"], "Task"], [8, 12, ["Deep", "Nested", "Adversarial", "Learning"], "Method"], [17, 21, ["Multi", "-", "Human", "Parsing"], "Task"], [26, 28, ["perceptual", "tasks"], "Task"], [29, 30, ["detection"], "Task"], [31, 33, ["instance", "segmentation"], "Task"], [34, 36, ["human", "parsing"], "Task"], [42, 48, ["visually", "understanding", "humans", "in", "crowded", "scenes"], "Task"], [51, 54, ["group", "behavior", "analysis"], "Task"], [55, 59, ["person", "re", "-", "identification"], "Task"], [60, 62, ["autonomous", "driving"], "Task"], [114, 117, ["M", "ulti", "-"], "Task"], [119, 121, ["P", "arsing"], "Method"], [122, 123, ["MHP"], "Task"], [126, 128, ["algorithm", "development"], "Task"], [129, 130, ["evaluation"], "Task"], [142, 147, ["understanding", "humans", "in", "crowded", "scenes"], "Task"], [148, 149, ["MHP"], "Task"], [197, 203, ["N", "ested", "A", "dversarial", "N", "etwork"], "Method"], [204, 205, ["NAN"], "Method"], [208, 212, ["multi", "-", "human", "parsing"], "Task"], [213, 214, ["NAN"], "Method"], [224, 225, ["GAN"], "Method"], [233, 236, ["semantic", "saliency", "prediction"], "Method"], [237, 241, ["instance", "-", "agnostic", "parsing"], "Method"], [242, 246, ["instance", "-", "aware", "clustering"], "Method"], [271, 272, ["NAN"], "Method"], [285, 286, ["MHP"], "Task"], [303, 307, ["multi", "-", "human", "parsing"], "Task"], [317, 322, ["intelligent", "human", "-", "computer", "interaction"], "Task"], [333, 335, ["perceptual", "tasks"], "Task"], [336, 337, ["detection"], "Task"], [360, 362, ["instance", "segmentation"], "Task"], [383, 385, ["human", "parsing"], "Task"], [406, 410, ["deep", "learning", "based", "methods"], "Method"], [416, 418, ["perceptual", "tasks"], "Task"], [429, 431, ["evaluation", "purposes"], "Task"], [445, 448, ["visually", "understanding", "humans"], "Task"], [461, 465, ["humans", "in", "visual", "scenes"], "Task"], [536, 540, ["Multi", "-", "human", "parsing"], "Task"], [543, 546, ["understanding", "of", "humans"], "Task"], [554, 558, ["real", "-", "world", "applications"], "Task"], [561, 564, ["group", "behavior", "analysis"], "Task"], [565, 569, ["person", "re", "-", "identification"], "Task"], [570, 573, ["e", "-", "commerce"], "Task"], [574, 576, ["image", "editing"], "Task"], [577, 579, ["video", "surveillance"], "Task"], [580, 582, ["autonomous", "driving"], "Task"], [583, 585, ["virtual", "reality"], "Task"], [618, 620, ["MHP", "v1.0"], "Material"], [644, 648, ["large", "-", "scale", "benchmark"], "Task"], [649, 656, ["M", "ulti", "-", "H", "uman", "P", "arsing"], "Task"], [657, 659, ["MHP", "v2.0"], "Material"], [674, 680, ["holistically", "understanding", "humans", "in", "crowded", "scenes"], "Task"], [684, 686, ["MHP", "v2.0"], "Material"], [729, 735, ["N", "ested", "A", "dversarial", "N", "etwork"], "Method"], [736, 737, ["NAN"], "Method"], [759, 761, ["instance", "localization"], "Task"], [762, 764, ["human", "parsing"], "Task"], [765, 767, ["result", "refinement"], "Task"], [770, 771, ["NAN"], "Method"], [772, 774, ["semantic", "categories"], "Task"], [789, 790, ["NAN"], "Method"], [800, 801, ["GAN"], "Method"], [809, 812, ["semantic", "saliency", "prediction"], "Method"], [813, 817, ["instance", "-", "agnostic", "parsing"], "Method"], [818, 822, ["instance", "-", "aware", "clustering"], "Method"], [824, 827, ["sub", "-", "task"], "Task"], [852, 857, ["multi", "-", "task", "learning", "applications"], "Task"], [881, 883, ["adversarial", "strategy"], "Method"], [898, 903, ["nested", "adversarial", "learning", "”", "structure"], "Method"], [910, 914, ["B", "ackpro", "P", "agation"], "Method"], [915, 916, ["BP"], "Method"], [918, 919, ["NAN"], "Method"], [941, 942, ["accuracy"], "Metric"], [943, 944, ["speed"], "Metric"], [953, 954, ["MHP"], "Task"], [965, 967, ["MHP", "v1.0"], "Material"], [968, 973, ["PASCAL", "-", "Person", "-", "Part"], "Material"], [974, 977, ["Buffy", "benchmark", "datasets"], "Material"], [984, 985, ["NAN"], "Method"], [986, 990, ["multi", "-", "human", "parsing"], "Task"], [1011, 1018, ["large", "-", "scale", "benchmark", "and", "evaluation", "server"], "Metric"], [1020, 1026, ["understanding", "of", "humans", "in", "crowded", "scenes"], "Task"], [1048, 1054, ["N", "ested", "A", "dversarial", "N", "etwork"], "Method"], [1055, 1056, ["NAN"], "Method"], [1059, 1063, ["multi", "-", "human", "parsing"], "Task"], [1084, 1085, ["MHP"], "Task"], [1096, 1098, ["MHP", "v1.0"], "Material"], [1099, 1104, ["PASCAL", "-", "Person", "-", "Part"], "Material"], [1105, 1108, ["Buffy", "benchmark", "datasets"], "Material"], [1112, 1113, ["NAN"], "Method"], [1114, 1116, ["understanding", "humans"], "Task"], [1135, 1138, ["Human", "Parsing", "Datasets"], "Material"], [1146, 1148, ["human", "parsing"], "Task"], [1158, 1160, ["Buffy", "dataset"], "Material"], [1165, 1167, ["human", "parsing"], "Task"], [1168, 1170, ["instance", "segmentation"], "Task"], [1183, 1185, ["Fashionista", "dataset"], "Material"], [1190, 1192, ["human", "parsing"], "Task"], [1219, 1225, ["PASCAL", "-", "VOC", "-", "2010", "dataset"], "Material"], [1233, 1235, ["human", "parsing"], "Task"], [1243, 1245, ["ATR", "dataset"], "Material"], [1250, 1252, ["human", "parsing"], "Task"], [1275, 1277, ["Microsoft", "COCO"], "Material"], [1322, 1331, ["fine", "-", "grained", "understanding", "of", "humans", "in", "visual", "scenes"], "Task"], [1335, 1336, ["LIP"], "Material"], [1351, 1355, ["instance", "-", "agnostic", "setting"], "Method"], [1373, 1377, ["multi", "-", "human", "parsing"], "Task"], [1395, 1398, ["MHP", "v2.0", "dataset"], "Material"], [1439, 1441, ["MHP", "v1.0"], "Material"], [1443, 1445, ["MHP", "v2.0"], "Material"], [1456, 1459, ["Human", "Parsing", "Approaches"], "Method"], [1468, 1470, ["human", "parsing"], "Task"], [1488, 1492, ["proposal", "-", "free", "network"], "Method"], [1493, 1495, ["instance", "segmentation"], "Task"], [1517, 1525, ["self", "-", "supervised", "structure", "-", "sensitive", "learning", "approach"], "Method"], [1546, 1551, ["single", "frame", "video", "parsing", "method"], "Method"], [1553, 1555, ["frame", "parsing"], "Method"], [1556, 1559, ["optical", "flow", "estimation"], "Method"], [1560, 1562, ["temporal", "fusion"], "Method"], [1564, 1566, ["unified", "network"], "Method"], [1573, 1579, ["self", "-", "supervised", "neural", "aggregation", "network"], "Method"], [1612, 1616, ["Mask", "R", "-", "CNN"], "Method"], [1641, 1644, ["bounding", "box", "recognition"], "Task"], [1652, 1654, ["instance", "segmentation"], "Task"], [1656, 1659, ["discriminative", "loss", "function"], "Method"], [1669, 1671, ["convolutional", "network"], "Method"], [1688, 1692, ["post", "-", "processing", "step"], "Method"], [1714, 1718, ["human", "-", "centric", "analysis"], "Task"], [1786, 1788, ["region", "assembly"], "Method"], [1798, 1800, ["human", "detector"], "Method"], [1802, 1807, ["category", "-", "level", "segmentation", "module"], "Method"], [1826, 1827, ["localization"], "Task"], [1828, 1830, ["human", "parsing"], "Task"], [1831, 1833, ["result", "refinement"], "Task"], [1839, 1840, ["NAN"], "Method"], [1842, 1846, ["multi", "-", "human", "parsing"], "Task"], [1850, 1853, ["forward", "-", "pass"], "Method"], [1864, 1867, ["post", "-", "processing"], "Task"], [1891, 1893, ["MHP", "v2.0"], "Material"], [1904, 1911, ["semantic", "understanding", "of", "humans", "in", "crowded", "scenes"], "Task"], [1959, 1961, ["MHP", "v2.0"], "Material"], [1990, 1992, ["MHP", "v2.0"], "Material"], [2013, 2016, ["MHP", "v2.0", "dataset"], "Material"], [2025, 2029, ["multi", "-", "human", "parsing"], "Task"], [2049, 2051, ["Image", "Collection"], "Task"], [2052, 2053, ["Annotation"], "Task"], [2138, 2141, ["Automated", "scrapping", "software"], "Method"], [2190, 2192, ["Internet", "scraping"], "Task"], [2199, 2201, ["manual", "annotation"], "Task"], [2282, 2284, ["annotation", "tool"], "Method"], [2286, 2292, ["multi", "-", "scale", "superpixels", "of", "images"], "Task"], [2298, 2299, ["annotation"], "Task"], [2612, 2614, ["manual", "inspection"], "Task"], [2677, 2680, ["MHP", "v2.0", "dataset"], "Material"], [2731, 2733, ["random", "selection"], "Method"], [2785, 2788, ["semantic", "category", "number"], "Metric"], [2845, 2849, ["humans", "in", "crowded", "scenes"], "Task"], [2944, 2948, ["instance", "-", "level", "analysis"], "Task"], [2951, 2953, ["clothing", "categories"], "Task"], [3082, 3084, ["MHP", "v2.0"], "Material"], [3109, 3112, ["MHP", "v2.0", "dataset"], "Material"], [3128, 3129, ["MHP"], "Task"], [3145, 3149, ["human", "-", "centric", "analysis"], "Task"], [3166, 3170, ["Deep", "Nested", "Adversarial", "Networks"], "Method"], [3182, 3188, ["N", "ested", "A", "dversarial", "N", "etwork"], "Method"], [3189, 3190, ["NAN"], "Method"], [3204, 3207, ["semantic", "saliency", "prediction"], "Method"], [3208, 3212, ["instance", "-", "agnostic", "parsing"], "Method"], [3213, 3217, ["instance", "-", "aware", "clustering"], "Method"], [3223, 3224, ["NAN"], "Method"], [3228, 3230, ["human", "parsing"], "Task"], [3234, 3237, ["forward", "-", "pass"], "Method"], [3248, 3251, ["post", "-", "processing"], "Task"], [3262, 3265, ["Semantic", "Saliency", "Prediction"], "Task"], [3275, 3279, ["multi", "-", "human", "parsing"], "Task"], [3290, 3295, ["human", "-", "centric", "analysis", "model"], "Method"], [3327, 3333, ["GAN", "-", "based", "sub", "-", "net"], "Method"], [3337, 3339, ["training", "complexity"], "Metric"], [3343, 3345, ["empirical", "performance"], "Metric"], [3355, 3358, ["semantic", "saliency", "maps"], "Task"], [3397, 3400, ["semantic", "saliency", "prediction"], "Method"], [3402, 3408, ["binary", "pixel", "-", "wise", "labelling", "problem"], "Task"], [3430, 3431, ["FCNs"], "Method"], [3435, 3441, ["image", "-", "to", "-", "image", "applications"], "Task"], [3445, 3447, ["FCN", "backbone"], "Method"], [3448, 3451, ["FCN", "-", "8s"], "Method"], [3458, 3461, ["semantic", "saliency", "prediction"], "Method"], [3500, 3502, ["RGB", "image"], "Material"], [3522, 3525, ["semantic", "saliency", "map"], "Method"], [3573, 3578, ["refining", "realism", "and", "alleviating", "artifacts"], "Task"], [3582, 3586, ["emantic", "s", "aliency", "loss"], "Metric"], [3587, 3592, ["pixel", "-", "wise", "image", "labelling"], "Task"], [3602, 3609, ["pixel", "-", "wise", "cross", "-", "entropy", "loss"], "Metric"], [3644, 3650, ["C", "onvolutional", "N", "eural", "N", "etwork"], "Method"], [3651, 3652, ["CNN"], "Method"], [3689, 3693, ["Instance", "-", "Agnostic", "Parsing"], "Method"], [3704, 3706, ["RGB", "image"], "Material"], [3718, 3723, ["instance", "-", "agnostic", "parsing", "map"], "Method"], [3737, 3741, ["instance", "-", "aware", "clustering"], "Method"], [3754, 3758, ["instance", "-", "agnostic", "parsing"], "Method"], [3760, 3766, ["multi", "-", "class", "dense", "classification", "problem"], "Task"], [3785, 3788, ["skip", "-", "net"], "Method"], [3789, 3791, ["recognition", "tasks"], "Task"], [3795, 3798, ["skip", "-", "net"], "Method"], [3799, 3802, ["WS", "-", "ResNet"], "Method"], [3805, 3809, ["FCN", "-", "based", "architecture"], "Method"], [3818, 3822, ["non", "-", "linear", "transformation"], "Method"], [3823, 3827, ["instance", "-", "agnostic", "parsing"], "Method"], [3844, 3845, ["prediction"], "Task"], [3849, 3850, ["accuracy"], "Metric"], [3852, 3856, ["speed", "trade", "-", "off"], "Metric"], [3870, 3872, ["human", "parsing"], "Task"], [3878, 3881, ["skip", "-", "net"], "Method"], [3918, 3923, ["smooth", "and", "accurate", "parsing", "results"], "Task"], [3928, 3933, ["instance", "-", "agnostic", "parsing", "map"], "Method"], [3956, 3961, ["g", "lobal", "p", "arsing", "loss"], "Method"], [3962, 3965, ["semantic", "part", "labelling"], "Task"], [3969, 3976, ["pixel", "-", "wise", "cross", "-", "entropy", "loss"], "Metric"], [3997, 4001, ["hinged", "gradient", "backpropagation", "route"], "Method"], [4024, 4027, ["information", "flow", "consistency"], "Task"], [4037, 4039, ["CNN", "backbone"], "Method"], [4062, 4066, ["Instance", "-", "Aware", "Clustering"], "Method"], [4077, 4079, ["RGB", "image"], "Material"], [4093, 4098, ["instance", "-", "aware", "clustering", "map"], "Method"], [4101, 4104, ["semantic", "parsing", "mask"], "Method"], [4151, 4155, ["instance", "-", "aware", "clustering"], "Method"], [4175, 4178, ["region", "proposal", "generation"], "Method"], [4183, 4185, ["backbone", "architecture"], "Method"], [4196, 4200, ["location", "-", "sensitive", "learning"], "Task"], [4236, 4239, ["location", "prediction", "accuracy"], "Task"], [4244, 4251, ["pixel", "-", "wise", "instance", "location", "prediction", "branch"], "Method"], [4263, 4264, ["MSFU"], "Method"], [4294, 4297, ["instance", "number", "regression"], "Task"], [4334, 4339, ["i", "nstance", "l", "ocation", "loss"], "Metric"], [4340, 4346, ["pixel", "-", "wise", "instance", "location", "regression"], "Method"], [4349, 4354, ["i", "nstance", "n", "umber", "loss"], "Metric"], [4355, 4358, ["instance", "number", "regression"], "Task"], [4435, 4436, ["normalization"], "Task"], [4442, 4443, ["loss"], "Metric"], [4471, 4476, ["instance", "-", "aware", "clustering", "maps"], "Method"], [4482, 4484, ["computational", "overhead"], "Metric"], [4506, 4509, ["information", "flow", "consistency"], "Task"], [4519, 4521, ["CNN", "backbone"], "Method"], [4547, 4548, ["Inference"], "Task"], [4551, 4552, ["NAN"], "Method"], [4574, 4578, ["multi", "-", "human", "parsing"], "Task"], [4592, 4594, ["network", "convergence"], "Task"], [4597, 4599, ["objective", "function"], "Metric"], [4600, 4601, ["NAN"], "Method"], [4605, 4606, ["NAN"], "Method"], [4620, 4624, ["nested", "adversarial", "learning", "strategy"], "Method"], [4625, 4627, ["BP", "algorithm"], "Method"], [4642, 4647, ["instance", "-", "agnostic", "parsing", "map"], "Method"], [4664, 4671, ["off", "-", "the", "-", "shelf", "clustering", "method"], "Method"], [4674, 4679, ["instance", "-", "aware", "clustering", "map"], "Method"], [4696, 4697, ["NAN"], "Method"], [4706, 4711, ["understanding", "humans", "in", "crowded", "scenes"], "Task"], [4717, 4721, ["multi", "-", "human", "parsing"], "Task"], [4724, 4725, ["MHP"], "Task"], [4736, 4738, ["MHP", "v1.0"], "Material"], [4739, 4746, ["PASCAL", "-", "Person", "-", "Part", "benchmark", "datasets"], "Material"], [4750, 4754, ["instance", "-", "agnostic", "parsing"], "Method"], [4755, 4759, ["instance", "-", "aware", "clustering"], "Method"], [4762, 4765, ["Buffy", "benchmark", "dataset"], "Material"], [4770, 4771, ["NAN"], "Method"], [4788, 4790, ["RGB", "image"], "Material"], [4792, 4795, ["semantic", "saliency", "prediction"], "Method"], [4799, 4800, ["discriminator"], "Method"], [4813, 4818, ["instance", "-", "agnostic", "parsing", "prediction"], "Method"], [4828, 4829, ["discriminator"], "Method"], [4884, 4885, ["generator"], "Method"], [4888, 4891, ["FCN", "-", "8s"], "Method"], [4899, 4901, ["convolutional", "layer"], "Method"], [4907, 4912, ["PASCAL", "-", "VOC", "-", "2011"], "Material"], [4924, 4927, ["WS", "-", "ResNet"], "Method"], [4958, 4960, ["linear", "classifier"], "Method"], [4965, 4967, ["convolutional", "layers"], "Method"], [4974, 4975, ["ImageNet"], "Material"], [4976, 4981, ["PASCAL", "-", "VOC", "-", "2012"], "Material"], [4996, 4998, ["backbone", "architecture"], "Method"], [5020, 5027, ["pixel", "-", "wise", "instance", "location", "map", "prediction"], "Task"], [5028, 5031, ["instance", "number", "prediction"], "Task"], [5038, 5039, ["MSFU"], "Method"], [5041, 5043, ["convolutional", "layers"], "Method"], [5054, 5056, ["convolutional", "layer"], "Method"], [5060, 5065, ["multi", "-", "scale", "feature", "aggregation"], "Method"], [5068, 5070, ["convolutional", "layer"], "Method"], [5074, 5076, ["location", "regression"], "Task"], [5099, 5102, ["global", "pooling", "layer"], "Method"], [5104, 5112, ["hidden", "512", "-", "way", "fully", "-", "connected", "layer"], "Method"], [5115, 5122, ["1", "-", "way", "fully", "-", "connected", "layer"], "Method"], [5123, 5126, ["instance", "number", "regression"], "Task"], [5141, 5144, ["VGG", "-", "16"], "Method"], [5158, 5160, ["input", "adaption"], "Task"], [5184, 5185, ["ImageNet"], "Material"], [5216, 5223, ["off", "-", "the", "-", "shelf", "clustering", "method"], "Method"], [5226, 5231, ["instance", "-", "aware", "clustering", "map"], "Method"], [5233, 5235, ["dropout", "ratio"], "Metric"], [5257, 5259, ["learning", "rate"], "Metric"], [5279, 5281, ["learning", "rate"], "Metric"], [5312, 5314, ["TensorFlow", "platform"], "Method"], [5319, 5321, ["Adam", "("], "Method"], [5324, 5330, ["NVIDIA", "GeForce", "GTX", "TITAN", "X", "GPUs"], "Method"], [5351, 5353, ["testing", "time"], "Metric"], [5356, 5358, ["running", "time"], "Metric"], [5365, 5371, ["NVIDIA", "GeForce", "GTX", "TITAN", "X", "GPU"], "Method"], [5381, 5382, ["NAN"], "Method"], [5420, 5423, ["region", "proposal", "preprocessing"], "Method"], [5430, 5432, ["Evaluation", "Metrics"], "Metric"], [5443, 5445, ["p", "art"], "Method"], [5448, 5457, ["P", "ercentage", "of", "C", "orrectly", "parsed", "semantic", "P", "arts"], "Method"], [5458, 5459, ["PCP"], "Method"], [5471, 5475, ["A", "verage", "P", "recision"], "Metric"], [5477, 5479, ["r", "egion"], "Method"], [5483, 5485, ["instance", "segmentation"], "Task"], [5526, 5530, ["human", "-", "centric", "analysis"], "Task"], [5598, 5599, ["PCP"], "Method"], [5603, 5606, ["human", "pose", "estimation"], "Task"], [5610, 5612, ["parsing", "quality"], "Metric"], [5650, 5651, ["PCP"], "Method"], [5685, 5686, ["PCP"], "Method"], [5697, 5698, ["PCP"], "Method"], [5701, 5706, ["human", "-", "centric", "evaluation", "metric"], "Metric"], [5712, 5715, ["MHP", "v2.0", "Benchmark"], "Material"], [5716, 5719, ["MHP", "v2.0", "dataset"], "Material"], [5739, 5741, ["MHP", "v1.0"], "Material"], [5746, 5748, ["understanding", "humans"], "Task"], [5838, 5840, ["Component", "Analysis"], "Method"], [5850, 5851, ["NAN"], "Method"], [5857, 5861, ["multi", "-", "human", "parsing"], "Task"], [5875, 5879, ["Mask", "R", "-", "CNN"], "Method"], [5880, 5883, ["MH", "-", "Parser"], "Method"], [5886, 5888, ["network", "structures"], "Method"], [5940, 5941, ["MSFU"], "Method"], [5945, 5946, ["NAN"], "Method"], [5971, 5978, ["ground", "truth", "instance", "-", "agnostic", "parsing", "maps"], "Method"], [6033, 6034, ["@IoU=0.5"], "Metric"], [6036, 6037, ["PCP@IoU=0.5"], "Method"], [6067, 6068, ["NAN"], "Method"], [6071, 6076, ["baselines", "Mask", "R", "-", "CNN"], "Method"], [6077, 6080, ["MH", "-", "Parser"], "Method"], [6104, 6108, ["Mask", "R", "-", "CNN"], "Method"], [6115, 6118, ["MH", "-", "Parser"], "Method"], [6122, 6124, ["instance", "localization"], "Task"], [6125, 6127, ["human", "parsing"], "Task"], [6128, 6130, ["result", "refinement"], "Task"], [6132, 6133, ["complexity"], "Metric"], [6141, 6142, ["NAN"], "Method"], [6155, 6159, ["deep", "nested", "adversarial", "learning"], "Method"], [6181, 6182, ["NAN"], "Method"], [6196, 6199, ["w", "/", "o"], "Method"], [6229, 6232, ["semantic", "saliency", "prediction"], "Method"], [6250, 6254, ["human", "-", "centic", "processing"], "Task"], [6296, 6298, ["adversarial", "learning"], "Method"], [6306, 6309, ["w", "/", "o"], "Method"], [6310, 6311, ["NAN"], "Method"], [6329, 6333, ["Nested", "adversarial", "learning", "strategy"], "Method"], [6342, 6345, ["information", "flow", "consistency"], "Task"], [6360, 6361, ["NAN"], "Method"], [6376, 6377, ["MSFU"], "Method"], [6399, 6403, ["w", "/", "o", "MSFU"], "Method"], [6404, 6405, ["NAN"], "Method"], [6432, 6433, ["NAN"], "Method"], [6452, 6455, ["semantic", "saliency", "prediction"], "Method"], [6479, 6480, ["NAN"], "Method"], [6484, 6490, ["instance", "-", "agnostic", "parsing", "network", "architecture"], "Method"], [6497, 6501, ["multi", "-", "human", "parsing"], "Task"], [6510, 6511, ["NAN"], "Method"], [6530, 6534, ["instance", "-", "aware", "clustering"], "Method"], [6539, 6543, ["multi", "-", "human", "parsing"], "Task"], [6544, 6545, ["subsubsection"], "Method"], [6546, 6548, ["Quantitative", "Comparison"], "Metric"], [6554, 6555, ["NAN"], "Method"], [6568, 6569, ["@IoU=0.5"], "Metric"], [6571, 6572, ["PCP@IoU=0.5"], "Method"], [6574, 6575, ["MHP"], "Task"], [6645, 6646, ["NAN"], "Method"], [6666, 6667, ["NAN"], "Method"], [6696, 6697, ["NAN"], "Method"], [6712, 6713, ["NAN"], "Method"], [6735, 6736, ["NAN"], "Method"], [6750, 6751, ["NAN"], "Method"], [6764, 6766, ["resource", "consumption"], "Metric"], [6779, 6782, ["MH", "-", "Parser"], "Method"], [6784, 6787, ["img", "/", "s"], "Method"], [6795, 6798, ["post", "-", "processing"], "Method"], [6800, 6801, ["CRF"], "Method"], [6806, 6808, ["Qualitative", "Comparison"], "Metric"], [6820, 6821, ["NAN"], "Method"], [6833, 6835, ["ground", "truths"], "Metric"], [6837, 6840, ["MHP", "v2.0", "dataset"], "Material"], [6843, 6847, ["Mask", "R", "-", "CNN"], "Method"], [6859, 6863, ["instance", "-", "aware", "clustering"], "Method"], [6872, 6875, ["MH", "-", "Parser"], "Method"], [6883, 6884, ["NAN"], "Method"], [6887, 6891, ["multi", "-", "human", "parsing"], "Task"], [6908, 6912, ["instance", "-", "agnostic", "parsing"], "Method"], [6919, 6920, ["NAN"], "Method"], [6932, 6934, ["network", "structure"], "Method"], [6936, 6938, ["training", "strategy"], "Method"], [6942, 6946, ["Mask", "R", "-", "CNN"], "Method"], [6954, 6957, ["MH", "-", "Parser"], "Method"], [6960, 6965, ["fine", "-", "grained", "parsing", "results"], "Task"], [6979, 6980, ["NAN"], "Method"], [6989, 6990, ["NAN"], "Method"], [7039, 7041, ["MHP", "v2.0"], "Material"], [7061, 7064, ["MHP", "v1.0", "Benchmark"], "Material"], [7165, 7166, ["NAN"], "Method"], [7179, 7180, ["@IoU=0.5"], "Metric"], [7182, 7183, ["PCP@IoU=0.5"], "Method"], [7200, 7207, ["nested", "adversarial", "learning", "of", "semantic", "saliency", "prediction"], "Method"], [7208, 7212, ["instance", "-", "agnostic", "parsing"], "Method"], [7213, 7217, ["instance", "-", "aware", "clustering"], "Method"], [7235, 7239, ["multi", "-", "human", "parsing"], "Task"], [7241, 7242, ["NAN"], "Method"], [7268, 7269, ["NAN"], "Method"], [7298, 7303, ["PASCAL", "-", "VOC", "-", "2010"], "Material"], [7309, 7313, ["PASCAL", "object", "detection", "task"], "Task"], [7402, 7406, ["multi", "-", "human", "parsing"], "Task"], [7419, 7420, ["NAN"], "Method"], [7433, 7434, ["@IoU="], "Metric"], [7472, 7473, ["NAN"], "Method"], [7489, 7491, ["ground", "truths"], "Metric"], [7501, 7505, ["human", "-", "centric", "analysis"], "Task"], [7511, 7513, ["Buffy", "Benchmark"], "Material"], [7514, 7516, ["Buffy", "dataset"], "Material"], [7521, 7523, ["human", "parsing"], "Task"], [7524, 7526, ["instance", "segmentation"], "Task"], [7566, 7570, ["F", "orward", "(", "F"], "Metric"], [7572, 7574, ["B", "ackward"], "Metric"], [7575, 7576, ["B"], "Metric"], [7587, 7590, ["instance", "segmentation", "evaluation"], "Task"], [7603, 7604, ["NAN"], "Method"], [7617, 7621, ["F", "and", "B", "scores"], "Metric"], [7623, 7625, ["Buffy", "dataset"], "Material"], [7641, 7642, ["NAN"], "Method"], [7654, 7655, ["NAN"], "Method"], [7662, 7664, ["F", "score"], "Metric"], [7666, 7668, ["B", "score"], "Metric"], [7676, 7680, ["instance", "-", "agnostic", "parsing"], "Method"], [7681, 7685, ["instance", "-", "aware", "clustering"], "Method"], [7687, 7688, ["NAN"], "Method"], [7707, 7715, ["fine", "-", "grained", "understanding", "humans", "in", "crowded", "scenes"], "Task"], [7734, 7736, ["MHP", "v2.0"], "Material"], [7757, 7762, ["understanding", "humans", "in", "crowded", "scenes"], "Task"], [7763, 7765, ["MHP", "v2.0"], "Material"], [7786, 7792, ["N", "ested", "A", "dversarial", "N", "etwork"], "Method"], [7793, 7794, ["NAN"], "Method"], [7819, 7821, ["MHP", "v2.0"], "Material"], [7830, 7833, ["MHP", "v2.0", "dataset"], "Material"], [7844, 7849, ["real", "-", "world", "application", "scenario"], "Task"], [7921, 7922, ["CSC"], "Method"], [7973, 7974, ["ECRA"], "Method"], [97, 99, ["human", "parsing"], "Task"], [532, 534, ["human", "parsing"], "Task"], [670, 672, ["human", "parsing"], "Task"], [745, 747, ["human", "parsing"], "Task"], [834, 836, ["human", "parsing"], "Task"], [1203, 1208, ["PASCAL", "-", "Person", "-", "Part"], "Material"], [1265, 1266, ["LIP"], "Material"], [1284, 1286, ["human", "parsing"], "Task"], [1363, 1364, ["MHP"], "Task"], [1423, 1425, ["human", "parsing"], "Task"], [1437, 1438, ["LIP"], "Material"], [1624, 1625, ["CNN"], "Method"], [1728, 1730, ["human", "parsing"], "Task"], [1745, 1748, ["MH", "-", "Parser"], "Method"], [1870, 1874, ["Multi", "-", "Human", "Parsing"], "Task"], [2799, 2800, ["MHP"], "Task"], [3160, 3162, ["human", "parsing"], "Task"], [3195, 3196, ["GAN"], "Method"], [3456, 3457, ["NAN"], "Method"], [3507, 3510, ["semantic", "saliency", "map"], "Method"], [3666, 3667, ["GAN"], "Method"], [3813, 3814, ["NAN"], "Method"], [4083, 4087, ["instance", "-", "agnostic", "parsing"], "Method"], [4194, 4195, ["NAN"], "Method"], [4562, 4563, ["GAN"], "Method"], [4638, 4639, ["NAN"], "Method"], [4819, 4823, ["instance", "-", "aware", "clustering"], "Method"], [5437, 5441, ["A", "verage", "P", "recision"], "Metric"], [5464, 5466, ["human", "parsing"], "Task"], [5681, 5682, ["PCP"], "Method"], [5689, 5690, ["PCP"], "Method"], [5731, 5733, ["human", "parsing"], "Task"], [5828, 5829, ["PCP"], "Method"], [6039, 6040, ["MHP"], "Task"], [6102, 6103, ["PCP"], "Method"], [6203, 6206, ["semantic", "saliency", "prediction"], "Method"], [6210, 6211, ["NAN"], "Method"], [6276, 6277, ["NAN"], "Method"], [6385, 6389, ["instance", "-", "aware", "clustering"], "Method"], [6503, 6504, ["NAN"], "Method"], [6739, 6741, ["human", "parsing"], "Task"], [6913, 6917, ["instance", "-", "aware", "clustering"], "Method"], [7065, 7066, ["MHP"], "Task"], [7073, 7075, ["human", "parsing"], "Task"], [7147, 7148, ["PCP"], "Method"], [7185, 7186, ["MHP"], "Task"], [7230, 7231, ["PCP"], "Method"], [7278, 7283, ["PASCAL", "-", "Person", "-", "Part"], "Material"], [7285, 7290, ["PASCAL", "-", "Person", "-", "Part"], "Material"], [7362, 7367, ["PASCAL", "-", "Person", "-", "Part"], "Material"], [7437, 7442, ["PASCAL", "-", "Person", "-", "Part"], "Material"], [7468, 7470, ["human", "parsing"], "Task"], [7745, 7747, ["human", "parsing"], "Task"], [7840, 7842, ["human", "parsing"], "Task"], [7877, 7879, ["human", "parsing"], "Task"], [7898, 7900, ["human", "parsing"], "Task"]]} {"doc_id": "27a99c21a1324f087b2f144adc119f04137dfd87", "text": ["document", ":", "Deep", "Fried", "Convnets", "The", "fully", "-", "connected", "layers", "of", "deep", "convolutional", "neural", "networks", "typically", "contain", "over", "90", "%", "of", "the", "network", "parameters", ".", "Reducing", "the", "number", "of", "parameters", "while", "preserving", "predictive", "performance", "is", "critically", "important", "for", "training", "big", "models", "in", "distributed", "systems", "and", "for", "deployment", "in", "embedded", "devices", ".", "In", "this", "paper", ",", "we", "introduce", "a", "novel", "Adaptive", "Fastfood", "transform", "to", "reparameterize", "the", "matrix", "-", "vector", "multiplication", "of", "fully", "connected", "layers", ".", "Reparameterizing", "a", "fully", "connected", "layer", "with", "inputs", "and", "outputs", "with", "the", "Adaptive", "Fastfood", "transform", "reduces", "the", "storage", "and", "computational", "costs", "costs", "from", "to", "and", "respectively", ".", "Using", "the", "Adaptive", "Fastfood", "transform", "in", "convolutional", "networks", "results", "in", "what", "we", "call", "a", "deep", "fried", "convnet", ".", "These", "convnets", "are", "end", "-", "to", "-", "end", "trainable", ",", "and", "enable", "us", "to", "attain", "substantial", "reductions", "in", "the", "number", "of", "parameters", "without", "affecting", "prediction", "accuracy", "on", "the", "MNIST", "and", "ImageNet", "datasets", ".", "section", ":", "Introduction", "In", "recent", "years", "we", "have", "witnessed", "an", "explosion", "of", "applications", "of", "convolutional", "neural", "networks", "with", "millions", "and", "billions", "of", "parameters", ".", "Reducing", "this", "vast", "number", "of", "parameters", "would", "improve", "the", "efficiency", "of", "training", "in", "distributed", "architectures", ".", "It", "would", "also", "allow", "for", "the", "deployment", "of", "state", "-", "of", "-", "the", "-", "art", "convolutional", "neural", "networks", "on", "embedded", "mobile", "applications", ".", "These", "train", "and", "test", "time", "considerations", "are", "both", "of", "great", "importance", ".", "A", "standard", "convolutional", "network", "is", "composed", "of", "two", "types", "of", "layers", ",", "each", "with", "very", "different", "properties", ".", "Convolutional", "layers", ",", "which", "contain", "a", "small", "fraction", "of", "the", "network", "parameters", ",", "represent", "most", "of", "the", "computational", "effort", ".", "In", "contrast", ",", "fully", "connected", "layers", "contain", "the", "vast", "majority", "of", "the", "parameters", "but", "are", "comparatively", "cheap", "to", "evaluate", ".", "This", "imbalance", "between", "memory", "and", "computation", "suggests", "that", "the", "efficiency", "of", "these", "two", "types", "of", "layers", "should", "be", "addressed", "in", "different", "ways", ".", "and", "both", "describe", "methods", "for", "minimizing", "computational", "cost", "of", "evaluating", "a", "network", "at", "test", "time", "by", "approximating", "the", "learned", "convolutional", "filters", "with", "separable", "approximations", ".", "These", "approaches", "realize", "speed", "gains", "at", "test", "time", "but", "do", "not", "address", "the", "issue", "of", "training", ",", "since", "the", "approximations", "are", "made", "after", "the", "network", "has", "been", "fully", "trained", ".", "Additionally", ",", "neither", "approach", "achieves", "a", "substantial", "reduction", "in", "the", "number", "of", "parameters", ",", "since", "they", "both", "work", "with", "approximations", "of", "the", "convolutional", "layers", ",", "which", "represent", "only", "a", "small", "portion", "of", "the", "total", "number", "of", "parameters", ".", "Many", "other", "works", "have", "addressed", "the", "computational", "efficiency", "of", "convolutional", "networks", "in", "more", "specialized", "settings", ".", "In", "contrast", "to", "the", "above", "approaches", ",", "demonstrates", "that", "there", "is", "significant", "redundancy", "in", "the", "parameterization", "of", "several", "deep", "learning", "models", ",", "and", "exploits", "this", "to", "reduce", "the", "number", "of", "parameters", ".", "More", "specifically", ",", "their", "method", "represents", "the", "parameter", "matrix", "as", "a", "product", "of", "two", "low", "rank", "factors", ",", "and", "the", "training", "algorithm", "fixes", "one", "factor", "(", "called", "static", "parameters", ")", "and", "only", "updates", "the", "other", "factor", "(", "called", "dynamic", "parameters", ")", ".", "uses", "low", "-", "rank", "matrix", "factorization", "to", "reduce", "the", "size", "of", "the", "fully", "connected", "layers", "at", "train", "time", ".", "They", "demonstrate", "large", "improvements", "in", "reducing", "the", "number", "of", "parameters", "of", "the", "output", "softmax", "layer", ",", "but", "only", "modest", "improvements", "for", "the", "hidden", "fully", "connected", "layers", ".", "implements", "low", "-", "rank", "factorizations", "using", "the", "SVD", "after", "training", "the", "full", "model", ".", "In", "contrast", ",", "the", "methods", "advanced", "in", "and", "this", "paper", "apply", "both", "at", "train", "and", "test", "time", ".", "In", "this", "paper", "we", "show", "how", "the", "number", "of", "parameters", "required", "to", "represent", "a", "deep", "convolutional", "neural", "network", "can", "be", "substantially", "reduced", "without", "sacrificing", "predictive", "performance", ".", "Our", "approach", "works", "by", "replacing", "the", "fully", "connected", "layers", "of", "the", "network", "with", "an", "Adaptive", "Fastfood", "transform", ",", "which", "is", "a", "generalization", "of", "the", "Fastfood", "transform", "for", "approximating", "kernels", ".", "Convolutional", "neural", "networks", "with", "Adaptive", "Fastfood", "transforms", ",", "which", "we", "refer", "to", "as", "deep", "fried", "convnets", ",", "are", "end", "-", "to", "-", "end", "trainable", "and", "achieve", "the", "same", "predictive", "performance", "as", "standard", "convolutional", "networks", "on", "ImageNet", "using", "approximately", "half", "the", "number", "of", "parameters", ".", "Several", "works", "have", "considered", "kernel", "methods", "in", "deep", "learning", ".", "The", "Doubly", "Stochastic", "Gradients", "method", "of", "showed", "that", "effective", "use", "of", "randomization", "can", "allow", "kernel", "methods", "to", "scale", "to", "extremely", "large", "data", "sets", ".", "However", ",", "the", "approach", "used", "fixed", "convolutional", "features", ",", "and", "can", "not", "jointly", "learn", "the", "kernel", "classifier", "and", "convolutional", "filters", ".", "showed", "how", "to", "learn", "a", "kernel", "function", "in", "an", "unsupervised", "manner", ".", "There", "have", "been", "other", "attempts", "to", "replace", "the", "fully", "connected", "layers", ".", "The", "Network", "in", "Network", "architecture", "of", "achieves", "state", "of", "the", "art", "results", "on", "several", "deep", "learning", "benchmarks", "by", "replacing", "the", "fully", "connected", "layers", "with", "global", "average", "pooling", ".", "A", "similar", "approach", "was", "used", "by", "to", "win", "the", "ILSVRC", "2014", "object", "detection", "competition", ".", "Although", "the", "global", "average", "pooling", "approach", "achieves", "impressive", "results", ",", "it", "has", "two", "significant", "drawbacks", ".", "First", ",", "feature", "transfer", "is", "more", "difficult", "with", "this", "approach", ".", "It", "is", "very", "common", "in", "practice", "to", "take", "a", "convolutional", "network", "trained", "on", "ImageNet", "and", "re", "-", "train", "the", "top", "layer", "on", "a", "different", "data", "set", ",", "re", "-", "using", "the", "features", "learned", "from", "ImageNet", "for", "the", "new", "task", "(", "potentially", "with", "fine", "-", "tuning", ")", ",", "and", "this", "is", "difficult", "with", "global", "average", "pooling", ".", "This", "deficiency", "is", "noted", "by", ",", "and", "motivates", "them", "to", "add", "an", "extra", "linear", "layer", "to", "the", "top", "of", "their", "network", "to", "enable", "them", "to", "more", "easily", "adapt", "and", "fine", "tune", "their", "network", "to", "other", "label", "sets", ".", "The", "second", "drawback", "of", "global", "average", "pooling", "is", "computation", ".", "Convolutional", "layers", "are", "much", "more", "expensive", "to", "evaluate", "than", "fully", "connected", "layers", ",", "so", "replacing", "fully", "connected", "layers", "with", "more", "convolutions", "can", "decrease", "model", "size", "but", "comes", "at", "the", "cost", "of", "increased", "evaluation", "time", ".", "In", "parallel", "or", "after", "the", "first", "(", "technical", "report", ")", "version", "of", "this", "work", ",", "several", "researchers", "have", "attempted", "to", "create", "sparse", "networks", "by", "applying", "pruning", "or", "sparsity", "regularizers", ".", "These", "approaches", "however", "require", "training", "the", "original", "full", "model", "and", ",", "consequently", ",", "do", "not", "enjoy", "the", "efficient", "training", "time", "benefits", "of", "the", "techniques", "proposed", "in", "this", "paper", ".", "Since", "then", ",", "hashing", "methods", "have", "also", "been", "advanced", "to", "reduce", "the", "number", "of", "parameters", ".", "Hashes", "have", "irregular", "memory", "access", "patterns", "and", ",", "consequently", ",", "good", "performance", "on", "large", "GPU", "-", "based", "platforms", "is", "yet", "to", "be", "demonstrated", ".", "Finally", ",", "distillation", "also", "offers", "a", "way", "of", "compressing", "neural", "networks", ",", "as", "a", "post", "-", "processing", "step", ".", "section", ":", "The", "Adaptive", "Fastfood", "Transform", "Large", "dense", "matrices", "are", "the", "main", "building", "block", "of", "fully", "connected", "neural", "network", "layers", ".", "In", "propagating", "the", "signal", "from", "the", "-", "th", "layer", "with", "activations", "to", "the", "-", "th", "layer", "with", "activations", ",", "we", "have", "to", "compute", "The", "storage", "and", "computational", "costs", "of", "this", "matrix", "multiplication", "step", "are", "both", ".", "The", "storage", "cost", "in", "particular", "can", "be", "prohibitive", "for", "many", "applications", ".", "Our", "proposed", "solution", "is", "to", "reparameterize", "the", "matrix", "of", "parameters", "with", "an", "Adaptive", "Fastfood", "transform", ",", "as", "follows", "In", "Section", "[", "reference", "]", ",", "we", "will", "provide", "background", "and", "intuitions", "behind", "this", "design", ".", "For", "now", "it", "suffices", "to", "state", "that", "the", "storage", "requirements", "of", "this", "reparameterization", "are", "and", "the", "computational", "cost", "is", ".", "We", "will", "also", "show", "in", "the", "experimental", "section", "that", "these", "theoretical", "savings", "are", "mirrored", "in", "practice", "by", "significant", "reductions", "in", "the", "number", "of", "parameters", "without", "increased", "prediction", "errors", ".", "To", "understand", "these", "claims", ",", "we", "need", "to", "describe", "the", "component", "modules", "of", "the", "Adaptive", "Fastfood", "transform", ".", "For", "simplicity", "of", "presentation", ",", "let", "us", "first", "assume", "that", ".", "Adaptive", "Fastfood", "has", "three", "types", "of", "module", ":", "and", "are", "diagonal", "matrices", "of", "parameters", ".", "In", "the", "original", "non", "-", "adaptive", "Fastfood", "formulation", "they", "are", "random", "matrices", ",", "as", "described", "further", "in", "Section", "[", "reference", "]", ".", "The", "computational", "and", "storage", "costs", "are", "trivially", ".", "is", "a", "random", "permutation", "matrix", ".", "It", "can", "be", "implemented", "as", "a", "lookup", "table", ",", "so", "the", "storage", "and", "computational", "costs", "are", "also", ".", "denotes", "the", "Walsh", "-", "Hadamard", "matrix", ",", "which", "is", "defined", "recursively", "as", "The", "Fast", "Hadamard", "Transform", ",", "a", "variant", "of", "Fast", "Fourier", "Transform", ",", "enables", "us", "to", "compute", "in", "time", ".", "In", "summary", ",", "the", "overall", "storage", "cost", "of", "the", "Adaptive", "Fastfood", "transform", "is", ",", "while", "the", "computational", "cost", "is", ".", "These", "are", "substantial", "theoretical", "improvements", "over", "the", "costs", "of", "ordinary", "fully", "connected", "layers", ".", "When", "the", "number", "of", "output", "units", "is", "larger", "than", "the", "number", "of", "inputs", ",", "we", "can", "perform", "Adaptive", "Fastfood", "transforms", "and", "stack", "them", "to", "attain", "the", "desired", "size", ".", "In", "doing", "so", ",", "the", "computational", "and", "storage", "costs", "become", "and", "respectively", ",", "as", "opposed", "to", "the", "more", "substantial", "costs", "for", "linear", "modules", ".", "The", "number", "of", "outputs", "can", "also", "be", "refined", "with", "pruning", ".", "subsection", ":", "Learning", "Fastfood", "by", "backpropagation", "The", "parameters", "of", "the", "Adaptive", "Fastfood", "transform", "(", "and", ")", "can", "be", "learned", "by", "standard", "error", "derivative", "backpropagation", ".", "Moreover", ",", "the", "backward", "pass", "can", "also", "be", "computed", "efficiently", "using", "the", "Fast", "Hadamard", "Transform", ".", "In", "particular", ",", "let", "us", "consider", "learning", "the", "-", "th", "layer", "of", "the", "network", ",", ".", "For", "simplicity", ",", "let", "us", "again", "assume", "that", "and", "that", ".", "Using", "backpropagation", ",", "assume", "we", "already", "have", ",", "where", "is", "the", "objective", "function", ",", "then", "Since", "is", "a", "diagonal", "matrix", ",", "we", "only", "need", "to", "calculate", "the", "derivative", "with", "respect", "to", "the", "diagonal", "entries", "and", "this", "step", "requires", "only", "operations", ".", "Proceeding", "in", "this", "way", ",", "denote", "the", "partial", "products", "by", "Then", "the", "gradients", "with", "respect", "to", "different", "parameters", "in", "the", "Fastfood", "layer", "can", "be", "computed", "recursively", "as", "follows", ":", "Note", "that", "the", "operations", "in", "and", "are", "simply", "applications", "of", "the", "Hadamard", "transform", ",", "since", ",", "and", "consequently", "can", "be", "computed", "in", "time", ".", "The", "operation", "in", "is", "an", "application", "of", "a", "permutation", "(", "the", "transpose", "of", "permutation", "matrix", "is", "a", "permutation", "matrix", ")", "and", "can", "be", "computed", "in", "time", ".", "All", "other", "operations", "are", "diagonal", "matrix", "multiplications", ".", "section", ":", "Intuitions", "behind", "Adaptive", "Fastfood", "The", "proposed", "Adaptive", "Fastfood", "transform", "may", "be", "understood", "either", "as", "a", "trainable", "type", "of", "structured", "random", "projection", "or", "as", "an", "approximation", "to", "the", "feature", "space", "of", "a", "learned", "kernel", ".", "Both", "views", "not", "only", "shed", "light", "on", "Adaptive", "Fastfood", "and", "competing", "techniques", ",", "but", "also", "open", "up", "room", "to", "innovate", "new", "techniques", "to", "reduce", "computation", "and", "memory", "in", "neural", "networks", ".", "subsection", ":", "A", "view", "from", "structured", "random", "projections", "Adaptive", "Fastfood", "is", "based", "on", "the", "Fastfood", "transform", ",", "in", "which", "the", "diagonal", "matrices", ",", "and", "have", "random", "entries", ".", "In", "the", "experiments", ",", "we", "will", "compare", "the", "performance", "of", "the", "existing", "random", "and", "proposed", "adaptive", "versions", "of", "Fastfood", "when", "used", "to", "replace", "fully", "connected", "layers", "in", "convolutional", "neural", "networks", ".", "The", "intriguing", "idea", "of", "constructing", "neural", "networks", "with", "random", "weights", "has", "been", "reasonably", "explored", "in", "the", "neural", "networks", "field", ".", "This", "idea", "is", "related", "to", "random", "projections", ",", "which", "have", "been", "deeply", "studied", "in", "theoretical", "computer", "science", ".", "In", "a", "random", "projection", ",", "the", "basic", "operation", "is", "of", "the", "form", "where", "is", "a", "random", "matrix", ",", "either", "Gaussian", "or", "binary", ".", "Importantly", ",", "the", "embeddings", "generated", "by", "these", "random", "projections", "approximately", "preserve", "metric", "information", ",", "as", "formalized", "by", "many", "variants", "of", "the", "celebrated", "Johnson", "-", "Lindenstrauss", "Lemma", ".", "The", "one", "shortcoming", "of", "random", "projections", "is", "that", "the", "cost", "of", "storing", "the", "matrix", "is", ".", "Using", "a", "sparse", "random", "matrix", "by", "itself", "to", "reduce", "this", "cost", "is", "often", "not", "a", "viable", "option", "because", "the", "variance", "of", "the", "estimates", "of", "can", "be", "very", "high", "for", "some", "inputs", ",", "for", "example", "when", "is", "also", "sparse", ".", "To", "see", "this", ",", "consider", "the", "extreme", "case", "of", "a", "very", "sparse", "input", ",", "then", "many", "of", "the", "products", "with", "will", "be", "zero", "and", "hence", "not", "help", "improve", "the", "estimates", "of", "metric", "properties", "of", "the", "embedding", "space", ".", "One", "popular", "option", "for", "reducing", "the", "storage", "and", "computational", "costs", "of", "random", "projections", "is", "to", "adopt", "random", "hash", "functions", "to", "replace", "the", "random", "matrix", "multiplication", ".", "For", "example", ",", "the", "count", "-", "sketch", "algorithm", "uses", "pairwise", "independent", "hash", "functions", "to", "carry", "this", "job", "very", "effectively", "in", "many", "applications", ".", "This", "technique", "is", "often", "referred", "to", "as", "the", "hashing", "trick", "in", "the", "machine", "learning", "literature", ".", "Hashes", "have", "irregular", "memory", "access", "patterns", ",", "so", "it", "is", "not", "clear", "how", "to", "get", "good", "performance", "on", "GPUs", "when", "following", "this", "approach", ",", "as", "pointed", "out", "in", ".", "Ailon", "and", "Chazelle", "introduced", "an", "alternative", "approach", "that", "is", "not", "only", "very", "efficient", ",", "but", "also", "preserves", "most", "of", "the", "desirable", "theoretical", "properties", "of", "random", "projections", ".", "Their", "idea", "was", "to", "replace", "the", "random", "matrix", "by", "a", "transform", "that", "mimics", "the", "properties", "of", "random", "matrices", ",", "but", "which", "can", "be", "stored", "efficiently", ".", "In", "particular", ",", "they", "proposed", "the", "following", "PHD", "transform", ":", "where", "is", "a", "sparse", "random", "matrix", "with", "Gaussian", "entries", ",", "is", "a", "Hadamard", "matrix", "and", "is", "a", "diagonal", "matrix", "with", "entries", "drawn", "independently", "with", "probability", ".", "The", "inclusion", "of", "the", "Hadamard", "transform", "avoids", "the", "problems", "of", "using", "a", "sparse", "random", "matrix", "by", "itself", ",", "but", "it", "is", "still", "efficient", "to", "compute", ".", "We", "can", "think", "of", "the", "original", "Fastfood", "transform", "as", "an", "alternative", "to", "this", ".", "Fastfood", "reduces", "the", "computation", "and", "storage", "of", "random", "projections", "to", "and", "respectively", ".", "In", "the", "original", "formulation", "and", "are", "diagonal", "random", "matrices", ",", "which", "are", "computed", "once", "and", "then", "stored", ".", "In", "contrast", ",", "in", "our", "proposed", "Adaptive", "Fastfood", "transform", ",", "the", "diagonal", "matrices", "are", "learned", "by", "backpropagation", ".", "By", "adapting", ",", "we", "are", "effectively", "implementing", "Automatic", "Relevance", "Determination", "on", "features", ".", "The", "matrix", "controls", "the", "bandwidth", "of", "the", "kernel", "and", "its", "spectral", "incoherence", ".", "Finally", ",", "represents", "different", "kernel", "types", ".", "For", "example", ",", "for", "the", "RBF", "kernel", "follows", "Chi", "-", "squared", "distribution", ".", "By", "adapting", ",", "we", "learn", "the", "correct", "kernel", "type", ".", "While", "we", "have", "introduced", "Fastfood", "in", "this", "section", ",", "it", "was", "originally", "proposed", "as", "a", "fast", "way", "of", "computing", "random", "features", "to", "approximate", "kernels", ".", "We", "expand", "on", "this", "perspective", "in", "the", "following", "section", ".", "subsection", ":", "A", "view", "from", "kernels", "There", "is", "a", "nice", "duality", "between", "inner", "products", "of", "features", "and", "kernels", ".", "This", "duality", "can", "be", "used", "to", "design", "neural", "network", "modules", "using", "kernels", "and", "vice", "-", "versa", ".", "For", "computational", "reasons", ",", "we", "often", "want", "to", "determine", "the", "features", "associated", "with", "a", "kernel", ".", "Working", "with", "features", "is", "preferable", "when", "the", "kernel", "matrix", "is", "dense", "and", "large", ".", "(", "Storing", "this", "matrix", "requires", "space", ",", "and", "computing", "it", "takes", "operations", ",", "where", "is", "the", "number", "of", "data", "points", "and", "is", "the", "dimension", ".", ")", "We", "might", "also", "want", "to", "design", "statistical", "methods", "using", "kernels", "and", "then", "map", "these", "designs", "to", "features", "that", "can", "be", "used", "as", "modules", "in", "neural", "networks", ".", "Unfortunately", ",", "one", "of", "the", "difficulties", "with", "this", "line", "of", "attack", "is", "that", "deriving", "features", "from", "kernels", "is", "far", "from", "trivial", "in", "general", ".", "An", "important", "fact", ",", "noted", "in", ",", "is", "that", "infinite", "kernel", "expansions", "can", "be", "approximated", "in", "an", "unbiased", "manner", "using", "randomly", "drawn", "features", ".", "For", "shift", "-", "invariant", "kernels", "this", "relies", "on", "a", "classical", "result", "from", "harmonic", "analysis", ",", "known", "as", "Bochner", "’s", "Lemma", ",", "which", "states", "that", "a", "continuous", "shift", "-", "invariant", "kernel", "on", "is", "positive", "definite", "if", "and", "only", "if", "is", "the", "Fourier", "transform", "of", "a", "non", "-", "negative", "measure", ".", "This", "measure", ",", "known", "as", "the", "spectral", "density", ",", "in", "turn", "implies", "the", "existence", "of", "a", "probability", "density", "such", "that", "where", "the", "imaginary", "part", "is", "dropped", "since", "both", "the", "kernel", "and", "distribution", "are", "real", ".", "We", "can", "apply", "Monte", "Carlo", "methods", "to", "approximate", "the", "above", "expectation", ",", "and", "hence", "approximate", "the", "kernel", "with", "an", "inner", "product", "of", "stacked", "cosine", "and", "sine", "features", ".", "Specifically", ",", "suppose", "we", "sample", "vectors", "from", "and", "collect", "them", "in", "a", "matrix", ".", "The", "kernel", "can", "then", "be", "approximated", "as", "the", "inner", "-", "product", "of", "the", "following", "random", "features", ":", "That", "is", ",", "is", "the", "neural", "network", "module", ",", "consisting", "of", "a", "linear", "layer", "and", "entry", "-", "wise", "nonlinearities", "(", "cosine", "and", "sine", "in", "the", "above", "equation", ")", ",", "that", "corresponds", "to", "a", "particular", "implicit", "kernel", "function", ".", "Approximating", "a", "given", "kernel", "function", "with", "random", "features", "requires", "the", "specification", "of", "a", "sampling", "distribution", ".", "Such", "distributions", "have", "been", "derived", "for", "many", "popular", "kernels", ".", "For", "example", ",", "if", "we", "want", "the", "implicit", "kernel", "to", "be", "a", "squared", "exponential", "kernel", ",", "we", "know", "that", "the", "distribution", "must", "be", "Gaussian", ":", ".", "In", "other", "words", ",", "if", "we", "draw", "the", "rows", "of", "from", "this", "Gaussian", "distribution", "and", "use", "equation", "(", "[", "reference", "]", ")", "to", "implement", "a", "neural", "module", ",", "we", "are", "implicitly", "approximating", "a", "squared", "exponential", "kernel", ".", "As", "another", "example", "of", "the", "mapping", "between", "kernels", "and", "random", "features", ",", "introduced", "the", "rotationally", "invariant", "arc", "-", "cosine", "kernel", "where", "is", "the", "angle", "between", "and", ".", "Then", "by", "choosing", "to", "be", "a", "random", "Gaussian", "matrix", ",", "they", "showed", "that", "this", "kernel", "can", "be", "approximated", "with", "Rectified", "Linear", "Unit", "(", "ReLU", ")", "features", ":", "The", "Fastfood", "transform", "was", "introduced", "to", "replace", "in", "Equation", "[", "reference", "]", "with", ",", "thus", "decreasing", "the", "computational", "and", "storage", "costs", ".", "section", ":", "Deep", "Fried", "Convolutional", "Networks", "We", "propose", "to", "greatly", "reduce", "the", "number", "of", "parameters", "of", "the", "fully", "connected", "layers", "by", "replacing", "them", "with", "an", "Adaptive", "Fastfood", "transform", "followed", "by", "a", "nonlinearity", ".", "We", "call", "this", "new", "architecture", "a", "deep", "fried", "convolutional", "network", ".", "An", "illustration", "of", "this", "architecture", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "In", "principle", ",", "we", "could", "also", "apply", "the", "Adaptive", "Fastfood", "transform", "to", "the", "softmax", "classifier", ".", "However", ",", "reducing", "the", "memory", "cost", "of", "this", "layer", "is", "already", "well", "studied", ";", "for", "example", ",", "show", "that", "low", "-", "rank", "matrix", "factorization", "can", "be", "applied", "during", "training", "to", "reduce", "the", "size", "of", "the", "softmax", "layer", "substantially", ".", "Importantly", ",", "they", "also", "show", "that", "training", "a", "low", "rank", "factorization", "for", "the", "internal", "layers", "performs", "poorly", ",", "which", "agrees", "with", "the", "results", "of", ".", "For", "this", "reason", ",", "we", "focus", "our", "attention", "on", "reducing", "the", "size", "of", "the", "internal", "layers", ".", "section", ":", "MNIST", "Experiment", "The", "first", "problem", "we", "study", "is", "the", "classical", "MNIST", "optical", "character", "recognition", "task", ".", "This", "simple", "task", "serves", "as", "an", "easy", "proof", "of", "concept", "for", "our", "method", ",", "and", "contrasting", "the", "results", "in", "this", "section", "with", "our", "later", "experiments", "gives", "insights", "into", "the", "behavior", "of", "the", "Adaptive", "Fastfood", "transform", "at", "different", "scales", ".", "As", "a", "reference", "model", "we", "use", "the", "Caffe", "implementation", "of", "the", "LeNet", "convolutional", "network", ".", "It", "achieves", "an", "error", "rate", "of", "on", "the", "MNIST", "dataset", ".", "We", "jointly", "train", "all", "layers", "of", "the", "deep", "fried", "network", "(", "including", "convolutional", "layers", ")", "from", "scratch", ".", "We", "compare", "both", "the", "adaptive", "and", "non", "-", "adaptive", "Fastfood", "transforms", "using", "1024", "and", "2048", "features", ".", "For", "the", "non", "-", "adaptive", "transforms", "we", "report", "the", "best", "performance", "achieved", "by", "varying", "the", "standard", "deviation", "of", "the", "random", "Gaussian", "matrix", "over", "the", "set", ",", "and", "for", "the", "adaptive", "variant", "we", "learn", "these", "parameters", "by", "backpropagation", "as", "described", "in", "Section", "[", "reference", "]", ".", "The", "results", "of", "the", "MNIST", "experiment", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Because", "the", "width", "of", "the", "deep", "fried", "network", "is", "substantially", "larger", "than", "the", "reference", "model", ",", "we", "also", "experimented", "with", "adding", "dropout", "in", "the", "model", ",", "which", "increased", "performance", "in", "the", "deep", "fried", "case", ".", "Deep", "fried", "networks", "are", "able", "to", "obtain", "high", "accuracy", "using", "only", "a", "small", "fraction", "of", "of", "parameters", "of", "the", "original", "network", "(", "11", "times", "reduction", "in", "the", "best", "case", ")", ".", "Interestingly", ",", "we", "see", "no", "benefit", "from", "adaptation", "in", "this", "experiment", ",", "with", "the", "more", "powerful", "adaptive", "models", "performing", "equivalently", "or", "worse", "than", "their", "non", "-", "adaptive", "counterparts", ";", "however", ",", "this", "should", "be", "contrasted", "with", "the", "ImageNet", "results", "reported", "in", "the", "following", "sections", ".", "section", ":", "Imagenet", "Experiments", "We", "now", "examine", "how", "deep", "fried", "networks", "behave", "in", "a", "more", "realistic", "setting", "with", "a", "much", "larger", "dataset", "and", "many", "more", "classes", ".", "Specifically", ",", "we", "use", "the", "ImageNet", "ILSVRC", "-", "2012", "dataset", "which", "has", "1.2", "M", "training", "examples", "and", "50", "K", "validation", "examples", "distributed", "across", "1000", "classes", ".", "We", "use", "the", "the", "Caffe", "ImageNet", "model", "as", "the", "reference", "model", "in", "these", "experiments", ".", "This", "model", "is", "a", "modified", "version", "of", "AlexNet", ",", "and", "achieves", "top", "-", "1", "error", "on", "the", "ILSVRC", "-", "2012", "validation", "set", ".", "The", "initial", "layers", "of", "this", "model", "are", "a", "cascade", "of", "convolution", "and", "pooling", "layers", "with", "interspersed", "normalization", ".", "The", "last", "several", "layers", "of", "the", "network", "take", "the", "form", "of", "an", "MLP", "and", "follow", "a", "9216–4096–4096–1000", "architecture", ".", "The", "final", "layer", "is", "a", "logistic", "regression", "layer", "with", "1000", "output", "classes", ".", "All", "layers", "of", "this", "network", "use", "the", "ReLU", "nonlinearity", ",", "and", "dropout", "is", "used", "in", "the", "fully", "connected", "layers", "to", "prevent", "overfitting", ".", "There", "are", "total", "of", "58", ",", "649", ",", "184", "parameters", "in", "the", "reference", "model", ",", "of", "which", "58", ",", "621", ",", "952", "are", "in", "the", "fully", "connected", "layers", "and", "only", "27", ",", "232", "are", "in", "the", "convolutional", "layers", ".", "The", "parameters", "of", "fully", "connected", "layer", "take", "up", "of", "the", "total", "number", "of", "parameters", ".", "We", "show", "that", "the", "Adaptive", "Fastfood", "transform", "can", "be", "used", "to", "substantially", "reduce", "the", "number", "of", "parameters", "in", "this", "model", ".", "subsection", ":", "Fixed", "feature", "extractor", "Previous", "work", "on", "applying", "kernel", "methods", "to", "ImageNet", "has", "focused", "on", "building", "models", "on", "features", "extracted", "from", "the", "convolutional", "layers", "of", "a", "pre", "-", "trained", "network", ".", "This", "setting", "is", "less", "general", "than", "training", "a", "network", "from", "scratch", "but", "does", "mirror", "the", "common", "use", "case", "where", "a", "convolutional", "network", "is", "first", "trained", "on", "ImageNet", "and", "used", "as", "a", "feature", "extractor", "for", "a", "different", "task", ".", "In", "order", "to", "compare", "our", "Adaptive", "Fastfood", "transform", "directly", "to", "this", "previous", "work", ",", "we", "extract", "features", "from", "the", "final", "convolutional", "layer", "of", "a", "pre", "-", "trained", "reference", "model", "and", "train", "an", "Adaptive", "Fastfood", "transform", "classifier", "using", "these", "features", ".", "Although", "the", "reference", "model", "uses", "two", "fully", "connected", "layers", ",", "we", "investigate", "replacing", "these", "with", "only", "a", "single", "Fastfood", "transform", ".", "We", "experiment", "with", "two", "sizes", "for", "this", "transform", ":", "Fastfood", "16", "and", "Fastfood", "32", "using", "16", ",", "384", "and", "32", ",", "768", "Fastfood", "features", "respectively", ".", "Since", "the", "Fastfood", "transform", "is", "a", "composite", "module", ",", "we", "can", "apply", "dropout", "between", "any", "of", "its", "layers", ".", "In", "the", "experiments", "reported", "here", ",", "we", "applied", "dropout", "after", "the", "matrix", "and", "after", "the", "matrix", ".", "We", "also", "applied", "dropout", "to", "the", "last", "convolutional", "layer", "(", "that", "is", ",", "before", "the", "matrix", ")", ".", "We", "also", "train", "an", "MLP", "with", "the", "same", "structure", "as", "the", "top", "layers", "of", "the", "reference", "model", "for", "comparison", ".", "In", "this", "setting", "it", "is", "important", "to", "compare", "against", "the", "re", "-", "trained", "MLP", "rather", "than", "the", "jointly", "trained", "reference", "model", ",", "as", "training", "on", "features", "extracted", "from", "fixed", "convolutional", "layers", "typically", "leads", "to", "lower", "performance", "than", "joint", "training", ".", "The", "results", "of", "the", "fixed", "feature", "experiment", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Following", "and", "we", "observe", "that", "training", "on", "ImageNet", "activations", "produces", "significantly", "lower", "performance", "than", "of", "the", "original", ",", "jointly", "trained", "network", ".", "Nonetheless", ",", "deep", "fried", "networks", "are", "able", "to", "outperform", "both", "the", "re", "-", "trained", "MLP", "model", "as", "well", "as", "the", "results", "in", "while", "using", "fewer", "parameters", ".", "In", "contrast", "with", "our", "MNIST", "experiment", ",", "here", "we", "find", "that", "the", "Adaptive", "Fastfood", "transform", "provides", "a", "significant", "performance", "boost", "over", "the", "non", "-", "adaptive", "version", ",", "improving", "top", "-", "1", "performance", "by", "4.5", "-", "6.5", "%", ".", "subsection", ":", "Jointly", "trained", "model", "Finally", ",", "we", "train", "a", "deep", "fried", "network", "from", "scratch", "on", "ImageNet", ".", "With", "16", ",", "384", "features", "in", "the", "Fastfood", "layer", "we", "lose", "less", "than", "0.3", "%", "top", "-", "1", "validation", "performance", ",", "but", "the", "number", "of", "parameters", "in", "the", "network", "is", "reduced", "from", "58.7", "M", "to", "16.4", "M", "which", "corresponds", "to", "a", "factor", "of", "3.6x", ".", "By", "further", "increasing", "the", "number", "of", "features", "to", "32", ",", "768", ",", "we", "are", "able", "to", "perform", "0.6", "%", "better", "than", "the", "reference", "model", "while", "using", "approximately", "half", "as", "many", "parameters", ".", "Results", "from", "this", "experiment", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Nearly", "all", "of", "the", "parameters", "of", "the", "deep", "fried", "network", "reside", "in", "the", "final", "softmax", "regression", "layer", ",", "which", "still", "uses", "a", "dense", "linear", "transformation", ",", "and", "accounts", "for", "more", "than", "99", "%", "of", "the", "parameters", "of", "the", "network", ".", "This", "is", "a", "side", "effect", "of", "the", "large", "number", "of", "classes", "in", "ImageNet", ".", "For", "a", "data", "set", "with", "fewer", "classes", "the", "advantage", "of", "deep", "fried", "convolutional", "networks", "would", "be", "even", "greater", ".", "Moreover", ",", "as", "shown", "by", ",", "the", "last", "layer", "often", "contains", "considerable", "redundancy", ".", "We", "also", "note", "that", "any", "of", "the", "techniques", "from", "could", "be", "applied", "to", "the", "final", "layer", "of", "a", "deep", "fried", "network", "to", "further", "reduce", "memory", "consumption", "at", "test", "time", ".", "We", "illustrate", "this", "with", "low", "-", "rank", "matrix", "factorization", "in", "the", "following", "section", ".", "section", ":", "Comparison", "with", "Post", "Processing", "In", "this", "section", "we", "provide", "a", "comparison", "to", "some", "existing", "works", "on", "reducing", "the", "number", "of", "parameters", "in", "a", "convolutional", "neural", "network", ".", "The", "techniques", "we", "compare", "against", "here", "are", "post", "-", "processing", "techniques", ",", "which", "start", "from", "a", "full", "trained", "model", "and", "attempt", "to", "compress", "it", ",", "whereas", "our", "method", "trains", "the", "compressed", "network", "from", "scratch", ".", "Matrix", "factorization", "is", "the", "most", "common", "method", "for", "compressing", "neural", "networks", ",", "and", "has", "proven", "to", "be", "very", "effective", ".", "Given", "the", "weight", "matrix", "of", "fully", "connected", "layers", ",", "we", "factorize", "it", "as", "where", "and", "and", "is", "a", "diagonal", "matrix", ".", "In", "order", "to", "reduce", "the", "parameters", ",", "we", "truncate", "all", "but", "the", "largest", "singular", "values", ",", "leading", "to", "the", "approximation", "where", "and", "and", "has", "been", "absorbed", "into", "the", "other", "two", "factors", ".", "If", "is", "sufficiently", "small", "then", "storing", "and", "is", "less", "expensive", "than", "storing", "directly", ",", "and", "this", "parameterization", "is", "still", "learnable", ".", "It", "has", "been", "shown", "that", "training", "a", "factorized", "representation", "directly", "leads", "to", "poor", "performance", "(", "although", "it", "does", "work", "when", "applied", "only", "to", "the", "final", "logistic", "regression", "layer", ")", ".", "However", ",", "first", "training", "a", "full", "model", ",", "then", "preforming", "an", "SVD", "of", "the", "weight", "matrices", "followed", "by", "a", "fine", "tuning", "phase", "preserves", "much", "of", "the", "performance", "of", "the", "original", "model", ".", "We", "compare", "our", "deep", "fried", "approach", "to", "SVD", "followed", "by", "fine", "tuning", ",", "and", "show", "that", "our", "approach", "achieves", "better", "performance", "per", "parameter", "in", "spite", "of", "training", "a", "compressed", "parameterization", "from", "scratch", ".", "We", "also", "compare", "against", "a", "post", "-", "processed", "version", "of", "our", "model", ",", "where", "we", "train", "a", "deep", "fried", "convnet", "and", "then", "apply", "SVD", "plus", "fine", "-", "tuning", "to", "the", "final", "softmax", "layer", ",", "which", "further", "reduces", "the", "number", "of", "parameters", ".", "Results", "of", "these", "post", "-", "processing", "experiments", "are", "shown", "in", "Table", "[", "reference", "]", ".", "For", "the", "SVD", "decomposition", "of", "each", "of", "the", "three", "fully", "connected", "layers", "in", "the", "reference", "model", "we", "set", "in", "SVD", "-", "half", "and", "in", "SVD", "-", "quarter", ".", "SVD", "-", "half", "-", "F", "and", "SVD", "-", "quarter", "-", "F", "mean", "that", "the", "model", "has", "been", "fine", "tuned", "after", "the", "decomposition", ".", "There", "is", "1", "%", "drop", "in", "accuracy", "for", "SVD", "-", "half", "and", "3.5", "%", "drop", "for", "SVD", "-", "quarter", ".", "Even", "though", "the", "increase", "in", "the", "error", "for", "the", "SVD", "can", "be", "mitigated", "by", "finetuning", "(", "the", "drop", "decreases", "to", "0.1", "%", "for", "SVD", "-", "half", "-", "F", "and", "1.3", "%", "for", "SVD", "-", "quarter", "-", "F", ")", ",", "deep", "fried", "convnets", "still", "perform", "better", "both", "in", "terms", "of", "the", "accuracy", "and", "the", "number", "of", "parameters", ".", "Applying", "a", "rank", "600", "SVD", "followed", "by", "fine", "tuning", "to", "the", "final", "softmax", "layer", "of", "the", "Adaptive", "Fastfood", "32", "model", "removes", "an", "additional", "12.5", "M", "parameters", "at", "the", "expense", "of", "0.7", "%", "top", "-", "1", "error", ".", "For", "reference", ",", "we", "also", "include", "the", "results", "of", "Collins", "and", "Kohli", ",", "who", "pre", "-", "train", "a", "full", "network", "and", "use", "a", "sparsity", "regularizer", "during", "fine", "-", "tuning", "to", "encourage", "connections", "in", "the", "fully", "connected", "layers", "to", "be", "zero", ".", "They", "are", "able", "to", "achieve", "a", "significant", "reduction", "in", "the", "number", "of", "parameters", "this", "way", ",", "however", "the", "performance", "of", "their", "compressed", "network", "suffers", "when", "compared", "to", "the", "reference", "model", ".", "Another", "drawback", "of", "this", "method", "is", "that", "using", "sparse", "weight", "matrices", "requires", "additional", "overhead", "to", "store", "the", "indexes", "of", "the", "non", "-", "zero", "values", ".", "The", "index", "storage", "takes", "up", "space", "and", "using", "sparse", "representation", "is", "better", "than", "using", "a", "dense", "matrix", "only", "when", "number", "of", "nonzero", "entries", "is", "small", ".", "section", ":", "Conclusion", "Many", "methods", "have", "been", "advanced", "to", "reduce", "the", "size", "of", "convolutional", "networks", "at", "test", "time", ".", "In", "contrast", "to", "this", "trend", ",", "the", "Adaptive", "Fastfood", "transform", "introduced", "in", "this", "paper", "is", "end", "-", "to", "-", "end", "differentiable", "and", "hence", "it", "enables", "us", "to", "attain", "reductions", "in", "the", "number", "of", "parameters", "even", "at", "train", "time", ".", "Deep", "fried", "convnets", "capitalize", "on", "the", "proposed", "Adaptive", "Fastfood", "transform", "to", "achieve", "a", "substantial", "reduction", "in", "the", "number", "of", "parameters", "without", "sacrificing", "predictive", "performance", "on", "MNIST", "and", "ImageNet", ".", "They", "also", "compare", "favorably", "against", "simple", "test", "-", "time", "low", "-", "rank", "matrix", "factorization", "schemes", ".", "Our", "experiments", "have", "also", "cast", "some", "light", "on", "the", "issue", "of", "random", "versus", "adaptive", "weights", ".", "The", "structured", "random", "transformations", "developed", "in", "the", "kernel", "literature", "perform", "very", "well", "on", "MNIST", "without", "any", "learning", ";", "however", ",", "when", "moving", "to", "ImageNet", ",", "the", "benefit", "of", "adaptation", "becomes", "clear", ",", "as", "it", "allows", "us", "to", "achieve", "substantially", "better", "performance", ".", "This", "is", "an", "important", "point", "which", "illustrates", "the", "importance", "of", "learning", "which", "would", "not", "have", "been", "visible", "from", "experiments", "only", "on", "small", "data", "sets", ".", "The", "Fastfood", "transform", "allows", "for", "a", "theoretical", "reduction", "in", "computation", "from", "to", ".", "However", ",", "the", "computation", "in", "convolutional", "neural", "networks", "is", "dominated", "by", "the", "convolutions", ",", "and", "hence", "deep", "fried", "convnets", "are", "not", "necessarily", "faster", "in", "practice", ".", "It", "is", "clear", "looking", "at", "out", "results", "on", "ImageNet", "in", "Table", "2", "that", "the", "remaining", "parameters", "are", "mostly", "in", "the", "output", "softmax", "layer", ".", "The", "comparative", "experiment", "in", "Section", "7", "showed", "that", "the", "matrix", "of", "parameters", "in", "the", "softmax", "can", "be", "easily", "compressed", "using", "the", "SVD", ",", "but", "many", "other", "methods", "could", "be", "used", "to", "achieve", "this", ".", "One", "avenue", "for", "future", "research", "involves", "replacing", "the", "softmax", "matrix", ",", "at", "train", "and", "test", "times", ",", "using", "the", "abundant", "set", "of", "techniques", "that", "have", "been", "proposed", "to", "solve", "this", "problem", ",", "including", "low", "-", "rank", "decomposition", ",", "Adaptive", "Fastfood", ",", "and", "pruning", ".", "The", "development", "of", "GPU", "optimized", "Fastfood", "transforms", "that", "can", "be", "used", "to", "replace", "linear", "layers", "in", "arbitrary", "neural", "models", "would", "also", "be", "of", "great", "value", "to", "the", "entire", "research", "community", "given", "the", "ubiquity", "of", "fully", "connected", "layers", "layers", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["Deep", "Fried", "Convnets"], "Method"], [6, 15, ["fully", "-", "connected", "layers", "of", "deep", "convolutional", "neural", "networks"], "Method"], [32, 34, ["predictive", "performance"], "Metric"], [39, 41, ["big", "models"], "Method"], [42, 44, ["distributed", "systems"], "Task"], [48, 50, ["embedded", "devices"], "Task"], [59, 62, ["Adaptive", "Fastfood", "transform"], "Method"], [65, 73, ["matrix", "-", "vector", "multiplication", "of", "fully", "connected", "layers"], "Method"], [76, 79, ["fully", "connected", "layer"], "Method"], [85, 88, ["Adaptive", "Fastfood", "transform"], "Method"], [90, 95, ["storage", "and", "computational", "costs", "costs"], "Metric"], [102, 105, ["Adaptive", "Fastfood", "transform"], "Method"], [106, 108, ["convolutional", "networks"], "Method"], [114, 117, ["deep", "fried", "convnet"], "Method"], [119, 120, ["convnets"], "Method"], [142, 144, ["prediction", "accuracy"], "Metric"], [146, 147, ["MNIST"], "Material"], [148, 150, ["ImageNet", "datasets"], "Material"], [165, 168, ["convolutional", "neural", "networks"], "Method"], [186, 187, ["training"], "Task"], [188, 190, ["distributed", "architectures"], "Task"], [206, 209, ["convolutional", "neural", "networks"], "Method"], [210, 213, ["embedded", "mobile", "applications"], "Task"], [215, 220, ["train", "and", "test", "time", "considerations"], "Metric"], [228, 230, ["convolutional", "network"], "Method"], [244, 246, ["Convolutional", "layers"], "Method"], [261, 263, ["computational", "effort"], "Metric"], [267, 270, ["fully", "connected", "layers"], "Method"], [313, 315, ["computational", "cost"], "Metric"], [325, 328, ["learned", "convolutional", "filters"], "Method"], [329, 331, ["separable", "approximations"], "Method"], [347, 348, ["training"], "Task"], [381, 386, ["approximations", "of", "the", "convolutional", "layers"], "Method"], [406, 408, ["computational", "efficiency"], "Metric"], [409, 411, ["convolutional", "networks"], "Method"], [434, 437, ["deep", "learning", "models"], "Method"], [468, 470, ["training", "algorithm"], "Method"], [491, 496, ["low", "-", "rank", "matrix", "factorization"], "Method"], [531, 535, ["hidden", "fully", "connected", "layers"], "Method"], [537, 541, ["low", "-", "rank", "factorizations"], "Method"], [543, 544, ["SVD"], "Method"], [547, 549, ["full", "model"], "Method"], [582, 586, ["deep", "convolutional", "neural", "network"], "Method"], [592, 594, ["predictive", "performance"], "Metric"], [601, 607, ["fully", "connected", "layers", "of", "the", "network"], "Method"], [609, 612, ["Adaptive", "Fastfood", "transform"], "Method"], [619, 621, ["Fastfood", "transform"], "Method"], [622, 624, ["approximating", "kernels"], "Task"], [625, 628, ["Convolutional", "neural", "networks"], "Method"], [629, 632, ["Adaptive", "Fastfood", "transforms"], "Method"], [638, 641, ["deep", "fried", "convnets"], "Method"], [657, 659, ["convolutional", "networks"], "Method"], [660, 661, ["ImageNet"], "Method"], [673, 675, ["kernel", "methods"], "Method"], [676, 678, ["deep", "learning"], "Task"], [680, 684, ["Doubly", "Stochastic", "Gradients", "method"], "Method"], [690, 691, ["randomization"], "Method"], [693, 695, ["kernel", "methods"], "Method"], [718, 720, ["kernel", "classifier"], "Method"], [721, 723, ["convolutional", "filters"], "Method"], [729, 731, ["kernel", "function"], "Method"], [733, 735, ["unsupervised", "manner"], "Method"], [749, 753, ["Network", "in", "Network", "architecture"], "Method"], [762, 765, ["deep", "learning", "benchmarks"], "Task"], [768, 771, ["fully", "connected", "layers"], "Method"], [772, 775, ["global", "average", "pooling"], "Method"], [785, 790, ["ILSVRC", "2014", "object", "detection", "competition"], "Task"], [793, 797, ["global", "average", "pooling", "approach"], "Method"], [809, 811, ["feature", "transfer"], "Task"], [827, 829, ["convolutional", "network"], "Method"], [831, 832, ["ImageNet"], "Material"], [852, 853, ["ImageNet"], "Material"], [860, 863, ["fine", "-", "tuning"], "Task"], [870, 873, ["global", "average", "pooling"], "Method"], [916, 919, ["global", "average", "pooling"], "Method"], [920, 921, ["computation"], "Task"], [922, 924, ["Convolutional", "layers"], "Method"], [937, 940, ["fully", "connected", "layers"], "Method"], [942, 943, ["convolutions"], "Method"], [945, 947, ["model", "size"], "Metric"], [954, 956, ["evaluation", "time"], "Metric"], [978, 980, ["sparse", "networks"], "Method"], [982, 986, ["pruning", "or", "sparsity", "regularizers"], "Method"], [994, 996, ["full", "model"], "Method"], [1005, 1007, ["training", "time"], "Metric"], [1019, 1021, ["hashing", "methods"], "Method"], [1032, 1033, ["Hashes"], "Method"], [1046, 1050, ["GPU", "-", "based", "platforms"], "Task"], [1058, 1059, ["distillation"], "Method"], [1064, 1067, ["compressing", "neural", "networks"], "Method"], [1070, 1074, ["post", "-", "processing", "step"], "Task"], [1078, 1081, ["Adaptive", "Fastfood", "Transform"], "Method"], [1090, 1095, ["fully", "connected", "neural", "network", "layers"], "Method"], [1120, 1124, ["storage", "and", "computational", "costs"], "Metric"], [1126, 1129, ["matrix", "multiplication", "step"], "Method"], [1133, 1135, ["storage", "cost"], "Metric"], [1156, 1159, ["Adaptive", "Fastfood", "transform"], "Method"], [1186, 1188, ["storage", "requirements"], "Metric"], [1190, 1191, ["reparameterization"], "Method"], [1194, 1196, ["computational", "cost"], "Metric"], [1224, 1226, ["prediction", "errors"], "Metric"], [1237, 1239, ["component", "modules"], "Method"], [1241, 1244, ["Adaptive", "Fastfood", "transform"], "Method"], [1256, 1258, ["Adaptive", "Fastfood"], "Method"], [1274, 1279, ["non", "-", "adaptive", "Fastfood", "formulation"], "Method"], [1294, 1298, ["computational", "and", "storage", "costs"], "Metric"], [1303, 1306, ["random", "permutation", "matrix"], "Method"], [1318, 1322, ["storage", "and", "computational", "costs"], "Metric"], [1327, 1331, ["Walsh", "-", "Hadamard", "matrix"], "Method"], [1338, 1341, ["Fast", "Hadamard", "Transform"], "Method"], [1345, 1348, ["Fast", "Fourier", "Transform"], "Method"], [1361, 1363, ["storage", "cost"], "Metric"], [1365, 1368, ["Adaptive", "Fastfood", "transform"], "Method"], [1372, 1374, ["computational", "cost"], "Metric"], [1386, 1389, ["fully", "connected", "layers"], "Method"], [1407, 1410, ["Adaptive", "Fastfood", "transforms"], "Method"], [1424, 1428, ["computational", "and", "storage", "costs"], "Metric"], [1440, 1442, ["linear", "modules"], "Method"], [1457, 1458, ["Fastfood"], "Task"], [1459, 1460, ["backpropagation"], "Method"], [1464, 1467, ["Adaptive", "Fastfood", "transform"], "Method"], [1475, 1478, ["error", "derivative", "backpropagation"], "Method"], [1491, 1494, ["Fast", "Hadamard", "Transform"], "Method"], [1503, 1509, ["-", "th", "layer", "of", "the", "network"], "Method"], [1523, 1524, ["backpropagation"], "Method"], [1603, 1605, ["Hadamard", "transform"], "Method"], [1647, 1650, ["diagonal", "matrix", "multiplications"], "Method"], [1659, 1662, ["Adaptive", "Fastfood", "transform"], "Method"], [1671, 1674, ["structured", "random", "projection"], "Method"], [1694, 1699, ["Adaptive", "Fastfood", "and", "competing", "techniques"], "Method"], [1711, 1714, ["computation", "and", "memory"], "Task"], [1715, 1717, ["neural", "networks"], "Method"], [1723, 1726, ["structured", "random", "projections"], "Method"], [1726, 1728, ["Adaptive", "Fastfood"], "Method"], [1732, 1734, ["Fastfood", "transform"], "Method"], [1761, 1763, ["adaptive", "versions"], "Method"], [1764, 1765, ["Fastfood"], "Method"], [1769, 1772, ["fully", "connected", "layers"], "Method"], [1773, 1776, ["convolutional", "neural", "networks"], "Method"], [1782, 1784, ["neural", "networks"], "Method"], [1793, 1796, ["neural", "networks", "field"], "Task"], [1811, 1814, ["theoretical", "computer", "science"], "Task"], [1817, 1819, ["random", "projection"], "Method"], [1860, 1864, ["Johnson", "-", "Lindenstrauss", "Lemma"], "Method"], [1964, 1971, ["storage", "and", "computational", "costs", "of", "random", "projections"], "Task"], [1974, 1977, ["random", "hash", "functions"], "Method"], [1980, 1983, ["random", "matrix", "multiplication"], "Method"], [1988, 1992, ["count", "-", "sketch", "algorithm"], "Method"], [1993, 1997, ["pairwise", "independent", "hash", "functions"], "Method"], [2019, 2022, ["machine", "learning", "literature"], "Task"], [2023, 2024, ["Hashes"], "Method"], [2112, 2114, ["PHD", "transform"], "Method"], [2145, 2147, ["Hadamard", "transform"], "Method"], [2173, 2175, ["Fastfood", "transform"], "Method"], [2181, 2182, ["Fastfood"], "Method"], [2218, 2221, ["Adaptive", "Fastfood", "transform"], "Method"], [2228, 2229, ["backpropagation"], "Method"], [2237, 2240, ["Automatic", "Relevance", "Determination"], "Task"], [2253, 2255, ["spectral", "incoherence"], "Method"], [2268, 2270, ["RBF", "kernel"], "Method"], [2271, 2275, ["Chi", "-", "squared", "distribution"], "Method"], [2338, 2339, ["kernels"], "Method"], [2347, 2350, ["neural", "network", "modules"], "Method"], [2351, 2352, ["kernels"], "Method"], [2358, 2360, ["computational", "reasons"], "Task"], [2371, 2372, ["kernel"], "Method"], [2419, 2421, ["statistical", "methods"], "Method"], [2422, 2423, ["kernels"], "Method"], [2437, 2439, ["neural", "networks"], "Method"], [2456, 2457, ["kernels"], "Method"], [2473, 2476, ["infinite", "kernel", "expansions"], "Method"], [2489, 2493, ["shift", "-", "invariant", "kernels"], "Method"], [2500, 2502, ["harmonic", "analysis"], "Method"], [2505, 2508, ["Bochner", "’s", "Lemma"], "Method"], [2575, 2578, ["Monte", "Carlo", "methods"], "Method"], [2588, 2589, ["kernel"], "Method"], [2591, 2599, ["inner", "product", "of", "stacked", "cosine", "and", "sine", "features"], "Method"], [2622, 2625, ["inner", "-", "product"], "Method"], [2636, 2639, ["neural", "network", "module"], "Method"], [2643, 2645, ["linear", "layer"], "Method"], [2682, 2684, ["sampling", "distribution"], "Method"], [2693, 2694, ["kernels"], "Method"], [2707, 2710, ["squared", "exponential", "kernel"], "Method"], [2733, 2735, ["Gaussian", "distribution"], "Method"], [2746, 2748, ["neural", "module"], "Method"], [2754, 2757, ["squared", "exponential", "kernel"], "Method"], [2772, 2778, ["rotationally", "invariant", "arc", "-", "cosine", "kernel"], "Method"], [2804, 2807, ["Rectified", "Linear", "Unit"], "Method"], [2808, 2809, ["ReLU"], "Method"], [2813, 2815, ["Fastfood", "transform"], "Method"], [2829, 2833, ["computational", "and", "storage", "costs"], "Metric"], [2836, 2840, ["Deep", "Fried", "Convolutional", "Networks"], "Method"], [2851, 2854, ["fully", "connected", "layers"], "Method"], [2859, 2862, ["Adaptive", "Fastfood", "transform"], "Method"], [2873, 2877, ["deep", "fried", "convolutional", "network"], "Method"], [2899, 2902, ["Adaptive", "Fastfood", "transform"], "Method"], [2904, 2906, ["softmax", "classifier"], "Method"], [2911, 2913, ["memory", "cost"], "Metric"], [2926, 2931, ["low", "-", "rank", "matrix", "factorization"], "Method"], [2935, 2936, ["training"], "Task"], [2942, 2944, ["softmax", "layer"], "Method"], [2954, 2957, ["low", "rank", "factorization"], "Method"], [3038, 3041, ["Adaptive", "Fastfood", "transform"], "Method"], [3052, 3059, ["Caffe", "implementation", "of", "the", "LeNet", "convolutional", "network"], "Method"], [3063, 3065, ["error", "rate"], "Metric"], [3078, 3081, ["deep", "fried", "network"], "Method"], [3083, 3085, ["convolutional", "layers"], "Method"], [3093, 3100, ["adaptive", "and", "non", "-", "adaptive", "Fastfood", "transforms"], "Method"], [3110, 3112, ["adaptive", "transforms"], "Method"], [3135, 3137, ["adaptive", "variant"], "Method"], [3142, 3143, ["backpropagation"], "Method"], [3170, 3173, ["deep", "fried", "network"], "Method"], [3186, 3187, ["dropout"], "Method"], [3196, 3199, ["deep", "fried", "case"], "Task"], [3200, 3203, ["Deep", "fried", "networks"], "Method"], [3208, 3209, ["accuracy"], "Metric"], [3238, 3239, ["adaptation"], "Method"], [3247, 3249, ["adaptive", "models"], "Method"], [3278, 3280, ["Imagenet", "Experiments"], "Task"], [3284, 3287, ["deep", "fried", "networks"], "Method"], [3308, 3313, ["ImageNet", "ILSVRC", "-", "2012", "dataset"], "Material"], [3333, 3336, ["Caffe", "ImageNet", "model"], "Method"], [3338, 3340, ["reference", "model"], "Method"], [3351, 3352, ["AlexNet"], "Method"], [3355, 3359, ["top", "-", "1", "error"], "Metric"], [3361, 3366, ["ILSVRC", "-", "2012", "validation", "set"], "Metric"], [3375, 3381, ["cascade", "of", "convolution", "and", "pooling", "layers"], "Method"], [3382, 3384, ["interspersed", "normalization"], "Method"], [3397, 3398, ["MLP"], "Method"], [3401, 3403, ["9216–4096–4096–1000", "architecture"], "Method"], [3409, 3412, ["logistic", "regression", "layer"], "Method"], [3428, 3429, ["dropout"], "Method"], [3476, 3478, ["convolutional", "layers"], "Method"], [3482, 3485, ["fully", "connected", "layer"], "Method"], [3498, 3501, ["Adaptive", "Fastfood", "transform"], "Method"], [3517, 3520, ["Fixed", "feature", "extractor"], "Method"], [3524, 3526, ["kernel", "methods"], "Method"], [3527, 3528, ["ImageNet"], "Task"], [3538, 3540, ["convolutional", "layers"], "Method"], [3542, 3546, ["pre", "-", "trained", "network"], "Method"], [3567, 3569, ["convolutional", "network"], "Method"], [3573, 3574, ["ImageNet"], "Material"], [3578, 3580, ["feature", "extractor"], "Method"], [3590, 3593, ["Adaptive", "Fastfood", "transform"], "Method"], [3605, 3607, ["convolutional", "layer"], "Method"], [3612, 3614, ["reference", "model"], "Method"], [3617, 3621, ["Adaptive", "Fastfood", "transform", "classifier"], "Method"], [3643, 3645, ["Fastfood", "transform"], "Method"], [3674, 3676, ["Fastfood", "transform"], "Method"], [3678, 3680, ["composite", "module"], "Method"], [3684, 3685, ["dropout"], "Method"], [3711, 3712, ["dropout"], "Method"], [3715, 3717, ["convolutional", "layer"], "Method"], [3730, 3731, ["MLP"], "Method"], [3756, 3760, ["re", "-", "trained", "MLP"], "Method"], [3763, 3767, ["jointly", "trained", "reference", "model"], "Method"], [3783, 3785, ["joint", "training"], "Method"], [3819, 3822, ["jointly", "trained", "network"], "Method"], [3825, 3828, ["deep", "fried", "networks"], "Method"], [3834, 3839, ["re", "-", "trained", "MLP", "model"], "Method"], [3862, 3865, ["Adaptive", "Fastfood", "transform"], "Method"], [3874, 3876, ["adaptive", "version"], "Method"], [3890, 3893, ["Jointly", "trained", "model"], "Method"], [3898, 3901, ["deep", "fried", "network"], "Method"], [3904, 3905, ["ImageNet"], "Material"], [3913, 3915, ["Fastfood", "layer"], "Method"], [3921, 3926, ["top", "-", "1", "validation", "performance"], "Metric"], [4002, 4005, ["deep", "fried", "network"], "Method"], [4009, 4012, ["softmax", "regression", "layer"], "Method"], [4047, 4048, ["ImageNet"], "Material"], [4059, 4063, ["deep", "fried", "convolutional", "networks"], "Method"], [4100, 4103, ["deep", "fried", "network"], "Method"], [4106, 4108, ["memory", "consumption"], "Metric"], [4116, 4121, ["low", "-", "rank", "matrix", "factorization"], "Task"], [4130, 4132, ["Post", "Processing"], "Method"], [4151, 4154, ["convolutional", "neural", "network"], "Method"], [4162, 4166, ["post", "-", "processing", "techniques"], "Method"], [4171, 4174, ["full", "trained", "model"], "Method"], [4185, 4187, ["compressed", "network"], "Method"], [4190, 4192, ["Matrix", "factorization"], "Method"], [4198, 4201, ["compressing", "neural", "networks"], "Task"], [4291, 4293, ["factorized", "representation"], "Method"], [4309, 4312, ["logistic", "regression", "layer"], "Method"], [4319, 4321, ["full", "model"], "Method"], [4325, 4330, ["SVD", "of", "the", "weight", "matrices"], "Method"], [4333, 4336, ["fine", "tuning", "phase"], "Method"], [4349, 4352, ["deep", "fried", "approach"], "Method"], [4353, 4354, ["SVD"], "Method"], [4356, 4358, ["fine", "tuning"], "Method"], [4396, 4399, ["deep", "fried", "convnet"], "Method"], [4402, 4407, ["SVD", "plus", "fine", "-", "tuning"], "Method"], [4410, 4412, ["softmax", "layer"], "Method"], [4424, 4427, ["post", "-", "processing"], "Task"], [4438, 4440, ["SVD", "decomposition"], "Method"], [4450, 4452, ["reference", "model"], "Method"], [4464, 4469, ["SVD", "-", "half", "-", "F"], "Method"], [4470, 4473, ["SVD", "-", "quarter"], "Method"], [4493, 4494, ["accuracy"], "Metric"], [4495, 4496, ["SVD"], "Method"], [4503, 4506, ["SVD", "-", "quarter"], "Method"], [4513, 4514, ["error"], "Metric"], [4516, 4517, ["SVD"], "Method"], [4539, 4544, ["SVD", "-", "quarter", "-", "F"], "Method"], [4546, 4549, ["deep", "fried", "convnets"], "Method"], [4557, 4558, ["accuracy"], "Metric"], [4566, 4569, ["rank", "600", "SVD"], "Method"], [4571, 4573, ["fine", "tuning"], "Method"], [4576, 4578, ["softmax", "layer"], "Method"], [4580, 4584, ["Adaptive", "Fastfood", "32", "model"], "Method"], [4596, 4600, ["top", "-", "1", "error"], "Metric"], [4610, 4611, ["Collins"], "Method"], [4612, 4613, ["Kohli"], "Method"], [4619, 4621, ["full", "network"], "Method"], [4624, 4626, ["sparsity", "regularizer"], "Method"], [4627, 4630, ["fine", "-", "tuning"], "Method"], [4663, 4665, ["compressed", "network"], "Method"], [4681, 4684, ["sparse", "weight", "matrices"], "Method"], [4699, 4701, ["index", "storage"], "Method"], [4706, 4708, ["sparse", "representation"], "Method"], [4737, 4739, ["convolutional", "networks"], "Method"], [4750, 4753, ["Adaptive", "Fastfood", "transform"], "Method"], [4782, 4785, ["Deep", "fried", "convnets"], "Method"], [4789, 4792, ["Adaptive", "Fastfood", "transform"], "Method"], [4804, 4806, ["predictive", "performance"], "Metric"], [4807, 4808, ["MNIST"], "Material"], [4809, 4810, ["ImageNet"], "Material"], [4817, 4826, ["test", "-", "time", "low", "-", "rank", "matrix", "factorization", "schemes"], "Method"], [4844, 4847, ["structured", "random", "transformations"], "Method"], [4850, 4852, ["kernel", "literature"], "Method"], [4856, 4857, ["MNIST"], "Material"], [4859, 4860, ["learning"], "Method"], [4866, 4867, ["ImageNet"], "Material"], [4871, 4872, ["adaptation"], "Task"], [4895, 4896, ["learning"], "Task"], [4911, 4913, ["Fastfood", "transform"], "Method"], [4916, 4920, ["theoretical", "reduction", "in", "computation"], "Task"], [4926, 4927, ["computation"], "Task"], [4928, 4931, ["convolutional", "neural", "networks"], "Method"], [4935, 4936, ["convolutions"], "Method"], [4939, 4942, ["deep", "fried", "convnets"], "Method"], [4957, 4958, ["ImageNet"], "Material"], [4970, 4972, ["softmax", "layer"], "Method"], [4994, 4995, ["SVD"], "Method"], [5040, 5044, ["low", "-", "rank", "decomposition"], "Method"], [5045, 5047, ["Adaptive", "Fastfood"], "Method"], [5049, 5050, ["pruning"], "Method"], [5054, 5058, ["GPU", "optimized", "Fastfood", "transforms"], "Method"], [5067, 5070, ["arbitrary", "neural", "models"], "Method"], [2990, 2991, ["MNIST"], "Material"], [3000, 3001, ["MNIST"], "Material"], [3068, 3069, ["MNIST"], "Material"], [3155, 3156, ["MNIST"], "Material"], [3424, 3425, ["ReLU"], "Method"], [3854, 3855, ["MNIST"], "Material"]]} {"doc_id": "27aa0f3ec934925265f93fac7ff1cd1d70ceb618", "text": ["Strong", "Baselines", "for", "Neural", "Semi", "-", "supervised", "Learning", "under", "Domain", "Shift", "section", ":", "Abstract", "Novel", "neural", "models", "have", "been", "proposed", "in", "recent", "years", "for", "learning", "under", "domain", "shift", ".", "Most", "models", ",", "however", ",", "only", "evaluate", "on", "a", "single", "task", ",", "on", "proprietary", "datasets", ",", "or", "compare", "to", "weak", "baselines", ",", "which", "makes", "comparison", "of", "models", "difficult", ".", "In", "this", "paper", ",", "we", "re", "-", "evaluate", "classic", "general", "-", "purpose", "bootstrapping", "approaches", "in", "the", "context", "of", "neural", "networks", "under", "domain", "shifts", "vs.", "recent", "neural", "approaches", "and", "propose", "a", "novel", "multi", "-", "task", "tri", "-", "training", "method", "that", "reduces", "the", "time", "and", "space", "complexity", "of", "classic", "tri", "-", "training", ".", "Extensive", "experiments", "on", "two", "benchmarks", "are", "negative", ":", "while", "our", "novel", "method", "establishes", "a", "new", "state", "-", "of", "-", "the", "-", "art", "for", "sentiment", "analysis", ",", "it", "does", "not", "fare", "consistently", "the", "best", ".", "More", "importantly", ",", "we", "arrive", "at", "the", "somewhat", "surprising", "conclusion", "that", "classic", "tri", "-", "training", ",", "with", "some", "additions", ",", "outperforms", "the", "state", "of", "the", "art", ".", "We", "conclude", "that", "classic", "approaches", "constitute", "an", "important", "and", "strong", "baseline", ".", "section", ":", "Introduction", "Deep", "neural", "networks", "(", "DNNs", ")", "excel", "at", "learning", "from", "labeled", "data", "and", "have", "achieved", "state", "of", "the", "art", "in", "a", "wide", "array", "of", "supervised", "NLP", "tasks", "such", "as", "dependency", "parsing", "[", "reference", "]", ",", "named", "entity", "recognition", "[", "reference", "]", ",", "and", "semantic", "role", "labeling", "[", "reference", "]", ".", "In", "contrast", ",", "learning", "from", "unlabeled", "data", ",", "especially", "under", "domain", "shift", ",", "remains", "a", "challenge", ".", "This", "is", "common", "in", "many", "real", "-", "world", "applications", "where", "the", "distribution", "of", "the", "training", "and", "test", "data", "differs", ".", "Many", "state", "-", "of", "-", "the", "-", "art", "domain", "adaptation", "approaches", "leverage", "task", "-", "specific", "characteristics", "such", "as", "sentiment", "words", "[", "reference", "][", "reference", "]", "or", "distributional", "features", "[", "reference", "][", "reference", "]", "which", "do", "not", "generalize", "to", "other", "tasks", ".", "Other", "approaches", "that", "are", "in", "theory", "more", "general", "only", "evaluate", "on", "proprietary", "datasets", "[", "reference", "]", "or", "on", "a", "single", "benchmark", "[", "reference", "]", ",", "which", "carries", "the", "risk", "of", "overfitting", "to", "the", "task", ".", "In", "addition", ",", "most", "models", "only", "compare", "against", "weak", "baselines", "and", ",", "strikingly", ",", "almost", "none", "considers", "evaluating", "against", "approaches", "from", "the", "extensive", "semi", "-", "supervised", "learning", "(", "SSL", ")", "literature", "[", "reference", "]", ".", "In", "this", "work", ",", "we", "make", "the", "argument", "that", "such", "algorithms", "make", "strong", "baselines", "for", "any", "task", "in", "line", "with", "recent", "efforts", "highlighting", "the", "usefulness", "of", "classic", "approaches", "[", "reference", "][", "reference", "]", ".", "We", "re", "-", "evaluate", "bootstrapping", "algorithms", "in", "the", "context", "of", "DNNs", ".", "These", "are", "general", "-", "purpose", "semi", "-", "supervised", "algorithms", "that", "treat", "the", "model", "as", "a", "black", "box", "and", "can", "thus", "be", "used", "easily", "-", "with", "a", "few", "additions", "-", "with", "the", "current", "generation", "of", "NLP", "models", ".", "Many", "of", "these", "methods", ",", "though", ",", "were", "originally", "developed", "with", "in", "-", "domain", "performance", "in", "mind", ",", "so", "their", "effectiveness", "in", "a", "domain", "adaptation", "setting", "remains", "unexplored", ".", "In", "particular", ",", "we", "re", "-", "evaluate", "three", "traditional", "bootstrapping", "methods", ",", "self", "-", "training", "[", "reference", "]", ",", "tri", "-", "training", "[", "reference", "]", ",", "and", "tritraining", "with", "disagreement", "[", "reference", "]", "for", "neural", "network", "-", "based", "approaches", "on", "two", "NLP", "tasks", "with", "different", "characteristics", ",", "namely", ",", "a", "sequence", "prediction", "and", "a", "classification", "task", "(", "POS", "tagging", "and", "sentiment", "analysis", ")", ".", "We", "evaluate", "the", "methods", "across", "multiple", "domains", "on", "two", "wellestablished", "benchmarks", ",", "without", "taking", "any", "further", "task", "-", "specific", "measures", ",", "and", "compare", "to", "the", "best", "results", "published", "in", "the", "literature", ".", "We", "make", "the", "somewhat", "surprising", "observation", "that", "classic", "tri", "-", "training", "outperforms", "task", "-", "agnostic", "state", "-", "of", "-", "the", "-", "art", "semi", "-", "supervised", "learning", "[", "reference", "]", "and", "recent", "neural", "adaptation", "approaches", "[", "reference", "][", "reference", "]", ".", "In", "addition", ",", "we", "propose", "multi", "-", "task", "tri", "-", "training", ",", "which", "reduces", "the", "main", "deficiency", "of", "tri", "-", "training", ",", "namely", "its", "time", "and", "space", "complexity", ".", "It", "establishes", "a", "new", "state", "of", "the", "art", "on", "unsupervised", "domain", "adaptation", "for", "sentiment", "analysis", "but", "it", "is", "outperformed", "by", "classic", "tri", "-", "training", "for", "POS", "tagging", ".", "Contributions", "Our", "contributions", "are", ":", "a", ")", "We", "propose", "a", "novel", "multi", "-", "task", "tri", "-", "training", "method", ".", "b", ")", "We", "show", "that", "tri", "-", "training", "can", "serve", "as", "a", "strong", "and", "robust", "semi", "-", "supervised", "learning", "baseline", "for", "the", "current", "generation", "of", "NLP", "models", ".", "c", ")", "We", "perform", "an", "extensive", "evaluation", "of", "bootstrapping", "1", "algorithms", "compared", "to", "state", "-", "of", "-", "the", "-", "art", "approaches", "on", "two", "benchmark", "datasets", ".", "d", ")", "We", "shed", "light", "on", "the", "task", "and", "data", "characteristics", "that", "yield", "the", "best", "performance", "for", "each", "model", ".", "section", ":", "Neural", "bootstrapping", "methods", "We", "first", "introduce", "three", "classic", "bootstrapping", "methods", ",", "self", "-", "training", ",", "tri", "-", "training", ",", "and", "tri", "-", "training", "with", "disagreement", "and", "detail", "how", "they", "can", "be", "used", "with", "neural", "networks", ".", "For", "in", "-", "depth", "details", "we", "refer", "the", "reader", "to", "[", "reference", "][", "reference", "][", "reference", "]", ")", ".", "We", "introduce", "our", "novel", "multitask", "tri", "-", "training", "method", "in", "§", "2.3", ".", "section", ":", "Self", "-", "training", "Self", "-", "training", "[", "reference", "][", "reference", "]", ")", "is", "one", "of", "the", "earliest", "and", "simplest", "bootstrapping", "approaches", ".", "In", "essence", ",", "it", "leverages", "the", "model", "'s", "own", "predictions", "on", "unlabeled", "data", "to", "obtain", "additional", "information", "that", "can", "be", "used", "during", "training", ".", "Typically", "the", "most", "confident", "predictions", "are", "taken", "at", "face", "value", ",", "as", "detailed", "next", ".", "Self", "-", "training", "trains", "a", "model", "m", "on", "a", "labeled", "training", "set", "L", "and", "an", "unlabeled", "data", "set", "U", ".", "At", "each", "iteration", ",", "the", "model", "provides", "predictions", "m", "(", "x", ")", "in", "the", "form", "of", "a", "probability", "distribution", "over", "classes", "for", "all", "unlabeled", "examples", "x", "in", "U", ".", "If", "the", "probability", "assigned", "to", "the", "most", "likely", "class", "is", "higher", "than", "a", "predetermined", "threshold", "τ", ",", "x", "is", "added", "to", "the", "labeled", "examples", "with", "p", "(", "x", ")", "=", "arg", "max", "m", "(", "x", ")", "as", "pseudo", "-", "label", ".", "This", "instantiation", "is", "the", "most", "widely", "used", "and", "shown", "in", "Algorithm", "1", ".", "Calibration", "It", "is", "well", "-", "known", "that", "output", "probabilities", "in", "neural", "networks", "are", "poorly", "calibrated", "[", "reference", "]", ".", "Using", "a", "fixed", "threshold", "τ", "is", "thus", "Algorithm", "1", "Self", "-", "training", "[", "reference", "]", "L", "←", "L", "∪", "{", "(", "x", ",", "p", "(", "x", ")", ")", "}", "6", ":", "until", "no", "more", "predictions", "are", "confident", "not", "the", "best", "choice", ".", "While", "the", "absolute", "confidence", "value", "is", "inaccurate", ",", "we", "can", "expect", "that", "the", "relative", "order", "of", "confidences", "is", "more", "robust", ".", "For", "this", "reason", ",", "we", "select", "the", "top", "n", "unlabeled", "examples", "that", "have", "been", "predicted", "with", "the", "highest", "confidence", "after", "every", "epoch", "and", "add", "them", "to", "the", "labeled", "data", ".", "This", "is", "one", "of", "the", "many", "variants", "for", "self", "-", "training", ",", "called", "throttling", "[", "reference", "]", ".", "We", "empirically", "confirm", "that", "this", "outperforms", "the", "classic", "selection", "in", "our", "experiments", ".", "Online", "learning", "In", "contrast", "to", "many", "classic", "algorithms", ",", "DNNs", "are", "trained", "online", "by", "default", ".", "We", "compare", "training", "setups", "and", "find", "that", "training", "until", "convergence", "on", "labeled", "data", "and", "then", "training", "until", "convergence", "using", "self", "-", "training", "performs", "best", ".", "Classic", "self", "-", "training", "has", "shown", "mixed", "success", ".", "In", "parsing", "it", "proved", "successful", "only", "with", "small", "datasets", "[", "reference", "]", "or", "when", "a", "generative", "component", "is", "used", "together", "with", "a", "reranker", "in", "high", "-", "data", "conditions", "[", "reference", "][", "reference", "]", ".", "Some", "success", "was", "achieved", "with", "careful", "task", "-", "specific", "data", "selection", "[", "reference", "]", ",", "while", "others", "report", "limited", "success", "on", "a", "variety", "of", "NLP", "tasks", "[", "reference", "][", "reference", "][", "reference", "]", ".", "Its", "main", "downside", "is", "that", "the", "model", "is", "not", "able", "to", "correct", "its", "own", "mistakes", "and", "errors", "are", "amplified", ",", "an", "effect", "that", "is", "increased", "under", "domain", "shift", ".", "section", ":", "Tri", "-", "training", "Tri", "-", "training", "[", "reference", "]", ")", "is", "a", "classic", "method", "that", "reduces", "the", "bias", "of", "predictions", "on", "unlabeled", "data", "by", "utilizing", "the", "agreement", "of", "three", "independently", "trained", "models", ".", "Tri", "-", "training", "(", "cf", ".", "Algorithm", "2", ")", "first", "trains", "three", "models", "m", "1", ",", "m", "2", ",", "and", "m", "3", "on", "bootstrap", "samples", "of", "the", "labeled", "data", "L.", "An", "unlabeled", "data", "point", "is", "added", "to", "the", "training", "set", "of", "a", "model", "m", "i", "if", "the", "other", "two", "models", "m", "j", "and", "m", "k", "agree", "on", "its", "label", ".", "Training", "stops", "when", "the", "classifiers", "do", "not", "change", "anymore", ".", "Tri", "-", "training", "with", "disagreement", "[", "reference", "]", "Algorithm", "2", "Tri", "-", "training", "[", "reference", "]", "1", ":", "for", "i", "∈", "{", "1", "..", "3", "}", "do", "2", ":", "for", "i", "∈", "{", "1", "..", "3", "}", "do", "6", ":", "for", "x", "∈", "U", "do", "8", ":", "10", ":", "until", "none", "of", "m", "i", "changes", "11", ":", "apply", "majority", "vote", "over", "m", "i", "is", "based", "on", "the", "intuition", "that", "a", "model", "should", "only", "be", "strengthened", "in", "its", "weak", "points", "and", "that", "the", "labeled", "data", "should", "not", "be", "skewed", "by", "easy", "data", "points", ".", "In", "order", "to", "achieve", "this", ",", "it", "adds", "a", "simple", "modification", "to", "the", "original", "algorithm", "(", "altering", "line", "8", "in", "Algorithm", "2", ")", ",", "requiring", "that", "for", "an", "unlabeled", "data", "point", "on", "which", "m", "j", "and", "m", "k", "agree", ",", "the", "other", "model", "m", "i", "disagrees", "on", "the", "prediction", ".", "Tri", "-", "training", "with", "disagreement", "is", "more", "data", "-", "efficient", "than", "tritraining", "and", "has", "achieved", "competitive", "results", "on", "part", "-", "of", "-", "speech", "tagging", "[", "reference", "]", ".", "Sampling", "unlabeled", "data", "Both", "tri", "-", "training", "and", "tri", "-", "training", "with", "disagreement", "can", "be", "very", "expensive", "in", "their", "original", "formulation", "as", "they", "require", "to", "produce", "predictions", "for", "each", "of", "the", "three", "models", "on", "all", "unlabeled", "data", "samples", ",", "which", "can", "be", "in", "the", "millions", "in", "realistic", "applications", ".", "We", "thus", "propose", "to", "sample", "a", "number", "of", "unlabeled", "examples", "at", "every", "epoch", ".", "For", "all", "traditional", "bootstrapping", "approaches", "we", "sample", "10k", "candidate", "instances", "in", "each", "epoch", ".", "For", "the", "neural", "approaches", "we", "use", "a", "linearly", "growing", "candidate", "sampling", "scheme", "proposed", "by", "[", "reference", "]", ",", "increasing", "the", "candidate", "pool", "size", "as", "the", "models", "become", "more", "accurate", ".", "Confidence", "thresholding", "Similar", "to", "selftraining", ",", "we", "can", "introduce", "an", "additional", "requirement", "that", "pseudo", "-", "labeled", "examples", "are", "only", "added", "if", "the", "probability", "of", "the", "prediction", "of", "at", "least", "one", "model", "is", "higher", "than", "some", "threshold", "τ", ".", "We", "did", "not", "find", "this", "to", "outperform", "prediction", "without", "threshold", "for", "traditional", "tri", "-", "training", ",", "but", "thresholding", "proved", "essential", "for", "our", "method", "(", "§", "2.3", ")", ".", "The", "most", "important", "condition", "for", "tri", "-", "training", "and", "tri", "-", "training", "with", "disagreement", "is", "that", "the", "models", "are", "diverse", ".", "Typically", ",", "bootstrap", "samples", "are", "used", "to", "create", "this", "diversity", "[", "reference", "][", "reference", "]", ".", "However", ",", "training", "separate", "models", "on", "bootstrap", "samples", "of", "a", "potentially", "large", "amount", "of", "training", "data", "is", "expensive", "and", "takes", "a", "lot", "of", "time", ".", "This", "drawback", "motivates", "our", "approach", ".", "section", ":", "Multi", "-", "task", "tri", "-", "training", "In", "order", "to", "reduce", "both", "the", "time", "and", "space", "complexity", "of", "tri", "-", "training", ",", "we", "propose", "Multi", "-", "task", "Tritraining", "(", "MT", "-", "Tri", ")", ".", "MT", "-", "Tri", "leverages", "insights", "from", "multi", "-", "task", "learning", "(", "MTL", ")", "[", "reference", "]", "to", "share", "knowledge", "across", "models", "and", "accelerate", "training", ".", "Rather", "than", "storing", "and", "training", "each", "model", "separately", ",", "we", "propose", "to", "share", "the", "parameters", "of", "the", "models", "and", "train", "them", "jointly", "using", "MTL", ".", "2", "All", "models", "thus", "collaborate", "on", "learning", "a", "joint", "representation", ",", "which", "improves", "convergence", ".", "The", "output", "softmax", "layers", "are", "model", "-", "specific", "and", "are", "only", "updated", "for", "the", "input", "of", "the", "respective", "model", ".", "We", "show", "the", "model", "in", "Figure", "1", "(", "as", "instantiated", "for", "POS", "tagging", ")", ".", "As", "the", "models", "leverage", "a", "joint", "representation", ",", "we", "need", "to", "ensure", "that", "the", "features", "used", "for", "prediction", "in", "the", "softmax", "layers", "of", "the", "different", "models", "are", "as", "diverse", "as", "possible", ",", "so", "that", "the", "models", "can", "still", "learn", "from", "each", "other", "'s", "predictions", ".", "In", "contrast", ",", "if", "the", "parameters", "in", "all", "output", "softmax", "layers", "were", "the", "same", ",", "the", "method", "would", "degenerate", "to", "self", "-", "training", ".", "To", "guarantee", "diversity", ",", "we", "introduce", "an", "orthogonality", "constraint", "[", "reference", "]", "as", "an", "additional", "loss", "term", ",", "which", "we", "define", "as", "follows", ":", "where", "|", "·", "2", "F", "is", "the", "squared", "Frobenius", "norm", "and", "W", "m", "1", "and", "W", "m", "2", "are", "the", "softmax", "output", "parameters", "of", "the", "two", "source", "and", "pseudo", "-", "labeled", "output", "layers", "m", "1", "and", "m", "2", ",", "respectively", ".", "The", "orthogonality", "constraint", "encourages", "the", "models", "not", "to", "rely", "on", "the", "same", "features", "for", "prediction", ".", "As", "enforcing", "pairwise", "orthogonality", "between", "three", "matrices", "is", "not", "possible", ",", "we", "only", "enforce", "orthogonality", "between", "the", "softmax", "output", "layers", "of", "m", "1", "and", "m", "2", ",", "3", "while", "m", "3", "is", "gradually", "trained", "to", "be", "more", "target", "-", "specific", ".", "We", "parameterize", "L", "orth", "by", "γ=0.01", "following", ".", "We", "do", "not", "further", "tune", "γ", ".", "More", "formally", ",", "let", "us", "illustrate", "the", "model", "by", "taking", "the", "sequence", "prediction", "task", "(", "Figure", "1", ")", "as", "illustration", ".", "Given", "an", "utterance", "with", "labels", "y", "1", ",", "..", ",", "y", "n", ",", "our", "Multi", "-", "task", "Tri", "-", "training", "loss", "consists", "of", "three", "task", "-", "specific", "(", "m", "1", ",", "m", "2", ",", "m", "3", ")", "tagging", "loss", "functions", "(", "where", "h", "is", "the", "uppermost", "Bi", "-", "LSTM", "encoding", ")", ":", "In", "contrast", "to", "classic", "tri", "-", "training", ",", "we", "can", "train", "the", "multi", "-", "task", "model", "with", "its", "three", "model", "-", "specific", "outputs", "jointly", "and", "without", "bootstrap", "sampling", "on", "the", "labeled", "source", "domain", "data", "until", "convergence", ",", "as", "the", "orthogonality", "constraint", "enforces", "different", "representations", "between", "models", "m", "1", "and", "m", "2", ".", "From", "this", "point", ",", "we", "can", "leverage", "the", "pair", "-", "wise", "agreement", "of", "two", "output", "layers", "to", "add", "pseudo", "-", "labeled", "examples", "as", "training", "data", "to", "the", "third", "model", ".", "We", "train", "the", "third", "output", "layer", "m", "3", "only", "on", "pseudo", "-", "labeled", "target", "instances", "in", "order", "to", "make", "tri", "-", "training", "more", "robust", "to", "a", "domain", "shift", ".", "For", "the", "final", "prediction", ",", "majority", "voting", "of", "all", "three", "output", "layers", "is", "used", ",", "which", "resulted", "in", "the", "best", "instantiation", ",", "together", "with", "confidence", "thresholding", "(", "τ", "=", "0.9", ",", "except", "for", "highresource", "POS", "where", "τ", "=", "0.8", "performed", "slightly", "better", ")", ".", "We", "also", "experimented", "with", "using", "a", "domainadversarial", "loss", "[", "reference", "]", "on", "the", "jointly", "learned", "representation", ",", "but", "found", "this", "not", "to", "help", ".", "The", "full", "pseudo", "-", "code", "is", "given", "in", "Algorithm", "3", ".", "Computational", "complexity", "The", "motivation", "for", "MT", "-", "Tri", "was", "to", "reduce", "the", "space", "and", "time", "complexity", "of", "tri", "-", "training", ".", "We", "thus", "give", "an", "estimate", "of", "its", "efficiency", "gains", ".", "MT", "-", "Tri", "is~3×", "more", "spaceefficient", "than", "regular", "tri", "-", "training", ";", "tri", "-", "training", "stores", "one", "set", "of", "parameters", "for", "each", "of", "the", "three", "models", ",", "while", "MT", "-", "Tri", "only", "stores", "one", "set", "of", "parameters", "(", "we", "use", "three", "output", "layers", ",", "but", "these", "make", "up", "a", "comparatively", "small", "part", "of", "the", "total", "parameter", "budget", ")", ".", "In", "terms", "of", "time", "efficiency", ",", "tri", "-", "training", "first", "[", "reference", "]", "We", "also", "tried", "enforcing", "orthogonality", "on", "a", "hidden", "layer", "rather", "than", "the", "output", "layer", ",", "but", "this", "did", "not", "help", ".", "10", ":", "until", "end", "condition", "is", "met", "11", ":", "apply", "majority", "vote", "over", "m", "i", "requires", "to", "train", "each", "of", "the", "models", "from", "scratch", ".", "The", "actual", "tri", "-", "training", "takes", "about", "the", "same", "time", "as", "training", "from", "scratch", "and", "requires", "a", "separate", "forward", "pass", "for", "each", "model", ",", "effectively", "training", "three", "independent", "models", "simultaneously", ".", "In", "contrast", ",", "MT", "-", "Tri", "only", "necessitates", "one", "forward", "pass", "as", "well", "as", "the", "evaluation", "of", "the", "two", "additional", "output", "layers", "(", "which", "takes", "a", "negligible", "amount", "of", "time", ")", "and", "requires", "about", "as", "many", "epochs", "as", "tri", "-", "training", "until", "convergence", "(", "see", "Table", "3", ",", "second", "column", ")", "while", "adding", "fewer", "unlabeled", "examples", "per", "epoch", "(", "see", "Section", "3.4", ")", ".", "In", "our", "experiments", ",", "MT", "-", "Tri", "trained", "about", "5", "-", "6×", "faster", "than", "traditional", "tri", "-", "training", ".", "MT", "-", "Tri", "can", "be", "seen", "as", "a", "self", "-", "ensembling", "technique", ",", "where", "different", "variations", "of", "a", "model", "are", "used", "to", "create", "a", "stronger", "ensemble", "prediction", ".", "Recent", "approaches", "in", "this", "line", "are", "snapshot", "ensembling", ")", "that", "ensembles", "models", "converged", "to", "different", "minima", "during", "a", "training", "run", ",", "asymmetric", "tri", "-", "training", "[", "reference", "]", ")", "(", "ASYM", ")", "that", "leverages", "agreement", "on", "two", "models", "as", "information", "for", "the", "third", ",", "and", "temporal", "ensembling", "[", "reference", "]", ",", "which", "ensembles", "predictions", "of", "a", "model", "at", "different", "epochs", ".", "We", "tried", "to", "compare", "to", "temporal", "ensembling", "in", "our", "experiments", ",", "but", "were", "not", "able", "to", "obtain", "consistent", "results", ".", "[", "reference", "]", "We", "compare", "to", "the", "closest", "most", "recent", "method", ",", "asymmetric", "tritraining", "[", "reference", "]", ".", "It", "differs", "from", "ours", "in", "two", "aspects", ":", "a", ")", "ASYM", "leverages", "only", "pseudolabels", "from", "data", "points", "on", "which", "m", "1", "and", "m", "2", "agree", ",", "and", "b", ")", "it", "uses", "only", "one", "task", "(", "m", "3", ")", "as", "final", "predictor", ".", "In", "essence", ",", "our", "formulation", "of", "MT", "-", "Tri", "is", "closer", "to", "the", "original", "tri", "-", "training", "formulation", "(", "agreements", "on", "two", "provide", "pseudo", "-", "labels", "to", "the", "third", ")", "thereby", "incorporating", "more", "diversity", ".", "[", "reference", "]", "for", "POS", "tagging", "(", "above", ")", "and", "the", "Amazon", "Reviews", "dataset", "[", "reference", "]", "for", "sentiment", "analysis", "(", "below", ")", ".", "section", ":", "Experiments", "In", "order", "to", "ascertain", "which", "methods", "are", "robust", "across", "different", "domains", ",", "we", "evaluate", "on", "two", "widely", "used", "unsupervised", "domain", "adaptation", "datasets", "for", "two", "tasks", ",", "a", "sequence", "labeling", "and", "a", "classification", "task", ",", "cf", ".", "Table", "1", "for", "data", "statistics", ".", "section", ":", "POS", "tagging", "For", "POS", "tagging", "we", "use", "the", "SANCL", "2012", "shared", "task", "dataset", "[", "reference", "]", "and", "compare", "to", "the", "top", "results", "in", "both", "low", "and", "high", "-", "data", "conditions", "[", "reference", "][", "reference", "]", ".", "Both", "are", "strong", "baselines", ",", "as", "the", "FLORS", "tagger", "has", "been", "developed", "for", "this", "challenging", "dataset", "and", "it", "is", "based", "on", "contextual", "distributional", "features", "(", "excluding", "the", "word", "'s", "identity", ")", ",", "and", "hand", "-", "crafted", "suffix", "and", "shape", "features", "(", "including", "some", "languagespecific", "morphological", "features", ")", ".", "We", "want", "to", "gauge", "to", "what", "extent", "we", "can", "adopt", "a", "nowadays", "fairly", "standard", "(", "but", "more", "lexicalized", ")", "general", "neural", "tagger", ".", "Our", "POS", "tagging", "model", "is", "a", "state", "-", "of", "-", "the", "-", "art", "Bi", "-", "LSTM", "tagger", "[", "reference", "]", "with", "word", "and", "100", "-", "dim", "character", "embeddings", ".", "Word", "embeddings", "are", "initialized", "with", "the", "100", "-", "dim", "Glove", "embeddings", "[", "reference", "]", ".", "The", "BiLSTM", "has", "one", "hidden", "layer", "with", "100", "dimensions", ".", "The", "base", "POS", "model", "is", "trained", "on", "WSJ", "with", "early", "stopping", "on", "the", "WSJ", "development", "set", ",", "using", "patience", "2", ",", "Gaussian", "noise", "with", "σ", "=", "0.2", "and", "word", "dropout", "with", "p", "=", "0.25", "[", "reference", "]", ".", "Regarding", "data", ",", "the", "source", "domain", "is", "the", "Ontonotes", "4.0", "release", "of", "the", "Penn", "treebank", "Wall", "Street", "Journal", "(", "WSJ", ")", "annotated", "for", "48", "fine", "-", "grained", "POS", "tags", ".", "This", "amounts", "to", "30", ",", "060", "labeled", "sentences", ".", "We", "use", "100", ",", "000", "WSJ", "sentences", "from", "1988", "as", "unlabeled", "data", ",", "following", "[", "reference", "]", ".", "[", "reference", "]", "As", "target", "data", ",", "we", "use", "the", "five", "SANCL", "domains", "(", "answers", ",", "emails", ",", "newsgroups", ",", "reviews", ",", "weblogs", ")", ".", "We", "restrict", "the", "amount", "of", "unlabeled", "data", "for", "each", "SANCL", "domain", "to", "the", "first", "100k", "sentences", ",", "and", "do", "not", "do", "any", "pre", "-", "processing", ".", "We", "consider", "the", "development", "set", "of", "ANSWERS", "as", "our", "only", "target", "dev", "set", "to", "set", "hyperparameters", ".", "This", "may", "result", "in", "suboptimal", "per", "-", "domain", "settings", "but", "better", "resembles", "an", "unsupervised", "adaptation", "scenario", ".", "section", ":", "Sentiment", "analysis", "For", "sentiment", "analysis", ",", "we", "evaluate", "on", "the", "Amazon", "reviews", "dataset", "[", "reference", "]", ".", "Reviews", "with", "1", "to", "3", "stars", "are", "ranked", "as", "negative", ",", "while", "reviews", "with", "4", "or", "5", "stars", "are", "ranked", "as", "positive", ".", "The", "dataset", "consists", "of", "four", "domains", ",", "yielding", "12", "adaptation", "scenarios", ".", "We", "use", "the", "same", "pre", "-", "processing", "and", "architecture", "as", "used", "in", "[", "reference", "][", "reference", "]", ":", "5", ",", "000", "-", "dimensional", "tf", "-", "idf", "weighted", "unigram", "and", "bigram", "features", "as", "input", ";", "2k", "labeled", "source", "samples", "and", "2k", "unlabeled", "target", "samples", "for", "training", ",", "200", "labeled", "target", "samples", "for", "validation", ",", "and", "between", "3k", "-", "6k", "samples", "for", "testing", ".", "The", "model", "is", "an", "MLP", "with", "one", "hidden", "layer", "with", "50", "dimensions", ",", "sigmoid", "activations", ",", "and", "a", "softmax", "output", ".", "We", "compare", "against", "the", "Variational", "Fair", "Autoencoder", "(", "VFAE", ")", "[", "reference", "]", "model", "and", "domain", "-", "adversarial", "neural", "networks", "(", "DANN", ")", "[", "reference", "]", ".", "section", ":", "Baselines", "Besides", "comparing", "to", "the", "top", "results", "published", "on", "both", "datasets", ",", "we", "include", "the", "following", "baselines", ":", "a", ")", "the", "task", "model", "trained", "on", "the", "source", "domain", ";", "b", ")", "self", "-", "training", "(", "Self", ")", ";", "c", ")", "tri", "-", "training", "(", "Tri", ")", ";", "d", ")", "tri", "-", "training", "with", "disagreement", "(", "Tri", "-", "D", ")", ";", "and", "e", ")", "asymmetric", "tri", "-", "training", "[", "reference", "]", ")", ".", "Our", "proposed", "model", "is", "multi", "-", "task", "tri", "-", "training", "(", "MTTri", ")", ".", "We", "implement", "our", "models", "in", "DyNet", ".", "Reporting", "single", "evaluation", "scores", "might", "result", "in", "biased", "results", "[", "reference", "]", ".", "Throughout", "the", "paper", ",", "we", "report", "mean", "accuracy", "and", "standard", "deviation", "over", "five", "runs", "for", "POS", "tagging", "and", "over", "ten", "runs", "for", "sentiment", "analysis", ".", "Significance", "is", "computed", "using", "bootstrap", "test", ".", "The", "code", "for", "all", "experiments", "is", "released", "at", ":", "https:", "//", "github.com", "/", "bplank", "/", "semi", "-", "supervised", "-", "baselines", ".", "section", ":", "Results", "section", ":", "Sentiment", "analysis", "We", "show", "results", "for", "sentiment", "analysis", "for", "all", "12", "domain", "adaptation", "scenarios", "in", "Figure", "2", ".", "For", "clarity", ",", "we", "also", "show", "the", "accuracy", "scores", "averaged", "across", "each", "target", "domain", "as", "well", "as", "a", "global", "macro", "average", "in", "Table", "2", "Self", "-", "training", "achieves", "surprisingly", "good", "results", "but", "is", "not", "able", "to", "compete", "with", "tri", "-", "training", ".", "Tritraining", "with", "disagreement", "is", "only", "slightly", "better", "than", "self", "-", "training", ",", "showing", "that", "the", "disagreement", "component", "might", "not", "be", "useful", "when", "there", "is", "a", "strong", "domain", "shift", ".", "Tri", "-", "training", "achieves", "the", "best", "average", "results", "on", "two", "target", "domains", "and", "clearly", "outperforms", "the", "state", "of", "the", "art", "on", "average", ".", "MT", "-", "Tri", "finally", "outperforms", "the", "state", "of", "the", "art", "on", "3", "/", "4", "domains", ",", "and", "even", "slightly", "traditional", "tritraining", ",", "resulting", "in", "the", "overall", "best", "method", ".", "This", "improvement", "is", "mainly", "due", "to", "the", "B", "-", ">E", "and", "D", "-", ">E", "scenarios", ",", "on", "which", "tri", "-", "training", "struggles", ".", "These", "domain", "pairs", "are", "among", "those", "with", "the", "highest", "Adistance", "[", "reference", "]", ",", "which", "highlights", "that", "tri", "-", "training", "has", "difficulty", "dealing", "with", "a", "strong", "shift", "in", "domain", ".", "Our", "method", "is", "able", "to", "mitigate", "this", "deficiency", "by", "training", "one", "of", "the", "three", "output", "layers", "only", "on", "pseudo", "-", "labeled", "target", "domain", "examples", ".", "In", "addition", ",", "MT", "-", "Tri", "is", "more", "efficient", "as", "it", "adds", "a", "smaller", "number", "of", "pseudo", "-", "labeled", "examples", "than", "tri", "-", "training", "at", "every", "epoch", ".", "For", "sentiment", "analysis", ",", "tri", "-", "training", "adds", "around", "1800", "-", "1950", "/", "2000", "unlabeled", "examples", "at", "every", "epoch", ",", "while", "MT", "-", "Tri", "only", "adds", "around", "100", "-", "300", "in", "early", "epochs", ".", "This", "shows", "that", "the", "orthogonality", "constraint", "is", "useful", "for", "inducing", "diversity", ".", "In", "addition", ",", "adding", "fewer", "examples", "poses", "a", "smaller", "risk", "of", "swamping", "the", "learned", "representations", "with", "useless", "signals", "and", "is", "more", "akin", "to", "fine", "-", "tuning", ",", "the", "standard", "method", "for", "supervised", "domain", "adaptation", "[", "reference", "]", ".", "We", "observe", "an", "asymmetry", "in", "the", "results", "between", "some", "of", "the", "domain", "pairs", ",", "e.g.", "B", "-", ">D", "and", "D", "-", ">B.", "We", "hypothesize", "that", "the", "asymmetry", "may", "be", "due", "to", "properties", "of", "the", "data", "and", "that", "the", "domains", "are", "relatively", "far", "apart", "e.g.", ",", "in", "terms", "of", "A", "-", "distance", ".", "In", "fact", ",", "asymmetry", "in", "these", "domains", "is", "already", "reflected", "Table", "4", ":", "Accuracy", "for", "POS", "tagging", "on", "the", "dev", "and", "test", "sets", "of", "the", "SANCL", "domains", ",", "models", "trained", "on", "full", "source", "data", "setup", ".", "Values", "for", "methods", "with", "*", "are", "from", "[", "reference", "]", ".", "in", "the", "results", "of", "[", "reference", "]", "and", "is", "corroborated", "in", "the", "results", "for", "asymmetric", "tri", "-", "training", "[", "reference", "]", "and", "our", "method", ".", "We", "note", "a", "weakness", "of", "this", "dataset", "is", "high", "variance", ".", "Existing", "approaches", "only", "report", "the", "mean", ",", "which", "makes", "an", "objective", "comparison", "difficult", ".", "For", "this", "reason", ",", "we", "believe", "it", "is", "essential", "to", "evaluate", "proposed", "approaches", "also", "on", "other", "tasks", ".", "POS", "tagging", "Results", "for", "tagging", "in", "the", "low", "-", "data", "regime", "(", "10", "%", "of", "WSJ", ")", "are", "given", "in", "Table", "3", ".", "Self", "-", "training", "does", "not", "work", "for", "the", "sequence", "prediction", "task", ".", "We", "report", "only", "the", "best", "instantiation", "(", "throttling", "with", "n=800", ")", ".", "Our", "results", "contribute", "to", "negative", "findings", "regarding", "self", "-", "training", "[", "reference", "][", "reference", "]", ".", "In", "the", "low", "-", "data", "setup", ",", "tri", "-", "training", "with", "disagreement", "works", "best", ",", "reaching", "an", "overall", "average", "accuracy", "of", "89.70", ",", "closely", "followed", "by", "classic", "tritraining", ",", "and", "significantly", "outperforming", "the", "baseline", "on", "4", "/", "5", "domains", ".", "The", "exception", "is", "newsgroups", ",", "a", "difficult", "domain", "with", "high", "OOV", "rate", "where", "none", "of", "the", "approches", "beats", "the", "baseline", "(", "see", "§", "3.4", ")", ".", "Our", "proposed", "MT", "-", "Tri", "is", "better", "than", "asymmetric", "tritraining", ",", "but", "falls", "below", "classic", "tri", "-", "training", ".", "It", "beats", "Table", "5", ":", "Accuracy", "scores", "on", "dev", "sets", "for", "OOV", "and", "unknown", "word", "-", "tag", "(", "UWT", ")", "tokens", ".", "the", "baseline", "significantly", "on", "only", "2", "/", "5", "domains", "(", "answers", "and", "emails", ")", ".", "The", "FLORS", "tagger", "[", "reference", "]", "fares", "better", ".", "Its", "contextual", "distributional", "features", "are", "particularly", "helpful", "on", "unknown", "word", "-", "tag", "combinations", "(", "see", "§", "3.4", ")", ",", "which", "is", "a", "limitation", "of", "the", "lexicalized", "generic", "bi", "-", "LSTM", "tagger", ".", "For", "the", "high", "-", "data", "setup", "(", "Table", "4", ")", "results", "are", "similar", ".", "Disagreement", ",", "however", ",", "is", "only", "favorable", "in", "the", "low", "-", "data", "setups", ";", "the", "effect", "of", "avoiding", "easy", "points", "no", "longer", "holds", "in", "the", "full", "data", "setup", ".", "Classic", "tritraining", "is", "the", "best", "method", ".", "In", "particular", ",", "traditional", "tri", "-", "training", "is", "complementary", "to", "word", "embedding", "initialization", ",", "pushing", "the", "non", "-", "pre", "-", "trained", "baseline", "to", "the", "level", "of", "SRC", "with", "Glove", "initalization", ".", "Tritraining", "pushes", "performance", "even", "further", "and", "results", "in", "the", "best", "model", ",", "significantly", "outperforming", "the", "baseline", "again", "in", "4", "/", "5", "cases", ",", "and", "reaching", "FLORS", "performance", "on", "weblogs", ".", "Multi", "-", "task", "tritraining", "is", "often", "slightly", "more", "effective", "than", "asymmetric", "tri", "-", "training", "[", "reference", "]", ";", "however", ",", "improvements", "for", "both", "are", "not", "robust", "across", "domains", ",", "sometimes", "performance", "even", "drops", ".", "The", "model", "likely", "is", "too", "simplistic", "for", "such", "a", "high", "-", "data", "POS", "setup", ",", "and", "exploring", "shared", "-", "private", "models", "might", "prove", "more", "fruitful", ".", "On", "the", "test", "sets", ",", "tri", "-", "training", "performs", "consistently", "the", "best", ".", "section", ":", "POS", "analysis", "We", "analyze", "POS", "tagging", "accuracy", "with", "respect", "to", "word", "frequency", "6", "and", "unseen", "word", "-", "tag", "combinations", "(", "UWT", ")", "on", "the", "dev", "sets", ".", "known", "tags", ",", "OOVs", "and", "unknown", "word", "-", "tag", "(", "UWT", ")", "rate", ".", "The", "SANCL", "dataset", "is", "overall", "very", "challenging", ":", "OOV", "rates", "are", "high", "(", "6.8", "-", "11", "%", "compared", "to", "2.3", "%", "in", "WSJ", ")", ",", "so", "is", "the", "unknown", "word", "-", "tag", "(", "UWT", ")", "rate", "(", "answers", "and", "emails", "contain", "2.91", "%", "and", "3.47", "%", "UWT", "compared", "to", "0.61", "%", "on", "WSJ", ")", "and", "almost", "all", "target", "domains", "even", "contain", "unknown", "tags", "[", "reference", "]", ")", "(", "unknown", "tags", ":", "ADD", ",", "GW", ",", "NFP", ",", "XX", ")", ",", "except", "for", "weblogs", ".", "Email", "is", "the", "domain", "with", "the", "highest", "OOV", "rate", "and", "highest", "unknown", "-", "tag", "-", "for", "-", "known", "-", "words", "rate", ".", "We", "plot", "accuracy", "with", "respect", "to", "word", "frequency", "on", "email", "in", "Figure", "3", ",", "analyzing", "how", "the", "three", "methods", "fare", "in", "comparison", "to", "the", "baseline", "on", "this", "difficult", "domain", ".", "Regarding", "OOVs", ",", "the", "results", "in", "Table", "5", "(", "second", "part", ")", "show", "that", "classic", "tri", "-", "training", "outperforms", "the", "source", "model", "(", "trained", "on", "only", "source", "data", ")", "on", "3", "/", "5", "domains", "in", "terms", "of", "OOV", "accuracy", ",", "except", "on", "two", "domains", "with", "high", "OOV", "rate", "(", "newsgroups", "and", "weblogs", ")", ".", "In", "general", ",", "we", "note", "that", "tri", "-", "training", "works", "best", "on", "OOVs", "and", "on", "low", "-", "frequency", "tokens", ",", "which", "is", "also", "shown", "in", "Figure", "3", "(", "leftmost", "bins", ")", ".", "Both", "other", "methods", "fall", "typically", "below", "the", "baseline", "in", "terms", "of", "OOV", "accuracy", ",", "but", "MT", "-", "Tri", "still", "outperforms", "Asym", "in", "4", "/", "5", "cases", ".", "Table", "5", "(", "last", "part", ")", "also", "shows", "that", "no", "bootstrapping", "method", "works", "well", "on", "unknown", "word", "-", "tag", "combinations", ".", "UWT", "tokens", "are", "very", "difficult", "to", "predict", "correctly", "using", "an", "unsupervised", "approach", ";", "the", "less", "lexicalized", "and", "more", "context", "-", "driven", "approach", "taken", "by", "FLORS", "is", "clearly", "superior", "for", "these", "cases", ",", "resulting", "in", "higher", "UWT", "accuracies", "for", "4", "/", "5", "domains", ".", "section", ":", "Related", "work", "Learning", "under", "Domain", "Shift", "There", "is", "a", "large", "body", "of", "work", "on", "domain", "adaptation", ".", "Studies", "on", "unsupervised", "domain", "adaptation", "include", "early", "work", "on", "bootstrapping", "[", "reference", "][", "reference", "]", ",", "shared", "feature", "representations", "[", "reference", "][", "reference", "]", "and", "instance", "weighting", "[", "reference", "]", ".", "Recent", "approaches", "include", "adversarial", "learning", "[", "reference", "]", "and", "fine", "-", "tuning", "[", "reference", "]", ".", "There", "is", "almost", "no", "work", "on", "bootstrapping", "approaches", "for", "recent", "neural", "NLP", ",", "in", "particular", "under", "domain", "shift", ".", "Tri", "-", "training", "is", "less", "studied", ",", "and", "only", "recently", "re", "-", "emerged", "in", "the", "vision", "community", "[", "reference", "]", ",", "albeit", "is", "not", "compared", "to", "classic", "tri", "-", "training", ".", "Neural", "network", "ensembling", "Related", "work", "on", "self", "-", "ensembling", "approaches", "includes", "snapshot", "ensembling", "or", "temporal", "ensembling", "[", "reference", "]", ".", "In", "general", ",", "the", "line", "between", "\"", "explicit", "\"", "and", "\"", "implicit", "\"", "ensembling", ",", "like", "dropout", "[", "reference", "]", "or", "temporal", "ensembling", "[", "reference", "]", ",", "is", "more", "fuzzy", ".", "As", "we", "noted", "earlier", "our", "multi", "-", "task", "learning", "setup", "can", "be", "seen", "as", "a", "form", "of", "self", "-", "ensembling", ".", "Multi", "-", "task", "learning", "in", "NLP", "Neural", "networks", "are", "particularly", "well", "-", "suited", "for", "MTL", "allowing", "for", "parameter", "sharing", "[", "reference", "]", ".", "Recent", "NLP", "conferences", "witnessed", "a", "\"", "tsunami", "\"", "of", "deep", "learning", "papers", "[", "reference", "]", ",", "followed", "by", "what", "we", "call", "a", "multi", "-", "task", "learning", "\"", "wave", "\"", ":", "MTL", "has", "been", "successfully", "applied", "to", "a", "wide", "range", "of", "NLP", "tasks", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "Related", "to", "it", "is", "the", "pioneering", "work", "on", "adversarial", "learning", "(", "DANN", ")", "[", "reference", "]", ".", "For", "sentiment", "analysis", "we", "found", "tri", "-", "training", "and", "our", "MT", "-", "Tri", "model", "to", "outperform", "DANN", ".", "Our", "MT", "-", "Tri", "model", "lends", "itself", "well", "to", "shared", "-", "private", "models", "such", "as", "those", "proposed", "recently", "[", "reference", "]", ",", "which", "extend", "upon", "[", "reference", "]", "by", "having", "separate", "source", "and", "target", "-", "specific", "encoders", ".", "section", ":", "Conclusions", "We", "re", "-", "evaluate", "a", "range", "of", "traditional", "generalpurpose", "bootstrapping", "algorithms", "in", "the", "context", "of", "neural", "network", "approaches", "to", "semi", "-", "supervised", "learning", "under", "domain", "shift", ".", "For", "the", "two", "examined", "NLP", "tasks", "classic", "tri", "-", "training", "works", "the", "best", "and", "even", "outperforms", "a", "recent", "state", "-", "of", "-", "the", "-", "art", "method", ".", "The", "drawback", "of", "tri", "-", "training", "it", "its", "time", "and", "space", "complexity", ".", "We", "therefore", "propose", "a", "more", "efficient", "multi", "-", "task", "tri", "-", "training", "model", ",", "which", "outperforms", "both", "traditional", "tri", "-", "training", "and", "recent", "alternatives", "in", "the", "case", "of", "sentiment", "analysis", ".", "For", "POS", "tagging", ",", "classic", "tri", "-", "training", "is", "superior", ",", "performing", "especially", "well", "on", "OOVs", "and", "low", "frequency", "tokens", ",", "which", "suggests", "it", "is", "less", "affected", "by", "error", "propagation", ".", "Overall", "we", "emphasize", "the", "importance", "of", "comparing", "neural", "approaches", "to", "strong", "baselines", "and", "reporting", "results", "across", "several", "runs", ".", "section", ":", "section", ":", "Acknowledgments", "We", "thank", "the", "anonymous", "reviewers", "for", "their", "valuable", "feedback", ".", "Sebastian", "is", "supported", "by", "Irish", "Research", "Council", "Grant", "Number", "EBPPG", "/", "2014", "/", "30", "and", "Science", "Foundation", "Ireland", "Grant", "Number", "SFI", "/", "12", "/", "RC", "/", "2289", ".", "Barbara", "is", "supported", "by", "NVIDIA", "corporation", "and", "thanks", "the", "Computing", "Center", "of", "the", "University", "of", "Groningen", "for", "HPC", "support", ".", "section", ":"], "entity": [[3, 8, ["Neural", "Semi", "-", "supervised", "Learning"], "Task"], [9, 11, ["Domain", "Shift"], "Task"], [15, 17, ["neural", "models"], "Method"], [24, 28, ["learning", "under", "domain", "shift"], "Task"], [67, 72, ["general", "-", "purpose", "bootstrapping", "approaches"], "Method"], [76, 78, ["neural", "networks"], "Method"], [83, 85, ["neural", "approaches"], "Method"], [89, 96, ["multi", "-", "task", "tri", "-", "training", "method"], "Method"], [99, 103, ["time", "and", "space", "complexity"], "Metric"], [105, 108, ["tri", "-", "training"], "Method"], [132, 134, ["sentiment", "analysis"], "Task"], [155, 158, ["tri", "-", "training"], "Method"], [185, 188, ["Deep", "neural", "networks"], "Method"], [189, 190, ["DNNs"], "Method"], [209, 212, ["supervised", "NLP", "tasks"], "Task"], [214, 216, ["dependency", "parsing"], "Task"], [220, 223, ["named", "entity", "recognition"], "Task"], [228, 231, ["semantic", "role", "labeling"], "Task"], [238, 239, ["learning"], "Task"], [245, 247, ["domain", "shift"], "Task"], [257, 261, ["real", "-", "world", "applications"], "Task"], [280, 283, ["domain", "adaptation", "approaches"], "Method"], [371, 375, ["semi", "-", "supervised", "learning"], "Method"], [376, 377, ["SSL"], "Method"], [421, 423, ["bootstrapping", "algorithms"], "Method"], [427, 428, ["DNNs"], "Method"], [434, 438, ["semi", "-", "supervised", "algorithms"], "Method"], [463, 465, ["NLP", "models"], "Method"], [489, 492, ["domain", "adaptation", "setting"], "Task"], [504, 506, ["bootstrapping", "methods"], "Method"], [507, 510, ["self", "-", "training"], "Method"], [514, 517, ["tri", "-", "training"], "Method"], [522, 523, ["tritraining"], "Method"], [524, 525, ["disagreement"], "Method"], [529, 534, ["neural", "network", "-", "based", "approaches"], "Method"], [536, 538, ["NLP", "tasks"], "Task"], [545, 547, ["sequence", "prediction"], "Task"], [549, 551, ["classification", "task"], "Task"], [552, 554, ["POS", "tagging"], "Task"], [555, 557, ["sentiment", "analysis"], "Task"], [575, 579, ["task", "-", "specific", "measures"], "Metric"], [599, 602, ["tri", "-", "training"], "Method"], [613, 617, ["semi", "-", "supervised", "learning"], "Method"], [622, 625, ["neural", "adaptation", "approaches"], "Method"], [636, 642, ["multi", "-", "task", "tri", "-", "training"], "Method"], [649, 652, ["tri", "-", "training"], "Method"], [655, 659, ["time", "and", "space", "complexity"], "Metric"], [669, 672, ["unsupervised", "domain", "adaptation"], "Task"], [673, 675, ["sentiment", "analysis"], "Task"], [681, 684, ["tri", "-", "training"], "Method"], [685, 687, ["POS", "tagging"], "Task"], [699, 706, ["multi", "-", "task", "tri", "-", "training", "method"], "Method"], [712, 715, ["tri", "-", "training"], "Method"], [722, 727, ["semi", "-", "supervised", "learning", "baseline"], "Method"], [732, 734, ["NLP", "models"], "Method"], [743, 746, ["bootstrapping", "1", "algorithms"], "Method"], [783, 786, ["Neural", "bootstrapping", "methods"], "Method"], [791, 793, ["bootstrapping", "methods"], "Method"], [794, 797, ["self", "-", "training"], "Method"], [798, 801, ["tri", "-", "training"], "Method"], [803, 806, ["tri", "-", "training"], "Method"], [807, 808, ["disagreement"], "Method"], [816, 818, ["neural", "networks"], "Method"], [856, 859, ["Self", "-", "training"], "Method"], [872, 874, ["bootstrapping", "approaches"], "Method"], [897, 898, ["training"], "Task"], [914, 917, ["Self", "-", "training"], "Method"], [919, 921, ["model", "m"], "Method"], [1017, 1018, ["Calibration"], "Task"], [1027, 1029, ["neural", "networks"], "Method"], [1045, 1048, ["Self", "-", "training"], "Method"], [1080, 1083, ["absolute", "confidence", "value"], "Metric"], [1137, 1140, ["self", "-", "training"], "Method"], [1142, 1143, ["throttling"], "Task"], [1155, 1156, ["selection"], "Method"], [1160, 1162, ["Online", "learning"], "Method"], [1169, 1170, ["DNNs"], "Method"], [1195, 1198, ["self", "-", "training"], "Method"], [1202, 1205, ["self", "-", "training"], "Method"], [1211, 1212, ["parsing"], "Task"], [1225, 1227, ["generative", "component"], "Method"], [1232, 1233, ["reranker"], "Method"], [1234, 1238, ["high", "-", "data", "conditions"], "Task"], [1253, 1255, ["data", "selection"], "Task"], [1268, 1270, ["NLP", "tasks"], "Task"], [1309, 1312, ["Tri", "-", "training"], "Method"], [1312, 1315, ["Tri", "-", "training"], "Method"], [1339, 1341, ["trained", "models"], "Method"], [1342, 1345, ["Tri", "-", "training"], "Method"], [1406, 1407, ["classifiers"], "Method"], [1412, 1415, ["Tri", "-", "training"], "Method"], [1422, 1425, ["Tri", "-", "training"], "Method"], [1555, 1558, ["Tri", "-", "training"], "Method"], [1559, 1560, ["disagreement"], "Method"], [1566, 1567, ["tritraining"], "Method"], [1573, 1579, ["part", "-", "of", "-", "speech", "tagging"], "Task"], [1587, 1590, ["tri", "-", "training"], "Method"], [1591, 1594, ["tri", "-", "training"], "Method"], [1595, 1596, ["disagreement"], "Task"], [1649, 1651, ["bootstrapping", "approaches"], "Method"], [1662, 1664, ["neural", "approaches"], "Method"], [1667, 1672, ["linearly", "growing", "candidate", "sampling", "scheme"], "Method"], [1690, 1692, ["Confidence", "thresholding"], "Method"], [1694, 1695, ["selftraining"], "Method"], [1735, 1736, ["prediction"], "Task"], [1740, 1743, ["tri", "-", "training"], "Method"], [1745, 1746, ["thresholding"], "Method"], [1761, 1764, ["tri", "-", "training"], "Method"], [1765, 1768, ["tri", "-", "training"], "Method"], [1838, 1842, ["time", "and", "space", "complexity"], "Metric"], [1843, 1846, ["tri", "-", "training"], "Method"], [1865, 1869, ["multi", "-", "task", "learning"], "Task"], [1870, 1871, ["MTL"], "Task"], [1882, 1883, ["training"], "Task"], [1907, 1908, ["MTL"], "Task"], [1917, 1919, ["joint", "representation"], "Method"], [1922, 1923, ["convergence"], "Task"], [1955, 1957, ["POS", "tagging"], "Task"], [1964, 1966, ["joint", "representation"], "Method"], [1976, 1977, ["prediction"], "Task"], [2024, 2027, ["self", "-", "training"], "Method"], [2107, 2108, ["prediction"], "Task"], [2176, 2179, ["sequence", "prediction", "task"], "Task"], [2231, 2236, ["uppermost", "Bi", "-", "LSTM", "encoding"], "Method"], [2242, 2245, ["tri", "-", "training"], "Method"], [2250, 2254, ["multi", "-", "task", "model"], "Method"], [2264, 2266, ["bootstrap", "sampling"], "Method"], [2304, 2306, ["output", "layers"], "Method"], [2339, 2342, ["tri", "-", "training"], "Method"], [2352, 2353, ["prediction"], "Task"], [2354, 2356, ["majority", "voting"], "Method"], [2373, 2375, ["confidence", "thresholding"], "Method"], [2428, 2430, ["Computational", "complexity"], "Metric"], [2440, 2444, ["space", "and", "time", "complexity"], "Metric"], [2445, 2448, ["tri", "-", "training"], "Method"], [2466, 2470, ["regular", "tri", "-", "training"], "Method"], [2471, 2474, ["tri", "-", "training"], "Method"], [2521, 2523, ["time", "efficiency"], "Metric"], [2524, 2528, ["tri", "-", "training", "first"], "Method"], [2562, 2564, ["majority", "vote"], "Method"], [2579, 2582, ["tri", "-", "training"], "Method"], [2595, 2597, ["forward", "pass"], "Method"], [2617, 2619, ["forward", "pass"], "Method"], [2646, 2649, ["tri", "-", "training"], "Method"], [2687, 2690, ["tri", "-", "training"], "Method"], [2699, 2703, ["self", "-", "ensembling", "technique"], "Method"], [2716, 2718, ["ensemble", "prediction"], "Task"], [2726, 2727, ["ensembling"], "Method"], [2741, 2744, ["tri", "-", "training"], "Method"], [2764, 2766, ["temporal", "ensembling"], "Method"], [2785, 2787, ["temporal", "ensembling"], "Method"], [2812, 2814, ["asymmetric", "tritraining"], "Method"], [2828, 2829, ["ASYM"], "Method"], [2874, 2878, ["tri", "-", "training", "formulation"], "Method"], [2899, 2901, ["POS", "tagging"], "Task"], [2906, 2909, ["Amazon", "Reviews", "dataset"], "Material"], [2913, 2915, ["sentiment", "analysis"], "Task"], [2949, 2951, ["sequence", "labeling"], "Task"], [2953, 2955, ["classification", "task"], "Task"], [2966, 2968, ["POS", "tagging"], "Task"], [2969, 2971, ["POS", "tagging"], "Task"], [2974, 2979, ["SANCL", "2012", "shared", "task", "dataset"], "Material"], [3009, 3011, ["FLORS", "tagger"], "Method"], [3070, 3072, ["neural", "tagger"], "Method"], [3074, 3077, ["POS", "tagging", "model"], "Method"], [3086, 3090, ["Bi", "-", "LSTM", "tagger"], "Method"], [3118, 3119, ["BiLSTM"], "Method"], [3121, 3123, ["hidden", "layer"], "Method"], [3129, 3131, ["POS", "model"], "Method"], [3134, 3135, ["WSJ"], "Material"], [3136, 3138, ["early", "stopping"], "Task"], [3140, 3143, ["WSJ", "development", "set"], "Material"], [3155, 3157, ["word", "dropout"], "Method"], [3173, 3176, ["Ontonotes", "4.0", "release"], "Material"], [3178, 3183, ["Penn", "treebank", "Wall", "Street", "Journal"], "Material"], [3184, 3185, ["WSJ"], "Material"], [3233, 3235, ["SANCL", "domains"], "Material"], [3256, 3258, ["SANCL", "domain"], "Material"], [3294, 3299, ["suboptimal", "per", "-", "domain", "settings"], "Task"], [3303, 3306, ["unsupervised", "adaptation", "scenario"], "Task"], [3309, 3311, ["Sentiment", "analysis"], "Task"], [3312, 3314, ["sentiment", "analysis"], "Task"], [3319, 3322, ["Amazon", "reviews", "dataset"], "Material"], [3365, 3368, ["pre", "-", "processing"], "Method"], [3384, 3392, ["tf", "-", "idf", "weighted", "unigram", "and", "bigram", "features"], "Method"], [3412, 3413, ["validation"], "Task"], [3427, 3428, ["MLP"], "Method"], [3430, 3432, ["hidden", "layer"], "Method"], [3448, 3451, ["Variational", "Fair", "Autoencoder"], "Method"], [3459, 3464, ["domain", "-", "adversarial", "neural", "networks"], "Method"], [3465, 3466, ["DANN"], "Method"], [3494, 3496, ["task", "model"], "Method"], [3504, 3507, ["self", "-", "training"], "Method"], [3513, 3516, ["tri", "-", "training"], "Method"], [3517, 3518, ["Tri"], "Method"], [3522, 3525, ["tri", "-", "training"], "Method"], [3526, 3527, ["disagreement"], "Method"], [3536, 3540, ["asymmetric", "tri", "-", "training"], "Method"], [3549, 3555, ["multi", "-", "task", "tri", "-", "training"], "Method"], [3556, 3557, ["MTTri"], "Method"], [3564, 3565, ["DyNet"], "Method"], [3568, 3570, ["evaluation", "scores"], "Metric"], [3585, 3587, ["mean", "accuracy"], "Metric"], [3594, 3596, ["POS", "tagging"], "Task"], [3601, 3603, ["sentiment", "analysis"], "Task"], [3608, 3610, ["bootstrap", "test"], "Method"], [3626, 3631, ["semi", "-", "supervised", "-", "baselines"], "Metric"], [3637, 3639, ["Sentiment", "analysis"], "Task"], [3643, 3645, ["sentiment", "analysis"], "Task"], [3648, 3651, ["domain", "adaptation", "scenarios"], "Task"], [3662, 3663, ["accuracy"], "Metric"], [3673, 3676, ["global", "macro", "average"], "Metric"], [3679, 3682, ["Self", "-", "training"], "Method"], [3693, 3696, ["tri", "-", "training"], "Method"], [3697, 3698, ["Tritraining"], "Method"], [3699, 3700, ["disagreement"], "Method"], [3705, 3708, ["self", "-", "training"], "Method"], [3712, 3714, ["disagreement", "component"], "Method"], [3726, 3729, ["Tri", "-", "training"], "Method"], [3769, 3770, ["tritraining"], "Method"], [3796, 3799, ["tri", "-", "training"], "Method"], [3818, 3821, ["tri", "-", "training"], "Method"], [3845, 3847, ["output", "layers"], "Method"], [3877, 3880, ["tri", "-", "training"], "Method"], [3885, 3887, ["sentiment", "analysis"], "Task"], [3888, 3891, ["tri", "-", "training"], "Method"], [3953, 3956, ["fine", "-", "tuning"], "Method"], [3961, 3964, ["supervised", "domain", "adaptation"], "Task"], [4033, 4034, ["Accuracy"], "Metric"], [4035, 4037, ["POS", "tagging"], "Task"], [4045, 4047, ["SANCL", "domains"], "Material"], [4081, 4085, ["asymmetric", "tri", "-", "training"], "Method"], [4135, 4137, ["POS", "tagging"], "Task"], [4139, 4140, ["tagging"], "Task"], [4150, 4151, ["WSJ"], "Material"], [4158, 4161, ["Self", "-", "training"], "Method"], [4166, 4169, ["sequence", "prediction", "task"], "Task"], [4189, 4192, ["self", "-", "training"], "Method"], [4200, 4204, ["low", "-", "data", "setup"], "Task"], [4205, 4208, ["tri", "-", "training"], "Method"], [4209, 4210, ["disagreement"], "Method"], [4215, 4218, ["overall", "average", "accuracy"], "Metric"], [4225, 4226, ["tritraining"], "Method"], [4248, 4250, ["OOV", "rate"], "Metric"], [4272, 4274, ["asymmetric", "tritraining"], "Method"], [4279, 4282, ["tri", "-", "training"], "Method"], [4288, 4290, ["Accuracy", "scores"], "Metric"], [4321, 4323, ["FLORS", "tagger"], "Method"], [4354, 4360, ["lexicalized", "generic", "bi", "-", "LSTM", "tagger"], "Method"], [4375, 4376, ["Disagreement"], "Task"], [4405, 4406, ["tritraining"], "Method"], [4415, 4418, ["tri", "-", "training"], "Method"], [4421, 4424, ["word", "embedding", "initialization"], "Task"], [4437, 4438, ["SRC"], "Method"], [4439, 4441, ["Glove", "initalization"], "Method"], [4442, 4443, ["Tritraining"], "Method"], [4482, 4486, ["asymmetric", "tri", "-", "training"], "Method"], [4515, 4520, ["high", "-", "data", "POS", "setup"], "Task"], [4523, 4527, ["shared", "-", "private", "models"], "Method"], [4537, 4540, ["tri", "-", "training"], "Method"], [4547, 4549, ["POS", "analysis"], "Method"], [4551, 4554, ["POS", "tagging", "accuracy"], "Metric"], [4589, 4591, ["SANCL", "dataset"], "Material"], [4596, 4598, ["OOV", "rates"], "Metric"], [4610, 4611, ["WSJ"], "Material"], [4634, 4635, ["UWT"], "Material"], [4640, 4641, ["WSJ"], "Material"], [4679, 4681, ["OOV", "rate"], "Metric"], [4689, 4693, ["known", "-", "words", "rate"], "Metric"], [4696, 4697, ["accuracy"], "Metric"], [4725, 4726, ["OOVs"], "Task"], [4739, 4742, ["tri", "-", "training"], "Method"], [4744, 4746, ["source", "model"], "Method"], [4761, 4763, ["OOV", "accuracy"], "Metric"], [4770, 4772, ["OOV", "rate"], "Metric"], [4784, 4787, ["tri", "-", "training"], "Method"], [4790, 4791, ["OOVs"], "Material"], [4821, 4823, ["OOV", "accuracy"], "Metric"], [4847, 4849, ["bootstrapping", "method"], "Method"], [4868, 4870, ["unsupervised", "approach"], "Method"], [4876, 4880, ["context", "-", "driven", "approach"], "Method"], [4882, 4883, ["FLORS"], "Method"], [4907, 4909, ["Domain", "Shift"], "Method"], [4917, 4919, ["domain", "adaptation"], "Task"], [4922, 4925, ["unsupervised", "domain", "adaptation"], "Task"], [4929, 4930, ["bootstrapping"], "Method"], [4936, 4939, ["shared", "feature", "representations"], "Method"], [4945, 4947, ["instance", "weighting"], "Method"], [4954, 4956, ["adversarial", "learning"], "Method"], [4960, 4963, ["fine", "-", "tuning"], "Method"], [4973, 4975, ["bootstrapping", "approaches"], "Method"], [4977, 4979, ["neural", "NLP"], "Task"], [4983, 4985, ["domain", "shift"], "Task"], [4986, 4989, ["Tri", "-", "training"], "Method"], [5001, 5003, ["vision", "community"], "Task"], [5013, 5016, ["tri", "-", "training"], "Method"], [5017, 5020, ["Neural", "network", "ensembling"], "Method"], [5023, 5027, ["self", "-", "ensembling", "approaches"], "Method"], [5028, 5030, ["snapshot", "ensembling"], "Task"], [5031, 5033, ["temporal", "ensembling"], "Method"], [5053, 5054, ["dropout"], "Method"], [5058, 5060, ["temporal", "ensembling"], "Method"], [5073, 5078, ["multi", "-", "task", "learning", "setup"], "Task"], [5085, 5088, ["self", "-", "ensembling"], "Task"], [5089, 5093, ["Multi", "-", "task", "learning"], "Method"], [5094, 5097, ["NLP", "Neural", "networks"], "Method"], [5103, 5104, ["MTL"], "Task"], [5106, 5108, ["parameter", "sharing"], "Task"], [5113, 5115, ["NLP", "conferences"], "Material"], [5134, 5138, ["multi", "-", "task", "learning"], "Method"], [5142, 5143, ["MTL"], "Task"], [5152, 5154, ["NLP", "tasks"], "Task"], [5178, 5180, ["adversarial", "learning"], "Method"], [5181, 5182, ["DANN"], "Task"], [5188, 5190, ["sentiment", "analysis"], "Task"], [5192, 5195, ["tri", "-", "training"], "Method"], [5203, 5204, ["DANN"], "Method"], [5214, 5218, ["shared", "-", "private", "models"], "Method"], [5240, 5242, ["specific", "encoders"], "Method"], [5254, 5257, ["generalpurpose", "bootstrapping", "algorithms"], "Method"], [5261, 5264, ["neural", "network", "approaches"], "Method"], [5265, 5269, ["semi", "-", "supervised", "learning"], "Task"], [5270, 5272, ["domain", "shift"], "Task"], [5277, 5279, ["NLP", "tasks"], "Task"], [5280, 5283, ["tri", "-", "training"], "Method"], [5303, 5306, ["tri", "-", "training"], "Method"], [5308, 5312, ["time", "and", "space", "complexity"], "Metric"], [5331, 5334, ["tri", "-", "training"], "Method"], [5341, 5343, ["sentiment", "analysis"], "Task"], [5345, 5347, ["POS", "tagging"], "Task"], [5349, 5352, ["tri", "-", "training"], "Method"], [5382, 5384, ["neural", "approaches"], "Method"], [843, 846, ["tri", "-", "training"], "Method"], [853, 856, ["Self", "-", "training"], "Method"], [1829, 1832, ["tri", "-", "training"], "Method"], [1852, 1853, ["Tritraining"], "Method"], [1856, 1857, ["Tri"], "Method"], [1861, 1862, ["Tri"], "Method"], [2203, 2204, ["Tri"], "Method"], [2435, 2436, ["Tri"], "Method"], [2461, 2462, ["Tri"], "Method"], [2489, 2490, ["Tri"], "Method"], [2613, 2614, ["Tri"], "Method"], [2678, 2679, ["Tri"], "Method"], [2693, 2694, ["Tri"], "Method"], [2868, 2869, ["Tri"], "Method"], [3209, 3210, ["WSJ"], "Material"], [3528, 3529, ["Tri"], "Method"], [3751, 3752, ["Tri"], "Method"], [3861, 3862, ["Tri"], "Method"], [3907, 3908, ["Tri"], "Method"], [4268, 4269, ["Tri"], "Method"], [4301, 4302, ["UWT"], "Material"], [4475, 4476, ["tritraining"], "Method"], [4567, 4568, ["UWT"], "Material"], [4584, 4585, ["UWT"], "Material"], [4621, 4622, ["UWT"], "Material"], [4827, 4828, ["Tri"], "Method"], [4858, 4859, ["UWT"], "Material"], [4893, 4894, ["UWT"], "Material"], [5199, 5200, ["Tri"], "Method"], [5208, 5209, ["Tri"], "Method"], [5322, 5325, ["tri", "-", "training"], "Method"]]} {"doc_id": "2a86bcdfb1d817ddb76ba202319f8267a36c0f62", "text": ["document", ":", "PCL", ":", "Proposal", "Cluster", "Learning", "for", "Weakly", "Supervised", "Object", "Detection", "Weakly", "Supervised", "Object", "Detection", "(", "WSOD", ")", ",", "using", "only", "image", "-", "level", "annotations", "to", "train", "object", "detectors", ",", "is", "of", "growing", "importance", "in", "object", "recognition", ".", "In", "this", "paper", ",", "we", "propose", "a", "novel", "deep", "network", "for", "WSOD", ".", "Unlike", "previous", "networks", "that", "transfer", "the", "object", "detection", "problem", "to", "an", "image", "classification", "problem", "using", "Multiple", "Instance", "Learning", "(", "MIL", ")", ",", "our", "strategy", "generates", "proposal", "clusters", "to", "learn", "refined", "instance", "classifiers", "by", "an", "iterative", "process", ".", "The", "proposals", "in", "the", "same", "cluster", "are", "spatially", "adjacent", "and", "associated", "with", "the", "same", "object", ".", "This", "prevents", "the", "network", "from", "concentrating", "too", "much", "on", "parts", "of", "objects", "instead", "of", "whole", "objects", ".", "We", "first", "show", "that", "instances", "can", "be", "assigned", "object", "or", "background", "labels", "directly", "based", "on", "proposal", "clusters", "for", "instance", "classifier", "refinement", ",", "and", "then", "show", "that", "treating", "each", "cluster", "as", "a", "small", "new", "bag", "yields", "fewer", "ambiguities", "than", "the", "directly", "assigning", "label", "method", ".", "The", "iterative", "instance", "classifier", "refinement", "is", "implemented", "online", "using", "multiple", "streams", "in", "convolutional", "neural", "networks", ",", "where", "the", "first", "is", "an", "MIL", "network", "and", "the", "others", "are", "for", "instance", "classifier", "refinement", "supervised", "by", "the", "preceding", "one", ".", "Experiments", "are", "conducted", "on", "the", "PASCAL", "VOC", ",", "ImageNet", "detection", ",", "and", "MS", "-", "COCO", "benchmarks", "for", "WSOD", ".", "Results", "show", "that", "our", "method", "outperforms", "the", "previous", "state", "of", "the", "art", "significantly", ".", "Object", "detection", ",", "weakly", "supervised", "learning", ",", "convolutional", "neural", "network", ",", "multiple", "instance", "learning", ",", "proposal", "cluster", ".", "section", ":", "Introduction", "Object", "detection", "is", "one", "of", "the", "most", "important", "problems", "in", "computer", "vision", "with", "many", "applications", ".", "Recently", ",", "due", "to", "the", "development", "of", "Convolutional", "Neural", "Network", "(", "CNN", ")", "and", "the", "availability", "of", "large", "scale", "datasets", "with", "detailed", "boundingbox", "-", "level", "annotations", ",", "there", "have", "been", "great", "leap", "forwards", "in", "object", "detection", ".", "However", ",", "it", "is", "very", "labor", "-", "intensive", "and", "time", "-", "consuming", "to", "collect", "detailed", "annotations", ",", "whereas", "acquiring", "images", "with", "only", "image", "-", "level", "annotations", "(", "i.e", ".", ",", "image", "tags", ")", "indicating", "whether", "an", "object", "class", "exists", "in", "an", "image", "or", "not", "is", "much", "easier", ".", "For", "example", ",", "we", "can", "use", "image", "search", "queries", "to", "search", "on", "the", "Internet", "(", "e.g", ".", ",", "Google", "and", "Flickr", ")", "to", "obtain", "a", "mass", "of", "images", "with", "such", "image", "-", "level", "annotations", ".", "This", "fact", "inspires", "us", "to", "explore", "methods", "for", "the", "Weakly", "Supervised", "Object", "Detection", "(", "WSOD", ")", "problem", ",", "i.e", ".", ",", "training", "object", "detectors", "with", "only", "image", "tag", "supervisions", ".", "Many", "previous", "methods", "follow", "the", "Multiple", "Instance", "Learning", "(", "MIL", ")", "pipeline", "for", "WSOD", ".", "They", "treat", "images", "as", "bags", "and", "proposals", "as", "instances", ";", "then", "instance", "classifiers", "(", "object", "detectors", ")", "are", "trained", "under", "MIL", "constraints", "(", "i.e", ".", ",", "a", "positive", "bag", "contains", "at", "least", "one", "positive", "instance", "and", "all", "instances", "in", "negative", "bags", "are", "negative", ")", ".", "In", "addition", ",", "inspired", "by", "the", "great", "success", "of", "CNN", ",", "recent", "efforts", "often", "combine", "MIL", "and", "CNN", "to", "obtain", "better", "WSOD", "performance", ".", "Some", "researches", "have", "shown", "that", "treating", "CNNs", "pre", "-", "trained", "on", "large", "scale", "datasets", "as", "off", "-", "the", "-", "shelf", "proposal", "feature", "extractors", "can", "obtain", "much", "better", "performance", "than", "traditional", "hand", "-", "designed", "features", ".", "Moreover", ",", "many", "recent", "works", "have", "achieved", "even", "better", "results", "for", "WSOD", "by", "an", "MIL", "network", "using", "standard", "end", "-", "to", "-", "end", "training", "or", "a", "variant", "of", "end", "-", "to", "-", "end", "training", ".", "See", "Section", "[", "reference", "]", "for", "this", "variant", "of", "end", "-", "to", "-", "end", "and", "how", "it", "differs", "from", "the", "standard", "one", ".", "We", "use", "the", "same", "strategy", "of", "training", "a", "variant", "of", "end", "-", "to", "-", "end", "MIL", "network", "inspired", "by", ".", "Although", "some", "promising", "results", "have", "been", "obtained", "by", "MIL", "networks", "for", "WSOD", ",", "they", "do", "not", "perform", "as", "well", "as", "fully", "supervised", "ones", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "a", ")", ",", "previous", "MIL", "networks", "integrate", "the", "MIL", "constraints", "into", "the", "network", "training", "by", "transferring", "the", "instance", "classification", "(", "object", "detection", ")", "problem", "to", "a", "bag", "classification", "(", "image", "classification", ")", "problem", ",", "where", "the", "final", "image", "scores", "are", "the", "aggregation", "of", "the", "proposal", "scores", ".", "However", ",", "there", "is", "a", "big", "gap", "between", "image", "classification", "and", "object", "detection", ".", "For", "classification", ",", "even", "parts", "of", "objects", "can", "contribute", "to", "correct", "results", "(", "e.g", ".", ",", "the", "red", "boxes", "in", "Fig", ".", "[", "reference", "]", ")", ",", "because", "important", "parts", "include", "many", "characteristics", "of", "the", "objects", ".", "Many", "proposals", "only", "cover", "parts", "of", "objects", ",", "and", "“", "seeing", "”", "proposals", "only", "of", "parts", "may", "be", "enough", "to", "roughly", "localize", "the", "objects", ".", "But", "this", "may", "not", "localize", "objects", "well", "enough", "considering", "the", "performance", "requirement", "of", "high", "Intersection", "-", "over", "-", "Union", "(", "IoU", ")", "between", "the", "resulting", "boxes", "and", "groundtruth", "boundingboxes", ":", "the", "top", "ranking", "proposals", "may", "only", "localize", "parts", "of", "objects", "instead", "of", "whole", "objects", ".", "Recall", "that", "for", "detection", ",", "the", "resulting", "boxes", "should", "not", "only", "give", "correct", "classification", ",", "but", "also", "localize", "objects", "and", "have", "enough", "overlap", "with", "groundtruth", "boundingboxes", "(", "e.g", ".", ",", "the", "green", "boxes", "in", "Fig", ".", "[", "reference", "]", ")", ".", "Before", "presenting", "our", "solution", "of", "the", "problem", "referred", "above", ",", "we", "first", "introduce", "the", "concept", "of", "proposal", "cluster", ".", "Object", "detection", "requires", "algorithms", "to", "generate", "multiple", "overlapping", "proposals", "closely", "surrounding", "objects", "to", "ensure", "high", "proposal", "recall", "(", "e.g", ".", ",", "for", "each", "object", ",", "there", "are", "tens", "of", "proposals", "on", "average", "from", "Selective", "Search", "which", "have", "IoU", "0.5", "with", "the", "groundtruth", "boundingbox", "on", "the", "PASCAL", "VOC", "dataset", ")", ".", "Object", "proposals", "in", "an", "image", "can", "be", "grouped", "into", "different", "spatial", "clusters", ".", "Except", "for", "one", "cluster", "for", "background", "proposals", ",", "each", "object", "cluster", "is", "associated", "with", "a", "single", "object", "and", "proposals", "in", "each", "cluster", "are", "spatially", "adjacent", ",", "as", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "For", "fully", "supervised", "object", "detection", "(", "i.e", ".", ",", "training", "object", "detectors", "using", "boundingbox", "-", "level", "annotations", ")", ",", "proposal", "clusters", "can", "be", "generated", "by", "treating", "the", "groundtruth", "boundingboxes", "as", "cluster", "centers", ".", "Then", "object", "detectors", "are", "trained", "according", "to", "the", "proposal", "clusters", "(", "e.g", ".", ",", "assigning", "all", "proposals", "the", "label", "of", "the", "corresponding", "object", "class", "for", "each", "cluster", ")", ".", "This", "alleviates", "the", "problem", "that", "detectors", "may", "only", "focus", "on", "parts", ".", "But", "in", "the", "weakly", "supervised", "scenario", ",", "it", "is", "difficult", "to", "generate", "proposal", "clusters", "because", "groundtruth", "boundingboxes", "that", "can", "be", "used", "as", "cluster", "centers", "are", "not", "provided", ".", "To", "cope", "with", "this", "difficulty", ",", "we", "suggest", "to", "find", "proposal", "clusters", "as", "follows", ".", "First", "we", "generate", "proposal", "cluster", "centers", "from", "those", "proposals", "which", "have", "high", "classification", "scores", "during", "training", ",", "because", "these", "top", "ranking", "proposals", "can", "always", "detect", "at", "least", "parts", "of", "objects", ".", "That", "is", ",", "for", "each", "image", ",", "after", "obtaining", "proposal", "scores", ",", "we", "select", "some", "proposals", "with", "high", "scores", "as", "cluster", "centers", ",", "and", "then", "proposal", "clusters", "are", "generated", "based", "on", "spatial", "overlaps", "with", "the", "cluster", "centers", ".", "Then", "the", "problem", "reduces", "to", "how", "to", "select", "proposals", "as", "centers", ",", "because", "many", "high", "scoring", "proposals", "may", "correspond", "to", "the", "same", "object", ".", "The", "most", "straightforward", "way", "is", "to", "choose", "the", "proposal", "with", "the", "highest", "score", "for", "each", "positive", "object", "class", "(", "i.e", ".", ",", "the", "object", "class", "exists", "in", "the", "image", ")", "as", "the", "center", ".", "But", "such", "a", "method", "ignores", "the", "fact", "that", "there", "may", "exist", "more", "than", "one", "object", "with", "the", "same", "object", "category", "in", "natural", "images", "(", "e.g", ".", ",", "the", "two", "motorbikes", "in", "Fig", ".", "[", "reference", "]", ")", ".", "Therefore", ",", "we", "propose", "a", "graph", "-", "based", "method", "to", "find", "cluster", "centers", ".", "More", "specifically", ",", "we", "build", "a", "graph", "of", "top", "ranking", "proposals", "according", "to", "the", "spatial", "similarity", "for", "each", "positive", "object", "class", ".", "In", "the", "graph", ",", "two", "proposals", "are", "connected", "if", "they", "have", "enough", "spatial", "overlaps", ".", "Then", "we", "greedily", "and", "iteratively", "choose", "the", "proposals", "which", "have", "most", "connections", "with", "others", "to", "estimate", "the", "centers", ".", "Although", "a", "cluster", "center", "proposal", "may", "only", "capture", "an", "object", "partially", ",", "its", "adjacent", "proposals", "(", "i.e", ".", ",", "other", "proposals", "in", "the", "cluster", ")", "can", "cover", "the", "whole", "object", ",", "or", "at", "worst", "contain", "larger", "parts", "of", "the", "object", ".", "Based", "on", "these", "proposal", "clusters", ",", "we", "propose", "two", "methods", "to", "refine", "instance", "classifiers", "(", "object", "detectors", ")", "during", "training", ".", "We", "first", "propose", "to", "assign", "proposals", "object", "labels", "directly", ".", "That", "is", ",", "for", "each", "cluster", ",", "we", "assign", "its", "proposals", "the", "label", "of", "its", "corresponding", "object", "class", ",", "as", "in", "Fig", ".", "[", "reference", "]", "(", "b", ")", ".", "Compared", "with", "the", "conventional", "MIL", "network", "in", "Fig", ".", "[", "reference", "]", "(", "a", ")", ",", "this", "strategy", "forces", "network", "to", "“", "see", "”", "larger", "parts", "of", "objects", "by", "assigning", "object", "labels", "to", "proposals", "that", "cover", "larger", "parts", "of", "objects", "directly", ",", "which", "fills", "the", "gap", "between", "classification", "and", "detection", "to", "some", "extent", ".", "While", "effective", ",", "this", "strategy", "still", "has", "potential", "ambiguities", ",", "because", "assigning", "the", "same", "object", "label", "to", "proposals", "that", "cover", "different", "parts", "of", "objects", "simultaneously", "may", "confuse", "the", "network", "and", "will", "hurt", "the", "discriminative", "power", "of", "the", "detector", ".", "To", "address", "this", "problem", ",", "we", "propose", "to", "treat", "each", "proposal", "cluster", "as", "a", "small", "new", "bag", "to", "train", "refined", "instance", "classifiers", ",", "as", "in", "Fig", ".", "[", "reference", "]", "(", "c", ")", ".", "Most", "of", "the", "proposals", "in", "these", "new", "bags", "should", "have", "relatively", "high", "classification", "scores", "because", "the", "cluster", "centers", "covers", "at", "least", "parts", "of", "objects", "and", "proposals", "in", "the", "same", "cluster", "are", "spatially", "adjacent", "(", "except", "for", "the", "background", "cluster", ")", ".", "In", "the", "same", "time", ",", "not", "all", "proposals", "in", "the", "bags", "should", "have", "high", "classification", "scores", ".", "Thus", "compared", "with", "the", "directly", "assigning", "label", "strategy", ",", "this", "strategy", "is", "more", "flexible", "and", "can", "reduce", "the", "ambiguities", "to", "some", "extent", ".", "We", "name", "our", "method", "Proposal", "Cluster", "Learning", "(", "PCL", ")", "because", "it", "learns", "refined", "instance", "classifiers", "based", "on", "proposal", "clusters", ".", "To", "implement", "our", "idea", "effectively", "and", "efficiently", ",", "we", "further", "propose", "an", "online", "training", "approach", ".", "Our", "network", "has", "multiple", "output", "streams", "as", "in", "Fig", ".", "[", "reference", "]", ".", "The", "first", "stream", "is", "a", "basic", "MIL", "network", "which", "aggregates", "proposal", "scores", "into", "final", "image", "scores", "to", "train", "basic", "instance", "classifiers", ",", "and", "the", "other", "streams", "refine", "the", "instance", "classifiers", "iteratively", ".", "During", "the", "forward", "process", "of", "training", ",", "proposal", "classification", "scores", "are", "obtained", "and", "proposal", "clusters", "are", "generated", "consequently", "for", "each", "stream", ".", "Then", "based", "on", "these", "proposal", "clusters", ",", "supervisions", "are", "generated", "to", "compute", "losses", "for", "the", "next", "stream", ".", "According", "to", "the", "losses", ",", "these", "refined", "classifiers", "are", "trained", "during", "back", "-", "propagation", ".", "Except", "for", "the", "first", "stream", "that", "is", "supervised", "by", "image", "labels", ",", "the", "other", "streams", "are", "supervised", "by", "the", "image", "labels", "as", "well", "as", "outputs", "from", "their", "preceding", "streams", ".", "As", "our", "method", "forces", "the", "network", "to", "“", "see", "”", "larger", "parts", "of", "objects", ",", "the", "detector", "can", "discover", "the", "whole", "object", "instead", "of", "parts", "gradually", "by", "performing", "refinement", "multiple", "times", "(", "i.e", ".", ",", "multiple", "output", "streams", ")", ".", "But", "at", "the", "start", "of", "training", ",", "all", "classifiers", "are", "almost", "untrained", ",", "which", "will", "result", "in", "very", "noisy", "proposal", "clusters", ",", "and", "so", "the", "training", "will", "deviate", "from", "the", "correct", "solutions", "a", "lot", ".", "Thus", "we", "design", "a", "weighted", "loss", "further", "by", "associating", "different", "proposals", "with", "different", "weights", "in", "different", "training", "iterations", ".", "After", "that", ",", "all", "training", "procedures", "can", "thus", "be", "integrated", "into", "a", "single", "end", "-", "to", "-", "end", "network", ".", "This", "can", "improve", "the", "performance", "benefiting", "from", "our", "PCL", "-", "based", "classifier", "refinement", "procedure", ".", "It", "is", "also", "very", "computational", "efficient", "in", "both", "training", "and", "testing", ".", "In", "addition", ",", "performance", "can", "be", "improved", "by", "sharing", "proposal", "features", "among", "different", "output", "streams", ".", "We", "elaborately", "conduct", "many", "experiments", "on", "the", "challenging", "PASCAL", "VOC", ",", "ImageNet", "detection", ",", "and", "MS", "-", "COCO", "datasets", "to", "confirm", "the", "effectiveness", "of", "our", "method", ".", "Our", "method", "achieves", "mAP", "and", "CorLoc", "on", "VOC", "2007", "which", "is", "more", "than", "absolute", "improvement", "compared", "with", "previous", "best", "performed", "methods", ".", "This", "paper", "is", "an", "extended", "version", "of", "our", "previous", "work", ".", "In", "particular", ",", "we", "give", "more", "analyses", "of", "our", "method", "and", "enrich", "literatures", "of", "most", "recent", "related", "works", ",", "making", "the", "manuscript", "more", "complete", ".", "In", "addition", ",", "we", "make", "two", "methodological", "improvements", ":", "the", "first", "one", "is", "to", "generate", "proposal", "clusters", "using", "graphs", "of", "top", "ranking", "proposals", "instead", "of", "using", "the", "highest", "scoring", "proposal", ",", "and", "the", "second", "one", "is", "to", "treat", "each", "proposal", "cluster", "as", "a", "small", "new", "bag", ".", "In", "addition", ",", "we", "provide", "more", "discussions", "of", "experimental", "results", ",", "and", "show", "the", "effectiveness", "of", "our", "method", "on", "the", "challenging", "ImageNet", "detection", "and", "MS", "-", "COCO", "datasets", ".", "The", "rest", "of", "our", "paper", "is", "organized", "as", "follows", ".", "In", "Section", "[", "reference", "]", ",", "some", "related", "works", "are", "introduced", ".", "In", "Section", "[", "reference", "]", ",", "the", "details", "of", "our", "method", "are", "described", ".", "Elaborate", "experiments", "and", "analyses", "are", "conducted", "in", "Section", "[", "reference", "]", ".", "Finally", ",", "conclusions", "and", "future", "directions", "are", "presented", "in", "Section", "[", "reference", "]", ".", "section", ":", "Related", "work", "subsection", ":", "Multiple", "instance", "learning", "MIL", ",", "first", "proposed", "for", "drug", "activity", "prediction", ",", "is", "a", "classical", "weakly", "supervised", "learning", "problem", ".", "Many", "variants", "have", "been", "proposed", "for", "MIL", ".", "In", "MIL", ",", "a", "set", "of", "bags", "are", "given", ",", "and", "each", "bag", "is", "associated", "with", "a", "collection", "of", "instances", ".", "It", "is", "natural", "to", "treat", "WSOD", "as", "an", "MIL", "problem", ".", "Then", "the", "problem", "turns", "into", "finding", "instance", "classifiers", "only", "given", "bag", "labels", ".", "Our", "method", "also", "follows", "the", "MIL", "strategy", "and", "makes", "several", "improvements", "to", "WSOD", ".", "In", "particular", ",", "we", "learn", "refined", "instance", "classifiers", "based", "on", "proposal", "clusters", "according", "to", "both", "instance", "scores", "and", "spatial", "relations", "in", "an", "online", "manner", ".", "MIL", "has", "many", "applications", "to", "computer", "vision", ",", "such", "as", "image", "classification", ",", "weakly", "supervised", "semantic", "segmentation", ",", "object", "detection", ",", "object", "tracking", ",", "etc", ".", "The", "strategy", "of", "treating", "proposal", "clusters", "as", "bags", "was", "partly", "inspired", "by", ",", "where", "proposes", "to", "train", "MIL", "for", "patches", "around", "groundtruth", "locations", "and", "proposes", "to", "train", "MIL", "for", "patches", "around", "predicted", "object", "locations", ".", "However", ",", "they", "require", "groundtruth", "locations", "for", "either", "all", "training", "samples", "or", "the", "beginning", "time", "frames", ",", "whereas", "WSOD", "does", "not", "have", "such", "annotations", ".", "Therefore", ",", "it", "is", "much", "harder", "to", "generate", "proposal", "clusters", "only", "guided", "by", "image", "-", "level", "supervisions", "for", "WSOD", ".", "In", "addition", ",", "we", "incorporate", "the", "strategy", "of", "treating", "proposal", "clusters", "as", "bags", "into", "the", "network", "training", "whereas", "do", "not", ".", "Oquab", "et", "al", ".", "also", "train", "a", "CNN", "network", "using", "the", "max", "pooing", "MIL", "strategy", "to", "localize", "objects", ".", "But", "their", "methods", "can", "only", "coarsely", "localize", "objects", "regardless", "of", "their", "sizes", "and", "aspect", "ratios", ",", "whereas", "our", "method", "can", "detect", "objects", "more", "accurately", ".", "subsection", ":", "Weakly", "supervised", "object", "detection", "WSOD", "has", "attracted", "great", "interests", "nowadays", "because", "the", "amount", "of", "data", "with", "image", "-", "level", "annotations", "is", "much", "bigger", "and", "is", "growing", "much", "faster", "than", "that", "with", "boundingbox", "-", "level", "annotations", ".", "Many", "methods", "are", "emerging", "for", "the", "WSOD", "problem", ".", "For", "example", ",", "Chum", "and", "Zisserman", "first", "initialize", "object", "locations", "by", "discriminative", "visual", "words", "and", "then", "introduce", "an", "exemplar", "model", "to", "measure", "similarity", "between", "image", "pairs", "for", "updating", "locations", ".", "Deselaers", "et", "al", ".", "propose", "to", "initialize", "boxes", "by", "objectness", "and", "use", "a", "CRF", "-", "based", "model", "to", "iteratively", "localize", "objects", ".", "Pandey", "and", "Lazebnik", "train", "a", "DPM", "model", "under", "weak", "supervisions", "for", "WSOD", ".", "Shi", "et", "al", ".", "use", "Bayesian", "latent", "topic", "models", "to", "jointly", "model", "different", "object", "classes", "and", "background", ".", "Song", "et", "al", ".", "develop", "a", "technology", "to", "discover", "frequent", "discriminative", "configurations", "of", "visual", "patterns", "for", "robust", "WSOD", ".", "Cinbis", "et", "al", ".", "iteratively", "train", "a", "multi", "-", "fold", "MIL", "to", "avoid", "the", "detector", "being", "locked", "onto", "inaccurate", "local", "optima", ".", "Wang", "et", "al", ".", "relax", "the", "MIL", "constraints", "into", "a", "derivable", "loss", "function", "to", "train", "detectors", "more", "efficient", ".", "Recently", ",", "with", "the", "revolution", "of", "CNNs", "in", "computer", "vision", ",", "many", "works", "also", "try", "to", "combine", "the", "WSOD", "with", "CNNs", ".", "Early", "works", "treat", "CNN", "models", "pre", "-", "trained", "on", "ImageNet", "as", "off", "-", "the", "-", "shelf", "feature", "extractors", ".", "They", "extract", "CNN", "features", "for", "each", "candidate", "regions", ",", "and", "then", "train", "their", "own", "detectors", "on", "top", "of", "these", "features", ".", "These", "methods", "have", "shown", "that", "CNN", "descriptors", "can", "boost", "performance", "against", "traditional", "hand", "-", "designed", "features", ".", "More", "recent", "efforts", "tend", "to", "train", "end", "-", "to", "-", "end", "networks", "for", "WSOD", ".", "They", "integrate", "the", "MIL", "constraints", "into", "the", "network", "training", "by", "aggregating", "proposal", "classification", "scores", "into", "final", "image", "classification", "scores", ",", "and", "then", "image", "-", "level", "supervision", "can", "be", "directly", "added", "to", "image", "classification", "scores", ".", "For", "example", ",", "Tang", "et", "al", ".", "propose", "to", "use", "max", "pooling", "for", "aggregation", ".", "Bilen", "and", "Vedaldi", "develop", "a", "weighted", "sum", "pooing", "strategy", ".", "Building", "on", ",", "Kantorov", "et", "al", ".", "argue", "that", "context", "information", "can", "improve", "the", "performance", ".", "Diba", "et", "al", ".", "show", "that", "weakly", "supervised", "segmentation", "map", "can", "be", "used", "as", "guidance", "to", "filter", "proposals", ",", "and", "jointly", "train", "the", "weakly", "supervised", "segmentation", "network", "and", "WSOD", "end", "-", "to", "-", "end", ".", "Our", "method", "is", "built", "on", "these", "networks", "and", "any", "of", "them", "can", "be", "chosen", "as", "our", "basic", "network", ".", "Our", "strategy", "proposes", "to", "learn", "refined", "instance", "classifiers", "based", "on", "proposal", "clusters", ",", "and", "propose", "a", "novel", "online", "approach", "to", "train", "our", "network", "effectively", "and", "efficiently", ".", "Experimental", "results", "show", "our", "strategies", "can", "boost", "the", "results", "significantly", ".", "In", "addition", "to", "the", "weighted", "sum", "pooing", ",", "also", "proposes", "a", "“", "spatial", "regulariser", "”", "that", "forces", "features", "of", "the", "highest", "scoring", "proposal", "and", "its", "spatially", "adjacent", "proposals", "to", "be", "the", "same", ".", "Unlike", "this", ",", "we", "show", "that", "finding", "proposal", "cluster", "centers", "using", "graph", "and", "treating", "proposal", "clusters", "as", "bags", "are", "more", "effective", ".", "The", "contemporary", "work", "uses", "a", "graph", "model", "to", "generate", "seed", "proposals", ".", "Their", "network", "training", "has", "many", "steps", ":", "first", ",", "an", "MIL", "network", "is", "trained", ";", "second", ",", "seed", "proposals", "are", "generated", "using", "the", "graph", ";", "third", ",", "based", "on", "these", "seed", "proposals", ",", "a", "Fast", "R", "-", "CNN", "like", "detector", "is", "trained", ".", "Our", "method", "differs", "from", "in", "many", "aspects", ":", "first", ",", "we", "propose", "to", "generate", "proposal", "clusters", "for", "each", "training", "iteration", "and", "thus", "our", "network", "is", "trained", "end", "-", "to", "-", "end", "instead", "of", "step", "-", "by", "-", "step", ",", "which", "is", "more", "efficient", "and", "can", "benefit", "from", "sharing", "proposal", "features", "among", "different", "streams", ";", "second", ",", "we", "propose", "to", "treat", "proposal", "clusters", "as", "bags", "for", "training", "better", "classifiers", ".", "As", "evidenced", "by", "experiments", ",", "our", "method", "obtains", "much", "better", "and", "more", "robust", "results", ".", "subsection", ":", "End", "-", "to", "-", "end", "and", "its", "variants", "In", "standard", "end", "-", "to", "-", "end", "training", ",", "the", "update", "requires", "optimizing", "losses", "w.r.t", ".", "all", "functions", "of", "network", "parameters", ".", "For", "example", ",", "the", "Fast", "R", "-", "CNN", "optimizes", "their", "classification", "loss", "and", "boundingbox", "regression", "loss", "w.r.t", ".", "proposal", "classification", "and", "feature", "extraction", "for", "fully", "supervised", "object", "detection", ".", "The", "MIL", "networks", "in", "optimize", "their", "MIL", "loss", "w.r.t", ".", "proposal", "classification", "and", "feature", "extraction", "for", "WSOD", ".", "Unlike", "the", "standard", "end", "-", "to", "-", "end", "training", ",", "there", "exists", "a", "variant", "of", "end", "-", "to", "-", "end", "training", ".", "The", "variant", "contains", "functions", "which", "depend", "on", "network", "parameters", ",", "but", "losses", "are", "not", "optimized", "w.r.t", ".", "all", "these", "functions", ".", "As", "we", "described", "in", "Section", "[", "reference", "]", ",", "the", "“", "spatial", "regulariser", "”", "in", "forces", "features", "of", "the", "highest", "scoring", "proposal", "and", "its", "spatially", "adjacent", "proposals", "to", "be", "the", "same", ".", "They", "use", "a", "function", "of", "network", "parameters", "to", "compute", "the", "highest", "scoring", "proposal", ",", "and", "do", "not", "optimize", "their", "losses", "w.r.t", ".", "this", "function", ".", "Diba", "et", "al", ".", "filter", "out", "background", "proposals", "using", "a", "function", "of", "network", "parameters", "and", "use", "these", "filtered", "proposals", "in", "their", "latter", "network", "computations", ".", "They", "also", "do", "not", "optimize", "their", "losses", "w.r.t", ".", "this", "function", ".", "Inspired", "by", ",", "we", "use", "this", "variant", "of", "end", "-", "to", "-", "end", "training", ".", "More", "precisely", ",", "we", "do", "not", "optimize", "our", "losses", "w.r.t", ".", "the", "generated", "supervisions", "for", "instance", "classifier", "refinement", ".", "subsection", ":", "Others", "There", "are", "many", "other", "important", "related", "works", "that", "do", "not", "focus", "on", "weakly", "supervised", "learning", "but", "should", "be", "discussed", ".", "Similar", "to", "other", "end", "-", "to", "-", "end", "MIL", "networks", ",", "our", "method", "is", "built", "on", "top", "of", "the", "Region", "of", "Interest", "(", "RoI", ")", "pooling", "layer", "or", "Spatial", "Pyramid", "Pooling", "(", "SPP", ")", "layer", "to", "share", "convolutional", "computations", "among", "different", "proposals", "for", "model", "acceleration", ".", "But", "both", "and", "require", "boundingbox", "-", "level", "annotations", "to", "train", "their", "detectors", ".", "The", "sharing", "proposal", "feature", "strategy", "in", "our", "network", "is", "similar", "to", "multi", "-", "task", "learning", ".", "Unlike", "the", "multi", "-", "task", "learning", "that", "each", "output", "stream", "has", "their", "own", "relatively", "independent", "external", "supervisions", "for", "different", "tasks", ",", "in", "our", "method", ",", "all", "streams", "have", "the", "same", "task", "and", "supervisions", "of", "later", "streams", "depend", "on", "the", "outputs", "from", "their", "preceding", "streams", ".", "section", ":", "Method", "The", "overall", "architecture", "of", "our", "method", "is", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "Given", "an", "image", ",", "about", "object", "proposals", "from", "Selective", "Search", "or", "EdgeBox", "are", "generated", ".", "During", "the", "forward", "process", "of", "training", ",", "the", "image", "and", "these", "proposals", "are", "fed", "into", "some", "convolutional", "(", "conv", ")", "layers", "with", "an", "SPP", "layer", "to", "produce", "a", "fixed", "-", "size", "conv", "feature", "map", "per", "-", "proposal", ".", "After", "that", ",", "proposal", "feature", "maps", "are", "fed", "into", "two", "fully", "connected", "(", "fc", ")", "layers", "to", "produce", "proposal", "features", ".", "These", "features", "are", "branched", "into", "different", "streams", ":", "the", "first", "one", "is", "an", "MIL", "network", "to", "train", "basic", "instance", "classifiers", "and", "the", "others", "refine", "the", "classifiers", "iteratively", ".", "For", "each", "stream", ",", "proposal", "classification", "scores", "are", "obtained", "and", "proposal", "clusters", "are", "generated", "consequently", ".", "Then", "based", "on", "these", "proposal", "clusters", ",", "supervisions", "are", "generated", "to", "compute", "losses", "for", "the", "next", "stream", ".", "During", "the", "back", "-", "propagation", "process", "of", "training", ",", "the", "network", "losses", "are", "optimized", "to", "train", "proposal", "features", "and", "classifiers", ".", "As", "shown", "in", "the", "figure", ",", "supervisions", "of", "the", "-", "st", "refined", "classifier", "depend", "on", "the", "output", "from", "the", "basic", "classifier", ",", "and", "supervisions", "of", "-", "th", "refined", "classifier", "depend", "on", "outputs", "from", "-", "th", "refined", "classifier", ".", "In", "this", "section", ",", "we", "will", "introduce", "our", "method", "of", "learning", "refined", "instance", "classifiers", "based", "on", "proposal", "clusters", "in", "detail", ".", "subsection", ":", "Notations", "Before", "presenting", "our", "method", ",", "we", "first", "introduce", "some", "of", "the", "mostly", "used", "notations", "as", "follows", ".", "We", "have", "proposals", "with", "boxes", "for", "an", "given", "image", "and", "proposal", "features", ",", "where", "is", "the", "-", "th", "proposal", "box", ".", "The", "number", "of", "refined", "instance", "classifiers", "is", "(", "i.e", ".", ",", "we", "refine", "instance", "classifier", "times", ")", ",", "and", "thus", "there", "are", "streams", ".", "The", "number", "of", "object", "classes", "is", ".", "and", "are", "the", "parameters", "of", "the", "basic", "instance", "classifier", "and", "the", "-", "th", "refined", "instance", "classifier", ",", "respectively", ".", "and", "are", "the", "predicted", "score", "matrices", "of", "the", "basic", "instance", "classifier", "and", "the", "-", "th", "refined", "instance", "classifier", ",", "respectively", ",", "where", "indicates", "the", "object", "classes", "and", "background", "class", ".", "We", "use", "later", "for", "simplification", ",", "dropping", "the", "dependence", "on", ".", "is", "the", "predicted", "score", "of", "the", "-", "th", "proposal", "for", "class", "from", "the", "-", "th", "instance", "classifier", ".", "is", "the", "image", "label", "vector", ",", "where", "or", "indicates", "the", "image", "with", "or", "without", "object", "class", ".", "is", "the", "supervision", "of", "the", "-", "th", "instance", "classifier", ",", "where", "is", "the", "image", "label", "vector", ".", "is", "the", "loss", "function", "to", "train", "the", "-", "th", "instance", "classifier", ".", "We", "compute", "proposal", "cluster", "centers", "for", "the", "-", "th", "refinement", ".", "The", "-", "th", "cluster", "center", "consists", "of", "a", "proposal", "box", ",", "an", "object", "label", "(", "indicates", "the", "-", "th", "object", "class", ")", ",", "and", "a", "confidence", "score", "indicating", "the", "confidence", "that", "covers", "at", "least", "part", "of", "an", "object", "of", "class", ".", "We", "have", "proposal", "clusters", "according", "to", "(", "for", "background", "and", "others", "for", "objects", ")", ".", "For", "object", "clusters", ",", "the", "-", "th", "cluster", "consists", "of", "proposal", "boxes", ",", "an", "object", "label", "that", "is", "the", "same", "as", "the", "cluster", "center", "label", ",", "and", "a", "confidence", "score", "that", "is", "the", "same", "as", "the", "cluster", "center", "score", ",", "where", "indicates", "the", "confidence", "that", "corresponds", "to", "an", "object", "of", "class", ".", "Unlike", "object", "clusters", ",", "the", "background", "cluster", "consists", "of", "proposals", "and", "a", "label", "indicating", "the", "background", ".", "The", "-", "th", "proposal", "consists", "of", "a", "proposal", "box", "and", "a", "confidence", "score", "indicating", "the", "confidence", "that", "is", "the", "background", ".", "subsection", ":", "Basic", "MIL", "network", "It", "is", "necessary", "to", "generate", "proposal", "scores", "and", "clusters", "to", "supervise", "refined", "instance", "classifiers", ".", "More", "specifically", ",", "the", "first", "refined", "classifier", "requires", "basic", "instance", "classifiers", "to", "generate", "proposal", "scores", "and", "clusters", ".", "Therefore", ",", "we", "first", "introduce", "our", "basic", "MIL", "network", "as", "the", "basic", "instance", "classifier", ".", "Our", "overall", "network", "is", "independent", "of", "the", "specific", "MIL", "methods", ",", "and", "thus", "any", "method", "that", "can", "be", "trained", "end", "-", "to", "-", "end", "could", "be", "used", ".", "There", "are", "many", "possible", "choices", ".", "Here", "we", "choose", "the", "method", "by", "Bilen", "and", "Vedaldi", "which", "proposes", "a", "weighted", "sum", "pooling", "strategy", "to", "obtain", "the", "instance", "classifier", ",", "because", "of", "its", "effectiveness", "and", "implementation", "convenience", ".", "To", "make", "our", "paper", "self", "-", "contained", ",", "we", "briefly", "introduce", "as", "follows", ".", "Given", "an", "input", "image", "and", "its", "proposal", "boxes", ",", "a", "set", "of", "proposal", "features", "are", "first", "generated", "by", "the", "network", ".", "Then", "as", "shown", "in", "the", "“", "Basic", "MIL", "network", "”", "block", "of", "Fig", ".", "[", "reference", "]", ",", "there", "are", "two", "branches", "which", "process", "the", "proposal", "features", "to", "produce", "two", "matrices", "(", "we", "use", "later", "for", "simplification", ",", "dropping", "the", "dependence", "on", ")", "of", "an", "input", "image", "by", "two", "fc", "layers", ",", "where", "and", "denote", "the", "parameters", "of", "the", "fc", "layer", "for", "and", "the", "parameters", "of", "the", "fc", "layer", "for", ",", "respectively", ".", "Then", "the", "two", "matrices", "are", "passed", "through", "two", "softmax", "layer", "along", "different", "directions", ":", "and", ".", "Let", "us", "denote", "by", ".", "The", "proposal", "scores", "are", "generated", "by", "element", "-", "wise", "product", ".", "Finally", ",", "the", "image", "score", "of", "the", "-", "th", "class", "is", "obtained", "by", "the", "sum", "over", "all", "proposals", ":", ".", "A", "simple", "interpretation", "of", "the", "two", "branches", "framework", "is", "as", "follows", ".", "is", "the", "probability", "of", "the", "-", "th", "proposal", "belonging", "to", "class", ".", "is", "the", "normalized", "weight", "that", "indicates", "the", "contribution", "of", "the", "-", "th", "proposal", "to", "image", "being", "classified", "to", "class", ".", "So", "is", "obtained", "by", "weighted", "sum", "pooling", "and", "falls", "in", "the", "range", "of", ".", "Given", "the", "image", "label", "vector", ".", "We", "train", "the", "basic", "instance", "classifier", "by", "optimizing", "the", "multi", "-", "class", "cross", "entropy", "loss", "Eq", ".", "(", "[", "reference", "]", ")", "w.r.t", ".", ".", "subsection", ":", "The", "overall", "training", "strategy", "To", "refine", "instance", "classifiers", "iteratively", ",", "we", "add", "multiple", "output", "streams", "in", "our", "network", "where", "each", "stream", "corresponds", "to", "a", "refined", "classifier", ",", "as", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "We", "integrate", "the", "basic", "MIL", "network", "and", "the", "classifier", "refinement", "into", "an", "end", "-", "to", "-", "end", "network", "to", "learn", "the", "refined", "classifier", "online", ".", "Unlike", "the", "basic", "instance", "classifier", ",", "for", "an", "input", "image", "the", "output", "score", "matrix", "of", "the", "-", "th", "refined", "classifier", "is", "a", "matrix", "and", "is", "obtained", "by", "passing", "the", "proposal", "features", "through", "a", "single", "fc", "layer", "(", "with", "parameters", ")", "as", "well", "as", "a", "softmax", "over", "-", "classes", "layer", ",", "i.e", ".", ",", ",", "as", "in", "the", "“", "Instance", "classifier", "refinement", "”", "blocks", "of", "Fig", ".", "[", "reference", "]", ".", "Notice", "that", "we", "use", "the", "same", "proposal", "features", "for", "all", "classifiers", ".", "We", "use", "later", "for", "simplification", ",", "dropping", "the", "dependence", "on", ".", "As", "we", "stated", "before", ",", "supervisions", "to", "train", "the", "-", "th", "instance", "classifier", "are", "generated", "based", "on", "proposal", "scores", "and", "image", "label", ".", "Thus", "we", "denote", "the", "supervisions", "by", ".", "Then", "we", "train", "our", "overall", "network", "by", "optimizing", "the", "loss", "Eq", ".", "(", "[", "reference", "]", ")", "w.r.t", ".", ".", "We", "do", "not", "optimize", "the", "loss", "w.r.t", ".", ",", "which", "means", "that", "the", "supervisions", "are", "only", "computed", "in", "the", "forward", "process", "and", "we", "do", "not", "compute", "their", "gradients", "to", "train", "our", "network", ".", "The", "loss", "for", "the", "-", "th", "refined", "instance", "classifier", "is", "defined", "in", "later", "Eq", ".", "(", "[", "reference", "]", ")", "/", "(", "[", "reference", "]", ")", "/", "(", "[", "reference", "]", ")", "which", "are", "loss", "functions", "with", "supervisions", "provided", "by", ".", "We", "will", "give", "details", "about", "how", "to", "get", "supervisions", "and", "loss", "functions", "in", "Section", "[", "reference", "]", ".", "[", "t", "]", "The", "overall", "training", "procedure", "(", "one", "iteration", ")", "[", "1", "]", "An", "image", ",", "its", "proposal", "boxes", ",", "and", "its", "image", "label", "vector", ";", "refinement", "times", ".", "An", "updated", "network", ".", "Feed", "the", "image", "and", "into", "the", "network", "to", "produce", "proposal", "score", "matrices", "(", "simplified", "as", "later", ")", ".", "Compute", "loss", "by", "Eq", ".", "(", "[", "reference", "]", ")", ",", "see", "Section", "[", "reference", "]", ".", "Generate", "supervisions", ",", "see", "Section", "[", "reference", "]", ".", "Compute", "loss", "by", "Eq", ".", "(", "[", "reference", "]", ")", "/", "(", "[", "reference", "]", ")", "/", "(", "[", "reference", "]", ")", ",", "see", "Section", "[", "reference", "]", ".", "Optimize", ",", "i.e", ".", ",", "Eq", ".", "(", "[", "reference", "]", ")", ",", "w.r.t", ".", "(", "not", "w.r.t", ".", ")", ".", "During", "the", "forward", "process", "of", "each", "Stochastic", "Gradient", "Descent", "(", "SGD", ")", "training", "iteration", ",", "we", "obtain", "a", "set", "of", "proposal", "scores", "of", "an", "input", "image", ".", "Accordingly", ",", "we", "generate", "the", "supervisions", "for", "the", "iteration", "to", "compute", "the", "loss", "Eq", ".", "(", "[", "reference", "]", ")", ".", "During", "the", "back", "-", "propagation", "process", "of", "each", "SGD", "training", "iteration", ",", "we", "optimize", "the", "loss", "Eq", ".", "(", "[", "reference", "]", ")", "w.r.t", ".", "proposal", "features", "and", "classifiers", ".", "We", "summarize", "this", "procedure", "in", "Algorithm", "[", "reference", "]", ".", "Note", "that", "we", "do", "not", "use", "an", "alternating", "training", "strategy", ",", "i.e", ".", ",", "fixing", "supervisions", "and", "training", "a", "complete", "model", ",", "fixing", "the", "model", "and", "updating", "supervisions", ".", "The", "reasons", "are", "that", ":", "1", ")", "it", "is", "very", "time", "-", "consuming", "because", "it", "requires", "training", "models", "multiple", "times", ";", "2", ")", "training", "different", "models", "in", "different", "refinement", "steps", "separately", "may", "harm", "the", "performance", "because", "it", "hinders", "the", "process", "to", "benefit", "from", "the", "shared", "proposal", "features", "(", "i.e", ".", ",", ")", ".", "subsection", ":", "Proposal", "cluster", "learning", "Here", "we", "will", "introduce", "our", "methods", "to", "learn", "refined", "instance", "classifiers", "based", "on", "proposal", "clusters", "(", "i.e", ".", ",", "proposal", "cluster", "learning", ")", ".", "Recall", "from", "Section", "[", "reference", "]", "that", "we", "have", "a", "set", "of", "proposals", "with", "boxes", ".", "For", "the", "-", "th", "refinement", ",", "our", "goal", "is", "to", "generate", "supervisions", "for", "the", "loss", "functions", "using", "the", "proposal", "scores", "and", "image", "label", "in", "each", "training", "iteration", ".", "We", "use", "later", "for", "simplification", ",", "dropping", "the", "dependence", "on", ".", "We", "do", "this", "in", "three", "steps", ".", "1", ")", "We", "find", "proposal", "cluster", "centers", "which", "are", "proposals", "corresponding", "to", "different", "objects", ".", "2", ")", "We", "group", "the", "remaining", "proposals", "into", "different", "clusters", ",", "where", "each", "cluster", "is", "associated", "with", "a", "cluster", "center", "or", "corresponds", "to", "the", "background", ".", "3", ")", "We", "generate", "the", "supervisions", "for", "the", "loss", "functions", ",", "enabling", "us", "to", "train", "the", "refined", "instance", "classifiers", ".", "For", "the", "first", "step", ",", "we", "compute", "proposal", "cluster", "centers", "based", "on", "and", ".", "The", "-", "th", "cluster", "center", "is", "defined", "in", "Section", "[", "reference", "]", ".", "We", "propose", "two", "algorithms", "to", "find", "in", "Section", "[", "reference", "]", "(", "1", ")", "and", "(", "2", ")", "(", "also", "Algorithm", "[", "reference", "]", "and", "Algorithm", "[", "reference", "]", ")", ",", "where", "the", "first", "one", "was", "proposed", "in", "the", "conference", "version", "paper", "and", "the", "second", "one", "is", "proposed", "in", "this", "paper", ".", "For", "the", "second", "step", ",", "according", "to", "the", "proposal", "cluster", "centers", ",", "proposal", "clusters", "are", "generated", "(", "for", "background", "and", "others", "for", "objects", ")", ".", "The", "-", "th", "object", "cluster", "and", "the", "background", "cluster", "are", "defined", "in", "Section", "[", "reference", "]", ".", "We", "use", "the", "different", "notation", "for", "the", "background", "cluster", "because", "background", "proposals", "are", "scattered", "in", "each", "image", ",", "and", "thus", "it", "is", "hard", "to", "determine", "a", "cluster", "center", "and", "accordingly", "a", "cluster", "score", ".", "The", "method", "to", "generate", "was", "proposed", "in", "the", "conference", "version", "paper", "and", "is", "described", "in", "Section", "[", "reference", "]", "(", "also", "Algorithm", "[", "reference", "]", ")", ".", "For", "the", "third", "step", ",", "supervisions", "to", "train", "the", "-", "th", "refined", "instance", "classifier", "are", "generated", "based", "on", "the", "proposal", "clusters", ".", "We", "use", "two", "strategies", "where", "are", "either", "proposal", "-", "level", "labels", "indicating", "whether", "a", "proposal", "belongs", "to", "an", "object", "class", ",", "or", "cluster", "-", "level", "labels", "that", "treats", "each", "proposal", "cluster", "as", "a", "bag", ".", "Subsequently", "these", "are", "used", "to", "compute", "the", "loss", "functions", ".", "We", "propose", "two", "approaches", "to", "do", "this", "as", "described", "in", "Section", "[", "reference", "]", "(", "1", ")", "and", "(", "2", ")", ",", "where", "the", "first", "one", "was", "proposed", "in", "the", "conference", "version", "paper", "and", "the", "second", "one", "is", "proposed", "in", "this", "paper", ".", "subsubsection", ":", "Finding", "proposal", "cluster", "centers", "In", "the", "following", "we", "introduce", "two", "algorithms", "to", "find", "proposal", "cluster", "centers", ".", "[", "t", "]", "Finding", "proposal", "cluster", "centers", "using", "the", "highest", "scoring", "proposal", "[", "1", "]", "Proposal", "boxes", ";", "image", "label", "vector", ";", "proposal", "score", "matrix", ".", "Proposal", "cluster", "centers", ".", "Initialize", ".", "Choose", "the", "-", "th", "proposal", "by", "Eq", ".", "(", "[", "reference", "]", ")", ".", ".", "(", "1", ")", "Finding", "proposal", "cluster", "centers", "using", "the", "highest", "scoring", "proposal", ".", "A", "solution", "for", "finding", "proposal", "cluster", "centers", "is", "to", "choose", "the", "highest", "scoring", "proposal", ",", "as", "in", "our", "conference", "version", "paper", ".", "As", "in", "Algorithm", "[", "reference", "]", ",", "suppose", "an", "image", "has", "object", "class", "label", "(", "i.e", ".", ",", ")", ".", "For", "the", "-", "th", "refinement", ",", "we", "first", "select", "the", "-", "th", "proposal", "which", "has", "the", "highest", "score", "by", "Eq", ".", "(", "[", "reference", "]", ")", ",", "where", "is", "the", "predicted", "score", "of", "the", "-", "th", "proposal", ",", "as", "defined", "in", "Section", "[", "reference", "]", ".", "Then", "this", "proposal", "is", "chosen", "as", "the", "cluster", "center", ",", "i.e", ".", ",", ",", "where", "is", "the", "box", "of", "the", "-", "th", "proposal", ".", "is", "chosen", "as", "the", "confidence", "score", "that", "the", "-", "th", "proposal", "covers", "at", "least", "part", "of", "an", "object", "of", "class", ",", "because", "is", "the", "predicted", "score", "of", "the", "-", "th", "proposal", "been", "categorized", "to", "class", ".", "Therefore", ",", "the", "highest", "scoring", "proposal", "can", "probably", "cover", "at", "least", "part", "of", "the", "object", "and", "thus", "be", "chosen", "as", "the", "cluster", "center", ".", "There", "is", "a", "potential", "problem", "that", "one", "proposal", "may", "be", "chosen", "as", "the", "cluster", "centers", "for", "multiple", "object", "classes", ".", "To", "avoid", "this", "problem", ",", "if", "one", "proposal", "corresponds", "to", "the", "cluster", "centers", "for", "multiple", "object", "classes", ",", "this", "proposal", "would", "be", "chosen", "as", "the", "cluster", "center", "only", "by", "the", "class", "with", "the", "highest", "predicted", "score", "and", "we", "re", "-", "choose", "cluster", "centers", "for", "other", "classes", ".", "(", "2", ")", "Finding", "proposal", "cluster", "centers", "using", "graphs", "of", "top", "ranking", "proposals", ".", "As", "stated", "in", "Section", "[", "reference", "]", ",", "although", "we", "can", "find", "good", "proposal", "cluster", "centers", "using", "the", "highest", "scoring", "proposal", ",", "this", "ignores", "that", "in", "natural", "images", "there", "are", "often", "more", "than", "one", "object", "for", "each", "category", ".", "Therefore", ",", "we", "propose", "a", "new", "method", "to", "find", "cluster", "centers", "using", "graphs", "of", "top", "ranking", "proposals", ".", "[", "t", "]", "Finding", "proposal", "cluster", "centers", "using", "graphs", "of", "top", "ranking", "proposals", "[", "1", "]", "Proposal", "boxes", ";", "image", "label", "vector", ";", "proposal", "score", "matrix", ".", "Proposal", "cluster", "centers", ".", "Initialize", ".", "Select", "top", "ranking", "proposals", "with", "indexes", ".", "Build", "a", "graph", "using", "the", "top", "ranking", "proposals", ".", "Set", ".", "Set", ".", ".", "Remove", "the", "-", "th", "proposal", "box", "from", ",", "or", ".", "is", "empty", ".", "More", "specifically", ",", "suppose", "an", "image", "has", "object", "class", "label", ".", "We", "first", "select", "the", "top", "ranking", "proposals", "with", "indexes", "for", "the", "-", "th", "refinement", ".", "Then", "we", "build", "an", "undirected", "unweighted", "graph", "of", "these", "proposals", "based", "on", "spatial", "similarity", ",", "where", "vertexes", "correspond", "to", "these", "top", "ranking", "proposals", ",", "and", "edges", "correspond", "to", "the", "connections", "between", "the", "vertexes", ".", "is", "determined", "according", "to", "the", "spatial", "similarity", "between", "two", "vertexes", "(", "i.e", ".", ",", "proposals", ")", "as", "in", "Eq", ".", "(", "[", "reference", "]", ")", ",", "where", "is", "the", "IoU", "between", "the", "-", "th", "and", "-", "th", "proposals", "and", "is", "a", "threshold", "(", "e.g", ".", ",", ")", ".", "Therefore", ",", "two", "vertexes", "are", "connected", "if", "they", "are", "spatially", "adjacent", ".", "After", "that", ",", "we", "greedily", "generate", "some", "cluster", "centers", "for", "class", "using", "this", "graph", ".", "That", "is", ",", "we", "iteratively", "select", "vertexes", "which", "have", "most", "connections", "to", "be", "the", "cluster", "centers", ",", "as", "in", "Algorithm", "[", "reference", "]", ".", "The", "number", "of", "cluster", "centers", "(", "i.e", ".", ",", ")", "changes", "for", "each", "image", "in", "each", "training", "iteration", "because", "the", "top", "ranking", "proposals", "change", ".", "See", "Section", "[", "reference", "]", "for", "some", "typical", "values", "of", ".", "We", "use", "the", "same", "method", "as", "in", "Section", "[", "reference", "]", "(", "1", ")", "to", "avoid", "one", "proposal", "been", "chosen", "as", "the", "cluster", "centers", "for", "multiple", "object", "classes", ".", "The", "reasons", "for", "this", "strategy", "are", "as", "follows", ".", "First", ",", "according", "to", "our", "observation", ",", "the", "top", "ranking", "proposals", "can", "always", "cover", "at", "least", "parts", "of", "objects", ",", "thus", "generating", "centers", "from", "these", "proposals", "encourages", "the", "selected", "centers", "to", "meet", "our", "requirements", ".", "Second", ",", "because", "these", "proposals", "cover", "objects", "well", ",", "better", "proposals", "(", "covering", "more", "parts", "of", "objects", ")", "should", "have", "more", "spatially", "overlapped", "proposals", "(", "i.e", ".", ",", "have", "more", "connections", ")", ".", "Third", ",", "these", "centers", "are", "spatially", "far", "apart", ",", "and", "thus", "different", "centers", "can", "correspond", "to", "different", "objects", ".", "This", "method", "also", "has", "the", "attractive", "characteristic", "that", "it", "can", "generate", "adaptive", "number", "of", "proposals", "for", "each", "object", "class", ",", "which", "is", "desirable", "because", "in", "natural", "images", "there", "are", "arbitrary", "number", "of", "objects", "per", "-", "class", ".", "We", "set", "the", "score", "of", "the", "-", "th", "proposal", "cluster", "center", "by", "(", "see", "the", "-", "th", "line", "in", "Algorithm", "[", "reference", "]", ")", "because", "if", "the", "adjacent", "proposals", "of", "a", "center", "proposal", "have", "high", "confidence", "to", "cover", "at", "least", "part", "of", "an", "object", "(", "i.e", ".", ",", "have", "high", "classification", "scores", ")", "the", "center", "proposal", "should", "also", "have", "such", "high", "confidence", ".", "There", "is", "an", "important", "issue", "for", "the", "graph", "-", "based", "method", ":", "how", "to", "select", "the", "top", "ranking", "proposals", "?", "A", "simple", "method", "is", "to", "select", "proposals", "whose", "scores", "exceed", "a", "threshold", ".", "But", "in", "our", "case", ",", "proposal", "scores", "change", "in", "each", "training", "iteration", ",", "and", "thus", "it", "is", "hard", "to", "determine", "a", "threshold", ".", "Instead", ",", "for", "each", "positive", "object", "class", ",", "we", "use", "the", "-", "means", "algorithm", "to", "divide", "proposal", "scores", "of", "an", "image", "into", "some", "clusters", ",", "and", "choose", "proposals", "in", "the", "cluster", "which", "has", "the", "highest", "score", "center", "to", "form", "the", "top", "ranking", "proposals", ".", "This", "method", "ensures", "that", "we", "can", "select", "the", "top", "ranking", "proposals", "although", "proposal", "scores", "change", "during", "training", ".", "Other", "choices", "are", "possible", ",", "but", "this", "method", "works", "well", "in", "experiments", ".", "[", "t", "]", "Generating", "proposal", "clusters", "[", "1", "]", "Proposal", "boxes", ";", "proposal", "cluster", "centers", ".", "Proposal", "clusters", ".", "Initialize", ".", "Set", "of", "to", "of", ",", ".", "Initialize", "and", "set", ".", "Compute", "IoUs", ".", "Choose", "the", "most", "spatially", "adjacent", "center", ".", ".", ".", "subsubsection", ":", "Generating", "proposal", "clusters", "After", "the", "cluster", "centers", "are", "found", ",", "we", "generate", "the", "proposal", "clusters", "as", "in", "our", "conference", "version", "paper", ".", "Except", "for", "the", "cluster", "for", "background", ",", "good", "proposal", "clusters", "require", "that", "proposals", "in", "the", "same", "cluster", "are", "associated", "with", "the", "same", "object", ",", "and", "thus", "proposals", "in", "the", "same", "cluster", "should", "be", "spatially", "adjacent", ".", "Specially", ",", "given", "the", "-", "th", "proposal", ",", "we", "compute", "a", "set", "of", "IoUs", ",", "where", "is", "the", "IoU", "between", "the", "-", "th", "proposal", "and", "the", "box", "of", "the", "-", "th", "cluster", "center", ".", "Then", "we", "assign", "the", "-", "th", "proposal", "to", "the", "-", "th", "object", "cluster", "if", "is", "larger", "than", "a", "threshold", "(", "e.g", ".", ",", ")", "and", "to", "the", "background", "cluster", "otherwise", ",", "where", "is", "the", "index", "of", "the", "most", "spatially", "adjacent", "cluster", "center", "as", "Eq", ".", "(", "[", "reference", "]", ")", ".", "The", "overall", "procedures", "to", "generate", "proposal", "clusters", "are", "summarized", "in", "Algorithm", "[", "reference", "]", ".", "We", "set", "the", "proposal", "scores", "for", "the", "background", "cluster", "to", "the", "scores", "of", "their", "most", "spatially", "adjacent", "centers", "as", "the", "10", "-", "the", "line", "in", "Algorithm", "[", "reference", "]", ",", "because", "if", "the", "cluster", "center", "has", "confidence", "that", "it", "covers", "an", "object", ",", "the", "proposal", "far", "away", "from", "should", "have", "confidence", "to", "be", "background", ".", "subsubsection", ":", "Learning", "refined", "instance", "classifiers", "To", "get", "supervisions", "and", "loss", "functions", "to", "learn", "the", "-", "th", "refined", "instance", "classifier", ",", "we", "design", "two", "approaches", "as", "follows", ".", "(", "1", ")", "Assigning", "proposals", "object", "labels", ".", "The", "most", "straightforward", "way", "to", "refine", "classifiers", "is", "to", "directly", "assign", "object", "labels", "to", "all", "proposals", "in", "object", "clusters", "because", "these", "proposals", "potentially", "correspond", "to", "whole", "objects", ",", "as", "in", "our", "conference", "version", "paper", ".", "As", "the", "cluster", "centers", "covers", "at", "least", "parts", "of", "objects", ",", "their", "adjacent", "proposals", "(", "i.e", ".", ",", "proposals", "in", "the", "cluster", ")", "can", "contain", "larger", "parts", "of", "objects", ".", "Accordingly", ",", "we", "can", "assign", "the", "cluster", "label", "to", "all", "proposals", "in", "the", "-", "th", "cluster", ".", "More", "specifically", ",", "the", "supervisions", "are", "proposal", "-", "level", "labels", ",", "i.e", ".", ",", ".", "is", "the", "label", "vector", "of", "the", "-", "th", "proposal", "for", "the", "-", "th", "refinement", ",", "where", "and", "if", "the", "-", "th", "proposal", "belongs", "to", "the", "-", "th", "clusters", ".", "Consequently", ",", "we", "use", "the", "standard", "softmax", "loss", "function", "to", "train", "the", "refined", "classifiers", "as", "in", "Eq", ".", "(", "[", "reference", "]", ")", ",", "where", "is", "the", "predicted", "score", "of", "the", "-", "th", "proposal", "as", "defined", "in", "Section", "[", "reference", "]", ".", "Through", "iterative", "instance", "classifier", "refinement", "(", "i.e", ".", ",", "multiple", "times", "of", "refinement", "as", "increase", ")", ",", "the", "detector", "detects", "larger", "parts", "of", "objects", "gradually", "by", "forcing", "the", "network", "to", "“", "see", "”", "larger", "parts", "of", "objects", ".", "Actually", ",", "the", "so", "learnt", "supervisions", "are", "very", "noisy", ",", "especially", "in", "the", "beginning", "of", "training", ".", "This", "results", "in", "unstable", "solutions", ".", "To", "solve", "this", "problem", ",", "we", "change", "the", "loss", "in", "Eq", ".", "(", "[", "reference", "]", ")", "to", "a", "weighted", "version", ",", "as", "in", "Eq", ".", "(", "[", "reference", "]", ")", ".", "is", "the", "loss", "weight", "that", "is", "the", "same", "as", "the", "cluster", "confidence", "score", "for", "object", "clusters", "or", "proposal", "confidence", "score", "for", "the", "background", "cluster", "if", "the", "-", "th", "proposal", "belongs", "to", "the", "-", "th", "cluster", ".", "From", "Algorithm", "[", "reference", "]", ",", "we", "can", "observe", "that", "is", "the", "same", "as", "the", "cluster", "center", "confidence", "score", ".", "The", "reasons", "for", "this", "strategy", "are", "as", "follows", ".", "In", "the", "beginning", "of", "training", ",", "although", "we", "can", "not", "obtain", "good", "proposal", "clusters", ",", "each", "is", "small", ",", "hence", "each", "is", "small", "and", "the", "loss", "is", "also", "small", ".", "As", "a", "consequence", ",", "the", "performance", "of", "the", "network", "will", "not", "decrease", "a", "lot", ".", "During", "the", "training", ",", "the", "top", "ranking", "proposals", "will", "cover", "objects", "well", ",", "and", "thus", "we", "can", "generate", "good", "proposal", "clusters", ".", "Then", "we", "can", "train", "satisfactory", "instance", "classifiers", ".", "(", "2", ")", "Treating", "clusters", "as", "bags", ".", "As", "we", "stressed", "before", ",", "although", "directly", "assigning", "proposals", "object", "labels", "can", "boost", "the", "results", ",", "it", "may", "confuse", "the", "network", "because", "we", "simultaneously", "assign", "the", "same", "label", "to", "different", "parts", "of", "objects", ".", "Focusing", "on", "this", ",", "we", "further", "propose", "to", "treat", "each", "proposal", "cluster", "as", "a", "small", "new", "bag", "and", "use", "the", "cluster", "label", "as", "the", "bag", "label", ".", "Thus", "the", "supervisions", "for", "the", "-", "th", "refinement", "are", "bag", "-", "level", "(", "cluster", "-", "level", ")", "labels", ",", "i.e", ".", ",", ".", "is", "the", "label", "of", "the", "-", "th", "bag", ",", "i.e", ".", ",", "the", "label", "of", "the", "-", "th", "proposal", "cluster", ",", "as", "defined", "in", "Section", "[", "reference", "]", ".", "Specially", ",", "for", "object", "clusters", ",", "we", "choose", "average", "MIL", "pooling", ",", "because", "these", "proposals", "should", "cover", "at", "least", "parts", "of", "objects", "and", "thus", "should", "have", "relatively", "high", "prediction", "scores", ".", "For", "the", "background", "cluster", ",", "we", "assign", "the", "background", "label", "to", "all", "proposals", "in", "the", "cluster", "according", "to", "the", "MIL", "constraints", "(", "all", "instances", "in", "negative", "bags", "are", "negative", ")", ".", "Then", "the", "loss", "function", "for", "refinement", "will", "be", "Eq", ".", "(", "[", "reference", "]", ")", ".", ",", ",", "and", "are", "the", "cluster", "confidence", "score", "of", "the", "-", "th", "object", "cluster", ",", "the", "number", "of", "proposals", "in", "the", "-", "th", "cluster", ",", "and", "the", "predicted", "score", "of", "the", "-", "th", "proposal", ",", "respectively", ",", "as", "defined", "in", "Section", "[", "reference", "]", ".", "and", "indicate", "that", "the", "-", "th", "proposal", "belongs", "to", "the", "-", "th", "object", "cluster", "and", "the", "background", "cluster", "respectively", ".", "Compared", "with", "the", "directly", "assigning", "label", "approach", ",", "this", "method", "tolerates", "some", "proposals", "to", "have", "low", "scores", ",", "which", "can", "reduce", "the", "ambiguities", "to", "some", "extent", ".", "subsection", ":", "Testing", "During", "testing", ",", "the", "proposal", "scores", "of", "refined", "instance", "classifiers", "are", "used", "as", "the", "final", "detection", "scores", ",", "as", "the", "blue", "arrows", "in", "Fig", ".", "[", "reference", "]", ".", "Here", "the", "mean", "output", "of", "all", "refined", "classifiers", "is", "chosen", ".", "The", "Non", "-", "Maxima", "Suppression", "(", "NMS", ")", "is", "used", "to", "filter", "out", "redundant", "detections", ".", "section", ":", "Experiments", "In", "this", "section", ",", "we", "first", "introduce", "our", "experimental", "setup", "including", "datasets", ",", "evaluation", "metrics", ",", "and", "implementation", "details", ".", "Then", "we", "conduct", "elaborate", "experiments", "to", "discuss", "the", "influence", "of", "different", "settings", ".", "Next", ",", "we", "compare", "our", "results", "with", "others", "to", "show", "the", "effectiveness", "of", "our", "method", ".", "After", "that", ",", "we", "show", "some", "qualitative", "results", "for", "further", "analyses", ".", "Finally", ",", "we", "give", "some", "runtime", "analyses", "of", "our", "method", ".", "Codes", "for", "reproducing", "our", "results", "are", "available", "at", ".", "subsection", ":", "Experimental", "setup", "subsubsection", ":", "Datasets", "and", "evaluation", "metrics", "We", "evaluate", "our", "method", "on", "four", "challenging", "datasets", ":", "the", "PASCAL", "VOC", "2007", "and", "2012", "datasets", ",", "the", "ImageNet", "detection", "dataset", ",", "and", "the", "MS", "-", "COCO", "dataset", ".", "Only", "image", "-", "level", "annotations", "are", "used", "to", "train", "our", "models", ".", "The", "PASCAL", "VOC", "2007", "and", "2012", "datasets", "have", "and", "images", "respectively", "for", "object", "classes", ".", "These", "two", "datasets", "are", "divided", "into", "train", ",", "val", ",", "and", "test", "sets", ".", "Here", "we", "choose", "the", "trainval", "set", "(", "images", "for", "2007", "and", "images", "for", "2012", ")", "to", "train", "our", "network", ".", "For", "testing", ",", "there", "are", "two", "metrics", "for", "evaluation", ":", "mAP", "and", "CorLoc", ".", "Following", "the", "standard", "PASCAL", "VOC", "protocol", ",", "Average", "Precision", "(", "AP", ")", "and", "the", "mean", "of", "AP", "(", "mAP", ")", "is", "the", "evaluation", "metric", "to", "test", "our", "model", "on", "the", "testing", "set", ".", "Correct", "Localization", "(", "CorLoc", ")", "is", "to", "test", "our", "model", "on", "the", "training", "set", "measuring", "the", "localization", "accuracy", ".", "All", "these", "two", "metrics", "are", "based", "on", "the", "PASCAL", "criterion", ",", "i.e", ".", ",", "IoU", "0.5", "between", "groundtruth", "boundingboxes", "and", "predicted", "boxes", ".", "The", "ImageNet", "detection", "dataset", "has", "hundreds", "of", "thousands", "of", "images", "with", "object", "classes", ".", "It", "is", "also", "divided", "into", "train", ",", "val", ",", "and", "test", "sets", ".", "Following", ",", "we", "split", "the", "val", "set", "into", "val1", "and", "val2", ",", "and", "randomly", "choose", "at", "most", "K", "images", "in", "the", "train", "set", "for", "each", "object", "class", "(", "we", "call", "it", "train", ")", ".", "We", "train", "our", "model", "on", "the", "mixture", "of", "train", "and", "val1", "sets", ",", "and", "test", "it", "on", "the", "val2", "set", ",", "which", "will", "lead", "to", "images", "for", "training", "and", "images", "for", "testing", ".", "We", "also", "use", "the", "mAP", "for", "evaluation", "on", "the", "ImageNet", ".", "The", "MS", "-", "COCO", "dataset", "has", "object", "classes", "and", "is", "divided", "into", "train", ",", "val", ",", "and", "test", "sets", ".", "Since", "the", "groundtruths", "on", "the", "test", "set", "are", "not", "released", ",", "we", "train", "our", "model", "on", "the", "MS", "-", "COCO", "2014", "train", "set", "(", "about", "K", "images", ")", "and", "test", "it", "on", "the", "val", "set", "(", "about", "K", "images", ")", ".", "For", "evaluation", ",", "we", "use", "two", "metrics", "mAP@0.5", "and", "mAP@", "[", ".5", ",", ".95", "]", "which", "are", "the", "standard", "PASCAL", "criterion", "(", "i.e", ".", ",", "IoU", "0.5", ")", "and", "the", "standard", "MS", "-", "COCO", "criterion", "(", "i.e", ".", ",", "computing", "the", "average", "of", "mAP", "for", "IoU", "[", "0.5", ":", "0.05", ":", "0.95", "]", ")", "respectively", ".", "subsubsection", ":", "Implementation", "details", "Our", "method", "is", "built", "on", "two", "pre", "-", "trained", "ImageNet", "networks", "VGG", "M", "and", "VGG16", ",", "each", "of", "which", "has", "some", "conv", "layers", "with", "max", "-", "pooling", "layers", "and", "three", "fc", "layers", ".", "We", "replace", "the", "last", "max", "-", "pooling", "layer", "by", "the", "SPP", "layer", ",", "and", "the", "last", "fc", "layer", "as", "well", "as", "the", "softmax", "loss", "layer", "by", "the", "layers", "described", "in", "Section", "[", "reference", "]", ".", "To", "increase", "the", "feature", "map", "size", "from", "the", "last", "conv", "layer", ",", "we", "replace", "the", "penultimate", "max", "-", "pooling", "layer", "and", "its", "subsequent", "conv", "layers", "by", "the", "dilated", "conv", "layers", ".", "The", "newly", "added", "layers", "are", "initialized", "using", "Gaussian", "distributions", "with", "-", "mean", "and", "standard", "deviations", ".", "Biases", "are", "initialized", "to", ".", "During", "training", ",", "the", "mini", "-", "batch", "size", "for", "SGD", "is", "set", "to", "be", ",", ",", "and", "for", "PASCAL", "VOC", ",", "ImageNet", ",", "and", "MS", "-", "COCO", ",", "respectively", ".", "The", "learning", "rate", "is", "set", "to", "for", "the", "first", "K", ",", "K", ",", "K", ",", "and", "K", "iterations", "for", "the", "PASCAL", "VOC", "2007", ",", "PASCAL", "VOC", "2012", ",", "ImageNet", ",", "and", "MS", "-", "COCO", "datasets", ",", "respectively", ".", "Then", "we", "decrease", "the", "learning", "rate", "to", "in", "the", "following", "K", ",", "K", ",", "K", ",", "and", "K", "iterations", "for", "the", "PASCAL", "VOC", "2007", ",", "PASCAL", "VOC", "2012", ",", "ImageNet", ",", "and", "MS", "-", "COCO", "datasets", ",", "respectively", ".", "The", "momentum", "and", "weight", "decay", "are", "set", "to", "be", "and", "respectively", ".", "Selective", "Search", ",", "EdgeBox", ",", "and", "MCG", "are", "adopted", "to", "generate", "about", "proposals", "per", "-", "image", "for", "the", "PASCAL", "VOC", ",", "ImageNet", ",", "and", "MS", "-", "COCO", "datasets", ",", "respectively", ".", "For", "data", "augmentation", ",", "we", "use", "five", "image", "scales", "(", "resize", "the", "shortest", "side", "to", "one", "of", "these", "scales", ")", "with", "horizontal", "flips", "for", "both", "training", "and", "testing", ".", "If", "not", "specified", ",", "the", "instance", "classifiers", "are", "refined", "three", "times", ",", "i.e", ".", ",", "in", "Section", "[", "reference", "]", ",", "so", "there", "are", "four", "output", "streams", ";", "the", "IoU", "threshold", "in", "Section", "[", "reference", "]", "(", "2", ")", "(", "also", "Eq", ".", "(", "[", "reference", "]", ")", ")", "is", "set", "to", ";", "the", "number", "of", "-", "means", "clusters", "in", "the", "last", "paragraph", "of", "Section", "[", "reference", "]", "(", "2", ")", "is", "set", "to", ";", "in", "Section", "[", "reference", "]", "(", "also", "the", "-", "th", "line", "of", "Algorithm", "[", "reference", "]", ")", "is", "set", "to", ".", "Similar", "to", "other", "works", ",", "we", "train", "a", "supervised", "object", "detector", "through", "choosing", "the", "top", "-", "scoring", "proposals", "given", "by", "our", "method", "as", "pseudo", "groundtruths", "to", "further", "improve", "our", "results", ".", "Here", "we", "train", "a", "Fast", "R", "-", "CNN", "(", "FRCNN", ")", "using", "the", "VGG16", "model", "and", "the", "same", "five", "image", "scales", "(", "horizontal", "flips", "only", "in", "training", ")", ".", "The", "same", "proposals", "are", "chosen", "to", "train", "and", "test", "the", "FRCNN", ".", "NMS", "(", "with", "IoU", "threshold", ")", "is", "applied", "to", "compute", "AP", ".", "Our", "experiments", "are", "implemented", "based", "on", "the", "Caffe", "deep", "learning", "framework", ",", "using", "Python", "and", "C", "++", ".", "The", "-", "means", "algorithm", "to", "produce", "top", "ranking", "proposals", "is", "implemented", "by", "scikit", "-", "learn", ".", "All", "of", "our", "experiments", "are", "running", "on", "an", "NVIDIA", "GTX", "TitanX", "Pascal", "GPU", "and", "Intel", "(", "R", ")", "i7", "-", "6850", "K", "CPU", "(", "3.60GHz", ")", ".", "subsection", ":", "Discussions", "We", "first", "conduct", "some", "experiments", "to", "discuss", "the", "influence", "of", "different", "components", "of", "our", "method", "(", "including", "instance", "classifier", "refinement", ",", "different", "proposal", "generation", "methods", ",", "different", "refinement", "strategies", ",", "and", "weighted", "loss", ")", "and", "different", "parameter", "settings", "(", "including", "the", "IoU", "threshold", "defined", "in", "Section", "[", "reference", "]", "(", "2", ")", ",", "the", "number", "of", "-", "means", "clusters", "described", "in", "Section", "[", "reference", "]", "(", "2", ")", ",", "the", "IoU", "threshold", "defined", "in", "Section", "[", "reference", "]", ",", "and", "multi", "-", "scale", "training", "and", "testing", ".", ")", "We", "also", "discuss", "the", "number", "of", "proposal", "cluster", "centers", ".", "Without", "loss", "of", "generality", ",", "we", "only", "perform", "experiments", "on", "the", "VOC", "2007", "dataset", "and", "use", "the", "VGG", "M", "model", ".", "subsubsection", ":", "The", "influence", "of", "instance", "classifier", "refinement", "As", "the", "five", "curves", "in", "Fig", ".", "[", "reference", "]", "show", ",", "we", "observe", "that", "compared", "with", "the", "basic", "MIL", "network", ",", "for", "both", "refinement", "methods", ",", "even", "refining", "instance", "classifier", "a", "single", "time", "boosts", "the", "performance", "a", "lot", ".", "This", "confirms", "the", "necessity", "of", "refinement", ".", "If", "we", "refine", "the", "classifier", "multiple", "times", ",", "the", "results", "are", "improved", "further", ".", "But", "when", "refinement", "is", "implemented", "too", "many", "times", ",", "the", "performance", "gets", "saturated", "(", "there", "are", "no", "obvious", "improvements", "from", "times", "to", "times", ")", ".", "This", "is", "because", "the", "network", "tends", "to", "converge", "so", "that", "the", "supervision", "of", "the", "-", "th", "time", "is", "similar", "to", "the", "-", "rd", "time", ".", "In", "the", "rest", "of", "this", "paper", "we", "only", "refine", "classifiers", "times", ".", "Notice", "that", "in", "Fig", ".", "[", "reference", "]", ",", "the", "“", "0", "time", "”", "is", "similar", "to", "the", "WSDDN", "using", "Selective", "Search", "as", "proposals", ".", "subsubsection", ":", "The", "influence", "of", "different", "proposal", "cluster", "generation", "methods", "We", "discuss", "the", "influence", "of", "different", "proposal", "cluster", "generation", "methods", ".", "As", "shown", "in", "the", "Fig", ".", "[", "reference", "]", "(", "green", "and", "purple", "solid", "curves", "for", "the", "highest", "scoring", "proposal", "based", "method", ",", "blue", "and", "red", "solid", "curves", "for", "the", "graph", "-", "based", "method", ")", ",", "for", "all", "refinement", "times", ",", "the", "graph", "-", "based", "method", "obtains", "better", "performance", ",", "because", "it", "can", "generate", "better", "cluster", "centers", ".", "Thus", "we", "choose", "the", "graph", "-", "based", "method", "in", "the", "rest", "of", "our", "paper", ".", "subsubsection", ":", "The", "influence", "of", "different", "refinement", "strategies", "We", "then", "show", "the", "influence", "of", "different", "refinement", "strategies", ".", "The", "directly", "assigning", "label", "method", "is", "replaced", "by", "treating", "clusters", "as", "bags", "(", "blue", "and", "green", "solid", "curves", ")", ".", "From", "Fig", ".", "[", "reference", "]", ",", "it", "is", "obvious", "that", "the", "results", "by", "treating", "clusters", "as", "bags", "are", "better", ".", "In", "addition", ",", "compared", "with", "the", "alternating", "training", "strategy", "(", "blue", "dashed", "curve", ")", ",", "our", "online", "training", "boosts", "the", "performance", "consistently", "and", "significantly", ",", "which", "confirms", "the", "necessity", "of", "sharing", "proposal", "features", ".", "Online", "training", "also", "reduces", "the", "training", "time", "a", "lot", ",", "because", "it", "only", "requires", "training", "a", "single", "model", "instead", "of", "training", "models", "for", "times", "refinement", "in", "the", "alternating", "strategy", ".", "In", "the", "rest", "of", "our", "paper", ",", "we", "only", "report", "results", "by", "the", "“", "PCL", "-", "OB", "-", "G", "”", "method", "in", "Fig", ".", "[", "reference", "]", "because", "it", "achieves", "the", "best", "performance", ".", "subsubsection", ":", "The", "influence", "of", "weighted", "loss", "We", "also", "study", "the", "influence", "of", "our", "weighted", "loss", "in", "Eq", ".", "(", "[", "reference", "]", ")", ".", "Note", "that", "Eq", ".", "(", "[", "reference", "]", ")", "can", "be", "easily", "changed", "to", "the", "unweighted", "version", "by", "simply", "setting", "and", "to", "be", ".", "Here", "we", "train", "a", "network", "using", "the", "unweighted", "loss", ".", "The", "results", "of", "the", "unweighted", "loss", "are", "mAP", "and", "CorLoc", ".", "We", "see", "that", "if", "we", "use", "the", "unweighted", "loss", ",", "the", "improvement", "from", "refinement", "is", "very", "scant", "and", "the", "performance", "is", "even", "worse", "than", "the", "alternating", "strategy", ".", "Using", "the", "weighted", "loss", "achieves", "much", "better", "performance", "(", "mAP", "and", "CorLoc", ")", ",", "which", "confirms", "our", "theory", "in", "Section", "[", "reference", "]", ".", "subsubsection", ":", "The", "influence", "of", "the", "IoU", "threshold", "Here", "we", "discuss", "the", "influence", "of", "the", "IoU", "threshold", "defined", "in", "Section", "[", "reference", "]", "(", "2", ")", "and", "Eq", ".", "(", "[", "reference", "]", ")", ".", "From", "Fig", ".", "[", "reference", "]", ",", "we", "see", "that", "setting", "to", "obtains", "the", "best", "performance", ".", "Therefore", ",", "we", "set", "to", "for", "the", "other", "experiments", ".", "subsubsection", ":", "The", "influence", "of", "the", "number", "of", "-", "means", "clusters", "In", "previous", "experiments", "we", "set", "the", "number", "of", "-", "means", "clusters", "described", "in", "the", "last", "paragraph", "of", "Section", "[", "reference", "]", "(", "2", ")", "to", "be", ".", "Here", "we", "set", "it", "to", "other", "numbers", "to", "explore", "its", "influence", ".", "The", "results", "from", "other", "numbers", "of", "-", "means", "clusters", "are", "mAP", "and", "CorLoc", "for", "clusters", ",", "and", "mAP", "and", "CorLoc", "for", "clusters", ",", "which", "are", "a", "little", "worse", "than", "the", "results", "from", "cluster", ".", "Therefore", ",", "we", "set", "the", "number", "of", "-", "means", "clusters", "to", "for", "the", "other", "experiments", ".", "subsubsection", ":", "The", "influence", "of", "the", "IoU", "threshold", "We", "also", "analyse", "the", "influence", "of", "defined", "in", "Section", "[", "reference", "]", "and", "the", "-", "th", "line", "of", "Algorithm", "[", "reference", "]", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "outperforms", "other", "choices", ".", "Therefore", ",", "we", "set", "to", "for", "the", "other", "experiments", ".", "subsubsection", ":", "The", "influence", "of", "multi", "-", "scale", "training", "and", "testing", "Previously", "our", "experiments", "are", "conducted", "based", "on", "five", "image", "scales", "for", "training", "and", "testing", ".", "Here", "we", "show", "the", "influence", "of", "this", "multi", "-", "scale", "setting", ".", "We", "train", "and", "test", "our", "method", "using", "a", "single", "image", "scale", "as", "the", "default", "scale", "setting", "of", "FRCNN", ".", "The", "single", "-", "scale", "results", "are", "mAP", "and", "CorLoc", "which", "are", "much", "worse", "than", "our", "multi", "-", "scale", "results", "(", "mAP", "and", "CorLoc", ")", ".", "Therefore", ",", "we", "use", "five", "image", "scales", "as", "many", "WSOD", "networks", ".", "subsubsection", ":", "The", "number", "of", "proposal", "cluster", "centers", "As", "we", "stated", "in", "Section", "[", "reference", "]", "(", "2", ")", ",", "the", "number", "of", "proposal", "cluster", "centers", "(", "i.e", ".", ",", ")", "changes", "for", "each", "image", "in", "each", "training", "iteration", ".", "Here", "we", "give", "some", "typical", "values", "of", ".", "In", "the", "beginning", "of", "training", ",", "the", "proposal", "scores", "are", "very", "noisy", "and", "thus", "the", "selected", "top", "ranking", "proposals", "to", "form", "graphs", "are", "scattered", "in", "images", ",", "which", "results", "in", "dozens", "of", "proposal", "cluster", "centers", "for", "each", "image", ".", "After", "some", "(", "about", "3", "K", ")", "training", "iterations", ",", "the", "proposal", "scores", "are", "more", "reliable", "and", "our", "method", "finds", "1", "3", "proposal", "cluster", "centers", "for", "each", "positive", "object", "class", ".", "To", "make", "the", "training", "more", "stable", "in", "the", "beginning", ",", "for", "each", "positive", "object", "class", "we", "empirically", "select", "at", "most", "five", "proposal", "cluster", "centers", "which", "have", "higher", "scores", ",", "and", "the", "number", "of", "selected", "proposal", "cluster", "centers", "does", "not", "influence", "the", "performance", "much", ".", "subsection", ":", "Comparison", "with", "other", "methods", "Here", "we", "compare", "our", "best", "performed", "strategy", "PCL", "-", "OB", "-", "G", ",", "i.e", ".", ",", "using", "graph", "-", "based", "method", "and", "treating", "clusters", "as", "bags", "to", "train", "the", "network", "online", ",", "with", "other", "methods", ".", "We", "first", "report", "our", "results", "for", "each", "class", "on", "VOC", "2007", "and", "2012", "in", "Table", "[", "reference", "]", ",", "Table", "[", "reference", "]", ",", "Table", "[", "reference", "]", ",", "and", "Table", "[", "reference", "]", ".", "It", "is", "obvious", "that", "our", "method", "outperforms", "other", "methods", "using", "single", "model", "VGG", "M", "or", "VGG16", "(", "PCL", "-", "OB", "-", "G", "+", "VGG", "M", "and", "PCL", "-", "OB", "-", "G", "+", "VGG16", "in", "tables", ".", ")", "Our", "single", "model", "results", "even", "better", "than", "others", "by", "combining", "multiple", "different", "models", "(", "e.g", ".", ",", "ensemble", "of", "models", ")", ".", "Specially", ",", "our", "method", "obtains", "much", "better", "results", "compared", "with", "other", "two", "methods", "also", "using", "the", "same", "basic", "MIL", "network", ".", "Importantly", ",", "also", "equips", "the", "weighted", "sum", "pooling", "with", "objectness", "measure", "of", "EdgeBox", "and", "the", "spatial", "regulariser", ",", "and", "adds", "context", "information", "into", "the", "network", ",", "both", "of", "which", "are", "more", "complicated", "than", "our", "basic", "MIL", "network", ".", "We", "believe", "that", "our", "performance", "can", "be", "improved", "by", "choosing", "better", "basic", "MIL", "networks", ",", "like", "the", "complete", "network", "in", "and", "using", "context", "information", ".", "As", "reimplementing", "their", "method", "completely", "is", "non", "-", "trivial", ",", "here", "we", "only", "choose", "the", "simplest", "architecture", "in", ".", "Even", "in", "this", "simplified", "case", ",", "our", "method", "achieves", "very", "promising", "results", ".", "Our", "results", "can", "also", "be", "improved", "by", "combing", "multiple", "models", ".", "As", "shown", "in", "the", "tables", ",", "there", "are", "little", "improvements", "from", "the", "ensemble", "of", "the", "VGG", "M", "and", "VGG16", "models", "(", "PCL", "-", "OB", "-", "G", "-", "Ens", ".", "in", "tables", ")", ".", "Here", "we", "do", "the", "ensemble", "by", "summing", "up", "the", "scores", "produced", "by", "the", "two", "models", ".", "Also", ",", "as", "mentioned", "in", "Section", "[", "reference", "]", ",", "similar", "to", ",", "we", "train", "a", "FRCNN", "detector", "using", "top", "-", "scoring", "proposals", "produced", "by", "PCL", "-", "OB", "-", "G", "-", "Ens", ".", "as", "groundtruths", "(", "PCL", "-", "OB", "-", "G", "-", "Ens.", "+", "FRCNN", "in", "tables", ")", ".", "As", "we", "can", "see", ",", "the", "performance", "is", "improved", "further", ".", "We", "then", "show", "results", "of", "our", "method", "on", "the", "large", "scale", "ImageNet", "detection", "dataset", "in", "Table", "[", "reference", "]", ".", "We", "observe", "similar", "phenomenon", "that", "our", "method", "outperforms", "other", "methods", "by", "a", "large", "margin", ".", "We", "finally", "report", "results", "of", "our", "method", "on", "MS", "-", "COCO", "in", "Table", "[", "reference", "]", ".", "Our", "method", "obtains", "better", "performance", "than", "the", "recent", "work", ".", "In", "particular", ",", "Ge", "et", "al", ".", "use", "the", "method", "proposed", "in", "our", "conference", "version", "paper", "as", "a", "basic", "component", ".", "We", "can", "expect", "to", "obtain", "better", "detection", "performance", "through", "replacing", "our", "conference", "version", "method", "in", "by", "our", "newly", "proposed", "method", "here", ",", "which", "we", "would", "like", "to", "explore", "in", "the", "future", ".", "subsection", ":", "Qualitative", "results", "We", "first", "show", "some", "proposal", "clusters", "generated", "by", "our", "method", "in", "Fig", ".", "[", "reference", "]", ".", "As", "we", "can", "see", ",", "the", "cluster", "centers", "contain", "at", "least", "parts", "of", "objects", "and", "are", "able", "to", "cover", "adaptive", "number", "of", "objects", "for", "each", "class", ".", "We", "then", "show", "qualitative", "comparisons", "among", "the", "WSDDN", ",", "the", "WSDDN", "+", "context", ",", "and", "our", "PCL", "method", ",", "both", "of", "which", "use", "the", "same", "basic", "MIL", "network", ".", "As", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "we", "can", "observe", "that", "for", "classes", "such", "as", "bike", ",", "car", ",", "cat", ",", "etc", ".", ",", "our", "method", "tends", "to", "provide", "more", "accurate", "detections", ",", "whereas", "other", "two", "methods", "sometimes", "fails", "by", "producing", "boxes", "that", "are", "overlarge", "or", "only", "contain", "parts", "of", "objects", "(", "the", "first", "four", "rows", "in", "Fig", ".", "[", "reference", "]", ")", ".", "But", "for", "some", "classes", "such", "as", "person", ",", "our", "method", "sometimes", "fails", "by", "only", "detecting", "parts", "of", "objects", "such", "as", "the", "head", "of", "person", "(", "the", "fifth", "row", "in", "Fig", ".", "[", "reference", "]", ")", ".", "Exploiting", "context", "information", "sometimes", "help", "the", "detection", "(", "as", "in", "WSDDN", "+", "context", ")", ",", "we", "believe", "our", "method", "can", "be", "further", "improved", "by", "incorporating", "context", "information", "into", "our", "framework", ".", "All", "these", "three", "methods", "(", "actually", "almost", "all", "weakly", "supervised", "object", "detection", "methods", ")", "suffers", "from", "two", "problems", ":", "producing", "boxes", "that", "not", "only", "contain", "the", "target", "object", "but", "also", "include", "their", "adjacent", "similar", "objects", ",", "or", "only", "detecting", "parts", "of", "object", "for", "objects", "with", "deformation", "(", "the", "last", "row", "in", "Fig", ".", "[", "reference", "]", ")", ".", "We", "finally", "visualize", "some", "success", "and", "failure", "detection", "results", "on", "VOC", "2007", "trainval", "by", "PCL", "-", "Ens.", "+", "FRCNN", ",", "as", "in", "Fig", ".", "[", "reference", "]", ".", "We", "observe", "similar", "phenomena", "as", "in", "Fig", ".", "[", "reference", "]", ".", "Our", "method", "is", "robust", "to", "the", "size", "and", "aspect", "of", "objects", ",", "especially", "for", "rigid", "objects", ".", "The", "main", "failures", "for", "these", "rigid", "objects", "are", "always", "due", "to", "overlarge", "boxes", "that", "not", "only", "contain", "objects", ",", "but", "also", "include", "adjacent", "similar", "objects", ".", "For", "non", "-", "rigid", "objects", "like", "“", "cat", "”", ",", "“", "dog", "”", ",", "and", "“", "person", "”", ",", "they", "often", "have", "great", "deformations", ",", "but", "their", "parts", "(", "e.g", ".", ",", "head", "of", "person", ")", "have", "much", "less", "deformation", ",", "so", "our", "detector", "is", "still", "inclined", "to", "find", "these", "parts", ".", "An", "ideal", "solution", "is", "yet", "wanted", "because", "there", "is", "still", "room", "for", "improvement", ".", "subsection", ":", "Runtime", "The", "runtime", "comparisons", "between", "our", "method", "and", "our", "basic", "MIL", "network", "are", "shown", "in", "Table", "[", "reference", "]", ",", "where", "the", "runtime", "of", "proposal", "generation", "is", "not", "considered", ".", "As", "we", "can", "see", ",", "although", "our", "method", "has", "more", "components", "than", "our", "basic", "MIL", "network", ",", "our", "method", "takes", "almost", "the", "same", "testing", "time", "as", "it", ".", "This", "is", "because", "all", "our", "output", "streams", "share", "the", "same", "proposal", "feature", "computations", ".", "The", "small", "extra", "training", "computations", "of", "our", "method", "mainly", "come", "from", "the", "procedures", "to", "find", "proposal", "cluster", "centers", "and", "generate", "proposal", "clusters", ".", "Although", "with", "small", "extra", "training", "computations", ",", "our", "method", "obtains", "much", "better", "detection", "results", "than", "the", "basic", "MIL", "network", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "propose", "to", "generate", "proposal", "clusters", "to", "learn", "refined", "instance", "classifiers", "for", "weakly", "supervised", "object", "detection", ".", "We", "propose", "two", "strategies", "for", "proposal", "cluster", "generation", "and", "classifier", "refinement", ",", "both", "of", "which", "can", "boost", "the", "performance", "significantly", ".", "The", "classifier", "refinement", "is", "implemented", "by", "multiple", "output", "streams", "corresponding", "to", "some", "instance", "classifiers", "in", "multiple", "instance", "learning", "networks", ".", "An", "online", "training", "algorithm", "is", "introduced", "to", "train", "the", "proposed", "network", "end", "-", "to", "-", "end", "for", "effectiveness", "and", "efficiency", ".", "Experiments", "show", "substantial", "and", "consistent", "improvements", "by", "our", "method", ".", "We", "observe", "that", "the", "most", "common", "failure", "cases", "of", "our", "algorithm", "are", "connected", "with", "the", "deformation", "of", "non", "-", "rigid", "objects", ".", "In", "the", "future", ",", "we", "will", "concentrate", "on", "this", "problem", ".", "In", "addition", ",", "we", "believe", "our", "learning", "algorithm", "has", "the", "potential", "to", "be", "applied", "in", "other", "weakly", "supervised", "visual", "learning", "tasks", "such", "as", "weakly", "supervised", "semantic", "segmentation", ".", "We", "will", "also", "explore", "how", "to", "apply", "our", "method", "to", "these", "related", "applications", ".", "section", ":", "Acknowledgements", "This", "work", "was", "supported", "by", "NSFC", "(", "No", ".", "61733007", ",", "No", ".", "61572207", ",", "No", ".", "61876212", ",", "No", ".", "61672336", ",", "No", ".", "61573160", ")", ",", "ONR", "with", "grant", "N00014", "-", "15", "-", "1", "-", "2356", ",", "Hubei", "Scientific", "and", "Technical", "Innovation", "Key", "Project", ",", "and", "the", "Program", "for", "HUST", "Academic", "Frontier", "Youth", "Team", ".", "The", "corresponding", "author", "of", "this", "paper", "is", "Xinggang", "Wang", ".", "bibliography", ":", "References", "[", "]", "Peng", "Tang", "received", "the", "B.S.", "degree", "in", "Electronics", "and", "Information", "Engineering", "from", "Huazhong", "University", "of", "Science", "and", "Technology", "(", "HUST", ")", "in", "2014", ".", "He", "is", "currently", "pursuing", "the", "Ph.D.", "degree", "in", "the", "School", "of", "Electronic", "Information", "and", "Communications", "at", "HUST", ",", "and", "visiting", "the", "Department", "of", "Computer", "Science", "at", "Johns", "Hopkins", "University", ".", "He", "was", "an", "intern", "at", "Microsoft", "Research", "Asia", "in", "2017", ".", "His", "research", "interests", "include", "image", "classification", "and", "object", "detection", "in", "images", "/", "videos", ".", "[", "]", "Xinggang", "Wang", "is", "an", "assistant", "professor", "of", "School", "of", "Electronics", "Information", "and", "Communications", "of", "Huazhong", "University", "of", "Science", "and", "Technology", "(", "HUST", ")", ".", "He", "received", "his", "Bachelor", "degree", "in", "communication", "and", "information", "system", "and", "Ph.D.", "degree", "in", "computer", "vision", "both", "from", "HUST", ".", "From", "May", "2010", "to", "July", "2011", ",", "he", "was", "with", "the", "Department", "of", "Computer", "and", "Information", "Science", ",", "Temple", "University", ",", "Philadelphia", ",", "PA", ".", ",", "as", "a", "visiting", "scholar", ".", "From", "February", "2013", "to", "September", "2013", ",", "he", "was", "with", "the", "University", "of", "California", ",", "Los", "Angeles", "(", "UCLA", ")", ",", "as", "a", "visiting", "graduate", "researcher", ".", "He", "is", "a", "reviewer", "of", "IEEE", "Trans", "on", "PAMI", ",", "IEEE", "Trans", "on", "Image", "Processing", ",", "IEEE", "Trans", ".", "on", "Cybernetics", ",", "Pattern", "Recognition", ",", "Computer", "Vision", "and", "Image", "Understanding", ",", "Neurocomputing", ",", "NIPS", ",", "ICML", ",", "CVPR", ",", "ICCV", "and", "ECCV", "etc", ".", "His", "research", "interests", "include", "computer", "vision", "and", "machine", "learning", ",", "especially", "object", "recognition", ".", "[", "]", "Song", "Bai", "received", "the", "B.S.", "and", "Ph.D.", "degree", "in", "Electronics", "and", "Information", "Engineering", "from", "Huazhong", "University", "of", "Science", "and", "Technology", "(", "HUST", ")", ",", "Wuhan", ",", "China", "in", "2013", "and", "2018", ",", "respectively", ".", "He", "was", "with", "University", "of", "Texas", "at", "San", "Antonio", "(", "UTSA", ")", "and", "Johns", "Hopkins", "University", "(", "JHU", ")", "as", "a", "research", "scholar", ".", "His", "research", "interests", "include", "image", "retrieval", "and", "classification", ",", "3D", "shape", "recognition", ",", "person", "re", "-", "identification", ",", "semantic", "segmentation", "and", "deep", "learning", ".", "More", "information", "can", "be", "found", "in", "his", "homepage", ":", ".", "[", "]", "Wei", "Shen", "received", "his", "B.S.", "and", "Ph.D.", "degree", "both", "in", "Electronics", "and", "Information", "Engineering", "from", "the", "Huazhong", "University", "of", "Science", "and", "Technology", "(", "HUST", ")", ",", "Wuhan", ",", "China", ",", "in", "2007", "and", "in", "2012", ".", "From", "April", "2011", "to", "November", "2011", ",", "he", "worked", "in", "Microsoft", "Research", "Asia", "as", "an", "intern", ".", "In", "2012", ",", "he", "joined", "School", "of", "Communication", "and", "Information", "Engineering", ",", "Shanghai", "University", "as", "an", "Assistant", "Professor", ".", "From", "2017", ",", "he", "became", "an", "Associate", "Professor", ".", "He", "is", "currently", "visiting", "Department", "of", "Computer", "Science", ",", "Johns", "Hopkins", "University", ".", "His", "current", "research", "interests", "include", "random", "forests", ",", "deep", "learning", ",", "object", "detection", "and", "segmentation", ".", "[", "]", "Xiang", "Bai", "received", "his", "B.S.", ",", "M.S.", ",", "and", "Ph.D.", "degrees", "from", "the", "Huazhong", "University", "of", "Science", "and", "Technology", "(", "HUST", ")", ",", "Wuhan", ",", "China", ",", "in", "2003", ",", "2005", ",", "and", "2009", ",", "respectively", ",", "all", "in", "electronics", "and", "information", "engineering", ".", "He", "is", "currently", "a", "Professor", "with", "the", "School", "of", "Electronic", "Information", "and", "Communications", ",", "HUST", ".", "He", "is", "also", "the", "Vice", "-", "director", "of", "the", "National", "Center", "of", "Anti", "-", "Counterfeiting", "Technology", ",", "HUST", ".", "His", "research", "interests", "include", "object", "recognition", ",", "shape", "analysis", ",", "scene", "text", "recognition", "and", "intelligent", "systems", ".", "He", "serves", "as", "an", "associate", "editor", "for", "Pattern", "Recognition", ",", "Pattern", "Recognition", "Letters", ",", "Neurocomputing", "and", "Frontiers", "of", "Computer", "Science", ".", "[", "]", "Wenyu", "Liu", "received", "the", "B.S.", "degree", "in", "Computer", "Science", "from", "Tsinghua", "University", ",", "Beijing", ",", "China", ",", "in", "1986", ",", "and", "the", "M.S.", "and", "Ph.D.", "degrees", ",", "both", "in", "Electronics", "and", "Information", "Engineering", ",", "from", "Huazhong", "University", "of", "Science", "and", "Technology", "(", "HUST", ")", ",", "Wuhan", ",", "China", ",", "in", "1991", "and", "2001", ",", "respectively", ".", "He", "is", "now", "a", "professor", "and", "associate", "dean", "of", "the", "School", "of", "Electronic", "Information", "and", "Communications", ",", "HUST", ".", "His", "current", "research", "areas", "include", "computer", "vision", ",", "multimedia", ",", "and", "machine", "learning", ".", "He", "is", "a", "senior", "member", "of", "IEEE", ".", "[", "]", "Alan", "Yuille", "received", "the", "B.A.", "degree", "in", "mathematics", "from", "the", "University", "of", "Cambridge", "in", "1976", ",", "and", "the", "Ph.D.", "degree", "in", "theoretical", "physics", "from", "Cambridge", "in", "1980", ".", "He", "then", "held", "a", "post", "-", "doctoral", "position", "with", "the", "Physics", "Department", ",", "University", "of", "Texas", ",", "Austin", ",", "and", "the", "Institute", "for", "Theoretical", "Physics", ",", "Santa", "Barbara", ".", "He", "then", "became", "a", "Research", "Scientists", "with", "the", "Artificial", "Intelligence", "Laboratory", ",", "MIT", ",", "from", "1982", "to", "1986", ",", "and", "followed", "this", "with", "a", "faculty", "position", "in", "the", "division", "of", "applied", "sciences", ",", "Harvard", ",", "from", "1986", "to", "1995", ",", "rising", "to", "the", "position", "of", "an", "associate", "professor", ".", "From", "1995", "to", "2002", ",", "he", "was", "a", "Senior", "Scientist", "with", "the", "Smith", "-", "Kettlewell", "Eye", "Research", "Institute", "in", "San", "Francisco", ".", "From", "2002", "to", "2016", ",", "he", "was", "a", "Full", "Professor", "with", "the", "Department", "of", "Statistics", ",", "UCLA", ",", "with", "joint", "appointments", "in", "Psychology", ",", "Computer", "Science", ",", "and", "Psychiatry", ".", "In", "2016", ",", "he", "became", "a", "Bloomberg", "Distinguished", "Professor", "of", "cognitive", "science", "and", "computer", "science", "with", "Johns", "Hopkins", "University", ".", "He", "received", "the", "Marr", "Prize", "and", "the", "Helmholtz", "Prize", "."], "entity": [[2, 3, ["PCL"], "Method"], [4, 7, ["Proposal", "Cluster", "Learning"], "Method"], [8, 12, ["Weakly", "Supervised", "Object", "Detection"], "Task"], [12, 16, ["Weakly", "Supervised", "Object", "Detection"], "Task"], [17, 19, ["WSOD", ")"], "Task"], [28, 30, ["object", "detectors"], "Method"], [36, 38, ["object", "recognition"], "Task"], [47, 49, ["deep", "network"], "Method"], [50, 51, ["WSOD"], "Task"], [63, 66, ["image", "classification", "problem"], "Task"], [67, 70, ["Multiple", "Instance", "Learning"], "Method"], [71, 72, ["MIL"], "Method"], [77, 79, ["proposal", "clusters"], "Method"], [82, 84, ["instance", "classifiers"], "Method"], [86, 88, ["iterative", "process"], "Method"], [137, 139, ["proposal", "clusters"], "Method"], [140, 143, ["instance", "classifier", "refinement"], "Task"], [162, 165, ["assigning", "label", "method"], "Method"], [167, 171, ["iterative", "instance", "classifier", "refinement"], "Method"], [194, 197, ["instance", "classifier", "refinement"], "Method"], [208, 210, ["PASCAL", "VOC"], "Material"], [211, 213, ["ImageNet", "detection"], "Material"], [220, 221, ["WSOD"], "Task"], [236, 238, ["Object", "detection"], "Task"], [239, 242, ["weakly", "supervised", "learning"], "Task"], [247, 250, ["multiple", "instance", "learning"], "Method"], [251, 253, ["proposal", "cluster"], "Task"], [257, 259, ["Object", "detection"], "Task"], [267, 269, ["computer", "vision"], "Task"], [280, 283, ["Convolutional", "Neural", "Network"], "Method"], [284, 285, ["CNN"], "Method"], [307, 309, ["object", "detection"], "Task"], [402, 406, ["Weakly", "Supervised", "Object", "Detection"], "Task"], [407, 408, ["WSOD"], "Task"], [415, 417, ["object", "detectors"], "Method"], [428, 431, ["Multiple", "Instance", "Learning"], "Method"], [432, 433, ["MIL"], "Method"], [436, 437, ["WSOD"], "Task"], [449, 451, ["instance", "classifiers"], "Method"], [452, 454, ["object", "detectors"], "Method"], [492, 493, ["CNN"], "Method"], [498, 499, ["MIL"], "Method"], [500, 501, ["CNN"], "Method"], [513, 514, ["CNNs"], "Method"], [527, 530, ["proposal", "feature", "extractors"], "Method"], [553, 554, ["WSOD"], "Task"], [560, 566, ["end", "-", "to", "-", "end", "training"], "Method"], [570, 576, ["end", "-", "to", "-", "end", "training"], "Method"], [586, 591, ["end", "-", "to", "-", "end"], "Task"], [628, 630, ["MIL", "networks"], "Method"], [631, 632, ["WSOD"], "Task"], [640, 643, ["fully", "supervised", "ones"], "Method"], [657, 659, ["MIL", "networks"], "Method"], [665, 667, ["network", "training"], "Method"], [670, 672, ["instance", "classification"], "Task"], [679, 681, ["bag", "classification"], "Task"], [682, 686, ["image", "classification", ")", "problem"], "Task"], [690, 692, ["image", "scores"], "Metric"], [708, 710, ["image", "classification"], "Task"], [711, 713, ["object", "detection"], "Task"], [715, 716, ["classification"], "Task"], [790, 795, ["Intersection", "-", "over", "-", "Union"], "Metric"], [796, 797, ["IoU"], "Metric"], [808, 810, ["ranking", "proposals"], "Method"], [824, 825, ["detection"], "Task"], [834, 835, ["classification"], "Task"], [878, 880, ["proposal", "cluster"], "Method"], [881, 883, ["Object", "detection"], "Task"], [896, 898, ["proposal", "recall"], "Metric"], [914, 916, ["Selective", "Search"], "Method"], [980, 984, ["fully", "supervised", "object", "detection"], "Task"], [989, 991, ["object", "detectors"], "Task"], [1013, 1015, ["object", "detectors"], "Method"], [1046, 1047, ["detectors"], "Task"], [1056, 1059, ["weakly", "supervised", "scenario"], "Task"], [1108, 1110, ["classification", "scores"], "Metric"], [1266, 1270, ["graph", "-", "based", "method"], "Method"], [1333, 1336, ["cluster", "center", "proposal"], "Method"], [1384, 1386, ["instance", "classifiers"], "Method"], [1387, 1389, ["object", "detectors"], "Method"], [1391, 1392, ["training"], "Task"], [1480, 1481, ["classification"], "Task"], [1482, 1483, ["detection"], "Task"], [1546, 1548, ["instance", "classifiers"], "Method"], [1572, 1574, ["classification", "scores"], "Metric"], [1615, 1617, ["classification", "scores"], "Metric"], [1623, 1626, ["assigning", "label", "strategy"], "Method"], [1645, 1648, ["Proposal", "Cluster", "Learning"], "Method"], [1649, 1650, ["PCL"], "Method"], [1655, 1657, ["instance", "classifiers"], "Method"], [1659, 1661, ["proposal", "clusters"], "Method"], [1674, 1677, ["online", "training", "approach"], "Method"], [1711, 1713, ["instance", "classifiers"], "Method"], [1720, 1722, ["instance", "classifiers"], "Method"], [1729, 1730, ["training"], "Task"], [1731, 1734, ["proposal", "classification", "scores"], "Metric"], [1771, 1772, ["classifiers"], "Method"], [1775, 1778, ["back", "-", "propagation"], "Method"], [1825, 1826, ["detector"], "Method"], [1857, 1858, ["classifiers"], "Method"], [1907, 1909, ["training", "procedures"], "Method"], [1916, 1922, ["end", "-", "to", "-", "end", "network"], "Method"], [1946, 1947, ["training"], "Task"], [1948, 1949, ["testing"], "Task"], [1974, 1976, ["PASCAL", "VOC"], "Material"], [1977, 1979, ["ImageNet", "detection"], "Material"], [1981, 1985, ["MS", "-", "COCO", "datasets"], "Material"], [1996, 1997, ["mAP"], "Metric"], [1998, 1999, ["CorLoc"], "Material"], [2000, 2002, ["VOC", "2007"], "Material"], [2119, 2121, ["ImageNet", "detection"], "Material"], [2122, 2126, ["MS", "-", "COCO", "datasets"], "Material"], [2195, 2198, ["Multiple", "instance", "learning"], "Method"], [2198, 2199, ["MIL"], "Method"], [2203, 2206, ["drug", "activity", "prediction"], "Task"], [2221, 2222, ["MIL"], "Method"], [2224, 2225, ["MIL"], "Method"], [2249, 2250, ["WSOD"], "Task"], [2261, 2263, ["instance", "classifiers"], "Method"], [2280, 2281, ["WSOD"], "Task"], [2288, 2290, ["instance", "classifiers"], "Method"], [2307, 2308, ["MIL"], "Method"], [2312, 2314, ["computer", "vision"], "Task"], [2317, 2319, ["image", "classification"], "Task"], [2320, 2324, ["weakly", "supervised", "semantic", "segmentation"], "Task"], [2325, 2327, ["object", "detection"], "Task"], [2328, 2330, ["object", "tracking"], "Task"], [2350, 2351, ["MIL"], "Method"], [2360, 2361, ["MIL"], "Method"], [2386, 2387, ["WSOD"], "Task"], [2411, 2412, ["WSOD"], "Task"], [2428, 2430, ["network", "training"], "Method"], [2450, 2452, ["localize", "objects"], "Task"], [2484, 2485, ["WSOD"], "Task"], [2522, 2524, ["WSOD", "problem"], "Task"], [2543, 2545, ["exemplar", "model"], "Method"], [2568, 2572, ["CRF", "-", "based", "model"], "Method"], [2574, 2576, ["localize", "objects"], "Task"], [2582, 2584, ["DPM", "model"], "Method"], [2585, 2587, ["weak", "supervisions"], "Method"], [2588, 2589, ["WSOD"], "Task"], [2595, 2599, ["Bayesian", "latent", "topic", "models"], "Method"], [2624, 2626, ["robust", "WSOD"], "Task"], [2641, 2642, ["detector"], "Method"], [2664, 2665, ["detectors"], "Method"], [2674, 2675, ["CNNs"], "Method"], [2676, 2678, ["computer", "vision"], "Task"], [2686, 2687, ["WSOD"], "Task"], [2688, 2689, ["CNNs"], "Method"], [2699, 2700, ["ImageNet"], "Material"], [2706, 2708, ["feature", "extractors"], "Method"], [2753, 2759, ["end", "-", "to", "-", "end", "networks"], "Method"], [2760, 2761, ["WSOD"], "Task"], [2769, 2771, ["network", "training"], "Method"], [2773, 2776, ["proposal", "classification", "scores"], "Metric"], [2778, 2781, ["image", "classification", "scores"], "Metric"], [2807, 2809, ["max", "pooling"], "Method"], [2810, 2811, ["aggregation"], "Task"], [2817, 2821, ["weighted", "sum", "pooing", "strategy"], "Method"], [2844, 2848, ["weakly", "supervised", "segmentation", "map"], "Method"], [2854, 2856, ["filter", "proposals"], "Task"], [2861, 2865, ["weakly", "supervised", "segmentation", "network"], "Method"], [2866, 2867, ["WSOD"], "Task"], [2898, 2900, ["instance", "classifiers"], "Method"], [2902, 2904, ["proposal", "clusters"], "Method"], [2909, 2911, ["online", "approach"], "Method"], [2934, 2937, ["weighted", "sum", "pooing"], "Method"], [2942, 2945, ["spatial", "regulariser", "”"], "Method"], [2974, 2975, ["graph"], "Method"], [2977, 2979, ["proposal", "clusters"], "Method"], [2990, 2992, ["graph", "model"], "Method"], [2998, 3000, ["network", "training"], "Method"], [3107, 3108, ["classifiers"], "Task"], [3136, 3142, ["end", "-", "to", "-", "end", "training"], "Task"], [3160, 3164, ["Fast", "R", "-", "CNN"], "Method"], [3166, 3168, ["classification", "loss"], "Metric"], [3169, 3172, ["boundingbox", "regression", "loss"], "Method"], [3174, 3176, ["proposal", "classification"], "Task"], [3177, 3179, ["feature", "extraction"], "Task"], [3180, 3184, ["fully", "supervised", "object", "detection"], "Task"], [3186, 3188, ["MIL", "networks"], "Method"], [3195, 3197, ["proposal", "classification"], "Task"], [3198, 3200, ["feature", "extraction"], "Task"], [3201, 3202, ["WSOD"], "Task"], [3206, 3212, ["end", "-", "to", "-", "end", "training"], "Method"], [3218, 3224, ["end", "-", "to", "-", "end", "training"], "Method"], [3257, 3259, ["spatial", "regulariser"], "Method"], [3325, 3327, ["network", "computations"], "Method"], [3348, 3354, ["end", "-", "to", "-", "end", "training"], "Task"], [3370, 3373, ["instance", "classifier", "refinement"], "Task"], [3389, 3392, ["weakly", "supervised", "learning"], "Task"], [3405, 3407, ["MIL", "networks"], "Method"], [3416, 3419, ["Region", "of", "Interest"], "Method"], [3420, 3421, ["RoI"], "Method"], [3425, 3428, ["Spatial", "Pyramid", "Pooling"], "Method"], [3429, 3430, ["SPP"], "Method"], [3440, 3442, ["model", "acceleration"], "Task"], [3457, 3461, ["sharing", "proposal", "feature", "strategy"], "Method"], [3467, 3471, ["multi", "-", "task", "learning"], "Task"], [3474, 3478, ["multi", "-", "task", "learning"], "Method"], [3555, 3556, ["training"], "Task"], [3566, 3567, ["convolutional"], "Method"], [3568, 3570, ["conv", ")"], "Method"], [3581, 3584, ["conv", "feature", "map"], "Method"], [3598, 3600, ["fully", "connected"], "Method"], [3601, 3602, ["fc"], "Method"], [3627, 3629, ["instance", "classifiers"], "Method"], [3634, 3635, ["classifiers"], "Method"], [3641, 3644, ["proposal", "classification", "scores"], "Metric"], [3673, 3677, ["back", "-", "propagation", "process"], "Method"], [3690, 3691, ["classifiers"], "Method"], [3703, 3705, ["refined", "classifier"], "Method"], [3711, 3713, ["basic", "classifier"], "Method"], [3715, 3716, ["supervisions"], "Method"], [3719, 3721, ["refined", "classifier"], "Method"], [3727, 3729, ["refined", "classifier"], "Method"], [3741, 3744, ["refined", "instance", "classifiers"], "Task"], [3746, 3748, ["proposal", "clusters"], "Method"], [3796, 3798, ["instance", "classifiers"], "Method"], [3829, 3832, ["basic", "instance", "classifier"], "Method"], [3834, 3839, ["-", "th", "refined", "instance", "classifier"], "Method"], [3845, 3848, ["predicted", "score", "matrices"], "Metric"], [3850, 3853, ["basic", "instance", "classifier"], "Method"], [3857, 3860, ["refined", "instance", "classifier"], "Method"], [3898, 3900, ["instance", "classifier"], "Method"], [3923, 3927, ["-", "th", "instance", "classifier"], "Method"], [3942, 3946, ["-", "th", "instance", "classifier"], "Method"], [3954, 3957, ["-", "th", "refinement"], "Method"], [4015, 4017, ["object", "clusters"], "Task"], [4042, 4044, ["confidence", "score"], "Metric"], [4067, 4069, ["object", "clusters"], "Method"], [4121, 4123, ["instance", "classifiers"], "Method"], [4130, 4131, ["classifier"], "Method"], [4133, 4135, ["instance", "classifiers"], "Method"], [4154, 4156, ["instance", "classifier"], "Method"], [4203, 4207, ["weighted", "sum", "pooling", "strategy"], "Method"], [4210, 4212, ["instance", "classifier"], "Method"], [4337, 4339, ["softmax", "layer"], "Method"], [4356, 4360, ["element", "-", "wise", "product"], "Method"], [4364, 4366, ["image", "score"], "Metric"], [4387, 4389, ["branches", "framework"], "Method"], [4429, 4432, ["weighted", "sum", "pooling"], "Method"], [4449, 4451, ["instance", "classifier"], "Method"], [4454, 4461, ["multi", "-", "class", "cross", "entropy", "loss", "Eq"], "Metric"], [4474, 4476, ["training", "strategy"], "Method"], [4478, 4480, ["instance", "classifiers"], "Method"], [4516, 4518, ["classifier", "refinement"], "Method"], [4520, 4526, ["end", "-", "to", "-", "end", "network"], "Method"], [4530, 4531, ["classifier"], "Method"], [4536, 4538, ["instance", "classifier"], "Method"], [4549, 4553, ["-", "th", "refined", "classifier"], "Method"], [4577, 4582, ["softmax", "over", "-", "classes", "layer"], "Method"], [4591, 4594, ["Instance", "classifier", "refinement"], "Method"], [4613, 4614, ["classifiers"], "Method"], [4635, 4639, ["-", "th", "instance", "classifier"], "Method"], [4665, 4667, ["loss", "Eq"], "Metric"], [4715, 4718, ["refined", "instance", "classifier"], "Method"], [4875, 4876, ["Optimize"], "Task"], [4902, 4905, ["Stochastic", "Gradient", "Descent"], "Method"], [4906, 4907, ["SGD"], "Method"], [4946, 4950, ["back", "-", "propagation", "process"], "Method"], [4969, 4971, ["proposal", "features"], "Method"], [4972, 4973, ["classifiers"], "Method"], [4991, 4994, ["alternating", "training", "strategy"], "Method"], [5003, 5005, ["complete", "model"], "Method"], [5068, 5071, ["Proposal", "cluster", "learning"], "Method"], [5079, 5082, ["refined", "instance", "classifiers"], "Method"], [5084, 5086, ["proposal", "clusters"], "Method"], [5090, 5093, ["proposal", "cluster", "learning"], "Method"], [5113, 5116, ["-", "th", "refinement"], "Task"], [5215, 5217, ["instance", "classifiers"], "Method"], [5323, 5327, ["-", "th", "object", "cluster"], "Method"], [5411, 5414, ["refined", "instance", "classifier"], "Method"], [5464, 5466, ["loss", "functions"], "Method"], [5510, 5511, ["subsubsection"], "Method"], [5513, 5516, ["proposal", "cluster", "centers"], "Method"], [5555, 5558, ["Proposal", "cluster", "centers"], "Method"], [5926, 5929, ["Proposal", "cluster", "centers"], "Method"], [6377, 6379, ["center", "proposal"], "Method"], [6393, 6397, ["graph", "-", "based", "method"], "Method"], [6453, 6456, ["-", "means", "algorithm"], "Method"], [6533, 6535, ["Proposal", "clusters"], "Method"], [6549, 6550, ["IoUs"], "Method"], [6560, 6561, ["subsubsection"], "Method"], [6563, 6565, ["proposal", "clusters"], "Method"], [6775, 6776, ["subsubsection"], "Method"], [6779, 6781, ["instance", "classifiers"], "Method"], [6792, 6795, ["refined", "instance", "classifier"], "Method"], [6817, 6818, ["classifiers"], "Method"], [6949, 6951, ["refined", "classifiers"], "Method"], [6980, 6984, ["iterative", "instance", "classifier", "refinement"], "Method"], [7059, 7061, ["weighted", "version"], "Method"], [7123, 7127, ["cluster", "center", "confidence", "score"], "Metric"], [7209, 7211, ["instance", "classifiers"], "Method"], [7286, 7289, ["-", "th", "refinement"], "Task"], [7336, 7338, ["object", "clusters"], "Task"], [7361, 7363, ["prediction", "scores"], "Metric"], [7400, 7401, ["refinement"], "Task"], [7416, 7419, ["cluster", "confidence", "score"], "Metric"], [7480, 7483, ["assigning", "label", "approach"], "Method"], [7514, 7516, ["instance", "classifiers"], "Method"], [7521, 7523, ["detection", "scores"], "Metric"], [7542, 7543, ["classifiers"], "Method"], [7547, 7551, ["Non", "-", "Maxima", "Suppression"], "Method"], [7552, 7553, ["NMS"], "Method"], [7576, 7577, ["datasets"], "Metric"], [7578, 7580, ["evaluation", "metrics"], "Metric"], [7654, 7656, ["evaluation", "metrics"], "Metric"], [7666, 7669, ["PASCAL", "VOC", "2007"], "Material"], [7670, 7672, ["2012", "datasets"], "Material"], [7698, 7701, ["PASCAL", "VOC", "2007"], "Material"], [7702, 7704, ["2012", "datasets"], "Material"], [7754, 7755, ["evaluation"], "Task"], [7756, 7757, ["mAP"], "Metric"], [7758, 7759, ["CorLoc"], "Material"], [7763, 7766, ["PASCAL", "VOC", "protocol"], "Metric"], [7767, 7769, ["Average", "Precision"], "Metric"], [7770, 7771, ["AP"], "Metric"], [7774, 7777, ["mean", "of", "AP"], "Metric"], [7778, 7779, ["mAP"], "Metric"], [7782, 7784, ["evaluation", "metric"], "Metric"], [7793, 7795, ["Correct", "Localization"], "Task"], [7796, 7797, ["CorLoc"], "Material"], [7809, 7811, ["localization", "accuracy"], "Metric"], [7820, 7822, ["PASCAL", "criterion"], "Metric"], [7933, 7934, ["mAP"], "Metric"], [7938, 7939, ["ImageNet"], "Material"], [8008, 8009, ["mAP@0.5"], "Metric"], [8010, 8013, ["mAP@", "[", ".5"], "Metric"], [8020, 8022, ["PASCAL", "criterion"], "Metric"], [8046, 8047, ["IoU"], "Metric"], [8075, 8076, ["VGG16"], "Method"], [8082, 8084, ["conv", "layers"], "Method"], [8085, 8089, ["max", "-", "pooling", "layers"], "Method"], [8098, 8102, ["max", "-", "pooling", "layer"], "Method"], [8144, 8149, ["penultimate", "max", "-", "pooling", "layer"], "Method"], [8152, 8154, ["conv", "layers"], "Method"], [8156, 8159, ["dilated", "conv", "layers"], "Method"], [8167, 8169, ["Gaussian", "distributions"], "Method"], [8182, 8183, ["training"], "Task"], [8185, 8189, ["mini", "-", "batch", "size"], "Metric"], [8190, 8191, ["SGD"], "Method"], [8199, 8201, ["PASCAL", "VOC"], "Material"], [8202, 8203, ["ImageNet"], "Material"], [8205, 8208, ["MS", "-", "COCO"], "Material"], [8212, 8214, ["learning", "rate"], "Metric"], [8231, 8234, ["PASCAL", "VOC", "2007"], "Material"], [8235, 8238, ["PASCAL", "VOC", "2012"], "Material"], [8239, 8240, ["ImageNet"], "Material"], [8242, 8246, ["MS", "-", "COCO", "datasets"], "Material"], [8253, 8255, ["learning", "rate"], "Metric"], [8270, 8273, ["PASCAL", "VOC", "2007"], "Material"], [8274, 8277, ["PASCAL", "VOC", "2012"], "Material"], [8278, 8279, ["ImageNet"], "Material"], [8281, 8285, ["MS", "-", "COCO", "datasets"], "Material"], [8300, 8302, ["Selective", "Search"], "Method"], [8303, 8304, ["EdgeBox"], "Method"], [8306, 8307, ["MCG"], "Method"], [8315, 8316, ["image"], "Material"], [8318, 8320, ["PASCAL", "VOC"], "Material"], [8332, 8334, ["data", "augmentation"], "Task"], [8365, 8367, ["instance", "classifiers"], "Method"], [8464, 8467, ["supervised", "object", "detector"], "Method"], [8491, 8495, ["Fast", "R", "-", "CNN"], "Method"], [8496, 8497, ["FRCNN"], "Method"], [8526, 8527, ["FRCNN"], "Method"], [8528, 8529, ["NMS"], "Method"], [8547, 8551, ["Caffe", "deep", "learning", "framework"], "Method"], [8553, 8554, ["Python"], "Material"], [8555, 8557, ["C", "++"], "Material"], [8559, 8562, ["-", "means", "algorithm"], "Method"], [8570, 8573, ["scikit", "-", "learn"], "Method"], [8582, 8587, ["NVIDIA", "GTX", "TitanX", "Pascal", "GPU"], "Material"], [8621, 8624, ["instance", "classifier", "refinement"], "Method"], [8626, 8629, ["proposal", "generation", "methods"], "Method"], [8631, 8633, ["refinement", "strategies"], "Method"], [8713, 8716, ["VOC", "2007", "dataset"], "Material"], [8728, 8731, ["instance", "classifier", "refinement"], "Method"], [8755, 8757, ["refinement", "methods"], "Method"], [8760, 8762, ["instance", "classifier"], "Method"], [8794, 8795, ["refinement"], "Method"], [8872, 8873, ["WSDDN"], "Method"], [8874, 8876, ["Selective", "Search"], "Method"], [8885, 8889, ["proposal", "cluster", "generation", "methods"], "Method"], [8895, 8899, ["proposal", "cluster", "generation", "methods"], "Method"], [8918, 8922, ["scoring", "proposal", "based", "method"], "Method"], [8930, 8934, ["graph", "-", "based", "method"], "Method"], [8938, 8940, ["refinement", "times"], "Metric"], [8942, 8946, ["graph", "-", "based", "method"], "Method"], [8962, 8966, ["graph", "-", "based", "method"], "Method"], [8979, 8981, ["refinement", "strategies"], "Method"], [8988, 8990, ["refinement", "strategies"], "Method"], [8993, 8996, ["assigning", "label", "method"], "Method"], [9038, 9041, ["alternating", "training", "strategy"], "Method"], [9048, 9050, ["online", "training"], "Method"], [9066, 9068, ["Online", "training"], "Task"], [9071, 9073, ["training", "time"], "Metric"], [9089, 9091, ["times", "refinement"], "Task"], [9093, 9095, ["alternating", "strategy"], "Method"], [9135, 9137, ["weighted", "loss"], "Method"], [9144, 9146, ["weighted", "loss"], "Metric"], [9186, 9188, ["unweighted", "loss"], "Method"], [9193, 9195, ["unweighted", "loss"], "Method"], [9196, 9197, ["mAP"], "Metric"], [9198, 9199, ["CorLoc"], "Material"], [9213, 9214, ["refinement"], "Task"], [9225, 9227, ["alternating", "strategy"], "Method"], [9230, 9232, ["weighted", "loss"], "Method"], [9237, 9238, ["mAP"], "Metric"], [9239, 9240, ["CorLoc"], "Material"], [9374, 9375, ["mAP"], "Metric"], [9376, 9377, ["CorLoc"], "Material"], [9381, 9382, ["mAP"], "Metric"], [9383, 9384, ["CorLoc"], "Material"], [9396, 9397, ["cluster"], "Method"], [9473, 9479, ["multi", "-", "scale", "training", "and", "testing"], "Task"], [9523, 9524, ["FRCNN"], "Method"], [9531, 9532, ["mAP"], "Metric"], [9533, 9534, ["CorLoc"], "Material"], [9545, 9546, ["mAP"], "Metric"], [9547, 9548, ["CorLoc"], "Material"], [9747, 9751, ["graph", "-", "based", "method"], "Method"], [9775, 9777, ["VOC", "2007"], "Material"], [9778, 9779, ["2012"], "Material"], [9816, 9817, ["VGG16"], "Method"], [9855, 9858, ["ensemble", "of", "models"], "Method"], [9886, 9889, ["weighted", "sum", "pooling"], "Method"], [9890, 9894, ["objectness", "measure", "of", "EdgeBox"], "Method"], [9896, 9898, ["spatial", "regulariser"], "Method"], [9936, 9938, ["complete", "network"], "Method"], [9984, 9986, ["multiple", "models"], "Method"], [10002, 10004, ["VGG", "M"], "Method"], [10005, 10006, ["VGG16"], "Method"], [10008, 10015, ["PCL", "-", "OB", "-", "G", "-", "Ens"], "Method"], [10061, 10068, ["PCL", "-", "OB", "-", "G", "-", "Ens"], "Method"], [10072, 10079, ["PCL", "-", "OB", "-", "G", "-", "Ens."], "Method"], [10080, 10081, ["FRCNN"], "Method"], [10139, 10142, ["MS", "-", "COCO"], "Material"], [10185, 10186, ["detection"], "Task"], [10190, 10193, ["conference", "version", "method"], "Method"], [10266, 10267, ["WSDDN"], "Method"], [10269, 10272, ["WSDDN", "+", "context"], "Method"], [10396, 10397, ["detection"], "Task"], [10400, 10403, ["WSDDN", "+", "context"], "Task"], [10485, 10487, ["failure", "detection"], "Metric"], [10652, 10653, ["runtime"], "Metric"], [10654, 10656, ["proposal", "generation"], "Task"], [10737, 10738, ["detection"], "Task"], [10756, 10758, ["proposal", "clusters"], "Method"], [10760, 10763, ["refined", "instance", "classifiers"], "Method"], [10774, 10777, ["proposal", "cluster", "generation"], "Task"], [10778, 10780, ["classifier", "refinement"], "Task"], [10791, 10793, ["classifier", "refinement"], "Method"], [10802, 10804, ["instance", "classifiers"], "Method"], [10805, 10809, ["multiple", "instance", "learning", "networks"], "Method"], [10811, 10814, ["online", "training", "algorithm"], "Method"], [10856, 10862, ["deformation", "of", "non", "-", "rigid", "objects"], "Task"], [10880, 10882, ["learning", "algorithm"], "Method"], [10890, 10895, ["weakly", "supervised", "visual", "learning", "tasks"], "Task"], [10897, 10901, ["weakly", "supervised", "semantic", "segmentation"], "Task"], [10998, 11002, ["Electronics", "and", "Information", "Engineering"], "Task"], [11026, 11030, ["Electronic", "Information", "and", "Communications"], "Task"], [11031, 11032, ["HUST"], "Material"], [11038, 11040, ["Computer", "Science"], "Task"], [11060, 11062, ["image", "classification"], "Task"], [11063, 11065, ["object", "detection"], "Task"], [11081, 11085, ["Electronics", "Information", "and", "Communications"], "Task"], [11102, 11106, ["communication", "and", "information", "system"], "Task"], [11107, 11109, ["Ph.D.", "degree"], "Task"], [11110, 11112, ["computer", "vision"], "Task"], [11194, 11195, ["Cybernetics"], "Task"], [11196, 11198, ["Pattern", "Recognition"], "Task"], [11199, 11201, ["Computer", "Vision"], "Task"], [11202, 11204, ["Image", "Understanding"], "Task"], [11205, 11206, ["Neurocomputing"], "Task"], [11207, 11208, ["NIPS"], "Material"], [11209, 11210, ["ICML"], "Material"], [11211, 11212, ["CVPR"], "Material"], [11213, 11214, ["ICCV"], "Material"], [11215, 11216, ["ECCV"], "Material"], [11222, 11224, ["computer", "vision"], "Task"], [11225, 11227, ["machine", "learning"], "Task"], [11229, 11231, ["object", "recognition"], "Task"], [11245, 11247, ["Information", "Engineering"], "Task"], [11296, 11298, ["image", "retrieval"], "Task"], [11299, 11300, ["classification"], "Task"], [11301, 11304, ["3D", "shape", "recognition"], "Task"], [11305, 11309, ["person", "re", "-", "identification"], "Task"], [11310, 11312, ["semantic", "segmentation"], "Task"], [11313, 11315, ["deep", "learning"], "Task"], [11338, 11342, ["Electronics", "and", "Information", "Engineering"], "Task"], [11388, 11392, ["Communication", "and", "Information", "Engineering"], "Task"], [11415, 11417, ["Computer", "Science"], "Task"], [11427, 11429, ["random", "forests"], "Task"], [11430, 11432, ["deep", "learning"], "Task"], [11479, 11483, ["electronics", "and", "information", "engineering"], "Task"], [11493, 11497, ["Electronic", "Information", "and", "Communications"], "Task"], [11498, 11499, ["HUST"], "Material"], [11512, 11516, ["Anti", "-", "Counterfeiting", "Technology"], "Task"], [11517, 11518, ["HUST"], "Material"], [11523, 11525, ["object", "recognition"], "Task"], [11526, 11528, ["shape", "analysis"], "Task"], [11529, 11532, ["scene", "text", "recognition"], "Task"], [11533, 11535, ["intelligent", "systems"], "Task"], [11543, 11545, ["Pattern", "Recognition"], "Task"], [11546, 11549, ["Pattern", "Recognition", "Letters"], "Task"], [11550, 11556, ["Neurocomputing", "and", "Frontiers", "of", "Computer", "Science"], "Task"], [11566, 11568, ["Computer", "Science"], "Task"], [11588, 11592, ["Electronics", "and", "Information", "Engineering"], "Task"], [11627, 11631, ["Electronic", "Information", "and", "Communications"], "Task"], [11632, 11633, ["HUST"], "Material"], [11639, 11641, ["computer", "vision"], "Task"], [11642, 11643, ["multimedia"], "Task"], [11645, 11647, ["machine", "learning"], "Task"], [58, 60, ["object", "detection"], "Task"], [178, 179, ["convolutional"], "Method"], [187, 188, ["MIL"], "Method"], [215, 218, ["MS", "-", "COCO"], "Material"], [243, 244, ["convolutional"], "Method"], [458, 459, ["MIL"], "Method"], [504, 505, ["WSOD"], "Task"], [556, 557, ["MIL"], "Method"], [615, 616, ["MIL"], "Method"], [661, 662, ["MIL"], "Method"], [673, 675, ["object", "detection"], "Task"], [918, 919, ["IoU"], "Metric"], [926, 928, ["PASCAL", "VOC"], "Material"], [1437, 1438, ["MIL"], "Method"], [1698, 1699, ["MIL"], "Method"], [1931, 1932, ["PCL"], "Method"], [2210, 2213, ["weakly", "supervised", "learning"], "Task"], [2252, 2253, ["MIL"], "Method"], [2273, 2274, ["MIL"], "Method"], [2441, 2442, ["CNN"], "Method"], [2447, 2448, ["MIL"], "Method"], [2482, 2484, ["object", "detection"], "Task"], [2637, 2638, ["MIL"], "Method"], [2655, 2656, ["MIL"], "Method"], [2693, 2694, ["CNN"], "Method"], [2711, 2712, ["CNN"], "Method"], [2735, 2736, ["CNN"], "Method"], [2765, 2766, ["MIL"], "Method"], [3007, 3008, ["MIL"], "Method"], [3034, 3035, ["CNN"], "Method"], [3191, 3192, ["MIL"], "Method"], [3434, 3435, ["convolutional"], "Method"], [3573, 3574, ["SPP"], "Method"], [3622, 3623, ["MIL"], "Method"], [4107, 4108, ["MIL"], "Method"], [4149, 4150, ["MIL"], "Method"], [4165, 4166, ["MIL"], "Method"], [4263, 4264, ["MIL"], "Method"], [4305, 4306, ["fc"], "Method"], [4315, 4316, ["fc"], "Method"], [4323, 4324, ["fc"], "Method"], [4512, 4513, ["MIL"], "Method"], [4567, 4568, ["fc"], "Method"], [4952, 4953, ["SGD"], "Method"], [6055, 6056, ["IoU"], "Metric"], [6638, 6639, ["IoU"], "Metric"], [7342, 7343, ["MIL"], "Method"], [7383, 7384, ["MIL"], "Method"], [7674, 7675, ["ImageNet"], "Material"], [7680, 7683, ["MS", "-", "COCO"], "Material"], [7739, 7740, ["2012"], "Material"], [7826, 7827, ["IoU"], "Metric"], [7836, 7837, ["ImageNet"], "Material"], [7941, 7944, ["MS", "-", "COCO"], "Material"], [7977, 7980, ["MS", "-", "COCO"], "Material"], [8026, 8027, ["IoU"], "Metric"], [8032, 8035, ["MS", "-", "COCO"], "Material"], [8044, 8045, ["mAP"], "Metric"], [8070, 8071, ["ImageNet"], "Material"], [8072, 8074, ["VGG", "M"], "Method"], [8091, 8092, ["fc"], "Method"], [8104, 8105, ["SPP"], "Method"], [8110, 8111, ["fc"], "Method"], [8321, 8322, ["ImageNet"], "Material"], [8324, 8327, ["MS", "-", "COCO"], "Material"], [8389, 8390, ["IoU"], "Metric"], [8500, 8501, ["VGG16"], "Method"], [8531, 8532, ["IoU"], "Metric"], [8538, 8539, ["AP"], "Metric"], [8645, 8646, ["IoU"], "Metric"], [8674, 8675, ["IoU"], "Metric"], [8719, 8721, ["VGG", "M"], "Method"], [8750, 8751, ["MIL"], "Method"], [9110, 9111, ["PCL"], "Method"], [9258, 9259, ["IoU"], "Metric"], [9267, 9268, ["IoU"], "Metric"], [9420, 9421, ["IoU"], "Metric"], [9559, 9560, ["WSOD"], "Task"], [9737, 9738, ["PCL"], "Method"], [9813, 9815, ["VGG", "M"], "Method"], [9818, 9819, ["PCL"], "Method"], [9824, 9826, ["VGG", "M"], "Method"], [9827, 9828, ["PCL"], "Method"], [9833, 9834, ["VGG16"], "Method"], [9878, 9879, ["MIL"], "Method"], [9916, 9917, ["MIL"], "Method"], [9931, 9932, ["MIL"], "Method"], [10052, 10053, ["FRCNN"], "Method"], [10107, 10108, ["ImageNet"], "Material"], [10275, 10276, ["PCL"], "Method"], [10285, 10286, ["MIL"], "Method"], [10431, 10433, ["object", "detection"], "Task"], [10489, 10491, ["VOC", "2007"], "Material"], [10493, 10494, ["PCL"], "Method"], [10497, 10498, ["FRCNN"], "Method"], [10640, 10641, ["MIL"], "Method"], [10674, 10675, ["MIL"], "Method"], [10742, 10743, ["MIL"], "Method"], [10766, 10768, ["object", "detection"], "Task"], [11362, 11363, ["2012"], "Material"], [11382, 11383, ["2012"], "Material"], [11433, 11435, ["object", "detection"], "Task"]]} {"doc_id": "2f04ba0f74df046b0080ca78e56898bd4847898b", "text": ["document", ":", "Aggregate", "Channel", "Features", "for", "Multi", "-", "view", "Face", "Detection", "Face", "detection", "has", "drawn", "much", "attention", "in", "recent", "decades", "since", "the", "seminal", "work", "by", "Viola", "and", "Jones", ".", "While", "many", "subsequences", "have", "improved", "the", "work", "with", "more", "powerful", "learning", "algorithms", ",", "the", "feature", "representation", "used", "for", "face", "detection", "still", "ca", "n’t", "meet", "the", "demand", "for", "effectively", "and", "efficiently", "handling", "faces", "with", "large", "appearance", "variance", "in", "the", "wild", ".", "To", "solve", "this", "bottleneck", ",", "we", "borrow", "the", "concept", "of", "channel", "features", "to", "the", "face", "detection", "domain", ",", "which", "extends", "the", "image", "channel", "to", "diverse", "types", "like", "gradient", "magnitude", "and", "oriented", "gradient", "histograms", "and", "therefore", "encodes", "rich", "information", "in", "a", "simple", "form", ".", "We", "adopt", "a", "novel", "variant", "called", "aggregate", "channel", "features", ",", "make", "a", "full", "exploration", "of", "feature", "design", ",", "and", "discover", "a", "multi", "-", "scale", "version", "of", "features", "with", "better", "performance", ".", "To", "deal", "with", "poses", "of", "faces", "in", "the", "wild", ",", "we", "propose", "a", "multi", "-", "view", "detection", "approach", "featuring", "score", "re", "-", "ranking", "and", "detection", "adjustment", ".", "Following", "the", "learning", "pipelines", "in", "Viola", "-", "Jones", "framework", ",", "the", "multi", "-", "view", "face", "detector", "using", "aggregate", "channel", "features", "shows", "competitive", "performance", "against", "state", "-", "of", "-", "the", "-", "art", "algorithms", "on", "AFW", "and", "FDDB", "testsets", ",", "while", "runs", "at", "42", "FPS", "on", "VGA", "images", ".", "section", ":", "Introduction", "Human", "face", "detection", "have", "long", "been", "one", "of", "the", "most", "fundamental", "problems", "in", "computer", "vision", "and", "human", "-", "computer", "interaction", ".", "In", "the", "past", "decade", ",", "the", "most", "influential", "work", "should", "be", "the", "face", "detection", "framework", "proposed", "by", "Viola", "and", "Jones", ".", "The", "Viola", "-", "Jones", "(", "abbreviated", "as", "VJ", "below", ")", "framework", "uses", "rectangular", "Haar", "-", "like", "features", "and", "learns", "the", "hypothesis", "using", "Adaboost", "algorithm", ".", "Combined", "with", "the", "attentional", "cascade", "structure", ",", "the", "VJ", "detector", "achieved", "real", "-", "time", "face", "detection", "at", "that", "time", ".", "Despite", "the", "great", "success", "of", "the", "VJ", "detector", ",", "the", "performance", "is", "still", "far", "from", "satisfactory", "due", "to", "the", "large", "appearance", "variance", "of", "faces", "in", "unconstrained", "settings", ".", "To", "handle", "faces", "in", "the", "wild", ",", "many", "subsequences", "of", "VJ", "framework", "merged", ".", "These", "methods", "mainly", "get", "the", "performance", "gains", "in", "two", "aspects", ",", "more", "complicated", "features", "and", "(", "or", ")", "more", "powerful", "learning", "algorithms", ".", "As", "the", "combination", "of", "boosting", "and", "cascade", "has", "been", "proven", "to", "be", "quite", "effective", "in", "face", "detection", ",", "the", "bottleneck", "lies", "in", "the", "feature", "representation", "since", "complicated", "features", "adopted", "in", "the", "above", "literatures", "bring", "about", "limited", "performance", "gains", "at", "the", "cost", "of", "large", "computation", "cost", ".", "Lately", "in", "another", "domain", "of", "pedestrian", "detection", ",", "a", "family", "of", "channel", "features", "has", "achieved", "record", "performances", ".", "Channel", "features", "compute", "registered", "maps", "of", "the", "original", "images", "like", "gradients", "and", "histograms", "of", "oriented", "gradients", "and", "then", "extract", "features", "on", "these", "extended", "channels", ".", "The", "classifier", "learning", "process", "follows", "the", "VJ", "framework", "pipeline", ".", "In", "this", "paper", ",", "we", "adopt", "a", "variant", "of", "channel", "features", "called", "aggregate", "channel", "features", ",", "which", "are", "extracted", "directly", "as", "pixel", "values", "on", "subsampled", "channels", ".", "Channel", "extension", "offers", "rich", "representation", "capacity", ",", "while", "simple", "feature", "form", "guarantees", "fast", "computation", ".", "With", "these", "two", "superiorities", ",", "the", "aggregate", "channel", "features", "break", "through", "the", "bottleneck", "in", "VJ", "framework", "and", "have", "the", "potential", "to", "make", "great", "advance", "in", "face", "detection", ".", "As", "we", "mainly", "concentrate", "our", "efforts", "to", "the", "feature", "representation", "rather", "than", "learning", "algorithms", "in", "this", "paper", ",", "we", "not", "only", "just", "adopt", "the", "aggregate", "channel", "features", "in", "face", "detection", ",", "but", "also", "try", "to", "explore", "the", "full", "potential", "of", "this", "novel", "representation", ".", "To", "do", "so", ",", "we", "make", "a", "deep", "and", "all", "-", "round", "investigation", "into", "the", "specific", "feature", "parameters", "concerning", "channel", "types", ",", "feature", "pool", "size", ",", "subsampling", "method", ",", "feature", "scale", "and", "so", "on", ",", "which", "gives", "insights", "into", "the", "feature", "design", "and", "hopefully", "provides", "helpful", "guidelines", "for", "practitioners", ".", "Through", "the", "deep", "exploration", ",", "we", "find", "that", ":", "1", ")", "multi", "-", "scaling", "the", "feature", "representation", "further", "enriches", "the", "representation", "capacity", "since", "original", "aggregate", "channel", "features", "have", "uniform", "feature", "scale", ";", "2", ")", "different", "combinations", "of", "channel", "types", "impact", "the", "performance", "greatly", ",", "while", "for", "face", "detection", "the", "color", "channel", "in", "LUV", "space", ",", "plus", "gradient", "magnitude", "channel", "and", "gradient", "histograms", "channels", "in", "RGB", "space", "show", "best", "result", ";", "3", ")", "multi", "-", "view", "detection", "is", "proven", "to", "be", "a", "good", "match", "with", "aggregate", "channel", "features", "as", "the", "representation", "naturally", "encodes", "the", "facial", "structure", "(", "Figure", "[", "reference", "]", ")", ".", "Although", "multi", "-", "view", "detection", "could", "effectively", "deal", "with", "diverse", "poses", ",", "additional", "issues", "come", "up", "as", "how", "to", "merge", "detections", "output", "by", "separately", "trained", "subview", "detectors", ",", "and", "how", "to", "deal", "with", "the", "offsets", "of", "location", "and", "scale", "between", "output", "detections", "and", "ground", "-", "truth", ".", "We", "solve", "these", "problems", "by", "carefully", "designed", "post", "-", "processing", "including", "score", "re", "-", "ranking", ",", "detection", "merging", "and", "bounding", "box", "adjustment", ".", "The", "detailed", "experimental", "exploration", "of", "aggregate", "channel", "features", ",", "along", "with", "our", "improvements", "on", "multi", "-", "view", "detection", ",", "leads", "to", "large", "performance", "gain", "in", "face", "detection", "in", "the", "wild", ".", "On", "two", "challenging", "face", "databases", ",", "AFW", "and", "FDDB", ",", "the", "proposed", "multi", "-", "view", "face", "detector", "shows", "competitive", "performance", "against", "state", "-", "of", "-", "the", "-", "art", "detectors", "in", "both", "detection", "accuracy", "and", "speed", ".", "The", "remaining", "parts", "of", "this", "paper", "are", "organized", "as", "follows", ".", "Section", "2", "revisits", "related", "work", "in", "face", "detection", ".", "Section", "3", "describes", "how", "we", "build", "the", "face", "detector", "using", "aggregate", "channel", "features", ".", "Section", "4", "addresses", "problems", "concerning", "multi", "-", "view", "face", "detection", ".", "Experimental", "results", "on", "AFW", "and", "FDDB", "are", "shown", "in", "section", "5", "and", "we", "conclude", "the", "paper", "in", "section", "6", ".", "section", ":", "Related", "work", "Face", "detection", "has", "drawn", "much", "attention", "since", "the", "early", "time", "of", "computer", "vision", ".", "Although", "many", "solutions", "had", "been", "put", "forward", ",", "it", "was", "not", "until", "Viola", "and", "Jones", "proposed", "their", "milestone", "work", "that", "face", "detection", "saw", "surprising", "progress", "in", "the", "past", "decades", ".", "The", "VJ", "face", "detector", "features", "in", "three", "aspects", ":", "fast", "feature", "computation", "via", "integral", "image", "representation", ",", "classifier", "learning", "using", "Adaboost", ",", "and", "the", "attentional", "cascade", "structure", ".", "One", "main", "drawback", "of", "the", "VJ", "framework", "is", "that", "the", "features", "have", "limited", "representation", "capacity", ",", "while", "the", "feature", "pool", "size", "is", "quite", "large", "to", "compensate", "for", "that", ".", "Typically", ",", "in", "a", "detection", "window", ",", "the", "number", "of", "Haar", "-", "like", "features", "is", "160", ",", "000", ".", "To", "address", "the", "problem", ",", "efforts", "are", "made", "in", "two", "directions", ".", "Some", "focus", "on", "more", "complicated", "features", "like", "HoG", ",", "SURF", ".", "Some", "aim", "to", "speed", "up", "the", "feature", "selection", "in", "a", "heuristic", "way", ".", "However", ",", "the", "problem", "has", "n’t", "been", "solved", "perfectly", ".", "In", "this", "paper", ",", "we", "mainly", "focus", "on", "the", "feature", "representation", "part", "and", "make", "a", "deep", "exploration", "into", "it", ",", "which", "is", "complementary", "to", "existing", "work", "on", "the", "learning", "algorithm", "and", "classifier", "structure", "in", "the", "VJ", "framework", ".", "Recently", "channel", "features", "have", "been", "proposed", "and", "shown", "record", "performance", "in", "pedestrian", "detection", ".", "Due", "to", "the", "channel", "extension", "to", "diverse", "types", "like", "gradients", "and", "local", "histograms", ",", "the", "features", "show", "richer", "representation", "capacity", "for", "classification", ".", "However", ",", "the", "features", "are", "extracted", "as", "rectangular", "sums", "at", "various", "locations", "and", "scales", "which", "we", "believe", "leads", "to", "a", "redundant", "feature", "pool", ".", "During", "preparation", "of", "this", "paper", ",", "Mathias", "independently", "discover", "the", "effectiveness", "of", "integral", "channel", "features", "in", "face", "detection", "domain", ".", "In", "this", "paper", ",", "we", "adopt", "a", "novel", "variant", "of", "channel", "features", "called", "aggregate", "channel", "features", ",", "which", "extract", "features", "directly", "as", "pixel", "values", "in", "extended", "channels", "without", "computing", "rectangular", "sums", "at", "various", "locations", "and", "scales", ".", "The", "feature", "has", "powerful", "representation", "capacity", "and", "the", "feature", "pool", "size", "is", "only", "several", "thousands", ".", "Through", "careful", "design", "in", "section", "3", "and", "implementation", "of", "multi", "-", "view", "detection", "in", "section", "4", ",", "the", "aggregate", "channel", "features", "based", "detector", "achieves", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "challenging", "databases", ".", "section", ":", "Proposed", "face", "detector", "In", "this", "section", ",", "we", "make", "a", "full", "exploration", "of", "the", "aggregate", "channel", "features", "in", "the", "context", "of", "face", "detection", ".", "We", "first", "give", "a", "brief", "introduction", "of", "the", "feature", "itself", ",", "including", "its", "computation", ",", "properties", "and", "advantages", "over", "traditional", "Haar", "-", "like", "features", "used", "in", "VJ", "framework", ".", "Then", "the", "detailed", "experimental", "investigation", "is", "described", "in", "two", "parts", ",", "feature", "design", "and", "training", "design", ".", "Before", "that", ",", "some", "guidelines", "concerning", "how", "we", "conduct", "the", "investigation", "are", "demonstrated", ".", "Each", "design", "part", "is", "divided", "into", "several", "separate", "experiments", "ended", "with", "a", "summary", "explaining", "the", "specific", "parameters", "used", "in", "our", "proposed", "face", "detector", ".", "Note", "that", "each", "experiment", "focuses", "on", "only", "one", "parameter", "and", "the", "others", "remain", "constant", ".", "Through", "the", "well", "-", "designed", "experiments", ",", "the", "proposed", "face", "detector", "based", "on", "aggregate", "channel", "features", "is", "built", "step", "by", "step", ".", "Issues", "concerning", "the", "implementation", "of", "multi", "-", "view", "face", "detection", "which", "further", "improves", "the", "performance", "are", "discussed", "in", "the", "next", "section", ".", "subsection", ":", "Feature", "description", "Channel", "extension", ":", "The", "basic", "structure", "of", "the", "aggregate", "channel", "features", "is", "channel", ".", "The", "application", "of", "channel", "has", "a", "long", "history", "since", "digital", "images", "were", "invented", ".", "The", "most", "common", "type", "of", "channel", "should", "be", "the", "color", "channels", "of", "the", "image", ",", "with", "Gray", "-", "scale", "and", "RGB", "being", "typical", "ones", ".", "Besides", "color", "channels", ",", "many", "different", "channel", "types", "have", "been", "invented", "to", "encode", "different", "types", "of", "information", "for", "more", "difficult", "problems", ".", "Generally", ",", "channels", "can", "be", "defined", "as", "a", "registered", "map", "of", "the", "original", "image", ",", "whose", "pixels", "are", "computed", "from", "corresponding", "patches", "of", "original", "pixels", ".", "Different", "channels", "can", "be", "computed", "with", "linear", "or", "non", "-", "linear", "transformation", "of", "the", "original", "image", ".", "To", "allow", "for", "sliding", "window", "detection", ",", "the", "transformations", "are", "constrained", "to", "be", "translationally", "invariant", ".", "Feature", "computation", ":", "Based", "on", "the", "definition", "of", "channels", ",", "the", "computation", "of", "aggregate", "channel", "features", "is", "quite", "simple", ".", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "given", "a", "color", "image", ",", "all", "defined", "channels", "are", "computed", "and", "subsampled", "by", "a", "pre", "-", "set", "factor", ".", "The", "aggregate", "pixels", "in", "all", "subsampled", "channels", "are", "then", "vectorized", "into", "a", "pixel", "look", "-", "up", "table", ".", "Note", "that", "an", "optional", "smoothing", "procedure", "can", "be", "done", "on", "each", "channel", "with", "a", "binomial", "filter", "both", "before", "computation", "and", "after", "subsampling", ".", "Classifier", "learning", ":", "The", "learning", "process", "is", "quite", "simple", ".", "Two", "changes", "are", "made", "compared", "with", "VJ", "framework", ".", "First", "is", "that", "weak", "classifier", "is", "changed", "from", "decision", "stump", "to", "depth", "-", "2", "decision", "tree", ".", "The", "more", "complex", "weak", "classifier", "shows", "stronger", "ability", "in", "seeking", "the", "discriminant", "intra", "and", "inter", "channel", "correlations", "for", "classification", ".", "Second", "difference", "is", "that", "soft", "-", "cascade", "structure", "is", "used", ".", "Unlike", "the", "attentional", "cascade", "structure", "in", "VJ", "framework", "which", "has", "several", "cascade", "stages", ",", "a", "single", "-", "stage", "classifier", "is", "trained", "on", "the", "whole", "training", "data", "and", "a", "threshold", "is", "then", "set", "after", "each", "weak", "classifier", "picked", "by", "Adaboost", ".", "These", "two", "changes", "lead", "to", "more", "efficient", "training", "and", "detection", ".", "Overall", "superiority", ":", "Compared", "with", "traditional", "Haar", "-", "like", "features", "used", "in", "VJ", "framework", ",", "aggregate", "channel", "features", "have", "the", "following", "differences", "and", "advantages", ":", "1", ")", "The", "image", "channels", "are", "extended", "to", "more", "types", "in", "order", "to", "encode", "diverse", "information", "like", "color", ",", "gradients", ",", "local", "histograms", "and", "so", "on", ",", "therefore", "possess", "richer", "representation", "capacity", ".", "2", ")", "Features", "are", "extracted", "directly", "as", "pixel", "values", "on", "downsampled", "channels", "rather", "than", "computing", "rectangular", "sums", "with", "various", "locations", "and", "scales", "using", "integral", "images", ",", "leading", "to", "a", "faster", "feature", "computation", "and", "smaller", "feature", "pool", "size", "for", "boosting", "learning", ".", "With", "the", "help", "of", "cascade", "structure", ",", "detection", "speed", "is", "accelerated", "more", ".", "3", ")", "Due", "to", "its", "structure", "consistence", "with", "the", "overall", "image", ",", "when", "coupled", "with", "boosting", "method", ",", "the", "boosted", "classifier", "naturally", "encodes", "structured", "pattern", "information", "from", "large", "training", "data", "(", "see", "Figure", "[", "reference", "]", "for", "an", "illustration", ")", ",", "which", "gives", "more", "accurate", "localization", "of", "faces", "in", "the", "image", ".", "subsection", ":", "Investigation", "guidelines", "All", "investigations", "are", "trained", "on", "the", "AFLW", "face", "database", "and", "tested", "on", "the", "Annotated", "Faces", "in", "the", "Wild", "(", "AFW", ")", "testset", ".", "To", "make", "it", "clear", ",", "there", "are", "in", "total", "positive", "samples", "and", "negative", "samples", "selected", "from", "AFLW", "which", "are", "kept", "constant", "in", "all", "investigations", ".", "Testset", "contains", "natural", "images", "with", "faces", "that", "vary", "a", "lot", "in", "pose", ",", "appearance", "and", "illumination", ".", "To", "alleviate", "the", "ground", "-", "truth", "offset", "caused", "by", "different", "annotation", "styles", "(", "Figure", "[", "reference", "]", ")", "in", "training", "and", "testing", "set", "and", "make", "the", "evaluation", "more", "comparable", ",", "a", "lower", "Jaccard", "index", "with", "threshold", "is", "adopted", "in", "comparative", "evaluation", ".", "Practically", "the", "lower", "threshold", "wo", "n’t", "cause", "errors", "being", "mistakenly", "corrected", ".", "Note", "that", "in", "final", "evaluation", "of", "the", "proposed", "face", "detector", "(", "section", "5", ")", ",", "the", "AFW", "testset", ",", "together", "with", "another", "face", "benchmark", "FDDB", "database", ",", "are", "used", "as", "testbed", "and", "the", "evaluation", "metric", "follows", "the", "database", "protocol", ".", "subsection", ":", "Feature", "design", "To", "fully", "exploit", "the", "power", "of", "aggregate", "channel", "features", "in", "face", "detection", "domain", ",", "a", "deep", "investigation", "into", "the", "design", "of", "the", "feature", "is", "done", "mainly", "on", "channel", "types", ",", "window", "size", ",", "subsampling", "method", "and", "feature", "scale", ".", "Results", "of", "comparative", "experiments", "are", "shown", "in", "Figure", "[", "reference", "]", ".", "Channel", "types", ":", "Three", "types", "of", "channels", "are", "used", ",", "which", "are", "color", "channel", "(", "Gray", "-", "scale", ",", "RGB", ",", "HSV", "and", "LUV", ")", ",", "gradient", "magnitude", ",", "and", "gradient", "histograms", ".", "The", "computation", "of", "the", "latter", "two", "channel", "types", "could", "be", "seen", "as", "a", "generalized", "version", "of", "HoG", "features", ".", "Specifically", ",", "gradient", "magnitude", "is", "the", "biggest", "response", "on", "all", "three", "color", "channels", ",", "and", "oriented", "gradient", "histograms", "follow", "the", "idea", "of", "HoG", "in", "that", ":", "1", ")", "rectangular", "cell", "size", "in", "HoG", "equals", "the", "subsampling", "factor", "in", "aggregated", "channel", "features", ";", "2", ")", "each", "orientation", "bin", "results", "in", "one", "feature", "channel", "(", "6", "orientation", "bins", "are", "used", "in", "this", "paper", ")", ".", "Figure", "[", "reference", "]", "(", "a", ")", "~", "(", "c", ")", "show", "how", "much", "each", "of", "these", "three", "types", "alone", "contributes", "to", "the", "performance", "of", "face", "detection", ".", "It", "can", "be", "seen", "that", "the", "gradient", "histograms", "contribute", "most", "to", "the", "performance", "among", "all", "three", "channel", "types", ".", "Figure", "[", "reference", "]", "(", "d", ")", "shows", "the", "performances", "of", "combinations", "of", "these", "three", "types", "computed", "on", "different", "color", "channels", ".", "Detection", "window", "size", ":", "Detection", "window", "size", "is", "the", "scale", "to", "which", "we", "resize", "all", "face", "and", "non", "-", "face", "samples", "and", "then", "train", "our", "detector", ".", "Larger", "window", "size", "includes", "more", "pixels", "in", "feature", "pool", "and", "thus", "may", "improve", "the", "face", "detection", "performance", ".", "On", "the", "other", "hand", ",", "too", "large", "window", "will", "miss", "some", "small", "faces", "and", "diminish", "the", "detection", "efficiency", ".", "Figure", "[", "reference", "]", "(", "e", ")", "shows", "comparison", "of", "window", "size", "ranging", "from", "to", "with", "a", "stride", "of", "pixels", ".", "Subsampling", ":", "The", "factor", "for", "subsampling", "can", "be", "regarded", "as", "the", "perceptive", "scale", "for", "that", "it", "controls", "the", "scale", "at", "which", "the", "aggregation", "is", "done", ".", "Changing", "the", "factor", "from", "large", "to", "small", "leads", "to", "the", "feature", "representation", "shifting", "from", "coarse", "to", "fine", "and", "the", "feature", "pool", "size", "getting", "bigger", ".", "Experiments", "on", "different", "subsampling", "factors", "are", "shown", "in", "Figure", "[", "reference", "]", "(", "f", ")", ".", "In", "original", "aggregate", "channel", "features", ",", "the", "way", "to", "do", "subsampling", "is", "average", "pooling", ".", "Following", "the", "idea", "in", "Convolutional", "Neural", "Networks", ",", "another", "two", "ways", "of", "subsampling", ",", "max", "pooling", "and", "stochastic", "pooling", "are", "tested", "in", "Figure", "[", "reference", "]", "(", "g", ")", ".", "Smoothing", ":", "As", "described", "in", "feature", "description", ",", "both", "pre", "and", "post", "smoothing", "is", "done", "in", "default", "setting", "of", "aggregate", "channel", "features", ".", "A", "binomial", "filter", "with", "a", "radius", "of", "is", "used", "for", "smoothing", ".", "The", "smoothing", "procedure", "also", "has", "a", "great", "influence", "on", "the", "scale", "of", "the", "feature", "representation", ".", "Concretely", ",", "pre", "-", "smoothing", "determines", "how", "far", "the", "local", "neighborhood", "is", "in", "which", "local", "correlations", "are", "encoded", "before", "channel", "computation", ",", "while", "post", "-", "smoothing", "determines", "the", "neighborhood", "size", "in", "which", "the", "computed", "channel", "features", "are", "integrated", "with", "each", "other", ".", "In", ",", "the", "former", "corresponds", "to", "the", "‘", "local", "scale", "’", "of", "the", "feature", ",", "while", "the", "latter", "represents", "the", "‘", "integration", "scale", "’", ".", "We", "vary", "the", "filter", "radius", "used", "in", "pre", "and", "post", "smoothing", "and", "find", "that", "both", "using", "a", "radius", "of", "gets", "the", "best", "results", ".", "Figure", "[", "reference", "]", "(", "h", ")", "~", "(", "i", ")", "present", "the", "comparative", "results", ".", "Multi", "-", "scale", ":", "In", "aggregate", "channel", "features", ",", "although", "hidden", "information", "at", "different", "scale", "could", "be", "extracted", "at", "a", "cost", "of", "more", "weak", "classifiers", ",", "it", "would", "be", "better", "to", "make", "the", "integrated", "channel", "features", "multi", "-", "scaled", "and", "thus", "make", "themselves", "more", "discriminant", ".", "Therefore", "the", "same", "or", "better", "classification", "performance", "can", "be", "achieved", "with", "fewer", "weak", "classifiers", ".", "In", "this", "part", ",", "we", "implement", "three", "multi", "-", "scale", "version", "of", "aggregate", "channel", "features", "in", "the", "aforementioned", "three", "kinds", "of", "scale", ",", "perceptive", "scale", "(", "subsampling", ")", ",", "local", "scale", "(", "pre", "-", "smoothing", ")", "and", "integration", "scale", "(", "post", "-", "smoothing", ")", "and", "compare", "their", "performaces", ".", "See", "results", "in", "Figure", "[", "reference", "]", "(", "j", ")", "~", "(", "l", ")", ".", "Summary", ":", "The", "color", "channel", ",", "gradient", "magnitude", "and", "gradient", "histograms", "prove", "themselves", "a", "good", "match", "in", "aggregate", "channel", "features", ".", "However", ",", "different", "choices", "of", "color", "channel", "used", "and", "on", "which", "gradients", "are", "computed", "have", "a", "great", "impact", "on", "performance", ".", "According", "to", "the", "experiments", ",", "LUV", "channel", "and", "gradient", "magnitude", "and", "6", "-", "bin", "histograms", "computed", "on", "RGB", "color", "space", "(", "in", "total", "10", "channels", ")", "are", "the", "best", "choice", "for", "face", "detection", ".", "Larger", "detection", "window", "size", "generally", "gets", "better", "performance", ",", "but", "will", "miss", "many", "small", "faces", "in", "testing", "and", "lead", "to", "inefficient", "detection", ".", "In", "this", "work", ",", "we", "set", "the", "size", "to", "as", "its", "optimal", "performance", ".", "A", "subsampling", "factor", "of", "is", "most", "reasonable", "according", "to", "the", "experiments", ",", "while", "different", "pooling", "methods", "show", "small", "differences", ".", "However", ",", "max", "pooling", "and", "stochastic", "pooling", "are", "much", "slower", "than", "average", "pooling", ",", "therefore", "the", "average", "pooling", "becomes", "the", "best", "match", "for", "the", "sake", "of", "efficiency", ".", "In", "this", "way", ",", "the", "resulting", "feature", "pool", "size", "of", "our", "face", "detector", "is", ",", "considerably", "smaller", "than", "that", "in", "VJ", "framework", ".", "As", "for", "multi", "-", "scale", "version", "of", "aggregate", "channel", "features", ",", "multi", "-", "local", "-", "scale", "with", "an", "additional", "scale", "of", "radius", "shows", "the", "best", "performance", ".", "The", "probable", "reason", "is", "that", "pre", "-", "smoothing", "controls", "the", "local", "scale", "of", "the", "neighborhood", "feature", "correlations", "and", "therefore", "matches", "the", "intuition", "inside", "multi", "-", "scale", "best", ".", "Compared", "with", "other", "fine", "-", "tuning", ",", "the", "multi", "-", "scale", "version", "has", "a", "notable", "performance", "gain", "for", "that", "it", "makes", "up", "for", "the", "scale", "uniformity", "caused", "by", "subsampling", "to", "some", "extent", ".", "One", "main", "drawback", "is", "that", "it", "doubles", "the", "feature", "pool", "size", "and", "as", "a", "result", "slows", "down", "the", "detection", "speed", "somewhat", ".", "Based", "on", "the", "trade", "-", "off", ",", "we", "implement", "two", "face", "detectors", "with", "different", "scale", "settings", ",", "one", "is", "single", "-", "scaled", "with", "faster", "speed", "and", "the", "other", "is", "multi", "-", "scaled", "with", "better", "accuracy", ".", "We", "evaluate", "and", "discuss", "the", "performances", "of", "these", "two", "versions", "in", "detail", "in", "section", "5", ".", "subsection", ":", "Training", "design", "Besides", "careful", "design", "of", "the", "aggregate", "channel", "features", ",", "experiments", "on", "the", "training", "process", "which", "is", "similar", "to", "that", "in", "VJ", "framework", "are", "also", "carried", "out", ".", "The", "differences", "are", "that", "the", "weak", "classifier", "is", "changed", "into", "depth", "-", "2", "decision", "tree", "and", "soft", "-", "cascade", "structure", "is", "used", ".", "Details", "of", "the", "training", "design", "are", "as", "follows", ".", "Number", "of", "weak", "classifiers", ":", "Given", "a", "feature", "pool", "size", "of", ",", "we", "vary", "the", "number", "of", "weak", "classifiers", "contained", "in", "the", "soft", "-", "cascade", ".", "In", "Figure", "[", "reference", "]", "performances", "of", "various", "numbers", "of", "weak", "classifiers", "ranging", "from", "to", "are", "displayed", ",", "which", "shows", "that", "apparently", "more", "classifiers", "generate", "better", "performance", ",", "and", "when", "the", "number", "gets", "larger", "the", "performance", "begins", "to", "saturate", ".", "Since", "more", "classifiers", "slow", "down", "the", "detection", "speed", ",", "there", "’s", "a", "trade", "-", "off", "between", "accuracy", "and", "speed", ".", "Searching", "for", "the", "saturate", "point", "as", "the", "optimal", "is", "significant", "during", "training", "in", "such", "framework", ".", "Training", "data", ":", "Empirically", ",", "more", "training", "data", "will", "get", "better", "performance", "given", "powerful", "representation", "capacity", ".", "In", "this", "case", ",", "AFLW", "database", "is", "used", "as", "the", "only", "positive", "training", "data", ".", "However", ",", "as", "images", "in", "AFLW", "database", "are", "very", "salient", "and", "the", "background", "has", "very", "less", "variance", ",", "negative", "samples", "cropped", "from", "the", "AFLW", "database", "ca", "n’t", "represent", "the", "real", "world", "scenario", "well", ",", "which", "limits", "the", "face", "detection", "performance", "in", "the", "wild", ".", "In", "this", "part", ",", "we", "further", "use", "PASCAL", "VOC", "database", "and", "randomly", "crop", "windows", "from", "images", "without", "person", "as", "the", "new", "negative", "samples", ".", "Experiments", "show", "that", "the", "new", "training", "data", "containing", "cluttered", "background", "significantly", "improve", "the", "performance", "with", ".", "Summary", ":", "Based", "on", "observations", "above", ",", "we", "choose", "as", "the", "number", "of", "weak", "classifiers", "contained", "in", "the", "soft", "cascade", ".", "As", "each", "weak", "classifier", "is", "a", "depth", "-", "2", "decision", "tree", ",", "it", "takes", "only", "two", "comparing", "operations", "to", "apply", "a", "weak", "classifier", ",", "which", "is", "quite", "fast", ".", "During", "training", ",", "as", "negative", "data", "is", "large", ",", "we", "adopt", "a", "standard", "Bootstrap", "procedure", "to", "sample", "hard", "negative", "samples", "from", "PASCAL", "VOC", "in", "the", "implementation", "of", "the", "proposed", "face", "detector", ".", "section", ":", "Multi", "-", "view", "detection", "Human", "faces", "in", "real", "world", "usually", "have", "highly", "varied", "poses", ".", "In", "AFLW", "database", ",", "the", "human", "pose", "is", "divided", "into", "three", "aspects", ":", "1", "in", "-", "plane", "rotation", "‘", "roll", "’", "and", "2", "out", "-", "of", "-", "plane", "rotations", "‘", "yaw", "’", "and", "‘", "pitch", "’", ".", "Because", "of", "this", "large", "variance", "in", "face", "pose", ",", "it", "is", "difficult", "to", "train", "a", "single", "view", "face", "detector", "to", "handle", "all", "the", "poses", "effectively", ".", "A", "multi", "-", "view", "detection", "is", "further", "examined", "in", "this", "part", ".", "Due", "to", "the", "adoption", "of", "soft", "-", "cascade", "structure", ",", "a", "multi", "-", "view", "version", "of", "face", "detector", "wo", "n’t", "cause", "too", "much", "computation", "burden", ".", "Typically", ",", "we", "divide", "the", "out", "-", "of", "-", "plane", "rotation", "¡", "°", "yaw¡±", "into", "different", "views", "and", "let", "the", "classifier", "itself", "tolerate", "the", "pose", "variance", "in", "the", "other", "two", "types", "of", "rotations", ".", "Adopting", "multi", "-", "view", "detection", "also", "brings", "about", "many", "troublesome", "issues", ".", "If", "handled", "improperly", ",", "the", "performance", "will", "differ", "greatly", ".", "First", ",", "detectors", "of", "different", "view", "will", "each", "produce", "a", "set", "of", "candidate", "positive", "windows", "followed", "with", "a", "set", "of", "confidence", "scores", ".", "For", "application", "purpose", ",", "we", "need", "to", "merge", "these", "detections", "from", "different", "views", "and", "also", "remove", "duplicated", "windows", ".", "A", "typical", "approach", "is", "Non", "-", "Maximum", "Suppression", "(", "NMS", ")", ".", "An", "issue", "rises", "on", "how", "to", "compare", "confidence", "scores", "from", "different", "classifiers", "and", "how", "to", "do", "window", "merging", "in", "the", "trade", "-", "off", "between", "high", "precision", "rate", "and", "high", "detection", "rate", ".", "Second", ",", "as", "for", "detection", "evaluation", ",", "usually", "the", "overlap", "of", "bounding", "boxes", "is", "used", "as", "the", "criterion", ".", "However", ",", "annotations", "in", "different", "data", "sets", "may", "not", "have", "a", "consistent", "style", "(", "Figure", "[", "reference", "]", "(", "a", ")", ")", ".", "This", "diversity", "suffers", "more", "in", "profile", "faces", ".", "Since", "our", "face", "detector", "is", "trained", "and", "tested", "on", "different", "data", "sets", ",", "this", "issue", "impacts", "the", "performance", "a", "lot", ".", "Third", ",", "detectors", "of", "different", "views", "need", "to", "be", "trained", "with", "different", "samples", "separately", ".", "How", "to", "divide", "the", "views", "therefore", "becomes", "another", "concerning", "problem", ".", "In", "this", "section", ",", "we", "address", "the", "above", "three", "issues", "successfully", "by", "careful", "designs", "and", "therefore", "fully", "exploit", "the", "advantage", "of", "multi", "-", "view", "detection", ".", "subsection", ":", "View", "partition", "In", "the", "scenario", "of", "detecting", "faces", "in", "the", "wild", ",", "pose", "variation", "caused", "by", "yaw", "is", "usually", "severer", "than", "pitch", "and", "roll", ".", "Therefore", "we", "divide", "the", "faces", "in", "AFLW", "database", "according", "to", "yaw", "angle", ".", "We", "have", "subviews", "which", "are", "horizontally", "symmetric", "(", "see", "Figure", ".", "[", "reference", "]", "(", "b", ")", ")", "because", "we", "flip", "each", "image", "in", "the", "training", "set", ".", "Specifically", ",", "there", "are", ",", ",", ",", ",", ",", "images", "in", "views", "from", "to", ".", "Benefitting", "from", "the", "symmetry", "of", "our", "model", ",", "we", "can", "only", "train", "three", "subview", "detectors", "of", "the", "right", "side", "for", "simplicity", ",", "and", "use", "these", "trained", "right", "-", "side", "detectors", "to", "generate", "the", "left", "-", "side", "detectors", ".", "Detections", "of", "all", "six", "detectors", "are", "then", "merged", "to", "get", "the", "final", "detections", ".", "Though", "multi", "-", "view", "detection", "significantly", "improves", "the", "detection", "performance", "(", "especially", "the", "recall", "rate", ")", ",", "the", "post", "-", "processing", "of", "detections", "from", "different", "detectors", "becomes", "a", "trouble", ".", "If", "handled", "improperly", ",", "the", "performance", "degrades", "a", "lot", ".", "subsection", ":", "Post", "-", "processing", "Difficulties", "in", "the", "post", "-", "processing", "of", "multi", "-", "view", "detection", "mainly", "reflect", "on", "the", "following", "aspects", ":", "1", ")", "different", "score", "distributions", "and", ";", "2", ")", "different", "bounding", "box", "styles", ".", "Concretely", ",", "as", "each", "subview", "detector", "is", "trained", "separately", ",", "their", "output", "confidence", "scores", "usually", "have", "different", "distributions", ".", "What", "’s", "more", ",", "due", "to", "the", "annotation", "rule", "in", "the", "AFLW", "database", "that", "the", "face", "’s", "nose", "is", "approximately", "at", "the", "center", "location", "of", "the", "bounding", "box", "ground", "-", "truth", ",", "as", "the", "subview", "changes", ",", "the", "bounding", "box", "shifts", ".", "This", "bounding", "box", "offset", "causes", "difficulty", "both", "in", "detection", "merging", "and", "final", "evaluation", "using", "Jaccard", "index", "metric", ".", "To", "solve", "these", "annoying", "issues", "and", "make", "the", "best", "use", "of", "multi", "-", "view", "detection", ",", "we", "introduce", "the", "following", "methods", "for", "post", "-", "processing", ".", "Score", "re", "-", "ranking", ":", "We", "propose", "the", "following", "three", "kinds", "of", "score", "re", "-", "ranking", ":", "1", ")", "normalizing", "scores", "of", "different", "views", "to", "[", "0", ",", "1", "]", ";", "2", ")", "defining", "a", "new", "score", "that", "has", "uniform", "distribution", "and", ";", "3", ")", "taking", "overlapping", "detections", "into", "consideration", ".", ":", "After", "training", "a", "classifier", ",", "calculate", "the", "output", "range", "of", "the", "classifier", "and", "use", "the", "range", "to", "do", "normalization", "later", "so", "that", "output", "score", "has", "a", "range", "of", "[", "0", ",", "1", "]", ".", ":", "Originally", ",", "each", "weak", "classifier", "in", "the", "soft", "-", "cascade", "owns", "a", "score", "and", "final", "score", "is", "the", "sum", "of", "all", "scores", ".", "Instead", ",", "we", "use", "the", "number", "of", "weak", "classifier", "that", "the", "image", "patch", "passed", "positively", "as", "the", "new", "score", ".", "Therefore", "the", "upper", "limit", "of", "the", "new", "score", "is", "in", "our", "case", ".", ":", "Given", "an", "image", ",", "multiple", "detections", "from", "multi", "-", "view", "detectors", "exist", "each", "with", "a", "score", ".", "For", "each", "detection", ",", "we", "first", "calculate", "the", "number", "of", "overlapped", "detection", "it", "has", "(", "overlap", "threshold", "is", ")", "and", "then", "multiply", "score", "of", "each", "detection", "with", "a", "factor", "of", "its", "overlapping", "number", "ranking", ".", ":", "Instead", "of", "using", "overlapping", "as", "a", "multiply", "factor", ",", "here", "we", "use", "the", "sum", "of", "overlapped", "detections", "’", "scores", "as", "the", "current", "detection", "’s", "new", "score", ".", "Detection", "merging", ":", "Apart", "from", "the", "version", "of", "Non", "Maximum", "Suppression", ",", "we", "also", "use", "the", "detection", "combination", "introduced", "in", ".", "It", "averages", "the", "locations", "of", "overlapped", "detections", "rather", "than", "suppresses", "them", ".", "Detection", "adjustment", ":", "As", "shown", "in", "Figure", "[", "reference", "]", "(", "a", ")", ",", "different", "databases", "have", "different", "annotation", "styles", "of", "ground", "-", "truth", ".", "Specifically", ",", "AFLW", "has", "square", "annotations", "with", "nose", "located", "approximately", "at", "the", "center", ".", "AFW", "uses", "tight", "rectangular", "bounding", "boxes", "as", "annotations", "with", "the", "eye", "-", "brow", "being", "the", "approximate", "upper", "bound", ".", "FDDB", "uses", "elliptical", "annotations", "bounding", "the", "whole", "head", ".", "As", "our", "detector", "is", "trained", "on", "AFLW", "and", "tested", "on", "AFW", "and", "FDDB", ",", "there", "exist", "offsets", "in", "both", "detection", "position", "and", "scale", ".", "According", "to", "observations", ",", "the", "offsets", "vary", "as", "face", "pose", "changes", ".", "Therefore", "we", "adopt", "a", "view", "-", "specific", "detection", "adjustment", "to", "alleviate", "the", "offsets", ".", "Note", "that", "the", "adjustment", "is", "constant", "for", "all", "images", "and", "faces", "in", "the", "same", "database", ",", "see", "Figure", "[", "reference", "]", "(", "b", ")", "for", "details", ".", "Summary", ":", "According", "to", "experimental", "results", "(", "Table", "[", "reference", "]", ")", ",", "seems", "to", "be", "the", "best", "score", "re", "-", "ranking", "method", ".", "The", "underlying", "reason", "may", "be", "that", "true", "positives", "usually", "have", "many", "overlapped", "detections", ",", "while", "the", "false", "positives", "would", "only", "get", "a", "few", "responses", ".", "Therefore", "leveraging", "this", "overlapping", "information", "in", "score", "re", "-", "ranking", "can", "reduce", "many", "false", "positives", ".", "However", ",", "in", "practice", ",", "overlap", "related", "methods", "and", "detection", "combination", "both", "cost", "much", "time", "to", "process", ",", "which", "is", "infeasible", "in", "a", "large", "majority", "of", "applications", ".", "We", "finally", "adopt", "score", "re", "-", "ranking", "combined", "with", "Non", "Maximum", "Suppression", "for", "the", "sake", "of", "detection", "speed", ".", "section", ":", "Experiments", "In", "this", "section", ",", "we", "compare", "our", "method", "with", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "AFW", "and", "FDDB", "databases", "which", "contain", "challenging", "faces", "in", "the", "wild", ".", "In", "AFW", ",", "we", "compare", "with", "three", "commercial", "systems", "(", "Google", "Picasa", ",", "Face.com", "and", "Face", "++", ")", "and", "five", "academic", "methods", "(", "Shen", ",", "Zhu", ",", "DPM", ",", "multiHOG", "and", "Kalal", ")", ".", "In", "FDDB", ",", "we", "compare", "with", "one", "commercial", "system", "(", "Olaworks", ")", "and", "six", "academic", "methods", "(", "Yan", ",", "Boosted", "Exemplar", ",", "SURF", "multiview", ",", "PEP", "-", "Adapt", ",", "XZJY", "and", "Zhu", ")", "listed", "on", "FDDB", "results", "page", ".", "subsection", ":", "Evaluation", "on", "benchmark", "face", "database", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "in", "AFW", ",", "our", "multi", "-", "scale", "detector", "achieves", "an", "ap", "value", "of", ",", "outperforming", "other", "academic", "methods", "by", "a", "large", "margin", ".", "When", "it", "comes", "to", "commercial", "systems", ",", "ours", "is", "better", "than", "Face.com", "and", "almost", "equal", "to", "Face", "++", "and", "Google", "Picasa", ".", "Note", "that", "most", "of", "our", "false", "positives", "on", "AFW", "database", "are", "faces", "that", "have", "n’t", "been", "annotated", "(", "small", ",", "seriously", "occluded", "or", "artificial", "faces", "like", "mask", "and", "cartoon", "character", ")", ".", "When", "evaluated", "on", "FDDB", "database", ",", "we", "follow", "the", "evaluation", "protocol", "in", "and", "report", "the", "average", "discrete", "and", "continuous", "ROC", "of", "the", "ten", "subfolders", ".", "For", "equality", ",", "we", "fix", "the", "number", "of", "false", "positives", "to", "(", "equivalent", "to", "an", "average", "of", "False", "Positive", "Per", "Image", ")", "and", "compare", "the", "true", "positive", "rate", ".", "In", "discrete", "score", "where", "evaluation", "metric", "is", "the", "same", "as", "in", "AFW", ",", "our", "detector", "achieves", ",", "which", "is", "a", "little", "better", "than", "Yan", ".", "Note", "that", "the", "ground", "-", "truth", "in", "FDDB", "are", "elliptical", "faces", ",", "therefore", "the", "evaluation", "metric", "of", "an", "overlap", "ratio", "bigger", "than", "can", "not", "reveal", "the", "true", "performance", "of", "the", "proposed", "detector", "well", ".", "When", "using", "continuous", "score", "which", "takes", "the", "overlap", "ratio", "as", "the", "score", ",", "our", "method", "gets", "true", "positive", "rate", "at", "FPPI", "for", "multi", "-", "scale", "version", ",", "surpassing", "other", "methods", "which", "output", "rectangular", "detections", "by", "a", "notable", "margin", "(", "the", "Yan", "detector", "outputs", "the", "same", "elliptical", "detections", "as", "the", "ground", "-", "truth", ",", "therefore", "having", "advantages", "with", "this", "metric", ")", ".", "Our", "detector", "using", "single", "-", "scale", "features", "performs", "a", "little", "worse", "with", "the", "benefit", "of", "faster", "detection", "speed", ".", "subsection", ":", "Discussion", "Training", "efficiency", ":", "We", "implement", "the", "method", "with", "Piotr", "’s", "MATLAB", "toolbox", "on", "a", "PC", "with", "Intel", "Core", "i7", "-", "3770", "CPU", "and", "16", "GB", "RAM", ".", "With", "positive", "images", "and", "negative", "images", "in", "total", "6", "views", ",", "the", "training", "process", "takes", "about", "mins", "for", "a", "single", "-", "scale", "subview", "detector", "containing", "weak", "classifiers", "and", "mins", "for", "multi", "-", "scale", "version", ".", "Note", "that", "we", "use", "much", "fewer", "training", "data", "than", "SURF", "multiview", "whilst", "still", "outperforming", "their", "performance", ".", "Comparative", "results", ":", "When", "inspecting", "detections", "of", "the", "proposed", "face", "detector", "and", "other", "algorithms", "on", "the", "testsets", ",", "some", "patterns", "can", "be", "found", "to", "explain", "why", "our", "detector", "outperforms", "others", ".", "One", "evident", "strength", "lies", "in", "detecting", "faces", "with", "extreme", "poses", ".", "Because", "we", "adopt", "multi", "-", "view", "detection", "and", "train", "each", "subview", "detector", "separately", ",", "our", "detector", "handles", "pose", "variations", "very", "well", ".", "Second", "is", "the", "outstanding", "illumination", "invariance", "of", "our", "detector", ",", "which", "is", "mainly", "owing", "to", "the", "extension", "of", "channel", "types", "to", "LUV", "color", "space", "and", "gradient", "-", "related", "channels", ".", "Detection", "speed", ":", "Due", "to", "the", "simple", "form", "of", "aggregate", "channel", "features", "and", "fast", "computation", "of", "feature", "pyramid", ",", "detection", "is", "quite", "efficient", ".", "For", "full", "yaw", "pose", "face", "detection", "in", "VGA", "image", ",", "the", "proposed", "detector", "using", "single", "-", "scale", "features", "runs", "at", "FPS", "on", "a", "single", "thread", "and", "FPS", "if", "threads", "are", "used", ".", "If", "only", "frontal", "faces", "are", "concerned", ",", "the", "detector", "runs", "at", "FPS", "and", "FPS", "after", "parallelization", ".", "When", "it", "comes", "to", "the", "proposed", "detector", "using", "multi", "-", "scale", "features", ",", "the", "above", "four", "indices", "reduce", "to", ",", ",", "and", "FPS", ".", "Considering", "the", "large", "performance", "gain", "and", "similar", "speed", ",", "the", "proposed", "method", "can", "replace", "Viola", "-", "Jones", "detector", "for", "face", "detection", "in", "the", "wild", ".", "section", ":", "Conclusion", "A", "novel", "feature", "representation", "called", "aggregate", "channel", "features", "possesses", "the", "merits", "of", "fast", "feature", "extraction", "and", "powerful", "representation", "capacity", ".", "In", "this", "paper", ",", "we", "successfully", "apply", "the", "feature", "representation", "to", "face", "detection", "domain", "through", "a", "deep", "investigation", "into", "the", "feature", "design", ",", "and", "propose", "a", "multi", "-", "scale", "version", "of", "feature", "which", "further", "enriches", "the", "representation", "capacity", ".", "Combined", "with", "our", "efforts", "into", "solving", "issues", "concerning", "multi", "-", "view", "detection", ",", "the", "proposed", "multi", "-", "view", "face", "detector", "shows", "state", "-", "of", "-", "the", "-", "art", "performance", "in", "both", "effectiveness", "and", "efficiency", "on", "faces", "in", "the", "wild", ".", "The", "proposed", "method", "appeals", "to", "real", "world", "application", "demands", "and", "has", "the", "potential", "to", "be", "embedded", "into", "low", "power", "devices", ".", "section", ":", "Acknowledgement", "This", "work", "was", "supported", "by", "the", "Chinese", "National", "Natural", "Science", "Foundation", "Projects", "#", "61105023", ",", "#", "61103156", ",", "#", "61105037", ",", "#", "61203267", ",", "#", "61375037", ",", "National", "Science", "and", "Technology", "Support", "Program", "Project", "#", "2013BAK02B01", ",", "Chinese", "Academy", "of", "Sciences", "Project", "No", ".", "KGZD", "-", "EW", "-", "102", "-", "2", ",", "and", "AuthenMetric", "R", "&", "D", "Funds", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["Aggregate", "Channel", "Features"], "Method"], [6, 11, ["Multi", "-", "view", "Face", "Detection"], "Task"], [39, 41, ["learning", "algorithms"], "Method"], [43, 45, ["feature", "representation"], "Method"], [118, 121, ["aggregate", "channel", "features"], "Method"], [127, 129, ["feature", "design"], "Method"], [133, 139, ["multi", "-", "scale", "version", "of", "features"], "Method"], [162, 166, ["score", "re", "-", "ranking"], "Method"], [172, 174, ["learning", "pipelines"], "Method"], [175, 179, ["Viola", "-", "Jones", "framework"], "Method"], [181, 186, ["multi", "-", "view", "face", "detector"], "Method"], [187, 190, ["aggregate", "channel", "features"], "Method"], [203, 204, ["AFW"], "Material"], [212, 213, ["FPS"], "Metric"], [214, 216, ["VGA", "images"], "Material"], [233, 235, ["computer", "vision"], "Task"], [236, 240, ["human", "-", "computer", "interaction"], "Task"], [284, 286, ["Adaboost", "algorithm"], "Method"], [290, 293, ["attentional", "cascade", "structure"], "Method"], [295, 297, ["VJ", "detector"], "Method"], [313, 315, ["VJ", "detector"], "Method"], [345, 347, ["VJ", "framework"], "Method"], [369, 371, ["learning", "algorithms"], "Method"], [376, 377, ["boosting"], "Method"], [378, 379, ["cascade"], "Method"], [395, 397, ["feature", "representation"], "Method"], [415, 417, ["computation", "cost"], "Metric"], [423, 425, ["pedestrian", "detection"], "Task"], [429, 431, ["channel", "features"], "Method"], [436, 438, ["Channel", "features"], "Method"], [462, 465, ["classifier", "learning", "process"], "Method"], [467, 470, ["VJ", "framework", "pipeline"], "Method"], [483, 486, ["aggregate", "channel", "features"], "Method"], [498, 500, ["Channel", "extension"], "Method"], [527, 529, ["VJ", "framework"], "Method"], [549, 551, ["feature", "representation"], "Method"], [553, 555, ["learning", "algorithms"], "Method"], [565, 568, ["aggregate", "channel", "features"], "Method"], [611, 613, ["subsampling", "method"], "Method"], [625, 627, ["feature", "design"], "Method"], [646, 649, ["multi", "-", "scaling"], "Method"], [650, 652, ["feature", "representation"], "Method"], [707, 711, ["multi", "-", "view", "detection"], "Task"], [738, 742, ["multi", "-", "view", "detection"], "Task"], [762, 764, ["subview", "detectors"], "Method"], [791, 794, ["post", "-", "processing"], "Method"], [795, 799, ["score", "re", "-", "ranking"], "Task"], [800, 802, ["detection", "merging"], "Task"], [803, 806, ["bounding", "box", "adjustment"], "Task"], [821, 825, ["multi", "-", "view", "detection"], "Task"], [844, 845, ["AFW"], "Material"], [846, 847, ["FDDB"], "Material"], [850, 855, ["multi", "-", "view", "face", "detector"], "Method"], [872, 873, ["speed"], "Metric"], [901, 903, ["face", "detector"], "Method"], [904, 907, ["aggregate", "channel", "features"], "Method"], [922, 923, ["AFW"], "Material"], [924, 925, ["FDDB"], "Material"], [954, 956, ["computer", "vision"], "Task"], [988, 992, ["VJ", "face", "detector", "features"], "Method"], [997, 999, ["feature", "computation"], "Task"], [1000, 1003, ["integral", "image", "representation"], "Method"], [1004, 1006, ["classifier", "learning"], "Method"], [1007, 1008, ["Adaboost"], "Method"], [1011, 1014, ["attentional", "cascade", "structure"], "Method"], [1020, 1022, ["VJ", "framework"], "Method"], [1092, 1094, ["feature", "selection"], "Task"], [1118, 1121, ["feature", "representation", "part"], "Method"], [1137, 1139, ["learning", "algorithm"], "Method"], [1140, 1142, ["classifier", "structure"], "Method"], [1144, 1146, ["VJ", "framework"], "Method"], [1148, 1150, ["channel", "features"], "Method"], [1158, 1160, ["pedestrian", "detection"], "Task"], [1164, 1166, ["channel", "extension"], "Method"], [1179, 1181, ["representation", "capacity"], "Metric"], [1182, 1183, ["classification"], "Task"], [1220, 1223, ["integral", "channel", "features"], "Method"], [1238, 1240, ["channel", "features"], "Method"], [1241, 1244, ["aggregate", "channel", "features"], "Method"], [1269, 1271, ["representation", "capacity"], "Method"], [1273, 1276, ["feature", "pool", "size"], "Metric"], [1290, 1294, ["multi", "-", "view", "detection"], "Task"], [1299, 1304, ["aggregate", "channel", "features", "based", "detector"], "Method"], [1320, 1322, ["face", "detector"], "Method"], [1363, 1367, ["Haar", "-", "like", "features"], "Method"], [1383, 1385, ["feature", "design"], "Method"], [1386, 1388, ["training", "design"], "Task"], [1424, 1426, ["face", "detector"], "Method"], [1451, 1453, ["face", "detector"], "Method"], [1488, 1490, ["Feature", "description"], "Method"], [1490, 1492, ["Channel", "extension"], "Method"], [1507, 1508, ["channel"], "Method"], [1573, 1575, ["registered", "map"], "Method"], [1597, 1603, ["linear", "or", "non", "-", "linear", "transformation"], "Method"], [1624, 1626, ["Feature", "computation"], "Task"], [1635, 1640, ["computation", "of", "aggregate", "channel", "features"], "Task"], [1693, 1695, ["smoothing", "procedure"], "Method"], [1703, 1705, ["binomial", "filter"], "Method"], [1710, 1711, ["subsampling"], "Method"], [1712, 1714, ["Classifier", "learning"], "Method"], [1716, 1718, ["learning", "process"], "Method"], [1728, 1730, ["VJ", "framework"], "Method"], [1734, 1736, ["weak", "classifier"], "Method"], [1751, 1753, ["weak", "classifier"], "Method"], [1766, 1767, ["classification"], "Task"], [1772, 1776, ["soft", "-", "cascade", "structure"], "Method"], [1781, 1784, ["attentional", "cascade", "structure"], "Method"], [1785, 1787, ["VJ", "framework"], "Method"], [1790, 1792, ["cascade", "stages"], "Method"], [1794, 1798, ["single", "-", "stage", "classifier"], "Method"], [1813, 1815, ["weak", "classifier"], "Method"], [1817, 1818, ["Adaboost"], "Method"], [1826, 1827, ["training"], "Task"], [1828, 1829, ["detection"], "Task"], [1836, 1840, ["Haar", "-", "like", "features"], "Method"], [1842, 1844, ["VJ", "framework"], "Method"], [1845, 1848, ["aggregate", "channel", "features"], "Method"], [1918, 1920, ["feature", "computation"], "Task"], [1922, 1925, ["feature", "pool", "size"], "Metric"], [1926, 1928, ["boosting", "learning"], "Method"], [1933, 1935, ["cascade", "structure"], "Method"], [1957, 1959, ["boosting", "method"], "Method"], [1961, 1963, ["boosted", "classifier"], "Method"], [1987, 1990, ["localization", "of", "faces"], "Task"], [2011, 2016, ["Annotated", "Faces", "in", "the", "Wild"], "Material"], [2017, 2018, ["AFW"], "Material"], [2037, 2038, ["AFLW"], "Material"], [2095, 2097, ["Jaccard", "index"], "Metric"], [2102, 2104, ["comparative", "evaluation"], "Metric"], [2125, 2127, ["face", "detector"], "Method"], [2150, 2152, ["evaluation", "metric"], "Metric"], [2159, 2161, ["Feature", "design"], "Method"], [2194, 2196, ["subsampling", "method"], "Method"], [2242, 2244, ["gradient", "histograms"], "Method"], [2261, 2263, ["HoG", "features"], "Method"], [2286, 2287, ["HoG"], "Method"], [2296, 2297, ["HoG"], "Method"], [2396, 2399, ["Detection", "window", "size"], "Metric"], [2400, 2403, ["Detection", "window", "size"], "Metric"], [2481, 2482, ["Subsampling"], "Method"], [2486, 2487, ["subsampling"], "Task"], [2503, 2504, ["aggregation"], "Task"], [2517, 2519, ["feature", "representation"], "Method"], [2558, 2559, ["subsampling"], "Method"], [2560, 2562, ["average", "pooling"], "Method"], [2567, 2570, ["Convolutional", "Neural", "Networks"], "Method"], [2575, 2576, ["subsampling"], "Method"], [2577, 2579, ["max", "pooling"], "Method"], [2580, 2582, ["stochastic", "pooling"], "Method"], [2593, 2594, ["Smoothing"], "Method"], [2598, 2600, ["feature", "description"], "Method"], [2602, 2606, ["pre", "and", "post", "smoothing"], "Method"], [2617, 2619, ["binomial", "filter"], "Method"], [2626, 2627, ["smoothing"], "Task"], [2629, 2631, ["smoothing", "procedure"], "Method"], [2641, 2643, ["feature", "representation"], "Method"], [2646, 2649, ["pre", "-", "smoothing"], "Method"], [2663, 2665, ["channel", "computation"], "Method"], [2667, 2670, ["post", "-", "smoothing"], "Method"], [2720, 2722, ["post", "smoothing"], "Task"], [2774, 2776, ["weak", "classifiers"], "Method"], [2802, 2804, ["classification", "performance"], "Metric"], [2809, 2811, ["weak", "classifiers"], "Method"], [2819, 2827, ["multi", "-", "scale", "version", "of", "aggregate", "channel", "features"], "Method"], [2838, 2839, ["subsampling"], "Method"], [2844, 2847, ["pre", "-", "smoothing"], "Method"], [2852, 2855, ["post", "-", "smoothing"], "Method"], [2973, 2974, ["detection"], "Task"], [3003, 3005, ["pooling", "methods"], "Method"], [3011, 3013, ["max", "pooling"], "Method"], [3014, 3016, ["stochastic", "pooling"], "Method"], [3020, 3022, ["average", "pooling"], "Method"], [3025, 3027, ["average", "pooling"], "Method"], [3043, 3046, ["feature", "pool", "size"], "Metric"], [3048, 3050, ["face", "detector"], "Method"], [3057, 3059, ["VJ", "framework"], "Method"], [3062, 3070, ["multi", "-", "scale", "version", "of", "aggregate", "channel", "features"], "Method"], [3071, 3076, ["multi", "-", "local", "-", "scale"], "Method"], [3094, 3095, ["smoothing"], "Method"], [3118, 3121, ["fine", "-", "tuning"], "Method"], [3123, 3127, ["multi", "-", "scale", "version"], "Method"], [3180, 3182, ["face", "detectors"], "Method"], [3204, 3205, ["accuracy"], "Metric"], [3238, 3240, ["training", "process"], "Task"], [3246, 3248, ["VJ", "framework"], "Method"], [3258, 3260, ["weak", "classifier"], "Method"], [3279, 3281, ["training", "design"], "Method"], [3287, 3289, ["weak", "classifiers"], "Method"], [3302, 3304, ["weak", "classifiers"], "Method"], [3307, 3310, ["soft", "-", "cascade"], "Method"], [3321, 3323, ["weak", "classifiers"], "Method"], [3353, 3354, ["classifiers"], "Method"], [3367, 3368, ["accuracy"], "Metric"], [3369, 3370, ["speed"], "Metric"], [3470, 3473, ["PASCAL", "VOC", "database"], "Material"], [3516, 3518, ["weak", "classifiers"], "Method"], [3521, 3523, ["soft", "cascade"], "Method"], [3526, 3528, ["weak", "classifier"], "Method"], [3540, 3542, ["comparing", "operations"], "Method"], [3545, 3547, ["weak", "classifier"], "Method"], [3554, 3555, ["training"], "Task"], [3566, 3568, ["Bootstrap", "procedure"], "Method"], [3574, 3576, ["PASCAL", "VOC"], "Material"], [3582, 3584, ["face", "detector"], "Method"], [3654, 3658, ["single", "view", "face", "detector"], "Method"], [3666, 3670, ["multi", "-", "view", "detection"], "Task"], [3682, 3686, ["soft", "-", "cascade", "structure"], "Method"], [3688, 3695, ["multi", "-", "view", "version", "of", "face", "detector"], "Method"], [3700, 3702, ["computation", "burden"], "Metric"], [3723, 3724, ["classifier"], "Method"], [3738, 3742, ["multi", "-", "view", "detection"], "Task"], [3805, 3809, ["Non", "-", "Maximum", "Suppression"], "Method"], [3810, 3811, ["NMS"], "Method"], [3824, 3825, ["classifiers"], "Method"], [3829, 3831, ["window", "merging"], "Task"], [3837, 3840, ["high", "precision", "rate"], "Metric"], [3849, 3851, ["detection", "evaluation"], "Task"], [3897, 3899, ["face", "detector"], "Method"], [3963, 3967, ["multi", "-", "view", "detection"], "Task"], [3970, 3972, ["View", "partition"], "Method"], [3976, 3981, ["detecting", "faces", "in", "the", "wild"], "Task"], [3982, 3984, ["pose", "variation"], "Task"], [4064, 4066, ["subview", "detectors"], "Method"], [4077, 4081, ["right", "-", "side", "detectors"], "Method"], [4084, 4088, ["left", "-", "side", "detectors"], "Method"], [4104, 4108, ["multi", "-", "view", "detection"], "Task"], [4111, 4112, ["detection"], "Task"], [4116, 4118, ["recall", "rate"], "Metric"], [4121, 4126, ["post", "-", "processing", "of", "detections"], "Task"], [4145, 4148, ["Post", "-", "processing"], "Task"], [4184, 4186, ["subview", "detector"], "Method"], [4192, 4194, ["confidence", "scores"], "Metric"], [4249, 4251, ["detection", "merging"], "Task"], [4253, 4254, ["evaluation"], "Task"], [4255, 4258, ["Jaccard", "index", "metric"], "Metric"], [4270, 4274, ["multi", "-", "view", "detection"], "Task"], [4281, 4284, ["post", "-", "processing"], "Task"], [4288, 4289, ["ranking"], "Task"], [4297, 4301, ["score", "re", "-", "ranking"], "Task"], [4340, 4341, ["classifier"], "Method"], [4375, 4377, ["weak", "classifier"], "Method"], [4402, 4404, ["weak", "classifier"], "Method"], [4436, 4440, ["multi", "-", "view", "detectors"], "Method"], [4509, 4511, ["Detection", "merging"], "Task"], [4517, 4520, ["Non", "Maximum", "Suppression"], "Method"], [4542, 4544, ["Detection", "adjustment"], "Task"], [4569, 4570, ["AFLW"], "Material"], [4581, 4582, ["AFW"], "Material"], [4600, 4601, ["FDDB"], "Material"], [4615, 4616, ["AFLW"], "Material"], [4619, 4620, ["AFW"], "Material"], [4621, 4622, ["FDDB"], "Material"], [4705, 4709, ["re", "-", "ranking", "method"], "Method"], [4741, 4745, ["score", "re", "-", "ranking"], "Task"], [4756, 4759, ["overlap", "related", "methods"], "Method"], [4782, 4786, ["score", "re", "-", "ranking"], "Method"], [4788, 4791, ["Non", "Maximum", "Suppression"], "Method"], [4819, 4820, ["AFW"], "Material"], [4821, 4823, ["FDDB", "databases"], "Material"], [4832, 4833, ["AFW"], "Material"], [4841, 4843, ["Google", "Picasa"], "Method"], [4844, 4845, ["Face.com"], "Method"], [4846, 4848, ["Face", "++"], "Method"], [4858, 4859, ["DPM"], "Method"], [4860, 4861, ["multiHOG"], "Method"], [4862, 4863, ["Kalal"], "Method"], [4866, 4867, ["FDDB"], "Material"], [4875, 4876, ["Olaworks"], "Method"], [4884, 4886, ["Boosted", "Exemplar"], "Method"], [4887, 4889, ["SURF", "multiview"], "Method"], [4890, 4893, ["PEP", "-", "Adapt"], "Method"], [4920, 4921, ["AFW"], "Material"], [4923, 4927, ["multi", "-", "scale", "detector"], "Method"], [4935, 4937, ["academic", "methods"], "Method"], [4953, 4954, ["Face.com"], "Method"], [4958, 4960, ["Face", "++"], "Method"], [4961, 4963, ["Google", "Picasa"], "Method"], [4969, 4971, ["false", "positives"], "Metric"], [5012, 5016, ["discrete", "and", "continuous", "ROC"], "Metric"], [5038, 5040, ["False", "Positive"], "Metric"], [5046, 5049, ["true", "positive", "rate"], "Metric"], [5051, 5053, ["discrete", "score"], "Metric"], [5054, 5056, ["evaluation", "metric"], "Metric"], [5061, 5062, ["AFW"], "Material"], [5089, 5091, ["evaluation", "metric"], "Metric"], [5093, 5095, ["overlap", "ratio"], "Metric"], [5125, 5128, ["true", "positive", "rate"], "Metric"], [5129, 5130, ["FPPI"], "Method"], [5131, 5135, ["multi", "-", "scale", "version"], "Task"], [5149, 5151, ["Yan", "detector"], "Method"], [5173, 5177, ["single", "-", "scale", "features"], "Method"], [5192, 5194, ["Training", "efficiency"], "Metric"], [5200, 5204, ["Piotr", "’s", "MATLAB", "toolbox"], "Method"], [5240, 5243, ["scale", "subview", "detector"], "Method"], [5244, 5246, ["weak", "classifiers"], "Method"], [5247, 5248, ["mins"], "Method"], [5249, 5253, ["multi", "-", "scale", "version"], "Method"], [5263, 5265, ["SURF", "multiview"], "Method"], [5280, 5282, ["face", "detector"], "Method"], [5316, 5320, ["multi", "-", "view", "detection"], "Task"], [5323, 5325, ["subview", "detector"], "Method"], [5365, 5367, ["Detection", "speed"], "Metric"], [5381, 5383, ["feature", "pyramid"], "Method"], [5384, 5385, ["detection"], "Task"], [5396, 5398, ["VGA", "image"], "Task"], [5403, 5407, ["single", "-", "scale", "features"], "Method"], [5444, 5445, ["detector"], "Method"], [5460, 5461, ["FPS"], "Metric"], [5476, 5480, ["Viola", "-", "Jones", "detector"], "Method"], [5492, 5494, ["feature", "representation"], "Method"], [5495, 5498, ["aggregate", "channel", "features"], "Method"], [5503, 5505, ["feature", "extraction"], "Method"], [5507, 5509, ["representation", "capacity"], "Method"], [5518, 5520, ["feature", "representation"], "Method"], [5530, 5532, ["feature", "design"], "Method"], [5536, 5542, ["multi", "-", "scale", "version", "of", "feature"], "Method"], [5557, 5561, ["multi", "-", "view", "detection"], "Task"], [5564, 5569, ["multi", "-", "view", "face", "detector"], "Method"], [5594, 5598, ["real", "world", "application", "demands"], "Task"], [12, 13, ["detection"], "Task"], [48, 49, ["detection"], "Task"], [84, 85, ["detection"], "Task"], [159, 160, ["detection"], "Task"], [167, 168, ["detection"], "Task"], [205, 206, ["FDDB"], "Material"], [222, 223, ["detection"], "Task"], [254, 255, ["detection"], "Task"], [302, 303, ["detection"], "Task"], [388, 389, ["detection"], "Task"], [539, 540, ["detection"], "Task"], [570, 571, ["detection"], "Task"], [682, 683, ["detection"], "Task"], [833, 834, ["detection"], "Task"], [869, 870, ["detection"], "Task"], [870, 871, ["accuracy"], "Metric"], [892, 893, ["detection"], "Task"], [917, 918, ["detection"], "Task"], [944, 945, ["detection"], "Task"], [978, 979, ["detection"], "Task"], [1048, 1049, ["detection"], "Task"], [1225, 1226, ["detection"], "Task"], [1341, 1342, ["detection"], "Task"], [1473, 1474, ["detection"], "Task"], [1613, 1614, ["detection"], "Task"], [1936, 1937, ["detection"], "Task"], [2004, 2005, ["AFLW"], "Material"], [2133, 2134, ["AFW"], "Material"], [2141, 2142, ["FDDB"], "Material"], [2172, 2173, ["detection"], "Task"], [2353, 2354, ["detection"], "Task"], [2438, 2439, ["detection"], "Task"], [2457, 2458, ["detection"], "Task"], [2950, 2951, ["detection"], "Task"], [2953, 2954, ["detection"], "Task"], [3166, 3167, ["detection"], "Task"], [3357, 3358, ["detection"], "Task"], [3408, 3409, ["AFLW"], "Material"], [3424, 3425, ["AFLW"], "Material"], [3442, 3443, ["AFLW"], "Material"], [3457, 3458, ["detection"], "Task"], [3590, 3591, ["detection"], "Task"], [3603, 3604, ["AFLW"], "Material"], [3842, 3843, ["detection"], "Task"], [4001, 4002, ["AFLW"], "Material"], [4158, 4159, ["detection"], "Task"], [4210, 4211, ["AFLW"], "Material"], [4448, 4449, ["detection"], "Task"], [4457, 4458, ["detection"], "Task"], [4471, 4472, ["detection"], "Task"], [4504, 4505, ["detection"], "Task"], [4525, 4526, ["detection"], "Task"], [4628, 4629, ["detection"], "Task"], [4652, 4653, ["detection"], "Task"], [4760, 4761, ["detection"], "Task"], [4795, 4796, ["detection"], "Task"], [4900, 4901, ["FDDB"], "Material"], [4972, 4973, ["AFW"], "Material"], [4999, 5000, ["FDDB"], "Material"], [5082, 5083, ["FDDB"], "Material"], [5116, 5118, ["overlap", "ratio"], "Metric"], [5186, 5187, ["detection"], "Task"], [5394, 5395, ["detection"], "Task"], [5482, 5483, ["detection"], "Task"], [5522, 5523, ["detection"], "Task"]]} {"doc_id": "2f56b1ac5b9faac9527b6814778925e9242cf5fd", "text": ["document", ":", "Training", "Region", "-", "based", "Object", "Detectors", "with", "Online", "Hard", "Example", "Mining", "The", "field", "of", "object", "detection", "has", "made", "significant", "advances", "riding", "on", "the", "wave", "of", "region", "-", "based", "ConvNets", ",", "but", "their", "training", "procedure", "still", "includes", "many", "heuristics", "and", "hyperparameters", "that", "are", "costly", "to", "tune", ".", "We", "present", "a", "simple", "yet", "surprisingly", "effective", "online", "hard", "example", "mining", "(", "OHEM", ")", "algorithm", "for", "training", "region", "-", "based", "ConvNet", "detectors", ".", "Our", "motivation", "is", "the", "same", "as", "it", "has", "always", "been", "–", "detection", "datasets", "contain", "an", "overwhelming", "number", "of", "easy", "examples", "and", "a", "small", "number", "of", "hard", "examples", ".", "Automatic", "selection", "of", "these", "hard", "examples", "can", "make", "training", "more", "effective", "and", "efficient", ".", "OHEM", "is", "a", "simple", "and", "intuitive", "algorithm", "that", "eliminates", "several", "heuristics", "and", "hyperparameters", "in", "common", "use", ".", "But", "more", "importantly", ",", "it", "yields", "consistent", "and", "significant", "boosts", "in", "detection", "performance", "on", "benchmarks", "like", "PASCAL", "VOC", "2007", "and", "2012", ".", "Its", "effectiveness", "increases", "as", "datasets", "become", "larger", "and", "more", "difficult", ",", "as", "demonstrated", "by", "the", "results", "on", "the", "MS", "COCO", "dataset", ".", "Moreover", ",", "combined", "with", "complementary", "advances", "in", "the", "field", ",", "OHEM", "leads", "to", "state", "-", "of", "-", "the", "-", "art", "results", "of", "78.9", "%", "and", "76.3", "%", "mAP", "on", "PASCAL", "VOC", "2007", "and", "2012", "respectively", ".", "section", ":", "Introduction", "Image", "classification", "and", "object", "detection", "are", "two", "fundamental", "computer", "vision", "tasks", ".", "Object", "detectors", "are", "often", "trained", "through", "a", "reduction", "that", "converts", "object", "detection", "into", "an", "image", "classification", "problem", ".", "This", "reduction", "introduces", "a", "new", "challenge", "that", "is", "not", "found", "in", "natural", "image", "classification", "tasks", ":", "the", "training", "set", "is", "distinguished", "by", "a", "large", "imbalance", "between", "the", "number", "of", "annotated", "objects", "and", "the", "number", "of", "background", "examples", "(", "image", "regions", "not", "belonging", "to", "any", "object", "class", "of", "interest", ")", ".", "In", "the", "case", "of", "sliding", "-", "window", "object", "detectors", ",", "such", "as", "the", "deformable", "parts", "model", "(", "DPM", ")", ",", "this", "imbalance", "may", "be", "as", "extreme", "as", "100", ",", "000", "background", "examples", "to", "every", "one", "object", ".", "The", "recent", "trend", "towards", "object", "-", "proposal", "-", "based", "detectors", "mitigates", "this", "issue", "to", "an", "extent", ",", "but", "the", "imbalance", "ratio", "may", "still", "be", "high", "(", ",", "70:1", ")", ".", "This", "challenge", "opens", "space", "for", "learning", "techniques", "that", "cope", "with", "imbalance", "and", "yield", "faster", "training", ",", "higher", "accuracy", ",", "or", "both", ".", "Unsurprisingly", ",", "this", "is", "not", "a", "new", "challenge", "and", "a", "standard", "solution", ",", "originally", "called", "bootstrapping", "(", "and", "now", "often", "called", "hard", "negative", "mining", ")", ",", "has", "existed", "for", "at", "least", "20", "years", ".", "Bootstrapping", "was", "introduced", "in", "the", "work", "of", "Sung", "and", "Poggio", "in", "the", "mid", "-", "1990", "’s", "(", "if", "not", "earlier", ")", "for", "training", "face", "detection", "models", ".", "Their", "key", "idea", "was", "to", "gradually", "grow", ",", "or", "bootstrap", ",", "the", "set", "of", "background", "examples", "by", "selecting", "those", "examples", "for", "which", "the", "detector", "triggers", "a", "false", "alarm", ".", "This", "strategy", "leads", "to", "an", "iterative", "training", "algorithm", "that", "alternates", "between", "updating", "the", "detection", "model", "given", "the", "current", "set", "of", "examples", ",", "and", "then", "using", "the", "updated", "model", "to", "find", "new", "false", "positives", "to", "add", "to", "the", "bootstrapped", "training", "set", ".", "The", "process", "typically", "commences", "with", "a", "training", "set", "consisting", "of", "all", "object", "examples", "and", "a", "small", ",", "random", "set", "of", "background", "examples", ".", "Bootstrapping", "has", "seen", "widespread", "use", "in", "the", "intervening", "decades", "of", "object", "detection", "research", ".", "Dalal", "and", "Triggs", "used", "it", "when", "training", "SVMs", "for", "pedestrian", "detection", ".", "Felzenszwalb", "later", "proved", "that", "a", "form", "of", "bootstrapping", "for", "SVMs", "converges", "to", "the", "global", "optimal", "solution", "defined", "on", "the", "entire", "dataset", ".", "Their", "algorithm", "is", "often", "referred", "to", "as", "hard", "negative", "mining", "and", "is", "frequently", "used", "when", "training", "SVMs", "for", "object", "detection", ".", "Bootstrapping", "was", "also", "successfully", "applied", "to", "a", "variety", "of", "other", "learning", "models", ",", "including", "shallow", "neural", "networks", "and", "boosted", "decision", "trees", ".", "Even", "modern", "detection", "methods", "based", "on", "deep", "convolutional", "neural", "networks", "(", "ConvNets", ")", ",", "such", "as", "R", "-", "CNN", "and", "SPPnet", ",", "still", "employ", "SVMs", "trained", "with", "hard", "negative", "mining", ".", "It", "may", "seem", "odd", "then", "that", "the", "current", "state", "-", "of", "-", "the", "-", "art", "object", "detectors", ",", "embodied", "by", "Fast", "R", "-", "CNN", "and", "its", "descendants", ",", "do", "not", "use", "bootstrapping", ".", "The", "underlying", "reason", "is", "a", "technical", "difficulty", "brought", "on", "by", "the", "shift", "towards", "purely", "online", "learning", "algorithms", ",", "particularly", "in", "the", "context", "of", "deep", "ConvNets", "trained", "with", "stochastic", "gradient", "descent", "(", "SGD", ")", "on", "millions", "of", "examples", ".", "Bootstrapping", ",", "and", "its", "variants", "in", "the", "literature", ",", "rely", "on", "the", "aforementioned", "alternation", "template", ":", "(", "a", ")", "for", "some", "period", "of", "time", "a", "fixed", "model", "is", "used", "to", "find", "new", "examples", "to", "add", "to", "the", "active", "training", "set", ";", "(", "b", ")", "then", ",", "for", "some", "period", "of", "time", "the", "model", "is", "trained", "on", "the", "fixed", "active", "training", "set", ".", "Training", "deep", "ConvNet", "detectors", "with", "SGD", "typically", "requires", "hundreds", "of", "thousands", "of", "SGD", "steps", "and", "freezing", "the", "model", "for", "even", "a", "few", "iterations", "at", "a", "time", "would", "dramatically", "slow", "progress", ".", "What", "is", "needed", ",", "instead", ",", "is", "a", "purely", "online", "form", "of", "hard", "example", "selection", ".", "In", "this", "paper", ",", "we", "propose", "a", "novel", "bootstrapping", "technique", "called", "online", "hard", "example", "mining", "(", "OHEM", ")", "for", "training", "state", "-", "of", "-", "the", "-", "art", "detection", "models", "based", "on", "deep", "ConvNets", ".", "The", "algorithm", "is", "a", "simple", "modification", "to", "SGD", "in", "which", "training", "examples", "are", "sampled", "according", "to", "a", "non", "-", "uniform", ",", "non", "-", "stationary", "distribution", "that", "depends", "on", "the", "current", "loss", "of", "each", "example", "under", "consideration", ".", "The", "method", "takes", "advantage", "of", "detection", "-", "specific", "problem", "structure", "in", "which", "each", "SGD", "mini", "-", "batch", "consists", "of", "only", "one", "or", "two", "images", ",", "but", "thousands", "of", "candidate", "examples", ".", "The", "candidate", "examples", "are", "subsampled", "according", "to", "a", "distribution", "that", "favors", "diverse", ",", "high", "loss", "instances", ".", "Gradient", "computation", "(", "backpropagation", ")", "is", "still", "efficient", "because", "it", "only", "uses", "a", "small", "subset", "of", "all", "candidates", ".", "We", "apply", "OHEM", "to", "the", "standard", "Fast", "R", "-", "CNN", "detection", "method", "and", "show", "three", "benefits", "compared", "to", "the", "baseline", "training", "algorithm", ":", "It", "removes", "the", "need", "for", "several", "heuristics", "and", "hyperparameters", "commonly", "used", "in", "region", "-", "based", "ConvNets", ".", "It", "yields", "a", "consistent", "and", "significant", "boosts", "in", "mean", "average", "precision", ".", "Its", "effectiveness", "increases", "as", "the", "training", "set", "becomes", "larger", "and", "more", "difficult", ",", "as", "demonstrated", "by", "results", "on", "the", "MS", "COCO", "dataset", ".", "Moreover", ",", "the", "gains", "from", "OHEM", "are", "complementary", "to", "recent", "improvements", "in", "object", "detection", ",", "such", "as", "multi", "-", "scale", "testing", "and", "iterative", "bounding", "-", "box", "regression", ".", "Combined", "with", "these", "tricks", ",", "OHEM", "gives", "state", "-", "of", "-", "the", "-", "art", "results", "of", "78.9", "%", "and", "76.3", "%", "mAP", "on", "PASCAL", "VOC", "2007", "and", "2012", ",", "respectively", ".", "section", ":", "Related", "work", "Object", "detection", "is", "one", "of", "the", "oldest", "and", "most", "fundamental", "problems", "in", "computer", "vision", ".", "The", "idea", "of", "dataset", "bootstrapping", ",", "typically", "called", "hard", "negative", "mining", "in", "recent", "work", ",", "appears", "in", "the", "training", "of", "most", "successful", "object", "detectors", ".", "Many", "of", "these", "approaches", "use", "SVMs", "as", "the", "detection", "scoring", "function", ",", "even", "after", "training", "a", "deep", "convolutional", "neural", "network", "(", "ConvNet", ")", "for", "feature", "extraction", ".", "One", "notable", "exception", "is", "the", "Fast", "R", "-", "CNN", "detector", "and", "its", "descendants", ",", "such", "as", "Faster", "R", "-", "CNN", ".", "Since", "these", "models", "do", "not", "use", "SVMs", ",", "and", "are", "trained", "purely", "online", "with", "SGD", ",", "existing", "hard", "example", "mining", "techniques", "can", "not", "be", "immediately", "applied", ".", "This", "work", "addresses", "that", "problem", "by", "introducing", "an", "online", "hard", "example", "mining", "algorithm", "that", "improves", "optimization", "and", "detection", "accuracy", ".", "We", "briefly", "review", "hard", "example", "mining", ",", "modern", "ConvNet", "-", "based", "object", "detection", ",", "and", "relationships", "to", "concurrent", "works", "using", "hard", "example", "selection", "for", "training", "deep", "networks", ".", "paragraph", ":", "Hard", "example", "mining", ".", "There", "are", "two", "hard", "example", "mining", "algorithms", "in", "common", "use", ".", "The", "first", "is", "used", "when", "optimizing", "SVMs", ".", "In", "this", "case", ",", "the", "training", "algorithm", "maintains", "a", "working", "set", "of", "examples", "and", "alternates", "between", "training", "an", "SVM", "to", "convergence", "on", "the", "working", "set", ",", "and", "updating", "the", "working", "set", "by", "removing", "some", "examples", "and", "adding", "others", "according", "to", "a", "specific", "rule", ".", "The", "rule", "removes", "examples", "that", "are", "“", "easy", "”", "in", "the", "sense", "that", "they", "are", "correctly", "classified", "beyond", "the", "current", "model", "’s", "margin", ".", "Conversely", ",", "the", "rule", "adds", "new", "examples", "that", "are", "hard", "in", "the", "sense", "that", "they", "violate", "the", "current", "model", "’s", "margin", ".", "Applying", "this", "rule", "leads", "to", "the", "global", "SVM", "solution", ".", "Importantly", ",", "the", "working", "set", "is", "usually", "a", "small", "subset", "of", "the", "entire", "training", "set", ".", "The", "second", "method", "is", "used", "for", "non", "-", "SVMs", "and", "has", "been", "applied", "to", "a", "variety", "of", "models", "including", "shallow", "neural", "networks", "and", "boosted", "decision", "trees", ".", "This", "algorithm", "usually", "starts", "with", "a", "dataset", "of", "positive", "examples", "and", "a", "random", "set", "of", "negative", "examples", ".", "The", "machine", "learning", "model", "is", "then", "trained", "to", "convergence", "on", "that", "dataset", "and", "subsequently", "applied", "to", "a", "larger", "dataset", "to", "harvest", "false", "positives", ".", "The", "false", "positives", "are", "then", "added", "to", "the", "training", "set", "and", "then", "the", "model", "is", "trained", "again", ".", "This", "process", "is", "usually", "iterated", "only", "once", "and", "does", "not", "have", "any", "convergence", "proofs", ".", "paragraph", ":", "ConvNet", "-", "based", "object", "detection", ".", "In", "the", "last", "three", "years", "significant", "gains", "have", "been", "made", "in", "object", "detection", ".", "These", "improvements", "were", "made", "possible", "by", "the", "successful", "application", "of", "deep", "ConvNets", "to", "ImageNet", "classification", ".", "The", "R", "-", "CNN", "and", "OverFeat", "detectors", "lead", "this", "wave", "with", "impressive", "results", "on", "PASCAL", "VOC", "and", "ImageNet", "detection", ".", "OverFeat", "is", "based", "on", "the", "sliding", "-", "window", "detection", "method", ",", "which", "is", "perhaps", "the", "most", "intuitive", "and", "oldest", "search", "method", "for", "detection", ".", "R", "-", "CNN", ",", "in", "contrast", ",", "uses", "region", "proposals", ",", "a", "method", "that", "was", "made", "popular", "by", "the", "selective", "search", "algorithm", ".", "Since", "R", "-", "CNN", ",", "there", "has", "been", "rapid", "progress", "in", "region", "-", "based", "ConvNets", ",", "including", "SPPnet", ",", "MR", "-", "CNN", ",", "and", "Fast", "R", "-", "CNN", ",", "which", "our", "work", "builds", "on", ".", "paragraph", ":", "Hard", "example", "selection", "in", "deep", "learning", ".", "There", "is", "recent", "work", "concurrent", "to", "our", "own", "that", "selects", "hard", "examples", "for", "training", "deep", "networks", ".", "Similar", "to", "our", "approach", ",", "all", "these", "methods", "base", "their", "selection", "on", "the", "current", "loss", "for", "each", "datapoint", ".", "independently", "selects", "hard", "positive", "and", "negative", "example", "from", "a", "larger", "set", "of", "random", "examples", "based", "on", "their", "loss", "to", "learn", "image", "descriptors", ".", "Given", "a", "positive", "pair", "of", "patches", ",", "finds", "hard", "negative", "patches", "from", "a", "large", "set", "using", "triplet", "loss", ".", "Akin", "to", "our", "approach", ",", "investigates", "online", "selection", "of", "hard", "examples", "for", "mini", "-", "batch", "SGD", "methods", ".", "Their", "selection", "is", "also", "based", "on", "loss", ",", "but", "the", "focus", "is", "on", "ConvNets", "for", "image", "classification", ".", "Complementary", "to", ",", "we", "focus", "on", "online", "hard", "example", "selection", "strategy", "for", "region", "-", "based", "object", "detectors", ".", "section", ":", "Overview", "of", "Fast", "R", "-", "CNN", "We", "first", "summarize", "the", "Fast", "R", "-", "CNN", "(", "FRCN", ")", "framework", ".", "FRCN", "takes", "as", "input", "an", "image", "and", "a", "set", "of", "object", "proposal", "regions", "of", "interest", "(", "RoIs", ")", ".", "The", "FRCN", "network", "itself", "can", "be", "divided", "into", "two", "sequential", "parts", ":", "a", "convolutional", "(", "conv", ")", "network", "with", "several", "convolution", "and", "max", "-", "pooling", "layers", "(", "Figure", "[", "reference", "]", ",", "“", "Convolutional", "Network", "”", ")", ";", "and", "an", "RoI", "network", "with", "an", "RoI", "-", "pooling", "layer", ",", "several", "fully", "-", "connected", "(", "fc", ")", "layers", "and", "two", "loss", "layers", "(", "Figure", "[", "reference", "]", ",", "“", "RoI", "Network", "”", ")", ".", "During", "inference", ",", "the", "conv", "network", "is", "applied", "to", "the", "given", "image", "to", "produce", "a", "conv", "feature", "map", ",", "size", "of", "which", "depends", "on", "the", "input", "image", "dimensions", ".", "Then", ",", "for", "each", "object", "proposal", ",", "the", "RoI", "-", "pooling", "layer", "projects", "the", "proposal", "onto", "the", "conv", "feature", "map", "and", "extracts", "a", "fixed", "-", "length", "feature", "vector", ".", "Each", "feature", "vector", "is", "fed", "into", "the", "fc", "layers", ",", "which", "finally", "give", "two", "outputs", ":", "(", "1", ")", "a", "softmax", "probability", "distribution", "over", "the", "object", "classes", "and", "background", ";", "and", "(", "2", ")", "regressed", "coordinates", "for", "bounding", "-", "box", "relocalization", ".", "There", "are", "several", "reasons", "for", "choosing", "FRCN", "as", "our", "base", "object", "detector", ",", "apart", "from", "it", "being", "a", "fast", "end", "-", "to", "-", "end", "system", ".", "Firstly", ",", "the", "basic", "two", "network", "setup", "(", "conv", "and", "RoI", ")", "is", "also", "used", "by", "other", "recent", "detectors", "like", "SPPnet", "and", "MR", "-", "CNN", ";", "therefore", ",", "our", "proposed", "algorithm", "is", "more", "broadly", "applicable", ".", "Secondly", ",", "though", "the", "basic", "setup", "is", "similar", ",", "FRCN", "also", "allows", "for", "training", "the", "entire", "conv", "network", ",", "as", "opposed", "to", "both", "SPPnet", "and", "MR", "-", "CNN", "which", "keep", "the", "conv", "network", "fixed", ".", "And", "finally", ",", "both", "SPPnet", "and", "MR", "-", "CNN", "require", "features", "from", "the", "RoI", "network", "to", "be", "cached", "for", "training", "a", "separate", "SVM", "classifier", "(", "using", "hard", "negative", "mining", ")", ".", "FRCN", "uses", "the", "RoI", "network", "itself", "to", "train", "the", "desired", "classifiers", ".", "In", "fact", ",", "shows", "that", "in", "the", "unified", "system", "using", "the", "SVM", "classifiers", "at", "later", "stages", "was", "unnecessary", ".", "subsection", ":", "Training", "Like", "most", "deep", "networks", ",", "FRCN", "is", "trained", "using", "stochastic", "gradient", "descent", "(", "SGD", ")", ".", "The", "loss", "per", "example", "RoI", "is", "the", "sum", "of", "a", "classification", "log", "loss", "that", "encourages", "predicting", "the", "correct", "object", "(", "or", "background", ")", "label", "and", "a", "localization", "loss", "that", "encourages", "predicting", "an", "accurate", "bounding", "box", "(", "see", "for", "details", ")", ".", "To", "share", "conv", "network", "computation", "between", "RoIs", ",", "SGD", "mini", "-", "batches", "are", "created", "hierarchically", ".", "For", "each", "mini", "-", "batch", ",", "images", "are", "first", "sampled", "from", "the", "dataset", ",", "and", "then", "RoIs", "are", "sampled", "from", "each", "image", ".", "Setting", "and", "works", "well", "in", "practice", ".", "The", "RoI", "sampling", "procedure", "uses", "several", "heuristics", ",", "which", "we", "describe", "briefly", "below", ".", "One", "contribution", "of", "this", "paper", "is", "to", "eliminate", "some", "of", "these", "heuristics", "and", "their", "hyperparameters", ".", "paragraph", ":", "Foreground", "RoIs", ".", "For", "an", "example", "RoI", "to", "be", "labeled", "as", "foreground", "(", "fg", ")", ",", "its", "intersection", "over", "union", "(", "IoU", ")", "overlap", "with", "a", "ground", "-", "truth", "bounding", "box", "should", "be", "at", "least", ".", "This", "is", "a", "fairly", "standard", "design", "choice", ",", "in", "part", "inspired", "by", "the", "evaluation", "protocol", "of", "the", "PASCAL", "VOC", "object", "detection", "benchmark", ".", "The", "same", "criterion", "is", "used", "in", "the", "SVM", "hard", "mining", "procedures", "of", "R", "-", "CNN", ",", "SPPnet", ",", "and", "MR", "-", "CNN", ".", "We", "use", "the", "same", "setting", ".", "paragraph", ":", "Background", "RoIs", ".", "A", "region", "is", "labeled", "background", "(", "bg", ")", "if", "its", "maximum", "IoU", "with", "ground", "truth", "is", "in", "the", "interval", "bg_lo", ",", "0.5", ")", ".", "A", "lower", "threshold", "of", "bg_lo", "is", "used", "by", "both", "FRCN", "and", "SPPnet", ",", "and", "is", "hypothesized", "in", "to", "crudely", "approximate", "hard", "negative", "mining", ";", "the", "assumption", "is", "that", "regions", "with", "some", "overlap", "with", "the", "ground", "truth", "are", "more", "likely", "to", "be", "the", "confusing", "or", "hard", "ones", ".", "We", "show", "in", "Section", "[", "reference", "]", "that", "although", "this", "heuristic", "helps", "convergence", "and", "detection", "accuracy", ",", "it", "is", "suboptimal", "because", "it", "ignores", "some", "infrequent", ",", "but", "important", ",", "difficult", "background", "regions", ".", "Our", "method", "removes", "the", "bg_lo", "threshold", ".", "paragraph", ":", "Balancing", "fg", "-", "bg", "RoIs", ":", "To", "handle", "the", "data", "imbalance", "described", "in", "Section", "[", "reference", "]", ",", "designed", "heuristics", "to", "rebalance", "the", "foreground", "-", "to", "-", "background", "ratio", "in", "each", "mini", "-", "batch", "to", "a", "target", "of", "by", "undersampling", "the", "background", "patches", "at", "random", ",", "thus", "ensuring", "that", "of", "a", "mini", "-", "batch", "is", "fg", "RoIs", ".", "We", "found", "that", "this", "is", "an", "important", "design", "decision", "for", "the", "training", "FRCN", ".", "Removing", "this", "ratio", "(", "randomly", "sampling", "RoIs", ")", ",", "or", "increasing", "it", ",", "decreases", "accuracy", "by", "points", "mAP", ".", "With", "our", "proposed", "method", ",", "we", "can", "remove", "this", "ratio", "hyperparameter", "with", "no", "ill", "effect", ".", "section", ":", "Our", "approach", "We", "propose", "a", "simple", "yet", "effective", "online", "hard", "example", "mining", "algorithm", "for", "training", "Fast", "R", "-", "CNN", "(", "or", "any", "Fast", "R", "-", "CNN", "style", "object", "detector", ")", ".", "We", "argue", "that", "the", "current", "way", "of", "creating", "mini", "-", "batches", "for", "SGD", "(", "Section", "[", "reference", "]", ")", "is", "inefficient", "and", "suboptimal", ",", "and", "we", "demonstrate", "that", "our", "approach", "leads", "to", "better", "training", "(", "lower", "training", "loss", ")", "and", "higher", "testing", "performance", "(", "mAP", ")", ".", "subsection", ":", "Online", "hard", "example", "mining", "Recall", "the", "alternating", "steps", "that", "define", "a", "hard", "example", "mining", "algorithm", ":", "(", "a", ")", "for", "some", "period", "of", "time", "a", "fixed", "model", "is", "used", "to", "find", "new", "examples", "to", "add", "to", "the", "active", "training", "set", ";", "(", "b", ")", "then", ",", "for", "some", "period", "of", "time", "the", "model", "is", "trained", "on", "the", "fixed", "active", "training", "set", ".", "In", "the", "context", "of", "SVM", "-", "based", "object", "detectors", ",", "such", "as", "the", "SVMs", "trained", "in", "R", "-", "CNN", "or", "SPPnet", ",", "step", "(", "a", ")", "inspects", "a", "variable", "number", "of", "images", "(", "often", "10", "’s", "or", "100", "’s", ")", "until", "the", "active", "training", "set", "reaches", "a", "threshold", "size", ",", "and", "then", "in", "step", "(", "b", ")", "the", "SVM", "is", "trained", "to", "convergence", "on", "the", "active", "training", "set", ".", "This", "process", "repeats", "until", "the", "active", "training", "set", "contains", "all", "support", "vectors", ".", "Applying", "an", "analogous", "strategy", "to", "FRCN", "ConvNet", "training", "slows", "learning", "because", "no", "model", "updates", "are", "made", "while", "selecting", "examples", "from", "the", "10", "’s", "or", "100", "’s", "of", "images", ".", "Our", "main", "observation", "is", "that", "these", "alternating", "steps", "can", "be", "combined", "with", "how", "FRCN", "is", "trained", "using", "online", "SGD", ".", "The", "key", "is", "that", "although", "each", "SGD", "iteration", "samples", "only", "a", "small", "number", "of", "images", ",", "each", "image", "contains", "thousands", "of", "example", "RoIs", "from", "which", "we", "can", "select", "the", "hard", "examples", "rather", "than", "a", "heuristically", "sampled", "subset", ".", "This", "strategy", "fits", "the", "alternation", "template", "to", "SGD", "by", "“", "freezing", "”", "the", "model", "for", "only", "one", "mini", "-", "batch", ".", "Thus", "the", "model", "is", "updated", "exactly", "as", "frequently", "as", "with", "the", "baseline", "SGD", "approach", "and", "therefore", "learning", "is", "not", "delayed", ".", "More", "specifically", ",", "the", "online", "hard", "example", "mining", "algorithm", "(", "OHEM", ")", "proceeds", "as", "follows", ".", "For", "an", "input", "image", "at", "SGD", "iteration", ",", "we", "first", "compute", "a", "conv", "feature", "map", "using", "the", "conv", "network", ".", "Then", "the", "RoI", "network", "uses", "this", "feature", "map", "and", "the", "all", "the", "input", "RoIs", ",", "instead", "of", "a", "sampled", "mini", "-", "batch", ",", "to", "do", "a", "forward", "pass", ".", "Recall", "that", "this", "step", "only", "involves", "RoI", "pooling", ",", "a", "few", "fc", "layers", ",", "and", "loss", "computation", "for", "each", "RoI.", "The", "loss", "represents", "how", "well", "the", "current", "network", "performs", "on", "each", "RoI.", "Hard", "examples", "are", "selected", "by", "sorting", "the", "input", "RoIs", "by", "loss", "and", "taking", "the", "examples", "for", "which", "the", "current", "network", "performs", "worst", ".", "Most", "of", "the", "forward", "computation", "is", "shared", "between", "RoIs", "via", "the", "conv", "feature", "map", ",", "so", "the", "extra", "computation", "needed", "to", "forward", "all", "RoIs", "is", "relatively", "small", ".", "Moreover", ",", "because", "only", "a", "small", "number", "of", "RoIs", "are", "selected", "for", "updating", "the", "model", ",", "the", "backward", "pass", "is", "no", "more", "expensive", "than", "before", ".", "However", ",", "there", "is", "a", "small", "caveat", ":", "co", "-", "located", "RoIs", "with", "high", "overlap", "are", "likely", "to", "have", "correlated", "losses", ".", "Moreover", ",", "these", "overlapping", "RoIs", "can", "project", "onto", "the", "same", "region", "in", "the", "conv", "feature", "map", ",", "because", "of", "resolution", "disparity", ",", "thus", "leading", "to", "loss", "double", "counting", ".", "To", "deal", "with", "these", "redundant", "and", "correlated", "regions", ",", "we", "use", "standard", "non", "-", "maximum", "suppression", "(", "NMS", ")", "to", "perform", "deduplication", "(", "the", "implementation", "from", ")", ".", "Given", "a", "list", "of", "RoIs", "and", "their", "losses", ",", "NMS", "works", "by", "iteratively", "selecting", "the", "RoI", "with", "the", "highest", "loss", ",", "and", "then", "removing", "all", "lower", "loss", "RoIs", "that", "have", "high", "overlap", "with", "the", "selected", "region", ".", "We", "use", "a", "relaxed", "IoU", "threshold", "of", "to", "suppress", "only", "highly", "overlapping", "RoIs", ".", "We", "note", "that", "the", "procedure", "described", "above", "does", "not", "need", "a", "fg", "-", "bg", "ratio", "for", "data", "balancing", ".", "If", "any", "class", "were", "neglected", ",", "its", "loss", "would", "increase", "until", "it", "has", "a", "high", "probability", "of", "being", "sampled", ".", "There", "can", "be", "images", "where", "the", "fg", "RoIs", "are", "easy", "(", "canonical", "view", "of", "a", "car", ")", ",", "so", "the", "network", "is", "free", "to", "use", "only", "bg", "regions", "in", "a", "mini", "-", "batch", ";", "and", "vice", "-", "versa", "when", "bg", "is", "trivial", "(", "sky", ",", "grass", ")", ",", "the", "mini", "-", "batch", "can", "be", "entirely", "fg", "regions", ".", "subsection", ":", "Implementation", "details", "There", "are", "many", "ways", "to", "implement", "OHEM", "in", "the", "FRCN", "detector", ",", "each", "with", "different", "trade", "-", "offs", ".", "An", "obvious", "way", "is", "to", "modify", "the", "loss", "layers", "to", "do", "the", "hard", "example", "selection", ".", "The", "loss", "layer", "can", "compute", "loss", "for", "all", "RoIs", ",", "sort", "them", "based", "on", "this", "loss", "to", "select", "hard", "RoIs", ",", "and", "finally", "set", "the", "loss", "of", "all", "non", "-", "hard", "RoIs", "to", ".", "Though", "straightforward", ",", "this", "implementation", "is", "inefficient", "as", "the", "RoI", "network", "still", "allocates", "memory", "and", "performs", "backward", "pass", "for", "all", "RoIs", ",", "even", "though", "most", "RoIs", "have", "loss", "and", "hence", "no", "gradient", "updates", "(", "a", "limitation", "of", "current", "deep", "learning", "toolboxes", ")", ".", "To", "overcome", "this", ",", "we", "propose", "the", "architecture", "presented", "in", "Figure", "[", "reference", "]", ".", "Our", "implementation", "maintains", "two", "copies", "of", "the", "RoI", "network", ",", "one", "of", "which", "is", "readonly", ".", "This", "implies", "that", "the", "readonly", "RoI", "network", "(", "Figure", "[", "reference", "]", "(", "a", ")", ")", "allocates", "memory", "only", "for", "forward", "pass", "of", "all", "RoIs", "as", "opposed", "to", "the", "standard", "RoI", "network", ",", "which", "allocates", "memory", "for", "both", "forward", "and", "backward", "passes", ".", "For", "an", "SGD", "iteration", ",", "given", "the", "conv", "feature", "map", ",", "the", "readonly", "RoI", "network", "performs", "a", "forward", "pass", "and", "computes", "loss", "for", "all", "input", "RoIs", "(", "Figure", "[", "reference", "]", ",", "green", "arrows", ")", ".", "Then", "the", "hard", "RoI", "sampling", "module", "uses", "the", "procedure", "described", "in", "Section", "[", "reference", "]", "to", "select", "hard", "examples", ",", "which", "are", "input", "to", "the", "regular", "RoI", "network", "(", "Figure", "[", "reference", "]", "(", "b", ")", ",", "red", "arrows", ")", ")", ".", "This", "network", "computes", "forward", "and", "backward", "passes", "only", "for", ",", "accumulates", "the", "gradients", "and", "passes", "them", "to", "the", "conv", "network", ".", "In", "practice", ",", "we", "use", "all", "RoIs", "from", "all", "images", "as", ",", "therefore", "the", "effective", "batch", "size", "for", "the", "readonly", "RoI", "network", "is", "and", "for", "the", "regular", "RoI", "network", "is", "the", "standard", "from", "Section", "[", "reference", "]", ".", "We", "implement", "both", "options", "described", "above", "using", "the", "Caffe", "framework", "(", "see", ")", ".", "Our", "implementation", "uses", "gradient", "accumulation", "with", "forward", "-", "backward", "passes", "of", "single", "image", "mini", "-", "batches", ".", "Following", "FRCN", ",", "we", "use", "(", "which", "results", "in", ")", "and", ".", "Under", "these", "settings", ",", "the", "proposed", "architecture", "(", "Figure", "[", "reference", "]", ")", "has", "similar", "memory", "footprint", "as", "the", "first", "option", ",", "but", "is", "faster", ".", "Unless", "specified", "otherwise", ",", "the", "architecture", "and", "settings", "described", "above", "will", "be", "used", "throughout", "this", "paper", ".", "section", ":", "Analyzing", "online", "hard", "example", "mining", "This", "section", "compares", "FRCN", "training", "with", "online", "hard", "example", "mining", "(", "OHEM", ")", "to", "the", "baseline", "heuristic", "sampling", "approach", ".", "We", "also", "compare", "FRCN", "with", "OHEM", "to", "a", "less", "efficient", "approach", "that", "uses", "all", "available", "example", "RoIs", "in", "each", "mini", "-", "batch", ",", "not", "just", "the", "hardest", "examples", ".", "subsection", ":", "Experimental", "setup", "We", "conduct", "experiments", "with", "two", "standard", "ConvNet", "architectures", ":", "VGG_CNN_M_1024", "(", "VGGM", ",", "for", "short", ")", "from", ",", "which", "is", "a", "wider", "version", "of", "AlexNet", ",", "and", "VGG16", "from", ".", "All", "experiments", "in", "this", "section", "are", "performed", "on", "the", "PASCAL", "VOC07", "dataset", ".", "Training", "is", "done", "on", "the", "trainval", "set", "and", "testing", "on", "the", "test", "set", ".", "Unless", "specified", "otherwise", ",", "we", "will", "use", "the", "default", "settings", "from", "FRCN", ".", "We", "train", "all", "methods", "with", "SGD", "for", "80k", "mini", "-", "batch", "iterations", ",", "with", "an", "initial", "learning", "rate", "of", "0.001", "and", "we", "decay", "the", "learning", "rate", "by", "0.1", "every", "30k", "iterations", ".", "The", "baseline", "numbers", "reported", "in", "Table", "[", "reference", "]", "(", "row", "1", "-", "2", ")", "were", "reproduced", "using", "our", "training", "schedule", "and", "are", "slightly", "higher", "than", "the", "ones", "reported", "in", ".", "subsection", ":", "OHEM", "vs.", "heuristic", "sampling", "Standard", "FRCN", ",", "reported", "in", "Table", "[", "reference", "]", "(", "rows", ")", ",", "uses", "as", "a", "heuristic", "for", "hard", "mining", "(", "Section", "[", "reference", "]", ")", ".", "To", "test", "the", "importance", "of", "this", "heuristic", ",", "we", "ran", "FRCN", "with", ".", "Table", "[", "reference", "]", "(", "rows", ")", "shows", "that", "for", "VGGM", ",", "mAP", "drops", "by", "points", ",", "whereas", "for", "VGG16", "it", "remains", "roughly", "the", "same", ".", "Now", "compare", "this", "to", "training", "FRCN", "with", "OHEM", "(", "rows", ")", ".", "OHEM", "improves", "mAP", "by", "points", "compared", "to", "FRCN", "with", "the", "heuristic", "for", "VGGM", ",", "and", "points", "without", "the", "heuristic", ".", "This", "result", "demonstrates", "the", "sub", "-", "optimality", "of", "these", "heuristics", "and", "the", "effectiveness", "of", "our", "hard", "mining", "approach", ".", "subsection", ":", "Robust", "gradient", "estimates", "One", "concern", "over", "using", "only", "images", "per", "batch", "is", "that", "it", "may", "cause", "unstable", "gradients", "and", "slow", "convergence", "because", "RoIs", "from", "an", "image", "may", "be", "highly", "correlated", ".", "FRCN", "reports", "that", "this", "was", "not", "a", "practical", "issue", "for", "their", "training", ".", "But", "this", "detail", "might", "raise", "concerns", "over", "our", "training", "procedure", "because", "we", "use", "examples", "with", "high", "loss", "from", "the", "same", "image", "and", "as", "a", "result", "they", "may", "be", "more", "highly", "correlated", ".", "To", "address", "this", "concern", ",", "we", "experiment", "with", "in", "order", "to", "increase", "correlation", "in", "an", "effort", "to", "break", "our", "method", ".", "As", "seen", "in", "Table", "[", "reference", "]", "(", "rows", ")", ",", "performance", "of", "the", "original", "FRCN", "drops", "by", "point", "with", ",", "but", "when", "using", "our", "training", "procedure", ",", "mAP", "remains", "approximately", "the", "same", ".", "This", "shows", "that", "OHEM", "is", "robust", "in", "case", "one", "needs", "fewer", "images", "per", "batch", "in", "order", "to", "reduce", "GPU", "memory", "usage", ".", "subsection", ":", "Why", "just", "hard", "examples", ",", "when", "you", "can", "use", "all", "?", "Online", "hard", "example", "mining", "is", "based", "on", "the", "hypothesis", "that", "it", "is", "important", "to", "consider", "all", "RoIs", "in", "an", "image", "and", "then", "select", "hard", "examples", "for", "training", ".", "But", "what", "if", "we", "train", "with", "all", "the", "RoIs", ",", "not", "just", "the", "hard", "ones", "?", "The", "easy", "examples", "will", "have", "low", "loss", ",", "and", "wo", "n’t", "contribute", "much", "to", "the", "gradient", ";", "training", "will", "automatically", "focus", "on", "the", "hard", "examples", ".", "To", "compare", "this", "option", ",", "we", "ran", "standard", "FRCN", "training", "with", "a", "large", "mini", "-", "batch", "size", "of", ",", "using", ",", "and", "with", "other", "hyperparameters", "fixed", ".", "Because", "this", "experiment", "uses", "a", "large", "mini", "-", "batch", ",", "it", "’s", "important", "to", "tune", "the", "learning", "rate", "to", "adjust", "for", "this", "change", ".", "We", "found", "optimal", "results", "by", "increasing", "it", "to", "for", "VGG16", "and", "for", "VGGM", ".", "The", "outcomes", "are", "reported", "in", "Table", "[", "reference", "]", "(", "rows", ")", ".", "Using", "these", "settings", ",", "mAP", "of", "both", "VGG16", "and", "VGGM", "increased", "by", "point", "compared", "to", ",", "but", "the", "improvement", "from", "our", "approach", "is", "still", "points", "over", "using", "all", "RoIs", ".", "Moreover", ",", "because", "we", "compute", "gradients", "with", "a", "smaller", "mini", "-", "batch", "size", "training", "is", "faster", ".", "Removing", "hard", "mining", "heuristic", "(", "Section", "[", "reference", "]", ")", "Fewer", "images", "per", "batch", "(", "Section", "[", "reference", "]", ")", "Bigger", "batch", ",", "High", "LR", "(", "Section", "[", "reference", "]", ")", "Our", "Approach", "subsection", ":", "Better", "optimization", "Finally", ",", "we", "analyze", "the", "training", "loss", "for", "the", "various", "FRCN", "training", "methods", "discussed", "above", ".", "It", "’s", "important", "to", "measure", "training", "loss", "in", "a", "way", "that", "does", "not", "depend", "on", "the", "sampling", "procedure", "and", "thus", "results", "in", "a", "valid", "comparison", "between", "methods", ".", "To", "achieve", "this", "goal", ",", "we", "take", "model", "snapshots", "from", "each", "method", "every", "20k", "steps", "of", "optimization", "and", "run", "them", "over", "the", "entire", "VOC07", "trainval", "set", "to", "compute", "the", "average", "loss", "over", "all", "RoIs", ".", "This", "measures", "the", "training", "set", "loss", "in", "a", "way", "that", "does", "not", "depend", "on", "the", "example", "sampling", "scheme", ".", "Figure", "[", "reference", "]", "shows", "the", "average", "loss", "per", "RoI", "for", "VGG16", "with", "the", "various", "hyperparameter", "settings", "discussed", "above", "and", "presented", "in", "Table", "[", "reference", "]", ".", "We", "see", "that", "results", "in", "the", "highest", "training", "loss", ",", "while", "using", "the", "heuristic", "results", "in", "a", "much", "lower", "training", "loss", ".", "Increasing", "the", "mini", "-", "batch", "size", "to", "and", "increasing", "the", "learning", "rate", "lowers", "the", "training", "loss", "below", "the", "heuristic", ".", "Our", "proposed", "online", "hard", "example", "mining", "method", "achieves", "the", "lowest", "training", "loss", "of", "all", "methods", ",", "validating", "our", "claims", "that", "OHEM", "leads", "to", "better", "training", "for", "FRCN", ".", "*", ":", "uses", "gradient", "accumulation", "over", "two", "forward", "/", "backward", "passes", "subsection", ":", "Computational", "cost", "OHEM", "adds", "reasonable", "computational", "and", "memory", "overhead", ",", "as", "reported", "in", "Table", "[", "reference", "]", ".", "OHEM", "costs", "0.09s", "per", "training", "iteration", "for", "VGGM", "network", "(", "0.43s", "for", "VGG16", ")", "and", "requires", "1", "G", "more", "memory", "(", "2.3", "G", "for", "VGG16", ")", ".", "Given", "that", "FRCN", "is", "a", "fast", "detector", "to", "train", ",", "the", "increase", "in", "training", "time", "is", "likely", "acceptable", "to", "most", "users", ".", ",", "2http:", "//", "host.robots.ox.ac.uk:8080", "/", "anonymous", "/", "H49PTT.html", ",", "3http:", "//", "host.robots.ox.ac.uk:8080", "/", "anonymous", "/", "LSANTB.html", ",", "4http:", "//", "host.robots.ox.ac.uk:8080", "/", "anonymous", "/", "R7EAMX.html", "section", ":", "PASCAL", "VOC", "and", "MS", "COCO", "results", "In", "this", "section", ",", "we", "evaluate", "our", "method", "on", "VOC", "2012", "as", "well", "as", "the", "more", "challenging", "MS", "COCO", "dataset", ".", "We", "demonstrate", "consistent", "and", "significant", "improvement", "in", "FRCN", "performance", "when", "using", "the", "proposed", "OHEM", "approach", ".", "Per", "-", "class", "results", "are", "also", "presented", "on", "VOC", "2007", "for", "comparison", "with", "prior", "work", ".", "paragraph", ":", "Experimental", "setup", ".", "We", "use", "VGG16", "for", "all", "experiments", ".", "When", "training", "on", "VOC07", "trainval", ",", "we", "use", "the", "SGD", "parameters", "as", "in", "Section", "[", "reference", "]", "and", "when", "using", "extra", "data", "(", "07", "+", "12", "and", "07", "++", "12", ",", "see", "Table", "[", "reference", "]", "and", "[", "reference", "]", ")", ",", "we", "use", "200k", "mini", "-", "batch", "iterations", ",", "with", "an", "initial", "learning", "rate", "of", "0.001", "and", "decay", "step", "size", "of", "40k", ".", "When", "training", "on", "MS", "COCO", ",", "we", "use", "240k", "mini", "-", "batch", "iterations", ",", "with", "an", "initial", "learning", "rate", "of", "0.001", "and", "decay", "step", "size", "of", "160k", ",", "owing", "to", "a", "larger", "epoch", "size", ".", "subsection", ":", "VOC", "2007", "and", "2012", "results", "Table", "[", "reference", "]", "shows", "that", "on", "VOC07", ",", "OHEM", "improves", "the", "mAP", "of", "FRCN", "from", "67.2", "%", "to", "69.9", "%", "(", "and", "70.0", "%", "to", "74.6", "%", "with", "extra", "data", ")", ".", "On", "VOC12", ",", "OHEM", "leads", "to", "an", "improvement", "of", "4.1", "points", "in", "mAP", "(", "from", "65.7", "%", "to", "69.8", "%", ")", ".", "With", "extra", "data", ",", "we", "achieve", "an", "mAP", "of", "71.9", "%", "as", "compared", "to", "68.4", "%", "mAP", "of", "FRCN", ",", "an", "improvement", "of", "3.5", "points", ".", "Interestingly", "the", "improvements", "are", "not", "uniform", "across", "categories", ".", "Bottle", ",", "chair", ",", "and", "tvmonitor", "show", "larger", "improvements", "that", "are", "consistent", "across", "the", "different", "PASCAL", "splits", ".", "Why", "these", "classes", "benefit", "the", "most", "is", "an", "interesting", "and", "open", "question", ".", "subsection", ":", "MS", "COCO", "results", "To", "test", "the", "benefit", "of", "using", "OHEM", "on", "a", "larger", "and", "more", "challenging", "dataset", ",", "we", "conduct", "experiments", "on", "MS", "COCO", "and", "report", "numbers", "from", "test", "-", "dev", "2015", "evaluation", "server", "(", "Table", "[", "reference", "]", ")", ".", "On", "the", "standard", "COCO", "evaluation", "metric", ",", "FRCN", "scores", "19.7", "%", "AP", ",", "and", "OHEM", "improves", "it", "to", "22.6", "%", "AP", ".", "Using", "the", "VOC", "overlap", "metric", "of", ",", "OHEM", "gives", "a", "6.6", "points", "boost", "in", "AP", ".", "It", "is", "also", "interesting", "to", "note", "that", "OHEM", "helps", "improve", "the", "AP", "of", "medium", "sized", "objects", "by", "4.9", "points", "on", "the", "strict", "COCO", "AP", "evaluation", "metric", ",", "which", "indicates", "that", "the", "proposed", "hard", "example", "mining", "approach", "is", "helpful", "when", "dealing", "with", "smaller", "sized", "objects", ".", "Note", "that", "FRCN", "with", "and", "without", "OHEM", "were", "trained", "on", "MS", "COCO", "train", "set", ".", "section", ":", "Adding", "bells", "and", "whistles", "We", "’", "ve", "demonstrated", "consistent", "gains", "in", "detection", "accuracy", "by", "applying", "OHEM", "to", "FRCN", "training", ".", "In", "this", "section", ",", "we", "show", "that", "these", "improvements", "are", "orthogonal", "to", "recent", "bells", "and", "whistles", "that", "enhance", "object", "detection", "accuracy", ".", "OHEM", "with", "the", "following", "two", "additions", "yields", "state", "-", "of", "-", "the", "-", "art", "results", "on", "VOC", "and", "competitive", "results", "on", "MS", "COCO", ".", "paragraph", ":", "Multi", "-", "scale", "(", "M", ")", ".", "We", "adopt", "the", "multi", "-", "scale", "strategy", "from", "SPPnet", "(", "and", "used", "by", "both", "FRCN", "and", "MR", "-", "CNN", ")", ".", "Scale", "is", "defined", "as", "the", "size", "of", "the", "shortest", "side", "(", ")", "of", "an", "image", ".", "During", "training", ",", "one", "scale", "is", "chosen", "at", "random", ",", "whereas", "at", "test", "time", "inference", "is", "run", "on", "all", "scales", ".", "For", "VGG16", "networks", ",", "we", "use", "for", "training", ",", "and", "during", "testing", ",", "with", "the", "max", "dimension", "capped", "at", "1000", ".", "The", "scales", "and", "caps", "were", "chosen", "because", "of", "GPU", "memory", "constraints", ".", "paragraph", ":", "Iterative", "bounding", "-", "box", "regression", "(", "B", ")", ".", "We", "adopt", "the", "iterative", "localization", "and", "bounding", "-", "box", "(", "bbox", ")", "voting", "scheme", "from", ".", "The", "network", "evaluates", "each", "proposal", "RoI", "to", "get", "scores", "and", "relocalized", "boxes", ".", "High", "-", "scoring", "boxes", "are", "the", "rescored", "and", "relocalized", ",", "yielding", "boxes", ".", "Union", "of", "and", "is", "used", "as", "the", "final", "set", "for", "post", "-", "processing", ",", "where", "is", "obtained", "using", "NMS", "on", "with", "an", "IoU", "threshold", "of", "0.3", "and", "weighted", "voting", "is", "performed", "on", "each", "box", "in", "using", "boxes", "in", "with", "an", "IoU", "of", "0.5", "with", "(", "see", "for", "details", ")", ".", "from", "the", "leaderboard", ",", "*", "trained", "on", "trainval", "set", "subsection", ":", "VOC", "2007", "and", "2012", "results", "We", "report", "the", "results", "on", "VOC", "benchmarks", "in", "Table", "[", "reference", "]", "and", "[", "reference", "]", ".", "On", "VOC07", ",", "FRCN", "with", "the", "above", "mentioned", "additions", "achieves", "72.4", "%", "mAP", "and", "OHEM", "improves", "it", "to", "75.1", "%", ",", "which", "is", "currently", "the", "highest", "reported", "score", "under", "this", "setting", "(", "07", "data", ")", ".", "When", "using", "extra", "data", "(", "07", "+", "12", ")", ",", "OHEM", "achieves", "78.9", "%", "mAP", ",", "surpassing", "the", "current", "state", "-", "of", "-", "the", "-", "art", "MR", "-", "CNN", "(", "78.2", "%", "mAP", ")", ".", "We", "note", "that", "MR", "-", "CNN", "uses", "selective", "search", "and", "edge", "boxes", "during", "training", ",", "whereas", "we", "only", "use", "selective", "search", "boxes", ".", "Our", "multi", "-", "scale", "implementation", "is", "also", "different", ",", "using", "fewer", "scales", "than", "MR", "-", "CNN", ".", "On", "VOC12", "(", "Table", "[", "reference", "]", ")", ",", "we", "consistently", "perform", "better", "than", "MR", "-", "CNN", ".", "When", "using", "extra", "data", ",", "we", "achieve", "state", "-", "of", "-", "the", "-", "art", "mAP", "of", "76.3", "%", "(", "73.9", "%", "mAP", "of", "MR", "-", "CNN", ")", ".", "paragraph", ":", "Ablation", "analysis", ".", "We", "now", "study", "in", "detail", "the", "impact", "of", "these", "two", "additions", "and", "whether", "OHEM", "is", "complementary", "to", "them", ",", "and", "report", "the", "analysis", "in", "Table", "[", "reference", "]", ".", "Baseline", "FRCN", "mAP", "improves", "from", "67.2", "%", "to", "68.6", "%", "when", "using", "multi", "-", "scale", "during", "both", "training", "and", "testing", "(", "we", "refer", "to", "this", "as", "M", ")", ".", "However", ",", "note", "that", "there", "is", "only", "a", "marginal", "benefit", "of", "using", "it", "at", "training", "time", ".", "Iterative", "bbox", "regression", "(", "B", ")", "further", "improves", "the", "FRCN", "mAP", "to", "72.4", "%", ".", "But", "more", "importantly", ",", "using", "OHEM", "improves", "it", "to", "75.1", "%", "mAP", ",", "which", "is", "state", "-", "of", "-", "the", "-", "art", "for", "methods", "trained", "on", "VOC07", "data", "(", "see", "Table", "[", "reference", "]", ")", ".", "In", "fact", ",", "using", "OHEM", "consistently", "results", "in", "higher", "mAP", "for", "all", "variants", "of", "these", "two", "additions", "(", "see", "Table", "[", "reference", "]", ")", ".", "Iterative", "bbox", "reg", ".", "(", "B", ")", "subsection", ":", "MS", "COCO", "results", "MS", "COCO", "test", "-", "dev", "2015", "evaluation", "server", "results", "are", "reported", "in", "Table", "[", "reference", "]", ".", "Using", "multi", "-", "scale", "improves", "the", "performance", "of", "our", "method", "to", "24.4", "%", "AP", "on", "the", "standard", "COCO", "metric", "and", "to", "44.4", "%", "AP", "on", "the", "VOC", "metric", ".", "This", "again", "shows", "the", "complementary", "nature", "of", "using", "multi", "-", "scale", "and", "OHEM", ".", "Finally", ",", "we", "train", "our", "method", "using", "the", "entire", "MS", "COCO", "trainval", "set", ",", "which", "further", "improves", "performance", "to", "25.5", "%", "AP", "(", "and", "45.9", "%", "AP", ")", ".", "In", "the", "2015", "MS", "COCO", "Detection", "Challenge", ",", "a", "variant", "of", "this", "approach", "finished", "place", "overall", ".", "section", ":", "Conclusion", "We", "presented", "an", "online", "hard", "example", "mining", "(", "OHEM", ")", "algorithm", ",", "a", "simple", "and", "effective", "method", "to", "train", "region", "-", "based", "ConvNet", "detectors", ".", "OHEM", "eliminates", "several", "heuristics", "and", "hyperparameters", "in", "common", "use", "by", "automatically", "selecting", "hard", "examples", ",", "thus", "simplifying", "training", ".", "We", "conducted", "extensive", "experimental", "analysis", "to", "demonstrate", "the", "effectiveness", "of", "the", "proposed", "algorithm", ",", "which", "leads", "to", "better", "training", "convergence", "and", "consistent", "improvements", "in", "detection", "accuracy", "on", "standard", "benchmarks", ".", "We", "also", "reported", "state", "-", "of", "-", "the", "-", "art", "results", "on", "PASCAL", "VOC", "2007", "and", "2012", "when", "using", "OHEM", "with", "other", "orthogonal", "additions", ".", "Though", "we", "used", "Fast", "R", "-", "CNN", "throughout", "this", "paper", ",", "OHEM", "can", "be", "used", "for", "training", "any", "region", "-", "based", "ConvNet", "detector", ".", "Our", "experimental", "analysis", "was", "based", "on", "the", "overall", "detection", "accuracy", ",", "however", "it", "will", "be", "an", "interesting", "future", "direction", "to", "study", "the", "impact", "of", "various", "training", "methodologies", "on", "individual", "category", "performance", ".", "paragraph", ":", "Acknowledgment", ".", "This", "project", "started", "as", "an", "intern", "project", "at", "Microsoft", "Research", "and", "continued", "at", "CMU", ".", "We", "thank", "Larry", "Zitnick", ",", "Ishan", "Misra", "and", "Sean", "Bell", "for", "many", "helpful", "discussions", ".", "AS", "was", "supported", "by", "the", "Microsoft", "Research", "PhD", "Fellowship", ".", "This", "work", "was", "also", "partially", "supported", "by", "ONR", "MURI", "N000141612007", ".", "We", "thank", "NVIDIA", "for", "donating", "GPUs", ".", "bibliography", ":", "References"], "entity": [[3, 8, ["Region", "-", "based", "Object", "Detectors"], "Method"], [9, 13, ["Online", "Hard", "Example", "Mining"], "Method"], [16, 18, ["object", "detection"], "Task"], [27, 31, ["region", "-", "based", "ConvNets"], "Method"], [55, 59, ["online", "hard", "example", "mining"], "Method"], [65, 70, ["region", "-", "based", "ConvNet", "detectors"], "Method"], [99, 101, ["Automatic", "selection"], "Task"], [113, 114, ["OHEM"], "Method"], [141, 143, ["detection", "performance"], "Metric"], [146, 149, ["PASCAL", "VOC", "2007"], "Material"], [184, 185, ["OHEM"], "Method"], [201, 202, ["mAP"], "Metric"], [203, 206, ["PASCAL", "VOC", "2007"], "Material"], [213, 215, ["Image", "classification"], "Task"], [216, 218, ["object", "detection"], "Task"], [221, 224, ["computer", "vision", "tasks"], "Task"], [225, 227, ["Object", "detectors"], "Method"], [232, 233, ["reduction"], "Method"], [235, 237, ["object", "detection"], "Task"], [239, 242, ["image", "classification", "problem"], "Task"], [254, 258, ["natural", "image", "classification", "tasks"], "Task"], [297, 302, ["sliding", "-", "window", "object", "detectors"], "Method"], [306, 309, ["deformable", "parts", "model"], "Method"], [310, 312, ["DPM", ")"], "Method"], [334, 340, ["object", "-", "proposal", "-", "based", "detectors"], "Method"], [349, 351, ["imbalance", "ratio"], "Metric"], [365, 367, ["learning", "techniques"], "Method"], [374, 375, ["training"], "Task"], [377, 378, ["accuracy"], "Metric"], [397, 398, ["bootstrapping"], "Method"], [403, 406, ["hard", "negative", "mining"], "Task"], [416, 417, ["Bootstrapping"], "Method"], [439, 442, ["face", "detection", "models"], "Method"], [466, 467, ["detector"], "Method"], [477, 480, ["iterative", "training", "algorithm"], "Method"], [485, 487, ["detection", "model"], "Method"], [536, 537, ["Bootstrapping"], "Method"], [546, 549, ["object", "detection", "research"], "Task"], [557, 558, ["SVMs"], "Method"], [559, 561, ["pedestrian", "detection"], "Task"], [569, 570, ["bootstrapping"], "Method"], [571, 572, ["SVMs"], "Method"], [591, 594, ["hard", "negative", "mining"], "Task"], [600, 601, ["SVMs"], "Method"], [602, 604, ["object", "detection"], "Task"], [605, 606, ["Bootstrapping"], "Method"], [615, 617, ["learning", "models"], "Method"], [619, 622, ["shallow", "neural", "networks"], "Method"], [623, 626, ["boosted", "decision", "trees"], "Method"], [629, 631, ["detection", "methods"], "Method"], [633, 637, ["deep", "convolutional", "neural", "networks"], "Method"], [638, 640, ["ConvNets", ")"], "Method"], [643, 646, ["R", "-", "CNN"], "Method"], [647, 648, ["SPPnet"], "Method"], [651, 652, ["SVMs"], "Method"], [654, 657, ["hard", "negative", "mining"], "Method"], [673, 675, ["object", "detectors"], "Method"], [678, 682, ["Fast", "R", "-", "CNN"], "Method"], [689, 690, ["bootstrapping"], "Method"], [705, 708, ["online", "learning", "algorithms"], "Method"], [714, 716, ["deep", "ConvNets"], "Method"], [718, 721, ["stochastic", "gradient", "descent"], "Method"], [722, 723, ["SGD"], "Method"], [729, 730, ["Bootstrapping"], "Method"], [742, 744, ["alternation", "template"], "Method"], [754, 756, ["fixed", "model"], "Method"], [792, 795, ["deep", "ConvNet", "detectors"], "Method"], [796, 797, ["SGD"], "Method"], [803, 805, ["SGD", "steps"], "Method"], [831, 833, ["online", "form"], "Method"], [834, 837, ["hard", "example", "selection"], "Task"], [846, 848, ["bootstrapping", "technique"], "Method"], [849, 853, ["online", "hard", "example", "mining"], "Method"], [865, 867, ["detection", "models"], "Method"], [869, 871, ["deep", "ConvNets"], "Method"], [879, 880, ["SGD"], "Method"], [914, 919, ["detection", "-", "specific", "problem", "structure"], "Task"], [922, 926, ["SGD", "mini", "-", "batch"], "Method"], [957, 959, ["Gradient", "computation"], "Method"], [960, 961, ["backpropagation"], "Method"], [978, 979, ["OHEM"], "Method"], [995, 998, ["baseline", "training", "algorithm"], "Method"], [1011, 1015, ["region", "-", "based", "ConvNets"], "Method"], [1024, 1027, ["mean", "average", "precision"], "Metric"], [1056, 1057, ["OHEM"], "Method"], [1063, 1065, ["object", "detection"], "Task"], [1073, 1078, ["iterative", "bounding", "-", "box", "regression"], "Method"], [1084, 1085, ["OHEM"], "Method"], [1100, 1101, ["mAP"], "Metric"], [1102, 1105, ["PASCAL", "VOC", "2007"], "Material"], [1114, 1116, ["Object", "detection"], "Task"], [1126, 1128, ["computer", "vision"], "Task"], [1132, 1134, ["dataset", "bootstrapping"], "Task"], [1137, 1140, ["hard", "negative", "mining"], "Task"], [1151, 1153, ["object", "detectors"], "Method"], [1159, 1160, ["SVMs"], "Method"], [1162, 1165, ["detection", "scoring", "function"], "Metric"], [1170, 1174, ["deep", "convolutional", "neural", "network"], "Method"], [1175, 1176, ["ConvNet"], "Method"], [1178, 1180, ["feature", "extraction"], "Task"], [1197, 1201, ["Faster", "R", "-", "CNN"], "Method"], [1208, 1209, ["SVMs"], "Method"], [1216, 1217, ["SGD"], "Method"], [1219, 1223, ["hard", "example", "mining", "techniques"], "Method"], [1237, 1242, ["online", "hard", "example", "mining", "algorithm"], "Method"], [1244, 1248, ["optimization", "and", "detection", "accuracy"], "Metric"], [1252, 1255, ["hard", "example", "mining"], "Task"], [1257, 1262, ["ConvNet", "-", "based", "object", "detection"], "Task"], [1269, 1272, ["hard", "example", "selection"], "Method"], [1274, 1276, ["deep", "networks"], "Method"], [1279, 1282, ["Hard", "example", "mining"], "Task"], [1287, 1290, ["example", "mining", "algorithms"], "Method"], [1300, 1301, ["SVMs"], "Method"], [1307, 1309, ["training", "algorithm"], "Method"], [1320, 1321, ["SVM"], "Method"], [1398, 1401, ["global", "SVM", "solution"], "Method"], [1424, 1427, ["non", "-", "SVMs"], "Method"], [1437, 1440, ["shallow", "neural", "networks"], "Method"], [1441, 1444, ["boosted", "decision", "trees"], "Method"], [1464, 1467, ["machine", "learning", "model"], "Method"], [1522, 1527, ["ConvNet", "-", "based", "object", "detection"], "Method"], [1539, 1541, ["object", "detection"], "Task"], [1552, 1554, ["deep", "ConvNets"], "Method"], [1555, 1557, ["ImageNet", "classification"], "Task"], [1559, 1565, ["R", "-", "CNN", "and", "OverFeat", "detectors"], "Method"], [1572, 1574, ["PASCAL", "VOC"], "Material"], [1575, 1577, ["ImageNet", "detection"], "Task"], [1578, 1579, ["OverFeat"], "Method"], [1583, 1588, ["sliding", "-", "window", "detection", "method"], "Method"], [1597, 1599, ["search", "method"], "Method"], [1600, 1601, ["detection"], "Task"], [1602, 1605, ["R", "-", "CNN"], "Method"], [1610, 1612, ["region", "proposals"], "Method"], [1621, 1624, ["selective", "search", "algorithm"], "Method"], [1626, 1629, ["R", "-", "CNN"], "Method"], [1636, 1640, ["region", "-", "based", "ConvNets"], "Task"], [1642, 1643, ["SPPnet"], "Method"], [1644, 1647, ["MR", "-", "CNN"], "Method"], [1649, 1653, ["Fast", "R", "-", "CNN"], "Method"], [1662, 1665, ["Hard", "example", "selection"], "Task"], [1666, 1668, ["deep", "learning"], "Task"], [1683, 1685, ["deep", "networks"], "Method"], [1725, 1727, ["image", "descriptors"], "Task"], [1744, 1746, ["triplet", "loss"], "Method"], [1753, 1758, ["online", "selection", "of", "hard", "examples"], "Task"], [1759, 1764, ["mini", "-", "batch", "SGD", "methods"], "Method"], [1771, 1772, ["loss"], "Task"], [1778, 1779, ["ConvNets"], "Method"], [1780, 1782, ["image", "classification"], "Task"], [1789, 1794, ["online", "hard", "example", "selection", "strategy"], "Method"], [1795, 1800, ["region", "-", "based", "object", "detectors"], "Task"], [1805, 1809, ["Fast", "R", "-", "CNN"], "Method"], [1813, 1817, ["Fast", "R", "-", "CNN"], "Method"], [1818, 1819, ["FRCN"], "Method"], [1822, 1823, ["FRCN"], "Method"], [1854, 1859, ["convolutional", "(", "conv", ")", "network"], "Method"], [1861, 1867, ["convolution", "and", "max", "-", "pooling", "layers"], "Method"], [1874, 1876, ["Convolutional", "Network"], "Method"], [1881, 1883, ["RoI", "network"], "Method"], [1885, 1889, ["RoI", "-", "pooling", "layer"], "Method"], [1891, 1898, ["fully", "-", "connected", "(", "fc", ")", "layers"], "Method"], [1900, 1902, ["loss", "layers"], "Method"], [1909, 1911, ["RoI", "Network"], "Method"], [1915, 1916, ["inference"], "Task"], [1918, 1920, ["conv", "network"], "Method"], [1951, 1955, ["RoI", "-", "pooling", "layer"], "Method"], [1979, 1981, ["fc", "layers"], "Method"], [2009, 2013, ["bounding", "-", "box", "relocalization"], "Task"], [2020, 2021, ["FRCN"], "Method"], [2024, 2026, ["object", "detector"], "Method"], [2037, 2039, ["end", "system"], "Method"], [2045, 2047, ["network", "setup"], "Method"], [2048, 2049, ["conv"], "Method"], [2050, 2051, ["RoI"], "Method"], [2058, 2059, ["detectors"], "Method"], [2060, 2061, ["SPPnet"], "Method"], [2062, 2065, ["MR", "-", "CNN"], "Method"], [2085, 2086, ["FRCN"], "Method"], [2092, 2094, ["conv", "network"], "Method"], [2107, 2109, ["conv", "network"], "Method"], [2115, 2116, ["SPPnet"], "Method"], [2117, 2120, ["MR", "-", "CNN"], "Method"], [2124, 2126, ["RoI", "network"], "Method"], [2133, 2135, ["SVM", "classifier"], "Method"], [2137, 2140, ["hard", "negative", "mining"], "Task"], [2142, 2143, ["FRCN"], "Method"], [2145, 2147, ["RoI", "network"], "Method"], [2152, 2153, ["classifiers"], "Method"], [2161, 2163, ["unified", "system"], "Method"], [2165, 2167, ["SVM", "classifiers"], "Method"], [2178, 2180, ["deep", "networks"], "Method"], [2181, 2182, ["FRCN"], "Method"], [2185, 2188, ["stochastic", "gradient", "descent"], "Method"], [2189, 2190, ["SGD"], "Method"], [2196, 2197, ["RoI"], "Method"], [2202, 2205, ["classification", "log", "loss"], "Method"], [2218, 2220, ["localization", "loss"], "Method"], [2235, 2238, ["conv", "network", "computation"], "Task"], [2280, 2283, ["RoI", "sampling", "procedure"], "Method"], [2328, 2331, ["intersection", "over", "union"], "Metric"], [2360, 2362, ["evaluation", "protocol"], "Metric"], [2377, 2381, ["SVM", "hard", "mining", "procedures"], "Method"], [2382, 2385, ["R", "-", "CNN"], "Method"], [2386, 2387, ["SPPnet"], "Method"], [2389, 2392, ["MR", "-", "CNN"], "Method"], [2432, 2433, ["bg_lo"], "Method"], [2437, 2438, ["FRCN"], "Method"], [2439, 2440, ["SPPnet"], "Method"], [2448, 2451, ["hard", "negative", "mining"], "Task"], [2487, 2488, ["convergence"], "Metric"], [2489, 2491, ["detection", "accuracy"], "Metric"], [2603, 2604, ["accuracy"], "Metric"], [2605, 2607, ["points", "mAP"], "Metric"], [2634, 2639, ["online", "hard", "example", "mining", "algorithm"], "Method"], [2641, 2645, ["Fast", "R", "-", "CNN"], "Method"], [2665, 2668, ["mini", "-", "batches"], "Method"], [2669, 2670, ["SGD"], "Method"], [2690, 2691, ["training"], "Metric"], [2693, 2695, ["training", "loss"], "Metric"], [2698, 2700, ["testing", "performance"], "Metric"], [2701, 2702, ["mAP"], "Metric"], [2706, 2710, ["Online", "hard", "example", "mining"], "Task"], [2717, 2721, ["hard", "example", "mining", "algorithm"], "Method"], [2731, 2733, ["fixed", "model"], "Method"], [2772, 2777, ["SVM", "-", "based", "object", "detectors"], "Method"], [2781, 2782, ["SVMs"], "Method"], [2784, 2787, ["R", "-", "CNN"], "Method"], [2788, 2789, ["SPPnet"], "Method"], [2826, 2827, ["SVM"], "Method"], [2859, 2860, ["learning"], "Task"], [2862, 2864, ["model", "updates"], "Method"], [2885, 2887, ["alternating", "steps"], "Method"], [2892, 2893, ["FRCN"], "Method"], [2896, 2898, ["online", "SGD"], "Method"], [2905, 2907, ["SGD", "iteration"], "Method"], [2944, 2945, ["SGD"], "Method"], [2969, 2972, ["baseline", "SGD", "approach"], "Method"], [2974, 2975, ["learning"], "Task"], [2983, 2988, ["online", "hard", "example", "mining", "algorithm"], "Method"], [2989, 2990, ["OHEM"], "Method"], [3000, 3002, ["SGD", "iteration"], "Task"], [3007, 3010, ["conv", "feature", "map"], "Method"], [3012, 3014, ["conv", "network"], "Method"], [3017, 3019, ["RoI", "network"], "Method"], [3021, 3023, ["feature", "map"], "Method"], [3041, 3043, ["forward", "pass"], "Method"], [3050, 3052, ["RoI", "pooling"], "Method"], [3055, 3057, ["fc", "layers"], "Method"], [3059, 3061, ["loss", "computation"], "Method"], [3102, 3104, ["forward", "computation"], "Method"], [3110, 3113, ["conv", "feature", "map"], "Method"], [3144, 3146, ["backward", "pass"], "Method"], [3200, 3203, ["loss", "double", "counting"], "Task"], [3216, 3220, ["non", "-", "maximum", "suppression"], "Method"], [3221, 3222, ["NMS"], "Method"], [3225, 3226, ["deduplication"], "Task"], [3241, 3242, ["NMS"], "Method"], [3299, 3301, ["data", "balancing"], "Task"], [3371, 3374, ["mini", "-", "batch"], "Method"], [3390, 3391, ["OHEM"], "Method"], [3415, 3418, ["hard", "example", "selection"], "Task"], [3420, 3422, ["loss", "layer"], "Method"], [3462, 3464, ["RoI", "network"], "Method"], [3491, 3494, ["deep", "learning", "toolboxes"], "Method"], [3518, 3520, ["RoI", "network"], "Method"], [3531, 3534, ["readonly", "RoI", "network"], "Method"], [3557, 3559, ["RoI", "network"], "Method"], [3572, 3574, ["SGD", "iteration"], "Method"], [3582, 3585, ["readonly", "RoI", "network"], "Method"], [3587, 3589, ["forward", "pass"], "Method"], [3608, 3612, ["hard", "RoI", "sampling", "module"], "Method"], [3631, 3634, ["regular", "RoI", "network"], "Method"], [3666, 3668, ["conv", "network"], "Method"], [3684, 3686, ["batch", "size"], "Metric"], [3688, 3691, ["readonly", "RoI", "network"], "Method"], [3695, 3698, ["regular", "RoI", "network"], "Method"], [3715, 3717, ["Caffe", "framework"], "Method"], [3724, 3726, ["gradient", "accumulation"], "Method"], [3727, 3737, ["forward", "-", "backward", "passes", "of", "single", "image", "mini", "-", "batches"], "Method"], [3739, 3740, ["FRCN"], "Method"], [3796, 3800, ["online", "hard", "example", "mining"], "Method"], [3806, 3810, ["online", "hard", "example", "mining"], "Method"], [3811, 3812, ["OHEM"], "Method"], [3815, 3819, ["baseline", "heuristic", "sampling", "approach"], "Method"], [3823, 3824, ["FRCN"], "Method"], [3825, 3826, ["OHEM"], "Method"], [3859, 3861, ["ConvNet", "architectures"], "Method"], [3862, 3863, ["VGG_CNN_M_1024"], "Method"], [3864, 3865, ["VGGM"], "Method"], [3877, 3878, ["AlexNet"], "Method"], [3880, 3881, ["VGG16"], "Method"], [3892, 3895, ["PASCAL", "VOC07", "dataset"], "Material"], [3921, 3922, ["FRCN"], "Method"], [3928, 3929, ["SGD"], "Method"], [3939, 3941, ["learning", "rate"], "Metric"], [3947, 3949, ["learning", "rate"], "Metric"], [3988, 3989, ["OHEM"], "Method"], [3990, 3992, ["heuristic", "sampling"], "Method"], [3993, 3994, ["FRCN"], "Method"], [4008, 4009, ["heuristic"], "Method"], [4010, 4012, ["hard", "mining"], "Task"], [4029, 4030, ["FRCN"], "Method"], [4042, 4043, ["VGGM"], "Method"], [4044, 4045, ["mAP"], "Metric"], [4063, 4064, ["FRCN"], "Method"], [4065, 4066, ["OHEM"], "Method"], [4070, 4071, ["OHEM"], "Method"], [4072, 4073, ["mAP"], "Metric"], [4077, 4078, ["FRCN"], "Method"], [4082, 4083, ["VGGM"], "Method"], [4105, 4108, ["hard", "mining", "approach"], "Method"], [4111, 4114, ["Robust", "gradient", "estimates"], "Method"], [4142, 4143, ["FRCN"], "Method"], [4223, 4224, ["FRCN"], "Method"], [4233, 4235, ["training", "procedure"], "Method"], [4236, 4237, ["mAP"], "Metric"], [4245, 4246, ["OHEM"], "Method"], [4277, 4281, ["Online", "hard", "example", "mining"], "Task"], [4390, 4392, ["learning", "rate"], "Metric"], [4407, 4408, ["VGG16"], "Method"], [4410, 4411, ["VGGM"], "Method"], [4429, 4430, ["mAP"], "Metric"], [4464, 4469, ["mini", "-", "batch", "size", "training"], "Method"], [4473, 4476, ["hard", "mining", "heuristic"], "Method"], [4508, 4509, ["optimization"], "Task"], [4514, 4516, ["training", "loss"], "Metric"], [4530, 4532, ["training", "loss"], "Metric"], [4541, 4543, ["sampling", "procedure"], "Method"], [4569, 4570, ["optimization"], "Task"], [4582, 4584, ["average", "loss"], "Metric"], [4591, 4594, ["training", "set", "loss"], "Metric"], [4603, 4606, ["example", "sampling", "scheme"], "Method"], [4613, 4617, ["average", "loss", "per", "RoI"], "Metric"], [4618, 4619, ["VGG16"], "Method"], [4641, 4643, ["training", "loss"], "Metric"], [4647, 4648, ["heuristic"], "Method"], [4653, 4655, ["training", "loss"], "Metric"], [4666, 4668, ["learning", "rate"], "Metric"], [4670, 4672, ["training", "loss"], "Metric"], [4686, 4688, ["training", "loss"], "Metric"], [4696, 4697, ["OHEM"], "Method"], [4700, 4701, ["training"], "Task"], [4702, 4703, ["FRCN"], "Method"], [4707, 4709, ["gradient", "accumulation"], "Method"], [4717, 4719, ["Computational", "cost"], "Metric"], [4719, 4720, ["OHEM"], "Method"], [4722, 4726, ["computational", "and", "memory", "overhead"], "Metric"], [4735, 4736, ["OHEM"], "Method"], [4742, 4744, ["VGGM", "network"], "Method"], [4747, 4748, ["VGG16"], "Method"], [4759, 4760, ["VGG16"], "Method"], [4764, 4765, ["FRCN"], "Method"], [4768, 4769, ["detector"], "Method"], [4775, 4777, ["training", "time"], "Metric"], [4799, 4800, ["LSANTB.html"], "Method"], [4810, 4812, ["PASCAL", "VOC"], "Material"], [4813, 4815, ["MS", "COCO"], "Material"], [4861, 4863, ["VOC", "2007"], "Material"], [4876, 4877, ["VGG16"], "Method"], [4884, 4886, ["VOC07", "trainval"], "Material"], [4926, 4930, ["mini", "-", "batch", "iterations"], "Method"], [4934, 4936, ["learning", "rate"], "Metric"], [4939, 4942, ["decay", "step", "size"], "Metric"], [4948, 4950, ["MS", "COCO"], "Material"], [4954, 4958, ["mini", "-", "batch", "iterations"], "Method"], [4962, 4964, ["learning", "rate"], "Metric"], [4967, 4970, ["decay", "step", "size"], "Metric"], [4994, 4995, ["VOC07"], "Material"], [4996, 4997, ["OHEM"], "Method"], [5021, 5022, ["VOC12"], "Material"], [5023, 5024, ["OHEM"], "Method"], [5032, 5033, ["mAP"], "Metric"], [5049, 5050, ["mAP"], "Metric"], [5058, 5059, ["mAP"], "Metric"], [5060, 5061, ["FRCN"], "Method"], [5110, 5112, ["MS", "COCO"], "Material"], [5119, 5120, ["OHEM"], "Method"], [5132, 5134, ["MS", "COCO"], "Material"], [5154, 5157, ["COCO", "evaluation", "metric"], "Metric"], [5162, 5163, ["AP"], "Metric"], [5165, 5166, ["OHEM"], "Method"], [5171, 5172, ["AP"], "Metric"], [5180, 5181, ["OHEM"], "Method"], [5187, 5188, ["AP"], "Metric"], [5196, 5197, ["OHEM"], "Method"], [5200, 5201, ["AP"], "Metric"], [5221, 5225, ["hard", "example", "mining", "approach"], "Method"], [5236, 5237, ["FRCN"], "Method"], [5240, 5241, ["OHEM"], "Method"], [5262, 5264, ["detection", "accuracy"], "Metric"], [5266, 5267, ["OHEM"], "Method"], [5289, 5292, ["object", "detection", "accuracy"], "Metric"], [5293, 5294, ["OHEM"], "Method"], [5309, 5310, ["VOC"], "Material"], [5314, 5316, ["MS", "COCO"], "Material"], [5334, 5335, ["SPPnet"], "Method"], [5340, 5341, ["FRCN"], "Method"], [5342, 5345, ["MR", "-", "CNN"], "Method"], [5377, 5378, ["inference"], "Task"], [5385, 5387, ["VGG16", "networks"], "Method"], [5391, 5392, ["training"], "Task"], [5419, 5424, ["Iterative", "bounding", "-", "box", "regression"], "Method"], [5431, 5442, ["iterative", "localization", "and", "bounding", "-", "box", "(", "bbox", ")", "voting", "scheme"], "Method"], [5480, 5483, ["post", "-", "processing"], "Task"], [5488, 5489, ["NMS"], "Method"], [5497, 5499, ["weighted", "voting"], "Method"], [5522, 5523, ["leaderboard"], "Method"], [5531, 5533, ["VOC", "2007"], "Material"], [5534, 5535, ["2012"], "Material"], [5554, 5555, ["VOC07"], "Material"], [5556, 5557, ["FRCN"], "Method"], [5565, 5566, ["mAP"], "Metric"], [5567, 5568, ["OHEM"], "Method"], [5599, 5600, ["OHEM"], "Method"], [5603, 5604, ["mAP"], "Metric"], [5615, 5618, ["MR", "-", "CNN"], "Method"], [5621, 5622, ["mAP"], "Metric"], [5627, 5630, ["MR", "-", "CNN"], "Method"], [5660, 5663, ["MR", "-", "CNN"], "Method"], [5665, 5666, ["VOC12"], "Material"], [5678, 5681, ["MR", "-", "CNN"], "Method"], [5696, 5697, ["mAP"], "Metric"], [5703, 5704, ["mAP"], "Metric"], [5705, 5708, ["MR", "-", "CNN"], "Method"], [5712, 5714, ["Ablation", "analysis"], "Method"], [5728, 5729, ["OHEM"], "Method"], [5744, 5747, ["Baseline", "FRCN", "mAP"], "Metric"], [5763, 5764, ["testing"], "Task"], [5790, 5793, ["Iterative", "bbox", "regression"], "Method"], [5810, 5811, ["OHEM"], "Method"], [5816, 5817, ["mAP"], "Metric"], [5831, 5833, ["VOC07", "data"], "Material"], [5845, 5846, ["OHEM"], "Method"], [5850, 5851, ["mAP"], "Metric"], [5866, 5869, ["Iterative", "bbox", "reg"], "Method"], [5875, 5877, ["MS", "COCO"], "Material"], [5878, 5884, ["MS", "COCO", "test", "-", "dev", "2015"], "Material"], [5908, 5909, ["AP"], "Metric"], [5912, 5914, ["COCO", "metric"], "Metric"], [5918, 5919, ["AP"], "Metric"], [5932, 5935, ["multi", "-", "scale"], "Method"], [5936, 5937, ["OHEM"], "Method"], [5959, 5960, ["AP"], "Metric"], [5964, 5965, ["AP"], "Metric"], [5990, 5994, ["online", "hard", "example", "mining"], "Method"], [5995, 5996, ["OHEM"], "Method"], [6006, 6011, ["region", "-", "based", "ConvNet", "detectors"], "Method"], [6012, 6013, ["OHEM"], "Method"], [6029, 6030, ["training"], "Task"], [6049, 6051, ["training", "convergence"], "Metric"], [6055, 6057, ["detection", "accuracy"], "Metric"], [6073, 6076, ["PASCAL", "VOC", "2007"], "Material"], [6077, 6078, ["2012"], "Material"], [6080, 6081, ["OHEM"], "Method"], [6083, 6085, ["orthogonal", "additions"], "Method"], [6089, 6093, ["Fast", "R", "-", "CNN"], "Method"], [6097, 6098, ["OHEM"], "Method"], [6104, 6109, ["region", "-", "based", "ConvNet", "detector"], "Method"], [6117, 6120, ["overall", "detection", "accuracy"], "Metric"], [6135, 6137, ["training", "methodologies"], "Method"], [60, 61, ["OHEM"], "Method"], [150, 151, ["2012"], "Material"], [170, 172, ["MS", "COCO"], "Material"], [207, 208, ["2012"], "Material"], [854, 855, ["OHEM"], "Method"], [982, 986, ["Fast", "R", "-", "CNN"], "Method"], [1047, 1049, ["MS", "COCO"], "Material"], [1068, 1071, ["multi", "-", "scale"], "Method"], [1106, 1107, ["2012"], "Material"], [1186, 1190, ["Fast", "R", "-", "CNN"], "Method"], [1842, 1843, ["FRCN"], "Method"], [2101, 2104, ["MR", "-", "CNN"], "Method"], [2364, 2366, ["PASCAL", "VOC"], "Material"], [2587, 2588, ["FRCN"], "Method"], [2648, 2652, ["Fast", "R", "-", "CNN"], "Method"], [2855, 2856, ["FRCN"], "Method"], [3393, 3394, ["FRCN"], "Method"], [3803, 3804, ["FRCN"], "Method"], [4355, 4356, ["FRCN"], "Method"], [4519, 4520, ["FRCN"], "Method"], [4576, 4577, ["VOC07"], "Material"], [4678, 4682, ["online", "hard", "example", "mining"], "Method"], [4825, 4826, ["VOC"], "Material"], [4826, 4827, ["2012"], "Material"], [4833, 4835, ["MS", "COCO"], "Material"], [4844, 4845, ["FRCN"], "Method"], [4850, 4851, ["OHEM"], "Method"], [4982, 4983, ["VOC"], "Material"], [4985, 4986, ["2012"], "Material"], [4999, 5000, ["mAP"], "Metric"], [5001, 5002, ["FRCN"], "Method"], [5158, 5159, ["FRCN"], "Method"], [5175, 5176, ["VOC"], "Material"], [5212, 5213, ["AP"], "Metric"], [5244, 5246, ["MS", "COCO"], "Material"], [5268, 5269, ["FRCN"], "Method"], [5329, 5332, ["multi", "-", "scale"], "Method"], [5541, 5542, ["VOC"], "Material"], [5648, 5651, ["multi", "-", "scale"], "Method"], [5756, 5759, ["multi", "-", "scale"], "Method"], [5799, 5800, ["FRCN"], "Method"], [5800, 5801, ["mAP"], "Metric"], [5896, 5899, ["multi", "-", "scale"], "Method"], [5921, 5922, ["VOC"], "Material"], [5947, 5949, ["MS", "COCO"], "Material"], [5970, 5972, ["MS", "COCO"], "Material"]]} {"doc_id": "2f97ee95cad6a1f13596b108072b846c6f747d4e", "text": ["document", ":", "The", "Microsoft", "2016", "Conversational", "Speech", "Recognition", "System", "We", "describe", "Microsoft", "’s", "conversational", "speech", "recognition", "system", ",", "in", "which", "we", "combine", "recent", "developments", "in", "neural", "-", "network", "-", "based", "acoustic", "and", "language", "modeling", "to", "advance", "the", "state", "of", "the", "art", "on", "the", "Switchboard", "recognition", "task", ".", "Inspired", "by", "machine", "learning", "ensemble", "techniques", ",", "the", "system", "uses", "a", "range", "of", "convolutional", "and", "recurrent", "neural", "networks", ".", "I", "-", "vector", "modeling", "and", "lattice", "-", "free", "MMI", "training", "provide", "significant", "gains", "for", "all", "acoustic", "model", "architectures", ".", "Language", "model", "rescoring", "with", "multiple", "forward", "and", "backward", "running", "RNNLMs", ",", "and", "word", "posterior", "-", "based", "system", "combination", "provide", "a", "20", "%", "boost", ".", "The", "best", "single", "system", "uses", "a", "ResNet", "architecture", "acoustic", "model", "with", "RNNLM", "rescoring", ",", "and", "achieves", "a", "word", "error", "rate", "of", "6.9", "%", "on", "the", "NIST", "2000", "Switchboard", "task", ".", "The", "combined", "system", "has", "an", "error", "rate", "of", "6.2", "%", ",", "representing", "an", "improvement", "over", "previously", "reported", "results", "on", "this", "benchmark", "task", ".", "W.Xiong", ",", "J.Droppo", ",", "X.Huang", ",", "F.Seide", ",", "M.Seltzer", ",", "A.Stolcke", ",", "D.YuandG.Zweig", "MicrosoftResearch", "Conversational", "speech", "recognition", ",", "convolutional", "neural", "networks", ",", "recurrent", "neural", "networks", ",", "VGG", ",", "ResNet", ",", "LACE", ",", "BLSTM", ".", "section", ":", "Introduction", "Recent", "years", "have", "seen", "a", "rapid", "reduction", "in", "speech", "recognition", "error", "rates", "as", "a", "result", "of", "careful", "engineering", "and", "optimization", "of", "convolutional", "and", "recurrent", "neural", "networks", ".", "While", "the", "basic", "structures", "have", "been", "well", "known", "for", "a", "long", "period", ",", "it", "is", "only", "recently", "that", "they", "have", "dominated", "the", "field", "as", "the", "best", "models", "for", "speech", "recognition", ".", "Surprisingly", ",", "this", "is", "the", "case", "for", "both", "acoustic", "modeling", "and", "language", "modeling", ".", "In", "comparison", "to", "standard", "feed", "-", "forward", "MLPs", "or", "DNNs", ",", "these", "acoustic", "models", "have", "the", "ability", "to", "model", "a", "large", "amount", "of", "acoustic", "context", "with", "temporal", "invariance", ",", "and", "in", "the", "case", "of", "convolutional", "models", ",", "with", "frequency", "invariance", "as", "well", ".", "In", "language", "modeling", ",", "recurrent", "models", "appear", "to", "improve", "over", "classical", "N", "-", "gram", "models", "through", "the", "generalization", "ability", "of", "continuous", "word", "representations", ".", "In", "the", "meantime", ",", "ensemble", "learning", "has", "become", "commonly", "used", "in", "several", "neural", "models", ",", "to", "improve", "robustness", "by", "reducing", "bias", "and", "variance", ".", "In", "this", "paper", ",", "we", "use", "ensembles", "of", "models", "extensively", ",", "as", "well", "as", "improvements", "to", "individual", "component", "models", ",", "to", "to", "advance", "the", "state", "-", "of", "-", "the", "-", "art", "in", "conversational", "telephone", "speech", "recognition", "(", "CTS", ")", ",", "which", "has", "been", "a", "benchmark", "speech", "recognition", "task", "since", "the", "1990s", ".", "The", "main", "features", "of", "this", "system", "are", ":", "An", "ensemble", "of", "two", "fundamental", "acoustic", "model", "architectures", ",", "convolutional", "neural", "nets", "(", "CNNs", ")", "and", "long", "-", "short", "-", "term", "memory", "nets", "(", "LSTMs", ")", ",", "with", "multiple", "variants", "of", "each", "An", "attention", "mechanism", "in", "the", "LACE", "CNN", "which", "differentially", "weights", "distant", "context", "Lattice", "-", "free", "MMI", "training", "The", "use", "of", "i", "-", "vector", "based", "adaptation", "in", "all", "models", "Language", "model", "(", "LM", ")", "rescoring", "with", "multiple", ",", "recurrent", "neural", "net", "LMs", "running", "in", "both", "forward", "and", "reverse", "direction", "Confusion", "network", "system", "combination", "coupled", "with", "search", "for", "best", "system", "subset", ",", "as", "necessitated", "by", "the", "large", "number", "of", "candidate", "systems", ".", "The", "remainder", "of", "this", "paper", "describes", "our", "system", "in", "detail", ".", "Section", "[", "reference", "]", "describes", "the", "CNN", "and", "LSTM", "models", ".", "Section", "[", "reference", "]", "describes", "our", "implementation", "of", "i", "-", "vector", "adaptation", ".", "Section", "[", "reference", "]", "presents", "out", "lattice", "-", "free", "MMI", "training", "process", ".", "Language", "model", "rescoring", "is", "a", "significant", "part", "of", "our", "system", ",", "and", "described", "in", "Section", "[", "reference", "]", ".", "Experimental", "results", "are", "presented", "in", "Section", "[", "reference", "]", ",", "followed", "by", "a", "discussion", "of", "related", "work", "and", "conclusions", ".", "section", ":", "Convolutional", "and", "LSTM", "Neural", "Networks", "We", "use", "three", "CNN", "variants", ".", "The", "first", "is", "the", "VGG", "architecture", "of", ".", "Compared", "to", "the", "networks", "used", "previously", "in", "image", "recognition", ",", "this", "network", "uses", "small", "(", "3x3", ")", "filters", ",", "is", "deeper", ",", "and", "applies", "up", "to", "five", "convolutional", "layers", "before", "pooling", ".", "The", "second", "network", "is", "modeled", "on", "the", "ResNet", "architecture", ",", "which", "adds", "highway", "connections", ",", "i.e.", "a", "linear", "transform", "of", "each", "layer", "’s", "input", "to", "the", "layer", "’s", "output", ".", "The", "only", "difference", "is", "that", "we", "move", "the", "Batch", "Normalization", "node", "to", "the", "place", "right", "before", "each", "ReLU", "activation", ".", "The", "last", "CNN", "variant", "is", "the", "LACE", "(", "layer", "-", "wise", "context", "expansion", "with", "attention", ")", "model", ".", "LACE", "is", "a", "TDNN", "variant", "in", "which", "each", "higher", "layer", "is", "a", "weighted", "sum", "of", "nonlinear", "transformations", "of", "a", "window", "of", "lower", "layer", "frames", ".", "In", "other", "words", ",", "each", "higher", "layer", "exploits", "broader", "context", "than", "lower", "layers", ".", "Lower", "layers", "focus", "on", "extracting", "simple", "local", "patterns", "while", "higher", "layers", "extract", "complex", "patterns", "that", "cover", "broader", "contexts", ".", "Since", "not", "all", "frames", "in", "a", "window", "carry", "the", "same", "importance", ",", "an", "attention", "mask", "is", "applied", ".", "The", "LACE", "model", "differs", "from", "the", "earlier", "TDNN", "models", "e.g.", "in", "the", "use", "of", "a", "learned", "attention", "mask", "and", "ResNet", "like", "linear", "pass", "-", "through", ".", "As", "illustrated", "in", "detail", "in", "Figure", "[", "reference", "]", ",", "the", "model", "is", "composed", "of", "4", "blocks", ",", "each", "with", "the", "same", "architecture", ".", "Each", "block", "starts", "with", "a", "convolution", "layer", "with", "stride", "2", "which", "sub", "-", "samples", "the", "input", "and", "increases", "the", "number", "of", "channels", ".", "This", "layer", "is", "followed", "by", "4", "RELU", "-", "convolution", "layers", "with", "jump", "links", "similar", "to", "those", "used", "in", "ResNet", ".", "Table", "[", "reference", "]", "compares", "the", "layer", "structure", "and", "parameters", "of", "the", "three", "CNN", "architectures", ".", "While", "our", "best", "performing", "models", "are", "convolutional", ",", "the", "use", "of", "long", "short", "-", "term", "memory", "networks", "is", "a", "close", "second", ".", "We", "use", "a", "bidirectional", "architecture", "without", "frame", "-", "skipping", ".", "The", "core", "model", "structure", "is", "the", "LSTM", "defined", "in", ".", "We", "found", "that", "using", "networks", "with", "more", "than", "six", "layers", "did", "not", "improve", "the", "word", "error", "rate", "on", "the", "development", "set", ",", "and", "chose", "512", "hidden", "units", ",", "per", "direction", ",", "per", "layer", ",", "as", "that", "provided", "a", "reasonable", "trade", "-", "off", "between", "training", "time", "and", "final", "model", "accuracy", ".", "Network", "parameters", "for", "different", "configurations", "of", "the", "LSTM", "architecture", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "section", ":", "Speaker", "Adaptive", "Modeling", "Speaker", "adaptive", "modeling", "in", "our", "system", "is", "based", "on", "conditioning", "the", "network", "on", "an", "i", "-", "vector", "characterization", "of", "each", "speaker", ".", "A", "100", "-", "dimensional", "i", "-", "vector", "is", "generated", "for", "each", "conversation", "side", ".", "For", "the", "LSTM", "system", ",", "the", "conversation", "-", "side", "i", "-", "vector", "is", "appended", "to", "each", "frame", "of", "input", ".", "For", "convolutional", "networks", ",", "this", "approach", "is", "inappropriate", "because", "we", "do", "not", "expect", "to", "see", "spatially", "contiguous", "patterns", "in", "the", "input", ".", "Instead", ",", "for", "the", "CNNs", ",", "we", "add", "a", "learnable", "weight", "matrix", "to", "each", "layer", ",", "and", "add", "to", "the", "activation", "of", "the", "layer", "before", "the", "nonlinearity", ".", "Thus", ",", "in", "the", "CNN", ",", "the", "i", "-", "vector", "essentially", "serves", "as", "an", "additional", "bias", "to", "each", "layer", ".", "Note", "that", "the", "i", "-", "vectors", "are", "estimated", "using", "MFCC", "features", ";", "by", "using", "them", "subsequently", "in", "systems", "based", "on", "log", "-", "filterbank", "features", ",", "we", "may", "benefit", "from", "a", "form", "of", "feature", "combination", ".", "section", ":", "Lattice", "-", "Free", "Sequence", "Training", "After", "standard", "cross", "-", "entropy", "training", ",", "we", "optimize", "the", "model", "parameters", "using", "the", "maximum", "mutual", "information", "(", "MMI", ")", "objective", "function", ".", "Denoting", "a", "word", "sequence", "by", "and", "its", "corresponding", "acoustic", "realization", "by", ",", "the", "training", "criterion", "is", "As", "noted", "in", ",", "the", "necessary", "gradient", "for", "use", "in", "backpropagation", "is", "a", "simple", "function", "of", "the", "posterior", "probability", "of", "a", "particular", "acoustic", "model", "state", "at", "a", "given", "time", ",", "as", "computed", "by", "summing", "over", "all", "possible", "word", "sequences", "in", "an", "unconstrained", "manner", ".", "As", "first", "done", "in", ",", "and", "more", "recently", "in", ",", "this", "can", "be", "accomplished", "with", "a", "straightforward", "alpha", "-", "beta", "computation", "over", "the", "finite", "state", "acceptor", "representing", "the", "decoding", "search", "space", ".", "In", ",", "the", "search", "space", "is", "taken", "to", "be", "an", "acceptor", "representing", "the", "composition", "for", "a", "unigram", "language", "model", "on", "words", ".", "In", ",", "a", "language", "model", "on", "phonemes", "is", "used", "instead", ".", "In", "our", "implementation", ",", "we", "use", "a", "mixed", "-", "history", "acoustic", "unit", "language", "model", ".", "In", "this", "model", ",", "the", "probability", "of", "transitioning", "into", "a", "new", "context", "-", "dependent", "phonetic", "state", "(", "senone", ")", "is", "conditioned", "both", "the", "senone", "and", "phone", "history", ".", "We", "found", "this", "model", "to", "perform", "better", "than", "either", "purely", "word", "-", "based", "or", "phone", "-", "based", "models", ".", "Based", "on", "a", "set", "of", "initial", "experiments", ",", "we", "developed", "the", "following", "procedure", ":", "Perform", "a", "forced", "alignment", "of", "the", "training", "data", "to", "select", "lexical", "variants", "and", "determine", "frame", "-", "aligned", "senone", "sequences", ".", "Compress", "consecutive", "framewise", "occurrences", "of", "a", "single", "senone", "into", "a", "single", "occurrence", ".", "Estimate", "an", "unsmoothed", ",", "variable", "-", "length", "N", "-", "gram", "language", "model", "from", "this", "data", ",", "where", "the", "history", "state", "consists", "of", "the", "previous", "phone", "and", "previous", "senones", "within", "the", "current", "phone", ".", "To", "illustrate", "this", ",", "consider", "the", "sample", "senone", "sequence", "{", "s", "_", "s2.1288", ",", "s", "_", "s3.1061", ",", "s", "_", "s4.1096", "}", ",", "{", "eh", "_", "s2.527", ",", "eh", "_", "s3.128", ",", "eh", "_", "s4.66", "}", ",", "{", "t", "_", "s2.729", ",", "t", "_", "s3.572", ",", "t", "_", "s4.748}.", "When", "predicting", "the", "state", "following", "eh", "_", "s4.66", "the", "history", "consists", "of", "(", "s", ",", "eh", "_", "s2.527", ",", "eh", "_", "s3.128", ",", "eh", "_", "s4.66", ")", ",", "and", "following", "t", "_", "s2.729", ",", "the", "history", "is", "(", "eh", ",", "t", "_", "s2.729", ")", ".", "We", "construct", "the", "denominator", "graph", "from", "this", "language", "model", ",", "and", "HMM", "transition", "probabilities", "as", "determined", "by", "transition", "-", "counting", "in", "the", "senone", "sequences", "found", "in", "the", "training", "data", ".", "Our", "approach", "not", "only", "largely", "reduces", "the", "complexity", "of", "building", "up", "the", "language", "model", "but", "also", "provides", "very", "reliable", "training", "performance", ".", "We", "have", "found", "it", "convenient", "to", "do", "the", "full", "computation", ",", "without", "pruning", ",", "in", "a", "series", "of", "matrix", "-", "vector", "operations", "on", "the", "GPU", ".", "The", "underlying", "acceptor", "is", "represented", "with", "a", "sparse", "matrix", ",", "and", "we", "maintain", "a", "dense", "likelihood", "vector", "for", "each", "time", "frame", ".", "The", "alpha", "and", "beta", "recursions", "are", "implemented", "with", "CUSPARSE", "level", "-", "2", "routines", ":", "sparse", "-", "matrix", ",", "dense", "vector", "multiplies", ".", "Run", "time", "is", "about", "100", "times", "faster", "than", "real", "time", ".", "As", "in", ",", "we", "use", "cross", "-", "entropy", "regularization", ".", "In", "all", "the", "lattice", "-", "free", "MMI", "(", "LFMMI", ")", "experiments", "mentioned", "below", "we", "use", "a", "trigram", "language", "model", ".", "Most", "of", "the", "gain", "is", "usually", "obtained", "after", "processing", "24", "to", "48", "hours", "of", "data", ".", "section", ":", "LM", "Rescoring", "and", "System", "Combination", "An", "initial", "decoding", "is", "done", "with", "a", "WFST", "decoder", ",", "using", "the", "architecture", "described", "in", ".", "We", "use", "an", "N", "-", "gram", "language", "model", "trained", "and", "pruned", "with", "the", "SRILM", "toolkit", ".", "The", "first", "-", "pass", "LM", "has", "approximately", "15.9", "million", "bigrams", ",", "trigrams", ",", "and", "4grams", ",", "and", "a", "vocabulary", "of", "30", ",", "500", "words", ".", "It", "gives", "a", "perplexity", "of", "69", "on", "the", "1997", "CTS", "evaluation", "transcripts", ".", "The", "initial", "decoding", "produces", "a", "lattice", "with", "the", "pronunciation", "variants", "marked", ",", "from", "which", "500", "-", "best", "lists", "are", "generated", "for", "rescoring", "purposes", ".", "Subsequent", "N", "-", "best", "rescoring", "uses", "an", "unpruned", "LM", "comprising", "145", "million", "N", "-", "grams", ".", "All", "N", "-", "gram", "LMs", "were", "estimated", "by", "a", "maximum", "entropy", "criterion", "as", "described", "in", ".", "subsection", ":", "RNNLM", "setup", "The", "N", "-", "best", "hypotheses", "are", "then", "rescored", "using", "a", "combination", "of", "the", "large", "N", "-", "gram", "LM", "and", "several", "RNNLMs", ",", "trained", "and", "evaluated", "using", "the", "CUED", "-", "RNNLM", "toolkit", ".", "Our", "RNNLM", "configuration", "has", "several", "distinctive", "features", ",", "as", "described", "below", ".", "1", ")", "We", "trained", "both", "standard", ",", "forward", "-", "predicting", "RNNLMs", "and", "backward", "RNNLMs", "that", "predict", "words", "in", "reverse", "temporal", "order", ".", "The", "log", "probabilities", "from", "both", "models", "are", "added", ".", "2", ")", "As", "is", "customary", ",", "the", "RNNLM", "probability", "estimates", "are", "interpolated", "at", "the", "word", "-", "level", "with", "corresponding", "N", "-", "gram", "LM", "probabilities", "(", "separately", "for", "the", "forward", "and", "backward", "models", ")", ".", "In", "addition", ",", "we", "trained", "a", "second", "RNNLM", "for", "each", "direction", ",", "obtained", "by", "starting", "with", "different", "random", "initial", "weights", ".", "The", "two", "RNNLMs", "and", "the", "N", "-", "gram", "LM", "for", "each", "direction", "are", "interpolated", "with", "weights", "of", "(", "0.375", ",", "0.375", ",", "0.25", ")", ".", "3", ")", "In", "order", "to", "make", "use", "of", "LM", "training", "data", "that", "is", "not", "fully", "matched", "to", "the", "target", "conversational", "speech", "domain", ",", "we", "start", "RNNLM", "training", "with", "the", "union", "of", "in", "-", "domain", "(", "here", ",", "CTS", ")", "and", "out", "-", "of", "-", "domain", "(", "e.g.", ",", "Web", ")", "data", ".", "Upon", "convergence", ",", "the", "network", "undergoes", "a", "second", "training", "phase", "using", "the", "in", "-", "domain", "data", "only", ".", "Both", "training", "phases", "use", "in", "-", "domain", "validation", "data", "to", "regulate", "the", "learning", "rate", "schedule", "and", "termination", ".", "Because", "the", "size", "of", "the", "out", "-", "of", "-", "domain", "data", "is", "a", "multiple", "of", "the", "in", "-", "domain", "data", ",", "a", "standard", "training", "on", "a", "simple", "union", "of", "the", "data", "would", "not", "yield", "a", "well", "-", "matched", "model", ",", "and", "have", "poor", "perplexity", "in", "the", "target", "domain", ".", "4", ")", "We", "found", "best", "results", "with", "an", "RNNLM", "configuration", "that", "had", "a", "second", ",", "non", "-", "recurrent", "hidden", "layer", ".", "This", "produced", "lower", "perplexity", "and", "word", "error", "than", "the", "standard", ",", "single", "-", "hidden", "-", "layer", "RNNLM", "architecture", ".", "The", "overall", "network", "architecture", "thus", "had", "two", "hidden", "layers", "with", "1000", "units", "each", ",", "using", "ReLU", "nonlinearities", ".", "Training", "used", "noise", "-", "contrastive", "estimation", "(", "NCE", ")", ".", "5", ")", "The", "RNNLM", "output", "vocabulary", "consists", "of", "all", "words", "occurring", "more", "than", "once", "in", "the", "in", "-", "domain", "training", "set", ".", "While", "the", "RNNLM", "estimates", "a", "probability", "for", "unknown", "words", ",", "we", "take", "a", "different", "approach", "in", "rescoring", ":", "The", "number", "of", "out", "-", "of", "-", "set", "words", "is", "recorded", "for", "each", "hypothesis", "and", "a", "penalty", "for", "them", "is", "estimated", "for", "them", "when", "optimizing", "the", "relative", "weights", "for", "all", "model", "scores", "(", "acoustic", ",", "LM", ",", "pronunciation", ")", ",", "using", "the", "SRILM", "nbest", "-", "optimize", "tool", ".", "subsection", ":", "Training", "data", "The", "4", "-", "gram", "language", "model", "for", "decoding", "was", "trained", "on", "the", "available", "CTS", "transcripts", "from", "the", "DARPA", "EARS", "program", ":", "Switchboard", "(", "3", "M", "words", ")", ",", "BBN", "Switchboard", "-", "2", "transcripts", "(", "850k", ")", ",", "Fisher", "(", "21", "M", ")", ",", "English", "CallHome", "(", "200k", ")", ",", "and", "the", "University", "of", "Washington", "conversational", "Web", "corpus", "(", "191", "M", ")", ".", "A", "separate", "model", "was", "trained", "from", "each", "source", "and", "interpolated", "with", "weights", "optimized", "on", "RT", "-", "03", "transcripts", ".", "For", "the", "unpruned", "large", "rescoring", "4", "-", "gram", ",", "an", "additional", "LM", "component", "was", "added", ",", "trained", "on", "133", "M", "word", "of", "LDC", "Broadcast", "News", "texts", ".", "The", "N", "-", "gram", "LM", "configuration", "is", "modeled", "after", "that", "described", "in", ",", "except", "that", "maxent", "smoothing", "was", "used", ".", "The", "RNNLMs", "were", "trained", "on", "Switchboard", "and", "Fisher", "transcripts", "as", "in", "-", "domain", "data", "(", "20", "M", "words", "for", "gradient", "computation", ",", "3", "M", "for", "validation", ")", ".", "To", "this", "we", "added", "62", "M", "words", "of", "UW", "Web", "data", "as", "out", "-", "of", "-", "domain", "data", ",", "for", "use", "in", "the", "two", "-", "phase", "training", "procedure", "described", "above", ".", "subsection", ":", "RNNLM", "performance", "Table", "[", "reference", "]", "gives", "perplexity", "and", "word", "error", "performance", "for", "various", "RNNLM", "setups", ",", "from", "simple", "to", "more", "complex", ".", "The", "acoustic", "model", "used", "was", "the", "ResNet", "CNN", ".", "As", "can", "be", "seen", ",", "each", "of", "the", "measures", "described", "earlier", "adds", "incremental", "gains", ",", "which", ",", "while", "small", "individually", ",", "add", "up", "to", "a", "9", "%", "relative", "error", "reduction", "over", "a", "plain", "RNNLM", ".", "subsection", ":", "System", "Combination", "The", "LM", "rescoring", "is", "carried", "out", "separately", "for", "each", "acoustic", "model", ".", "The", "rescored", "N", "-", "best", "lists", "from", "each", "subsystem", "are", "then", "aligned", "into", "a", "single", "confusion", "network", "using", "the", "SRILM", "nbest", "-", "rover", "tool", ".", "However", ",", "the", "number", "of", "potential", "candidate", "systems", "is", "too", "large", "to", "allow", "an", "all", "-", "out", "combination", ",", "both", "for", "practical", "reasons", "and", "due", "to", "overfitting", "issues", ".", "Instead", ",", "we", "perform", "a", "greedy", "search", ",", "starting", "with", "the", "single", "best", "system", ",", "and", "successively", "adding", "additional", "systems", ",", "to", "find", "a", "small", "set", "of", "systems", "that", "are", "maximally", "complementary", ".", "The", "RT", "-", "02", "Switchboard", "set", "was", "used", "for", "this", "search", "procedure", ".", "The", "relative", "weighting", "(", "for", "confusion", "-", "network", "mediated", "voting", ")", "of", "the", "different", "systems", "is", "optimized", "using", "an", "EM", "algorithm", ",", "using", "the", "same", "data", ",", "and", "smoothed", "hierarchically", "by", "interpolating", "each", "set", "of", "system", "weights", "with", "the", "preceding", "one", "in", "the", "search", ".", "section", ":", "Experimental", "Setup", "and", "Results", "subsection", ":", "Speech", "corpora", "We", "train", "with", "the", "commonly", "used", "English", "CTS", "(", "Switchboard", "and", "Fisher", ")", "corpora", ".", "Evaluation", "is", "carried", "out", "on", "the", "NIST", "2000", "CTS", "test", "set", ",", "which", "comprises", "both", "Switchboard", "(", "SWB", ")", "and", "CallHome", "(", "CH", ")", "subsets", ".", "The", "Switchboard", "-", "1", "portion", "of", "the", "NIST", "2002", "CTS", "test", "set", "was", "used", "for", "tuning", "and", "development", ".", "The", "acoustic", "training", "data", "is", "comprised", "by", "LDC", "corpora", "97S62", ",", "2004S13", ",", "2005S13", ",", "2004S11", "and", "2004S09", ";", "see", "for", "a", "full", "description", ".", "subsection", ":", "1", "-", "bit", "SGD", "Training", "All", "presented", "models", "are", "costly", "to", "train", ".", "To", "make", "training", "feasible", ",", "we", "parallelize", "training", "with", "our", "previously", "proposed", "1", "-", "bit", "SGD", "parallelization", "technique", ".", "This", "data", "-", "parallel", "method", "distributes", "minibatches", "over", "multiple", "worker", "nodes", ",", "and", "then", "aggregates", "the", "sub", "-", "gradients", ".", "While", "the", "necessary", "communication", "time", "would", "otherwise", "be", "prohibitive", ",", "the", "1", "-", "bit", "SGD", "method", "eliminates", "the", "bottleneck", "by", "two", "techniques", ":", "1", "-", "bit", "quantization", "of", "gradients", "and", "automatic", "minibatch", "-", "size", "scaling", ".", "In", ",", "we", "showed", "that", "gradient", "values", "can", "be", "quantized", "to", "just", "a", "single", "bit", ",", "if", "one", "carries", "over", "the", "quantization", "error", "from", "one", "minibatch", "to", "the", "next", ".", "Each", "time", "a", "sub", "-", "gradient", "is", "quantized", ",", "the", "quantization", "error", "is", "computed", "and", "remembered", ",", "and", "then", "added", "to", "the", "next", "minibatch", "’s", "sub", "-", "gradient", ".", "This", "reduces", "the", "required", "bandwidth", "32", "-", "fold", "with", "minimal", "loss", "in", "accuracy", ".", "Secondly", ",", "automatic", "minibatch", "-", "size", "scaling", "progressively", "decreases", "the", "frequency", "of", "model", "updates", ".", "At", "regular", "intervals", "(", "e.g.", "every", "72h", "of", "training", "data", ")", ",", "the", "trainer", "tries", "larger", "minibatch", "sizes", "on", "a", "small", "subset", "of", "data", "and", "picks", "the", "largest", "that", "maintains", "training", "loss", ".", "subsection", ":", "Acoustic", "Model", "Details", "Forty", "-", "dimensional", "log", "-", "filterbank", "features", "were", "extracted", "every", "10", "milliseconds", ",", "using", "a", "25", "-", "millisecond", "analysis", "window", ".", "The", "CNN", "models", "used", "window", "sizes", "as", "indicated", "in", "Table", "[", "reference", "]", ",", "and", "the", "LSTMs", "processed", "one", "frame", "of", "input", "at", "a", "time", ".", "The", "bulk", "of", "our", "models", "use", "three", "state", "left", "-", "to", "-", "right", "triphone", "models", "with", "9000", "tied", "states", ".", "Additionally", ",", "we", "have", "trained", "several", "models", "with", "27k", "tied", "states", ".", "The", "phonetic", "inventory", "includes", "special", "models", "for", "noise", ",", "vocalized", "-", "noise", ",", "laughter", "and", "silence", ".", "We", "use", "a", "30k", "-", "vocabulary", "derived", "from", "the", "most", "common", "words", "in", "the", "Switchboard", "and", "Fisher", "corpora", ".", "The", "decoder", "uses", "a", "statically", "compiled", "unigram", "graph", ",", "and", "dynamically", "applies", "the", "language", "model", "score", ".", "The", "unigram", "graph", "has", "about", "300k", "states", "and", "500k", "arcs", ".", "All", "acoustic", "models", "were", "trained", "using", "the", "open", "-", "source", "Computational", "Network", "Toolkit", "(", "CNTK", ")", ".", "Table", "[", "reference", "]", "shows", "the", "result", "of", "i", "-", "vector", "adaptation", "and", "LFMMI", "training", "on", "several", "of", "our", "systems", ".", "We", "achieve", "a", "5–8", "%", "relative", "improvement", "from", "i", "-", "vectors", ",", "including", "on", "CNN", "systems", ".", "The", "last", "row", "of", "Table", "[", "reference", "]", "shows", "the", "effect", "of", "LFMMI", "training", "on", "the", "different", "models", ".", "We", "see", "a", "consistent", "7–10", "%", "further", "relative", "reduction", "in", "error", "rate", "for", "all", "models", ".", "Considering", "the", "great", "increase", "in", "procedural", "simplicity", "of", "LFMMI", "over", "the", "previous", "practice", "of", "writing", "lattices", "and", "post", "-", "processing", "them", ",", "we", "consider", "LFMMI", "to", "be", "a", "significant", "advance", "in", "technology", ".", "subsection", ":", "Comparative", "System", "Performance", "Model", "performance", "for", "our", "individual", "models", "as", "well", "as", "relevant", "comparisons", "from", "the", "literature", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Out", "of", "the", "15", "models", "built", ",", "only", "models", "given", "non", "-", "zero", "weight", "in", "the", "final", "system", "combination", "are", "shown", ".", "section", ":", "Relation", "to", "Prior", "Work", "Compared", "to", "earlier", "applications", "of", "CNNs", "to", "speech", "recognition", ",", "our", "networks", "are", "much", "deeper", ",", "and", "use", "linear", "bypass", "connections", "across", "convolutional", "layers", ".", "They", "are", "similar", "in", "spirit", "to", "those", "studied", "more", "recently", "by", ".", "We", "improve", "on", "these", "architectures", "with", "the", "LACE", "model", ",", "which", "iteratively", "expands", "the", "effective", "window", "size", ",", "layer", "-", "by", "-", "layer", ",", "and", "adds", "an", "attention", "mask", "to", "differentially", "weight", "distant", "context", ".", "Our", "use", "of", "lattice", "-", "free", "MMI", "is", "distinctive", ",", "and", "extends", "previous", "work", "by", "proposing", "the", "use", "of", "a", "mixed", "triphone", "/", "phoneme", "history", "in", "the", "language", "model", ".", "On", "the", "language", "modeling", "side", ",", "we", "achieve", "a", "performance", "boost", "by", "combining", "multiple", "RNNLMs", "in", "both", "forward", "and", "backward", "directions", ",", "and", "by", "using", "a", "two", "-", "phase", "training", "regimen", "to", "get", "best", "results", "from", "out", "-", "of", "-", "domain", "data", ".", "For", "our", "best", "CNN", "system", ",", "RNNLM", "rescoring", "yields", "a", "relative", "word", "error", "reduction", "of", "20", "%", ",", "and", "a", "16", "%", "relative", "gain", "for", "the", "combined", "recognition", "system", ".", "(", "Elsewhere", "we", "report", "further", "improvements", ",", "using", "LSTM", "-", "based", "LMs", ".", ")", "section", ":", "Conclusions", "We", "have", "described", "Microsoft", "’s", "conversational", "speech", "recognition", "system", "for", "2016", ".", "The", "use", "of", "CNNs", "in", "the", "acoustic", "model", "has", "proved", "singularly", "effective", ",", "as", "has", "the", "use", "of", "RNN", "language", "models", ".", "Our", "best", "single", "system", "achieves", "an", "error", "rate", "of", "6.9", "%", "on", "the", "NIST", "2000", "Switchboard", "set", ".", "We", "believe", "this", "is", "the", "best", "performance", "reported", "to", "date", "for", "a", "recognition", "system", "not", "based", "on", "system", "combination", ".", "An", "ensemble", "of", "acoustic", "models", "advances", "the", "state", "of", "the", "art", "to", "6.2", "%", "on", "the", "Switchboard", "test", "data", ".", "Acknowledgments", ".", "We", "thank", "X.", "Chen", "from", "CUED", "for", "valuable", "assistance", "with", "the", "CUED", "-", "RNNLM", "toolkit", ",", "and", "ICSI", "for", "compute", "and", "data", "resources", ".", "bibliography", ":", "References"], "entity": [[5, 9, ["Conversational", "Speech", "Recognition", "System"], "Method"], [11, 17, ["Microsoft", "’s", "conversational", "speech", "recognition", "system"], "Method"], [25, 34, ["neural", "-", "network", "-", "based", "acoustic", "and", "language", "modeling"], "Method"], [49, 53, ["machine", "learning", "ensemble", "techniques"], "Method"], [60, 65, ["convolutional", "and", "recurrent", "neural", "networks"], "Method"], [66, 70, ["I", "-", "vector", "modeling"], "Method"], [71, 76, ["lattice", "-", "free", "MMI", "training"], "Method"], [81, 84, ["acoustic", "model", "architectures"], "Method"], [85, 88, ["Language", "model", "rescoring"], "Method"], [90, 95, ["forward", "and", "backward", "running", "RNNLMs"], "Method"], [97, 103, ["word", "posterior", "-", "based", "system", "combination"], "Method"], [115, 119, ["ResNet", "architecture", "acoustic", "model"], "Method"], [120, 122, ["RNNLM", "rescoring"], "Method"], [126, 129, ["word", "error", "rate"], "Metric"], [144, 146, ["error", "rate"], "Metric"], [175, 179, ["MicrosoftResearch", "Conversational", "speech", "recognition"], "Task"], [180, 183, ["convolutional", "neural", "networks"], "Method"], [184, 187, ["recurrent", "neural", "networks"], "Method"], [188, 189, ["VGG"], "Method"], [190, 191, ["ResNet"], "Method"], [192, 193, ["LACE"], "Method"], [194, 195, ["BLSTM"], "Method"], [207, 211, ["speech", "recognition", "error", "rates"], "Metric"], [218, 225, ["optimization", "of", "convolutional", "and", "recurrent", "neural", "networks"], "Method"], [254, 256, ["speech", "recognition"], "Task"], [265, 267, ["acoustic", "modeling"], "Task"], [268, 270, ["language", "modeling"], "Task"], [275, 279, ["feed", "-", "forward", "MLPs"], "Method"], [280, 281, ["DNNs"], "Method"], [283, 285, ["acoustic", "models"], "Method"], [305, 307, ["convolutional", "models"], "Method"], [315, 317, ["language", "modeling"], "Task"], [318, 320, ["recurrent", "models"], "Method"], [325, 329, ["N", "-", "gram", "models"], "Method"], [331, 333, ["generalization", "ability"], "Method"], [334, 337, ["continuous", "word", "representations"], "Method"], [342, 344, ["ensemble", "learning"], "Method"], [350, 352, ["neural", "models"], "Method"], [355, 356, ["robustness"], "Metric"], [358, 361, ["bias", "and", "variance"], "Metric"], [368, 371, ["ensembles", "of", "models"], "Method"], [379, 381, ["component", "models"], "Method"], [394, 398, ["conversational", "telephone", "speech", "recognition"], "Task"], [399, 401, ["CTS", ")"], "Task"], [407, 410, ["speech", "recognition", "task"], "Task"], [427, 430, ["acoustic", "model", "architectures"], "Method"], [431, 434, ["convolutional", "neural", "nets"], "Method"], [435, 436, ["CNNs"], "Method"], [438, 445, ["long", "-", "short", "-", "term", "memory", "nets"], "Method"], [446, 447, ["LSTMs"], "Method"], [455, 457, ["attention", "mechanism"], "Method"], [459, 461, ["LACE", "CNN"], "Method"], [464, 471, ["distant", "context", "Lattice", "-", "free", "MMI", "training"], "Method"], [474, 479, ["i", "-", "vector", "based", "adaptation"], "Method"], [482, 488, ["Language", "model", "(", "LM", ")", "rescoring"], "Method"], [491, 495, ["recurrent", "neural", "net", "LMs"], "Method"], [502, 505, ["Confusion", "network", "system"], "Method"], [541, 545, ["CNN", "and", "LSTM", "models"], "Method"], [554, 558, ["i", "-", "vector", "adaptation"], "Method"], [565, 571, ["lattice", "-", "free", "MMI", "training", "process"], "Method"], [572, 575, ["Language", "model", "rescoring"], "Method"], [613, 618, ["Convolutional", "and", "LSTM", "Neural", "Networks"], "Method"], [621, 623, ["CNN", "variants"], "Method"], [628, 630, ["VGG", "architecture"], "Method"], [639, 641, ["image", "recognition"], "Task"], [659, 661, ["convolutional", "layers"], "Method"], [662, 663, ["pooling"], "Method"], [671, 673, ["ResNet", "architecture"], "Method"], [716, 718, ["CNN", "variant"], "Method"], [720, 731, ["LACE", "(", "layer", "-", "wise", "context", "expansion", "with", "attention", ")", "model"], "Method"], [732, 733, ["LACE"], "Method"], [735, 737, ["TDNN", "variant"], "Method"], [744, 749, ["weighted", "sum", "of", "nonlinear", "transformations"], "Method"], [771, 773, ["Lower", "layers"], "Method"], [809, 811, ["LACE", "model"], "Method"], [815, 817, ["TDNN", "models"], "Method"], [863, 865, ["convolution", "layer"], "Method"], [887, 891, ["RELU", "-", "convolution", "layers"], "Method"], [899, 900, ["ResNet"], "Method"], [914, 916, ["CNN", "architectures"], "Method"], [923, 924, ["convolutional"], "Method"], [928, 934, ["long", "short", "-", "term", "memory", "networks"], "Method"], [942, 944, ["bidirectional", "architecture"], "Method"], [945, 948, ["frame", "-", "skipping"], "Method"], [950, 953, ["core", "model", "structure"], "Method"], [955, 956, ["LSTM"], "Method"], [963, 964, ["networks"], "Method"], [973, 976, ["word", "error", "rate"], "Metric"], [1002, 1004, ["training", "time"], "Metric"], [1006, 1008, ["model", "accuracy"], "Metric"], [1016, 1018, ["LSTM", "architecture"], "Method"], [1028, 1031, ["Speaker", "Adaptive", "Modeling"], "Method"], [1031, 1034, ["Speaker", "adaptive", "modeling"], "Method"], [1045, 1049, ["i", "-", "vector", "characterization"], "Method"], [1069, 1071, ["LSTM", "system"], "Method"], [1088, 1090, ["convolutional", "networks"], "Method"], [1113, 1114, ["CNNs"], "Method"], [1141, 1142, ["CNN"], "Method"], [1189, 1191, ["feature", "combination"], "Method"], [1194, 1199, ["Lattice", "-", "Free", "Sequence", "Training"], "Method"], [1201, 1205, ["cross", "-", "entropy", "training"], "Method"], [1213, 1221, ["maximum", "mutual", "information", "(", "MMI", ")", "objective", "function"], "Metric"], [1235, 1237, ["training", "criterion"], "Metric"], [1248, 1249, ["backpropagation"], "Method"], [1299, 1303, ["alpha", "-", "beta", "computation"], "Method"], [1305, 1308, ["finite", "state", "acceptor"], "Method"], [1330, 1333, ["unigram", "language", "model"], "Method"], [1339, 1341, ["language", "model"], "Method"], [1342, 1343, ["phonemes"], "Method"], [1354, 1361, ["mixed", "-", "history", "acoustic", "unit", "language", "model"], "Method"], [1400, 1408, ["word", "-", "based", "or", "phone", "-", "based", "models"], "Method"], [1425, 1427, ["forced", "alignment"], "Task"], [1458, 1468, ["unsmoothed", ",", "variable", "-", "length", "N", "-", "gram", "language", "model"], "Method"], [1590, 1592, ["language", "model"], "Method"], [1600, 1603, ["transition", "-", "counting"], "Method"], [1620, 1621, ["complexity"], "Metric"], [1625, 1627, ["language", "model"], "Method"], [1653, 1657, ["matrix", "-", "vector", "operations"], "Method"], [1684, 1688, ["alpha", "and", "beta", "recursions"], "Method"], [1691, 1696, ["CUSPARSE", "level", "-", "2", "routines"], "Method"], [1697, 1700, ["sparse", "-", "matrix"], "Method"], [1701, 1704, ["dense", "vector", "multiplies"], "Method"], [1705, 1707, ["Run", "time"], "Metric"], [1721, 1725, ["cross", "-", "entropy", "regularization"], "Method"], [1729, 1735, ["lattice", "-", "free", "MMI", "(", "LFMMI"], "Method"], [1742, 1745, ["trigram", "language", "model"], "Method"], [1764, 1766, ["LM", "Rescoring"], "Task"], [1767, 1769, ["System", "Combination"], "Task"], [1771, 1772, ["decoding"], "Task"], [1776, 1778, ["WFST", "decoder"], "Method"], [1788, 1793, ["N", "-", "gram", "language", "model"], "Method"], [1798, 1800, ["SRILM", "toolkit"], "Method"], [1802, 1806, ["first", "-", "pass", "LM"], "Method"], [1835, 1838, ["CTS", "evaluation", "transcripts"], "Material"], [1841, 1842, ["decoding"], "Method"], [1860, 1862, ["rescoring", "purposes"], "Task"], [1864, 1868, ["N", "-", "best", "rescoring"], "Method"], [1870, 1872, ["unpruned", "LM"], "Method"], [1880, 1884, ["N", "-", "gram", "LMs"], "Method"], [1888, 1891, ["maximum", "entropy", "criterion"], "Metric"], [1897, 1899, ["RNNLM", "setup"], "Method"], [1912, 1917, ["large", "N", "-", "gram", "LM"], "Method"], [1919, 1920, ["RNNLMs"], "Method"], [1926, 1930, ["CUED", "-", "RNNLM", "toolkit"], "Method"], [1932, 1934, ["RNNLM", "configuration"], "Method"], [1950, 1954, ["forward", "-", "predicting", "RNNLMs"], "Method"], [1955, 1957, ["backward", "RNNLMs"], "Method"], [1981, 1984, ["RNNLM", "probability", "estimates"], "Method"], [2002, 2006, ["forward", "and", "backward", "models"], "Method"], [2015, 2016, ["RNNLM"], "Method"], [2031, 2032, ["RNNLMs"], "Method"], [2034, 2038, ["N", "-", "gram", "LM"], "Method"], [2062, 2065, ["LM", "training", "data"], "Material"], [2079, 2081, ["RNNLM", "training"], "Method"], [2091, 2092, ["CTS"], "Material"], [2102, 2103, ["Web"], "Material"], [2114, 2116, ["training", "phase"], "Method"], [2136, 2139, ["learning", "rate", "schedule"], "Metric"], [2140, 2141, ["termination"], "Task"], [2179, 2181, ["matched", "model"], "Method"], [2199, 2201, ["RNNLM", "configuration"], "Method"], [2206, 2211, ["non", "-", "recurrent", "hidden", "layer"], "Method"], [2215, 2216, ["perplexity"], "Metric"], [2217, 2219, ["word", "error"], "Metric"], [2223, 2230, ["single", "-", "hidden", "-", "layer", "RNNLM", "architecture"], "Method"], [2233, 2235, ["network", "architecture"], "Method"], [2246, 2248, ["ReLU", "nonlinearities"], "Method"], [2249, 2250, ["Training"], "Method"], [2251, 2255, ["noise", "-", "contrastive", "estimation"], "Method"], [2256, 2257, ["NCE"], "Method"], [2262, 2265, ["RNNLM", "output", "vocabulary"], "Material"], [2283, 2284, ["RNNLM"], "Method"], [2297, 2298, ["rescoring"], "Task"], [2341, 2346, ["SRILM", "nbest", "-", "optimize", "tool"], "Method"], [2352, 2357, ["4", "-", "gram", "language", "model"], "Method"], [2358, 2359, ["decoding"], "Task"], [2364, 2366, ["CTS", "transcripts"], "Material"], [2368, 2371, ["DARPA", "EARS", "program"], "Material"], [2372, 2373, ["Switchboard"], "Material"], [2388, 2389, ["Fisher"], "Material"], [2427, 2431, ["RT", "-", "03", "transcripts"], "Material"], [2434, 2440, ["unpruned", "large", "rescoring", "4", "-", "gram"], "Task"], [2443, 2445, ["LM", "component"], "Method"], [2454, 2458, ["LDC", "Broadcast", "News", "texts"], "Material"], [2460, 2465, ["N", "-", "gram", "LM", "configuration"], "Method"], [2474, 2476, ["maxent", "smoothing"], "Method"], [2480, 2481, ["RNNLMs"], "Method"], [2484, 2485, ["Switchboard"], "Material"], [2486, 2487, ["Fisher"], "Material"], [2498, 2500, ["gradient", "computation"], "Task"], [2504, 2505, ["validation"], "Task"], [2533, 2535, ["training", "procedure"], "Method"], [2540, 2541, ["RNNLM"], "Method"], [2547, 2548, ["perplexity"], "Metric"], [2549, 2552, ["word", "error", "performance"], "Metric"], [2554, 2556, ["RNNLM", "setups"], "Method"], [2564, 2566, ["acoustic", "model"], "Method"], [2569, 2571, ["ResNet", "CNN"], "Method"], [2599, 2602, ["relative", "error", "reduction"], "Metric"], [2605, 2606, ["RNNLM"], "Method"], [2612, 2614, ["LM", "rescoring"], "Method"], [2620, 2622, ["acoustic", "model"], "Method"], [2638, 2640, ["confusion", "network"], "Method"], [2642, 2647, ["SRILM", "nbest", "-", "rover", "tool"], "Method"], [2674, 2676, ["overfitting", "issues"], "Task"], [2682, 2684, ["greedy", "search"], "Method"], [2720, 2722, ["search", "procedure"], "Method"], [2728, 2733, ["confusion", "-", "network", "mediated", "voting"], "Task"], [2742, 2744, ["EM", "algorithm"], "Method"], [2776, 2778, ["Speech", "corpora"], "Material"], [2784, 2786, ["English", "CTS"], "Material"], [2787, 2788, ["Switchboard"], "Material"], [2789, 2790, ["Fisher"], "Material"], [2799, 2804, ["NIST", "2000", "CTS", "test", "set"], "Material"], [2808, 2809, ["Switchboard"], "Material"], [2810, 2811, ["SWB"], "Material"], [2813, 2814, ["CallHome"], "Material"], [2815, 2816, ["CH"], "Material"], [2826, 2831, ["NIST", "2002", "CTS", "test", "set"], "Material"], [2834, 2835, ["tuning"], "Task"], [2836, 2837, ["development"], "Task"], [2839, 2842, ["acoustic", "training", "data"], "Material"], [2845, 2847, ["LDC", "corpora"], "Material"], [2865, 2870, ["1", "-", "bit", "SGD", "Training"], "Method"], [2884, 2886, ["parallelize", "training"], "Task"], [2890, 2896, ["1", "-", "bit", "SGD", "parallelization", "technique"], "Method"], [2898, 2902, ["data", "-", "parallel", "method"], "Method"], [2903, 2904, ["minibatches"], "Method"], [2920, 2922, ["communication", "time"], "Metric"], [2928, 2933, ["1", "-", "bit", "SGD", "method"], "Method"], [2940, 2946, ["1", "-", "bit", "quantization", "of", "gradients"], "Method"], [2947, 2952, ["automatic", "minibatch", "-", "size", "scaling"], "Method"], [3024, 3025, ["accuracy"], "Metric"], [3028, 3033, ["automatic", "minibatch", "-", "size", "scaling"], "Method"], [3038, 3040, ["model", "updates"], "Method"], [3076, 3078, ["Acoustic", "Model"], "Method"], [3101, 3103, ["CNN", "models"], "Method"], [3116, 3117, ["LSTMs"], "Method"], [3133, 3141, ["state", "left", "-", "to", "-", "right", "triphone", "models"], "Method"], [3189, 3190, ["Switchboard"], "Material"], [3191, 3193, ["Fisher", "corpora"], "Material"], [3195, 3196, ["decoder"], "Method"], [3207, 3210, ["language", "model", "score"], "Method"], [3212, 3214, ["unigram", "graph"], "Method"], [3223, 3225, ["acoustic", "models"], "Method"], [3229, 3235, ["open", "-", "source", "Computational", "Network", "Toolkit"], "Method"], [3236, 3237, ["CNTK"], "Method"], [3247, 3251, ["i", "-", "vector", "adaptation"], "Method"], [3252, 3254, ["LFMMI", "training"], "Method"], [3268, 3271, ["i", "-", "vectors"], "Metric"], [3274, 3276, ["CNN", "systems"], "Method"], [3289, 3291, ["LFMMI", "training"], "Method"], [3306, 3308, ["error", "rate"], "Metric"], [3317, 3319, ["procedural", "simplicity"], "Metric"], [3320, 3321, ["LFMMI"], "Method"], [3326, 3328, ["writing", "lattices"], "Task"], [3336, 3337, ["LFMMI"], "Method"], [3405, 3406, ["CNNs"], "Method"], [3407, 3409, ["speech", "recognition"], "Task"], [3422, 3424, ["convolutional", "layers"], "Method"], [3444, 3446, ["LACE", "model"], "Method"], [3475, 3479, ["lattice", "-", "free", "MMI"], "Method"], [3499, 3501, ["language", "model"], "Method"], [3504, 3507, ["language", "modeling", "side"], "Task"], [3516, 3517, ["RNNLMs"], "Method"], [3530, 3533, ["phase", "training", "regimen"], "Method"], [3548, 3550, ["CNN", "system"], "Method"], [3551, 3553, ["RNNLM", "rescoring"], "Method"], [3555, 3559, ["relative", "word", "error", "reduction"], "Metric"], [3572, 3574, ["recognition", "system"], "Method"], [3583, 3587, ["LSTM", "-", "based", "LMs"], "Method"], [3595, 3601, ["Microsoft", "’s", "conversational", "speech", "recognition", "system"], "Method"], [3607, 3608, ["CNNs"], "Method"], [3610, 3612, ["acoustic", "model"], "Method"], [3622, 3625, ["RNN", "language", "models"], "Method"], [3632, 3634, ["error", "rate"], "Metric"], [3656, 3658, ["recognition", "system"], "Method"], [3661, 3663, ["system", "combination"], "Method"], [3665, 3669, ["ensemble", "of", "acoustic", "models"], "Method"], [3691, 3692, ["CUED"], "Material"], [3697, 3701, ["CUED", "-", "RNNLM", "toolkit"], "Method"], [3703, 3704, ["ICSI"], "Method"], [3705, 3709, ["compute", "and", "data", "resources"], "Material"], [43, 44, ["Switchboard"], "Material"], [136, 137, ["Switchboard"], "Material"], [2380, 2381, ["Switchboard"], "Material"], [2395, 2396, ["CallHome"], "Material"], [2406, 2407, ["Web"], "Material"], [2516, 2517, ["Web"], "Material"], [2714, 2715, ["Switchboard"], "Material"], [2820, 2821, ["Switchboard"], "Material"], [3641, 3642, ["Switchboard"], "Material"], [3680, 3681, ["Switchboard"], "Material"]]} {"doc_id": "322a7dad274f440a92548faa8f2b2be666b2d01f", "text": ["document", ":", "Pyramid", "Scene", "Parsing", "Network", "Scene", "parsing", "is", "challenging", "for", "unrestricted", "open", "vocabulary", "and", "diverse", "scenes", ".", "In", "this", "paper", ",", "we", "exploit", "the", "capability", "of", "global", "context", "information", "by", "different", "-", "region", "-", "based", "context", "aggregation", "through", "our", "pyramid", "pooling", "module", "together", "with", "the", "proposed", "pyramid", "scene", "parsing", "network", "(", "PSPNet", ")", ".", "Our", "global", "prior", "representation", "is", "effective", "to", "produce", "good", "quality", "results", "on", "the", "scene", "parsing", "task", ",", "while", "PSPNet", "provides", "a", "superior", "framework", "for", "pixel", "-", "level", "prediction", ".", "The", "proposed", "approach", "achieves", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "various", "datasets", ".", "It", "came", "first", "in", "ImageNet", "scene", "parsing", "challenge", "2016", ",", "PASCAL", "VOC", "2012", "benchmark", "and", "Cityscapes", "benchmark", ".", "A", "single", "PSPNet", "yields", "the", "new", "record", "of", "mIoU", "accuracy", "85.4", "%", "on", "PASCAL", "VOC", "2012", "and", "accuracy", "80.2", "%", "on", "Cityscapes", ".", "section", ":", "Introduction", "Scene", "parsing", ",", "based", "on", "semantic", "segmentation", ",", "is", "a", "fundamental", "topic", "in", "computer", "vision", ".", "The", "goal", "is", "to", "assign", "each", "pixel", "in", "the", "image", "a", "category", "label", ".", "Scene", "parsing", "provides", "complete", "understanding", "of", "the", "scene", ".", "It", "predicts", "the", "label", ",", "location", ",", "as", "well", "as", "shape", "for", "each", "element", ".", "This", "topic", "is", "of", "broad", "interest", "for", "potential", "applications", "of", "automatic", "driving", ",", "robot", "sensing", ",", "to", "name", "a", "few", ".", "Difficulty", "of", "scene", "parsing", "is", "closely", "related", "to", "scene", "and", "label", "variety", ".", "The", "pioneer", "scene", "parsing", "task", "is", "to", "classify", "33", "scenes", "for", "2", ",", "688", "images", "on", "LMO", "dataset", ".", "More", "recent", "PASCAL", "VOC", "semantic", "segmentation", "and", "PASCAL", "context", "datasets", "include", "more", "labels", "with", "similar", "context", ",", "such", "as", "chair", "and", "sofa", ",", "horse", "and", "cow", ",", "etc", ".", "The", "new", "ADE20", "K", "dataset", "is", "the", "most", "challenging", "one", "with", "a", "large", "and", "unrestricted", "open", "vocabulary", "and", "more", "scene", "classes", ".", "A", "few", "representative", "images", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "To", "develop", "an", "effective", "algorithm", "for", "these", "datasets", "needs", "to", "conquer", "a", "few", "difficulties", ".", "State", "-", "of", "-", "the", "-", "art", "scene", "parsing", "frameworks", "are", "mostly", "based", "on", "the", "fully", "convolutional", "network", "(", "FCN", ")", ".", "The", "deep", "convolutional", "neural", "network", "(", "CNN", ")", "based", "methods", "boost", "dynamic", "object", "understanding", ",", "and", "yet", "still", "face", "challenges", "considering", "diverse", "scenes", "and", "unrestricted", "vocabulary", ".", "One", "example", "is", "shown", "in", "the", "first", "row", "of", "Fig", ".", "[", "reference", "]", ",", "where", "a", "boat", "is", "mistaken", "as", "a", "car", ".", "These", "errors", "are", "due", "to", "similar", "appearance", "of", "objects", ".", "But", "when", "viewing", "the", "image", "regarding", "the", "context", "prior", "that", "the", "scene", "is", "described", "as", "boathouse", "near", "a", "river", ",", "correct", "prediction", "should", "be", "yielded", ".", "Towards", "accurate", "scene", "perception", ",", "the", "knowledge", "graph", "relies", "on", "prior", "information", "of", "scene", "context", ".", "We", "found", "that", "the", "major", "issue", "for", "current", "FCN", "based", "models", "is", "lack", "of", "suitable", "strategy", "to", "utilize", "global", "scene", "category", "clues", ".", "For", "typical", "complex", "scene", "understanding", ",", "previously", "to", "get", "a", "global", "image", "-", "level", "feature", ",", "spatial", "pyramid", "pooling", "was", "widely", "employed", "where", "spatial", "statistics", "provide", "a", "good", "descriptor", "for", "overall", "scene", "interpretation", ".", "Spatial", "pyramid", "pooling", "network", "further", "enhances", "the", "ability", ".", "Different", "from", "these", "methods", ",", "to", "incorporate", "suitable", "global", "features", ",", "we", "propose", "pyramid", "scene", "parsing", "network", "(", "PSPNet", ")", ".", "In", "addition", "to", "traditional", "dilated", "FCN", "for", "pixel", "prediction", ",", "we", "extend", "the", "pixel", "-", "level", "feature", "to", "the", "specially", "designed", "global", "pyramid", "pooling", "one", ".", "The", "local", "and", "global", "clues", "together", "make", "the", "final", "prediction", "more", "reliable", ".", "We", "also", "propose", "an", "optimization", "strategy", "with", "deeply", "supervised", "loss", ".", "We", "give", "all", "implementation", "details", ",", "which", "are", "key", "to", "our", "decent", "performance", "in", "this", "paper", ",", "and", "make", "the", "code", "and", "trained", "models", "publicly", "available", ".", "Our", "approach", "achieves", "state", "-", "of", "-", "the", "-", "art", "performance", "on", "all", "available", "datasets", ".", "It", "is", "the", "champion", "of", "ImageNet", "scene", "parsing", "challenge", "2016", ",", "and", "arrived", "the", "1st", "place", "on", "PASCAL", "VOC", "2012", "semantic", "segmentation", "benchmark", ",", "and", "the", "1st", "place", "on", "urban", "scene", "Cityscapes", "data", ".", "They", "manifest", "that", "PSPNet", "gives", "a", "promising", "direction", "for", "pixel", "-", "level", "prediction", "tasks", ",", "which", "may", "even", "benefit", "CNN", "-", "based", "stereo", "matching", ",", "optical", "flow", ",", "depth", "estimation", ",", "etc", ".", "in", "follow", "-", "up", "work", ".", "Our", "main", "contributions", "are", "threefold", ".", "We", "propose", "a", "pyramid", "scene", "parsing", "network", "to", "embed", "difficult", "scenery", "context", "features", "in", "an", "FCN", "based", "pixel", "prediction", "framework", ".", "We", "develop", "an", "effective", "optimization", "strategy", "for", "deep", "ResNet", "based", "on", "deeply", "supervised", "loss", ".", "We", "build", "a", "practical", "system", "for", "state", "-", "of", "-", "the", "-", "art", "scene", "parsing", "and", "semantic", "segmentation", "where", "all", "crucial", "implementation", "details", "are", "included", ".", "section", ":", "Related", "Work", "In", "the", "following", ",", "we", "review", "recent", "advances", "in", "scene", "parsing", "and", "semantic", "segmentation", "tasks", ".", "Driven", "by", "powerful", "deep", "neural", "networks", ",", "pixel", "-", "level", "prediction", "tasks", "like", "scene", "parsing", "and", "semantic", "segmentation", "achieve", "great", "progress", "inspired", "by", "replacing", "the", "fully", "-", "connected", "layer", "in", "classification", "with", "the", "convolution", "layer", ".", "To", "enlarge", "the", "receptive", "field", "of", "neural", "networks", ",", "methods", "of", "used", "dilated", "convolution", ".", "Noh", "proposed", "a", "coarse", "-", "to", "-", "fine", "structure", "with", "deconvolution", "network", "to", "learn", "the", "segmentation", "mask", ".", "Our", "baseline", "network", "is", "FCN", "and", "dilated", "network", ".", "Other", "work", "mainly", "proceeds", "in", "two", "directions", ".", "One", "line", "is", "with", "multi", "-", "scale", "feature", "ensembling", ".", "Since", "in", "deep", "networks", ",", "higher", "-", "layer", "feature", "contains", "more", "semantic", "meaning", "and", "less", "location", "information", ".", "Combining", "multi", "-", "scale", "features", "can", "improve", "the", "performance", ".", "The", "other", "direction", "is", "based", "on", "structure", "prediction", ".", "The", "pioneer", "work", "used", "conditional", "random", "field", "(", "CRF", ")", "as", "post", "processing", "to", "refine", "the", "segmentation", "result", ".", "Following", "methods", "refined", "networks", "via", "end", "-", "to", "-", "end", "modeling", ".", "Both", "of", "the", "two", "directions", "ameliorate", "the", "localization", "ability", "of", "scene", "parsing", "where", "predicted", "semantic", "boundary", "fits", "objects", ".", "Yet", "there", "is", "still", "much", "room", "to", "exploit", "necessary", "information", "in", "complex", "scenes", ".", "To", "make", "good", "use", "of", "global", "image", "-", "level", "priors", "for", "diverse", "scene", "understanding", ",", "methods", "of", "extracted", "global", "context", "information", "with", "traditional", "features", "not", "from", "deep", "neural", "networks", ".", "Similar", "improvement", "was", "made", "under", "object", "detection", "frameworks", ".", "Liu", "proved", "that", "global", "average", "pooling", "with", "FCN", "can", "improve", "semantic", "segmentation", "results", ".", "However", ",", "our", "experiments", "show", "that", "these", "global", "descriptors", "are", "not", "representative", "enough", "for", "the", "challenging", "ADE20", "K", "data", ".", "Therefore", ",", "different", "from", "global", "pooling", "in", ",", "we", "exploit", "the", "capability", "of", "global", "context", "information", "by", "different", "-", "region", "-", "based", "context", "aggregation", "via", "our", "pyramid", "scene", "parsing", "network", ".", "section", ":", "Pyramid", "Scene", "Parsing", "Network", "We", "start", "with", "our", "observation", "and", "analysis", "of", "representative", "failure", "cases", "when", "applying", "FCN", "methods", "to", "scene", "parsing", ".", "They", "motivate", "proposal", "of", "our", "pyramid", "pooling", "module", "as", "the", "effective", "global", "context", "prior", ".", "Our", "pyramid", "scene", "parsing", "network", "(", "PSPNet", ")", "illustrated", "in", "Fig", ".", "[", "reference", "]", "is", "then", "described", "to", "improve", "performance", "for", "open", "-", "vocabulary", "object", "and", "stuff", "identification", "in", "complex", "scene", "parsing", ".", "subsection", ":", "Important", "Observations", "The", "new", "ADE20", "K", "dataset", "contains", "150", "stuff", "/", "object", "category", "labels", "(", ",", "wall", ",", "sky", ",", "and", "tree", ")", "and", "1", ",", "038", "image", "-", "level", "scene", "descriptors", "(", ",", "airport_terminal", ",", "bedroom", ",", "and", "street", ")", ".", "So", "a", "large", "amount", "of", "labels", "and", "vast", "distributions", "of", "scenes", "come", "into", "existence", ".", "Inspecting", "the", "prediction", "results", "of", "the", "FCN", "baseline", "provided", "in", ",", "we", "summarize", "several", "common", "issues", "for", "complex", "-", "scene", "parsing", ".", "paragraph", ":", "Mismatched", "Relationship", "Context", "relationship", "is", "universal", "and", "important", "especially", "for", "complex", "scene", "understanding", ".", "There", "exist", "co", "-", "occurrent", "visual", "patterns", ".", "For", "example", ",", "an", "airplane", "is", "likely", "to", "be", "in", "runway", "or", "fly", "in", "sky", "while", "not", "over", "a", "road", ".", "For", "the", "first", "-", "row", "example", "in", "Fig", ".", "[", "reference", "]", ",", "FCN", "predicts", "the", "boat", "in", "the", "yellow", "box", "as", "a", "“", "car", "”", "based", "on", "its", "appearance", ".", "But", "the", "common", "knowledge", "is", "that", "a", "car", "is", "seldom", "over", "a", "river", ".", "Lack", "of", "the", "ability", "to", "collect", "contextual", "information", "increases", "the", "chance", "of", "misclassification", ".", "paragraph", ":", "Confusion", "Categories", "There", "are", "many", "class", "label", "pairs", "in", "the", "ADE20", "K", "dataset", "that", "are", "confusing", "in", "classification", ".", "Examples", "are", "field", "and", "earth", ";", "mountain", "and", "hill", ";", "wall", ",", "house", ",", "building", "and", "skyscraper", ".", "They", "are", "with", "similar", "appearance", ".", "The", "expert", "annotator", "who", "labeled", "the", "entire", "dataset", ",", "still", "makes", "17.60", "%", "pixel", "error", "as", "described", "in", ".", "In", "the", "second", "row", "of", "Fig", ".", "[", "reference", "]", ",", "FCN", "predicts", "the", "object", "in", "the", "box", "as", "part", "of", "skyscraper", "and", "part", "of", "building", ".", "These", "results", "should", "be", "excluded", "so", "that", "the", "whole", "object", "is", "either", "skyscraper", "or", "building", ",", "but", "not", "both", ".", "This", "problem", "can", "be", "remedied", "by", "utilizing", "the", "relationship", "between", "categories", ".", "paragraph", ":", "Inconspicuous", "Classes", "Scene", "contains", "objects", "/", "stuff", "of", "arbitrary", "size", ".", "Several", "small", "-", "size", "things", ",", "like", "streetlight", "and", "signboard", ",", "are", "hard", "to", "find", "while", "they", "may", "be", "of", "great", "importance", ".", "Contrarily", ",", "big", "objects", "or", "stuff", "may", "exceed", "the", "receptive", "field", "of", "FCN", "and", "thus", "cause", "discontinuous", "prediction", ".", "As", "shown", "in", "the", "third", "row", "of", "Fig", ".", "[", "reference", "]", ",", "the", "pillow", "has", "similar", "appearance", "with", "the", "sheet", ".", "Overlooking", "the", "global", "scene", "category", "may", "fail", "to", "parse", "the", "pillow", ".", "To", "improve", "performance", "for", "remarkably", "small", "or", "large", "objects", ",", "one", "should", "pay", "much", "attention", "to", "different", "sub", "-", "regions", "that", "contain", "inconspicuous", "-", "category", "stuff", ".", "To", "summarize", "these", "observations", ",", "many", "errors", "are", "partially", "or", "completely", "related", "to", "contextual", "relationship", "and", "global", "information", "for", "different", "receptive", "fields", ".", "Thus", "a", "deep", "network", "with", "a", "suitable", "global", "-", "scene", "-", "level", "prior", "can", "much", "improve", "the", "performance", "of", "scene", "parsing", ".", "subsection", ":", "Pyramid", "Pooling", "Module", "With", "above", "analysis", ",", "in", "what", "follows", ",", "we", "introduce", "the", "pyramid", "pooling", "module", ",", "which", "empirically", "proves", "to", "be", "an", "effective", "global", "contextual", "prior", ".", "In", "a", "deep", "neural", "network", ",", "the", "size", "of", "receptive", "field", "can", "roughly", "indicates", "how", "much", "we", "use", "context", "information", ".", "Although", "theoretically", "the", "receptive", "field", "of", "ResNet", "is", "already", "larger", "than", "the", "input", "image", ",", "it", "is", "shown", "by", "Zhou", "that", "the", "empirical", "receptive", "field", "of", "CNN", "is", "much", "smaller", "than", "the", "theoretical", "one", "especially", "on", "high", "-", "level", "layers", ".", "This", "makes", "many", "networks", "not", "sufficiently", "incorporate", "the", "momentous", "global", "scenery", "prior", ".", "We", "address", "this", "issue", "by", "proposing", "an", "effective", "global", "prior", "representation", ".", "Global", "average", "pooling", "is", "a", "good", "baseline", "model", "as", "the", "global", "contextual", "prior", ",", "which", "is", "commonly", "used", "in", "image", "classification", "tasks", ".", "In", ",", "it", "was", "successfully", "applied", "to", "semantic", "segmentation", ".", "But", "regarding", "the", "complex", "-", "scene", "images", "in", "ADE20", "K", ",", "this", "strategy", "is", "not", "enough", "to", "cover", "necessary", "information", ".", "Pixels", "in", "these", "scene", "images", "are", "annotated", "regarding", "many", "stuff", "and", "objects", ".", "Directly", "fusing", "them", "to", "form", "a", "single", "vector", "may", "lose", "the", "spatial", "relation", "and", "cause", "ambiguity", ".", "Global", "context", "information", "along", "with", "sub", "-", "region", "context", "is", "helpful", "in", "this", "regard", "to", "distinguish", "among", "various", "categories", ".", "A", "more", "powerful", "representation", "could", "be", "fused", "information", "from", "different", "sub", "-", "regions", "with", "these", "receptive", "fields", ".", "Similar", "conclusion", "was", "drawn", "in", "classical", "work", "of", "scene", "/", "image", "classification", ".", "In", ",", "feature", "maps", "in", "different", "levels", "generated", "by", "pyramid", "pooling", "were", "finally", "flattened", "and", "concatenated", "to", "be", "fed", "into", "a", "fully", "connected", "layer", "for", "classification", ".", "This", "global", "prior", "is", "designed", "to", "remove", "the", "fixed", "-", "size", "constraint", "of", "CNN", "for", "image", "classification", ".", "To", "further", "reduce", "context", "information", "loss", "between", "different", "sub", "-", "regions", ",", "we", "propose", "a", "hierarchical", "global", "prior", ",", "containing", "information", "with", "different", "scales", "and", "varying", "among", "different", "sub", "-", "regions", ".", "We", "call", "it", "pyramid", "pooling", "module", "for", "global", "scene", "prior", "construction", "upon", "the", "final", "-", "layer", "-", "feature", "-", "map", "of", "the", "deep", "neural", "network", ",", "as", "illustrated", "in", "part", "(", "c", ")", "of", "Fig", ".", "[", "reference", "]", ".", "The", "pyramid", "pooling", "module", "fuses", "features", "under", "four", "different", "pyramid", "scales", ".", "The", "coarsest", "level", "highlighted", "in", "red", "is", "global", "pooling", "to", "generate", "a", "single", "bin", "output", ".", "The", "following", "pyramid", "level", "separates", "the", "feature", "map", "into", "different", "sub", "-", "regions", "and", "forms", "pooled", "representation", "for", "different", "locations", ".", "The", "output", "of", "different", "levels", "in", "the", "pyramid", "pooling", "module", "contains", "the", "feature", "map", "with", "varied", "sizes", ".", "To", "maintain", "the", "weight", "of", "global", "feature", ",", "we", "use", "convolution", "layer", "after", "each", "pyramid", "level", "to", "reduce", "the", "dimension", "of", "context", "representation", "to", "of", "the", "original", "one", "if", "the", "level", "size", "of", "pyramid", "is", ".", "Then", "we", "directly", "upsample", "the", "low", "-", "dimension", "feature", "maps", "to", "get", "the", "same", "size", "feature", "as", "the", "original", "feature", "map", "via", "bilinear", "interpolation", ".", "Finally", ",", "different", "levels", "of", "features", "are", "concatenated", "as", "the", "final", "pyramid", "pooling", "global", "feature", ".", "Noted", "that", "the", "number", "of", "pyramid", "levels", "and", "size", "of", "each", "level", "can", "be", "modified", ".", "They", "are", "related", "to", "the", "size", "of", "feature", "map", "that", "is", "fed", "into", "the", "pyramid", "pooling", "layer", ".", "The", "structure", "abstracts", "different", "sub", "-", "regions", "by", "adopting", "varying", "-", "size", "pooling", "kernels", "in", "a", "few", "strides", ".", "Thus", "the", "multi", "-", "stage", "kernels", "should", "maintain", "a", "reasonable", "gap", "in", "representation", ".", "Our", "pyramid", "pooling", "module", "is", "a", "four", "-", "level", "one", "with", "bin", "sizes", "of", ",", ",", "and", "respectively", ".", "For", "the", "type", "of", "pooling", "operation", "between", "max", "and", "average", ",", "we", "perform", "extensive", "experiments", "to", "show", "the", "difference", "in", "Section", "[", "reference", "]", ".", "subsection", ":", "Network", "Architecture", "With", "the", "pyramid", "pooling", "module", ",", "we", "propose", "our", "pyramid", "scene", "parsing", "network", "(", "PSPNet", ")", "as", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "Given", "an", "input", "image", "in", "Fig", ".", "[", "reference", "]", "(", "a", ")", ",", "we", "use", "a", "pretrained", "ResNet", "model", "with", "the", "dilated", "network", "strategy", "to", "extract", "the", "feature", "map", ".", "The", "final", "feature", "map", "size", "is", "of", "the", "input", "image", ",", "as", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "b", ")", ".", "On", "top", "of", "the", "map", ",", "we", "use", "the", "pyramid", "pooling", "module", "shown", "in", "(", "c", ")", "to", "gather", "context", "information", ".", "Using", "our", "4", "-", "level", "pyramid", ",", "the", "pooling", "kernels", "cover", "the", "whole", ",", "half", "of", ",", "and", "small", "portions", "of", "the", "image", ".", "They", "are", "fused", "as", "the", "global", "prior", ".", "Then", "we", "concatenate", "the", "prior", "with", "the", "original", "feature", "map", "in", "the", "final", "part", "of", "(", "c", ")", ".", "It", "is", "followed", "by", "a", "convolution", "layer", "to", "generate", "the", "final", "prediction", "map", "in", "(", "d", ")", ".", "To", "explain", "our", "structure", ",", "PSPNet", "provides", "an", "effective", "global", "contextual", "prior", "for", "pixel", "-", "level", "scene", "parsing", ".", "The", "pyramid", "pooling", "module", "can", "collect", "levels", "of", "information", ",", "more", "representative", "than", "global", "pooling", ".", "In", "terms", "of", "computational", "cost", ",", "our", "PSPNet", "does", "not", "much", "increase", "it", "compared", "to", "the", "original", "dilated", "FCN", "network", ".", "In", "end", "-", "to", "-", "end", "learning", ",", "the", "global", "pyramid", "pooling", "module", "and", "the", "local", "FCN", "feature", "can", "be", "optimized", "simultaneously", ".", "section", ":", "Deep", "Supervision", "for", "ResNet", "-", "Based", "FCN", "Deep", "pretrained", "networks", "lead", "to", "good", "performance", ".", "However", ",", "increasing", "depth", "of", "the", "network", "may", "introduce", "additional", "optimization", "difficulty", "as", "shown", "in", "for", "image", "classification", ".", "ResNet", "solves", "this", "problem", "with", "skip", "connection", "in", "each", "block", ".", "Latter", "layers", "of", "deep", "ResNet", "mainly", "learn", "residues", "based", "on", "previous", "ones", ".", "We", "contrarily", "propose", "generating", "initial", "results", "by", "supervision", "with", "an", "additional", "loss", ",", "and", "learning", "the", "residue", "afterwards", "with", "the", "final", "loss", ".", "Thus", ",", "optimization", "of", "the", "deep", "network", "is", "decomposed", "into", "two", ",", "each", "is", "simpler", "to", "solve", ".", "An", "example", "of", "our", "deeply", "supervised", "ResNet101", "model", "is", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "Apart", "from", "the", "main", "branch", "using", "softmax", "loss", "to", "train", "the", "final", "classifier", ",", "another", "classifier", "is", "applied", "after", "the", "fourth", "stage", ",", "i.e.", ",", "the", "res4b22", "residue", "block", ".", "Different", "from", "relay", "backpropagation", "that", "blocks", "the", "backward", "auxiliary", "loss", "to", "several", "shallow", "layers", ",", "we", "let", "the", "two", "loss", "functions", "pass", "through", "all", "previous", "layers", ".", "The", "auxiliary", "loss", "helps", "optimize", "the", "learning", "process", ",", "while", "the", "master", "branch", "loss", "takes", "the", "most", "responsibility", ".", "We", "add", "weight", "to", "balance", "the", "auxiliary", "loss", ".", "In", "the", "testing", "phase", ",", "we", "abandon", "this", "auxiliary", "branch", "and", "only", "use", "the", "well", "optimized", "master", "branch", "for", "final", "prediction", ".", "This", "kind", "of", "deeply", "supervised", "training", "strategy", "for", "ResNet", "-", "based", "FCN", "is", "broadly", "useful", "under", "different", "experimental", "settings", "and", "works", "with", "the", "pre", "-", "trained", "ResNet", "model", ".", "This", "manifests", "the", "generality", "of", "such", "a", "learning", "strategy", ".", "More", "details", "are", "provided", "in", "Section", "[", "reference", "]", ".", "section", ":", "Experiments", "Our", "proposed", "method", "is", "successful", "on", "scene", "parsing", "and", "semantic", "segmentation", "challenges", ".", "We", "evaluate", "it", "in", "this", "section", "on", "three", "different", "datasets", ",", "including", "ImageNet", "scene", "parsing", "challenge", "2016", ",", "PASCAL", "VOC", "2012", "semantic", "segmentation", "and", "urban", "scene", "understanding", "dataset", "Cityscapes", ".", "subsection", ":", "Implementation", "Details", "For", "a", "practical", "deep", "learning", "system", ",", "devil", "is", "always", "in", "the", "details", ".", "Our", "implementation", "is", "based", "on", "the", "public", "platform", "Caffe", ".", "Inspired", "by", ",", "we", "use", "the", "“", "poly", "”", "learning", "rate", "policy", "where", "current", "learning", "rate", "equals", "to", "the", "base", "one", "multiplying", ".", "We", "set", "base", "learning", "rate", "to", "0.01", "and", "power", "to", "0.9", ".", "The", "performance", "can", "be", "improved", "by", "increasing", "the", "iteration", "number", ",", "which", "is", "set", "to", "150", "K", "for", "ImageNet", "experiment", ",", "30", "K", "for", "PASCAL", "VOC", "and", "90", "K", "for", "Cityscapes", ".", "Momentum", "and", "weight", "decay", "are", "set", "to", "0.9", "and", "0.0001", "respectively", ".", "For", "data", "augmentation", ",", "we", "adopt", "random", "mirror", "and", "random", "resize", "between", "0.5", "and", "2", "for", "all", "datasets", ",", "and", "additionally", "add", "random", "rotation", "between", "-", "10", "and", "10", "degrees", ",", "and", "random", "Gaussian", "blur", "for", "ImageNet", "and", "PASCAL", "VOC", ".", "This", "comprehensive", "data", "augmentation", "scheme", "makes", "the", "network", "resist", "overfitting", ".", "Our", "network", "contains", "dilated", "convolution", "following", ".", "During", "the", "course", "of", "experiments", ",", "we", "notice", "that", "an", "appropriately", "large", "“", "cropsize", "”", "can", "yield", "good", "performance", "and", "“", "batchsize", "”", "in", "the", "batch", "normalization", "layer", "is", "of", "great", "importance", ".", "Due", "to", "limited", "physical", "memory", "on", "GPU", "cards", ",", "we", "set", "the", "“", "batchsize", "”", "to", "16", "during", "training", ".", "To", "achieve", "this", ",", "we", "modify", "Caffe", "from", "together", "with", "branch", "and", "make", "it", "support", "batch", "normalization", "on", "data", "gathered", "from", "multiple", "GPUs", "based", "on", "OpenMPI", ".", "For", "the", "auxiliary", "loss", ",", "we", "set", "the", "weight", "to", "0.4", "in", "experiments", ".", "subsection", ":", "ImageNet", "Scene", "Parsing", "Challenge", "2016", "paragraph", ":", "Dataset", "and", "Evaluation", "Metrics", "The", "ADE20", "K", "dataset", "is", "used", "in", "ImageNet", "scene", "parsing", "challenge", "2016", ".", "Different", "from", "other", "datasets", ",", "ADE20", "K", "is", "more", "challenging", "for", "the", "up", "to", "150", "classes", "and", "diverse", "scenes", "with", "a", "total", "of", "1", ",", "038", "image", "-", "level", "labels", ".", "The", "challenge", "data", "is", "divided", "into", "20K", "/", "2K", "/", "3", "K", "images", "for", "training", ",", "validation", "and", "testing", ".", "Also", ",", "it", "needs", "to", "parse", "both", "objects", "and", "stuff", "in", "the", "scene", ",", "which", "makes", "it", "more", "difficult", "than", "other", "datasets", ".", "For", "evaluation", ",", "both", "pixel", "-", "wise", "accuracy", "(", "Pixel", "Acc", ".", ")", "and", "mean", "of", "class", "-", "wise", "intersection", "over", "union", "(", "Mean", "IoU", ")", "are", "used", ".", "paragraph", ":", "Ablation", "Study", "for", "PSPNet", "To", "evaluate", "PSPNet", ",", "we", "conduct", "experiments", "with", "several", "settings", ",", "including", "pooling", "types", "of", "max", "and", "average", ",", "pooling", "with", "just", "one", "global", "feature", "or", "four", "-", "level", "features", ",", "with", "and", "without", "dimension", "reduction", "after", "the", "pooling", "operation", "and", "before", "concatenation", ".", "As", "listed", "in", "Table", "[", "reference", "]", ",", "average", "pooling", "works", "better", "than", "max", "pooling", "in", "all", "settings", ".", "Pooling", "with", "pyramid", "parsing", "outperforms", "that", "using", "global", "pooling", ".", "With", "dimension", "reduction", ",", "the", "performance", "is", "further", "enhanced", ".", "With", "our", "proposed", "PSPNet", ",", "the", "best", "setting", "yields", "results", "41.68", "/", "80.04", "in", "terms", "of", "Mean", "IoU", "and", "Pixel", "Acc", ".", "(", "%", ")", ",", "exceeding", "global", "average", "pooling", "of", "40.07", "/", "79.52", "as", "idea", "in", "Liu", "by", "1.61", "/", "0.52", ".", "And", "compared", "to", "the", "baseline", ",", "PSPNet", "outperforming", "it", "by", "4.45", "/", "2.03", "in", "terms", "of", "absolute", "improvement", "and", "11.95", "/", "2.60", "in", "terms", "of", "relative", "difference", ".", "paragraph", ":", "Ablation", "Study", "for", "Auxiliary", "Loss", "The", "introduced", "auxiliary", "loss", "helps", "optimize", "the", "learning", "process", "while", "not", "influencing", "learning", "in", "the", "master", "branch", ".", "We", "experiment", "with", "setting", "the", "auxiliary", "loss", "weight", "between", "0", "and", "1", "and", "show", "the", "results", "in", "Table", "[", "reference", "]", ".", "The", "baseline", "uses", "ResNet50", "-", "based", "FCN", "with", "dilated", "network", ",", "with", "the", "master", "branch", "’s", "softmax", "loss", "for", "optimization", ".", "Adding", "the", "auxiliary", "loss", "branch", ",", "=", "0.4", "yields", "the", "best", "performance", ".", "It", "outperforms", "the", "baseline", "with", "an", "improvement", "of", "1.41", "/", "0.94", "in", "terms", "of", "Mean", "IoU", "and", "Pixel", "Acc", ".", "(", "%", ")", ".", "We", "believe", "deeper", "networks", "will", "benefit", "more", "given", "the", "new", "augmented", "auxiliary", "loss", ".", "paragraph", ":", "Ablation", "Study", "for", "Pre", "-", "trained", "Model", "Deeper", "neural", "networks", "have", "been", "shown", "in", "previous", "work", "to", "be", "beneficial", "to", "large", "scale", "data", "classification", ".", "To", "further", "analyze", "PSPNet", ",", "we", "conduct", "experiments", "for", "different", "depths", "of", "pre", "-", "trained", "ResNet", ".", "We", "test", "four", "depths", "of", "{", "50", ",", "101", ",", "152", ",", "269}.", "As", "shown", "in", "Fig", ".", "[", "reference", "]", ",", "with", "the", "same", "setting", ",", "increasing", "the", "depth", "of", "ResNet", "from", "50", "to", "269", "can", "improve", "the", "score", "of", "(", "Mean", "IoU", "+", "Pixel", "Acc", ".", ")", "/", "2", "(", "%", ")", "from", "60.86", "to", "62.35", ",", "with", "1.49", "absolute", "improvement", ".", "Detailed", "scores", "of", "PSPNet", "pre", "-", "trained", "from", "different", "depth", "ResNet", "models", "are", "listed", "in", "Table", "[", "reference", "]", ".", "paragraph", ":", "More", "Detailed", "Performance", "Analysis", "We", "show", "our", "more", "detailed", "analysis", "on", "the", "validation", "set", "of", "ADE20", "K", "in", "Table", "[", "reference", "]", ".", "All", "our", "results", "except", "the", "last", "-", "row", "one", "use", "single", "-", "scale", "test", ".", "“", "ResNet269", "+", "DA", "+", "AL", "+", "PSP", "+", "MS", "”", "uses", "multi", "-", "scale", "testing", ".", "Our", "baseline", "is", "adapted", "from", "ResNet50", "with", "dilated", "network", ",", "which", "yields", "MeanIoU", "34.28", "and", "Pixel", "Acc", ".", "76.35", ".", "It", "already", "outperforms", "other", "prior", "systems", "possibly", "due", "to", "the", "powerful", "ResNet", ".", "Our", "proposed", "architecture", "makes", "further", "improvement", "compared", "to", "the", "baseline", ".", "Using", "data", "augmentation", ",", "our", "result", "exceeds", "the", "baseline", "by", "1.54", "/", "0.72", "and", "reaches", "35.82", "/", "77.07", ".", "Using", "the", "auxiliary", "loss", "can", "further", "improve", "it", "by", "1.41", "/", "0.94", "and", "reaches", "37.23", "/", "78.01", ".", "With", "PSPNet", ",", "we", "notice", "relatively", "more", "significant", "progress", "for", "improvement", "of", "4.45", "/", "2.03", ".", "The", "result", "reaches", "41.68", "/", "80.04", ".", "The", "difference", "from", "the", "baseline", "result", "is", "7.40", "/", "3.69", "in", "terms", "of", "absolute", "improvement", "and", "21.59", "/", "4.83", "(", "%", ")", "in", "terms", "of", "relativity", ".", "A", "deeper", "network", "of", "ResNet269", "yields", "even", "higher", "performance", "up", "to", "43.81", "/", "80.88", ".", "Finally", ",", "the", "multi", "-", "scale", "testing", "scheme", "moves", "the", "scores", "to", "44.94", "/", "81.69", ".", "paragraph", ":", "Results", "in", "Challenge", "Using", "the", "proposed", "architecture", ",", "our", "team", "came", "in", "the", "1st", "place", "in", "ImageNet", "scene", "parsing", "challenge", "2016", ".", "Table", "[", "reference", "]", "shows", "a", "few", "results", "in", "this", "competition", ".", "Our", "ensemble", "submission", "achieves", "score", "57.21", "%", "on", "the", "testing", "set", ".", "Our", "single", "-", "model", "yields", "score", "55.38", "%", ",", "which", "is", "even", "higher", "than", "a", "few", "other", "multi", "-", "model", "ensemble", "submissions", ".", "This", "score", "is", "lower", "than", "that", "on", "the", "validation", "set", "possibly", "due", "to", "the", "difference", "of", "data", "distributions", "between", "validation", "and", "testing", "sets", ".", "As", "shown", "in", "column", "(", "d", ")", "of", "Fig", ".", "[", "reference", "]", ",", "PSPNet", "solves", "the", "common", "problems", "in", "FCN", ".", "Fig", ".", "[", "reference", "]", "shows", "another", "few", "parsing", "results", "on", "validation", "set", "of", "ADE20K.", "Our", "results", "contain", "more", "accurate", "and", "detailed", "structures", "compared", "to", "the", "baseline", ".", "subsection", ":", "PASCAL", "VOC", "2012", "Our", "PSPNet", "also", "works", "satisfyingly", "on", "semantic", "segmentation", ".", "We", "carry", "out", "experiments", "on", "the", "PASCAL", "VOC", "2012", "segmentation", "dataset", ",", "which", "contains", "20", "object", "categories", "and", "one", "background", "class", ".", "Following", "the", "procedure", "of", ",", "we", "use", "augmented", "data", "with", "the", "annotation", "of", "resulting", "10", ",", "582", ",", "1", ",", "449", "and", "1", ",", "456", "images", "for", "training", ",", "validation", "and", "testing", ".", "Results", "are", "shown", "in", "Table", "[", "reference", "]", ",", "we", "compare", "PSPNet", "with", "previous", "best", "-", "performing", "methods", "on", "the", "testing", "set", "based", "on", "two", "settings", ",", "i.e.", ",", "with", "or", "without", "pre", "-", "training", "on", "MS", "-", "COCO", "dataset", ".", "Methods", "pre", "-", "trained", "with", "MS", "-", "COCO", "are", "marked", "by", "‘", "†", "’", ".", "For", "fair", "comparison", "with", "current", "ResNet", "based", "frameworks", "in", "scene", "parsing", "/", "semantic", "segmentation", "task", ",", "we", "build", "our", "architecture", "based", "on", "ResNet101", "while", "without", "post", "-", "processing", "like", "CRF", ".", "We", "evaluate", "PSPNet", "with", "several", "-", "scale", "input", "and", "use", "the", "average", "results", "following", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "PSPNet", "outperforms", "prior", "methods", "on", "both", "settings", ".", "Trained", "with", "only", "VOC", "2012", "data", ",", "we", "achieve", "82.6", "%", "accuracy", "–", "we", "get", "the", "highest", "accuracy", "on", "all", "20", "classes", ".", "When", "PSPNet", "is", "pre", "-", "trained", "with", "MS", "-", "COCO", "dataset", ",", "it", "reaches", "85.4", "%", "accuracy", "where", "19", "out", "of", "the", "20", "classes", "receive", "the", "highest", "accuracy", ".", "Intriguingly", ",", "our", "PSPNet", "trained", "with", "only", "VOC", "2012", "data", "outperforms", "existing", "methods", "trained", "with", "the", "MS", "-", "COCO", "pre", "-", "trained", "model", ".", "One", "may", "argue", "that", "our", "based", "classification", "model", "is", "more", "powerful", "than", "several", "prior", "methods", "since", "ResNet", "was", "recently", "proposed", ".", "To", "exhibit", "our", "unique", "contribution", ",", "we", "show", "that", "our", "method", "also", "outperforms", "state", "-", "of", "-", "the", "-", "art", "frameworks", "that", "use", "the", "same", "model", ",", "including", "FCRNs", ",", "LRR", ",", "and", "DeepLab", ".", "In", "this", "process", ",", "we", "even", "do", "not", "employ", "time", "-", "consuming", "but", "effective", "post", "-", "processing", ",", "such", "as", "CRF", ",", "as", "that", "in", ".", "Several", "examples", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "For", "“", "cows", "”", "in", "row", "one", ",", "our", "baseline", "model", "treats", "it", "as", "“", "horse", "”", "and", "“", "dog", "”", "while", "PSPNet", "corrects", "these", "errors", ".", "For", "“", "aeroplane", "”", "and", "“", "table", "”", "in", "the", "second", "and", "third", "rows", ",", "PSPNet", "finds", "missing", "parts", ".", "For", "“", "person", "”", ",", "“", "bottle", "”", "and", "“", "plant", "”", "in", "following", "rows", ",", "PSPNet", "performs", "well", "on", "these", "small", "-", "size", "-", "object", "classes", "in", "the", "images", "compared", "to", "the", "baseline", "model", ".", "More", "visual", "comparisons", "between", "PSPNet", "and", "other", "methods", "are", "included", "in", "Fig", ".", "[", "reference", "]", ".", "subsection", ":", "Cityscapes", "Cityscapes", "is", "a", "recently", "released", "dataset", "for", "semantic", "urban", "scene", "understanding", ".", "It", "contains", "5", ",", "000", "high", "quality", "pixel", "-", "level", "finely", "annotated", "images", "collected", "from", "50", "cities", "in", "different", "seasons", ".", "The", "images", "are", "divided", "into", "sets", "with", "numbers", "2", ",", "975", ",", "500", ",", "and", "1", ",", "525", "for", "training", ",", "validation", "and", "testing", ".", "It", "defines", "19", "categories", "containing", "both", "stuff", "and", "objects", ".", "Also", ",", "20", ",", "000", "coarsely", "annotated", "images", "are", "provided", "for", "two", "settings", "in", "comparison", ",", "i.e.", ",", "training", "with", "only", "fine", "data", "or", "with", "both", "the", "fine", "and", "coarse", "data", ".", "Methods", "trained", "using", "both", "fine", "and", "coarse", "data", "are", "marked", "with", "‘", "’", ".", "Detailed", "results", "are", "listed", "in", "Table", "[", "reference", "]", ".", "Our", "base", "model", "is", "ResNet101", "as", "in", "DeepLab", "for", "fair", "comparison", "and", "the", "testing", "procedure", "follows", "Section", "[", "reference", "]", ".", "Statistics", "in", "Table", "[", "reference", "]", "show", "that", "PSPNet", "outperforms", "other", "methods", "with", "notable", "advantage", ".", "Using", "both", "fine", "and", "coarse", "data", "for", "training", "makes", "our", "method", "yield", "80.2", "accuracy", ".", "Several", "examples", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "Detailed", "per", "-", "class", "results", "on", "testing", "set", "are", "shown", "in", "Table", "[", "reference", "]", ".", "section", ":", "Concluding", "Remarks", "We", "have", "proposed", "an", "effective", "pyramid", "scene", "parsing", "network", "for", "complex", "scene", "understanding", ".", "The", "global", "pyramid", "pooling", "feature", "provides", "additional", "contextual", "information", ".", "We", "have", "also", "provided", "a", "deeply", "supervised", "optimization", "strategy", "for", "ResNet", "-", "based", "FCN", "network", ".", "We", "hope", "the", "implementation", "details", "publicly", "available", "can", "help", "the", "community", "adopt", "these", "useful", "strategies", "for", "scene", "parsing", "and", "semantic", "segmentation", "and", "advance", "related", "techniques", ".", "section", ":", "Acknowledgements", "We", "would", "like", "to", "thank", "Gang", "Sun", "and", "Tong", "Xiao", "for", "their", "help", "in", "training", "the", "basic", "classification", "models", ",", "Qun", "Luo", "for", "technical", "support", ".", "This", "work", "is", "supported", "by", "a", "grant", "from", "the", "Research", "Grants", "Council", "of", "the", "Hong", "Kong", "SAR", "(", "project", "No", ".", "2150760", ")", ".", "bibliography", ":", "References"], "entity": [[2, 6, ["Pyramid", "Scene", "Parsing", "Network"], "Method"], [6, 8, ["Scene", "parsing"], "Task"], [33, 38, ["region", "-", "based", "context", "aggregation"], "Method"], [40, 43, ["pyramid", "pooling", "module"], "Method"], [47, 51, ["pyramid", "scene", "parsing", "network"], "Method"], [52, 53, ["PSPNet"], "Method"], [56, 59, ["global", "prior", "representation"], "Method"], [68, 71, ["scene", "parsing", "task"], "Task"], [73, 74, ["PSPNet"], "Method"], [79, 83, ["pixel", "-", "level", "prediction"], "Task"], [104, 109, ["ImageNet", "scene", "parsing", "challenge", "2016"], "Task"], [110, 114, ["PASCAL", "VOC", "2012", "benchmark"], "Material"], [115, 117, ["Cityscapes", "benchmark"], "Material"], [120, 121, ["PSPNet"], "Method"], [126, 128, ["mIoU", "accuracy"], "Metric"], [131, 134, ["PASCAL", "VOC", "2012"], "Material"], [135, 136, ["accuracy"], "Metric"], [139, 140, ["Cityscapes"], "Material"], [144, 146, ["Scene", "parsing"], "Task"], [149, 151, ["semantic", "segmentation"], "Task"], [157, 159, ["computer", "vision"], "Task"], [174, 176, ["Scene", "parsing"], "Task"], [208, 210, ["automatic", "driving"], "Task"], [211, 213, ["robot", "sensing"], "Task"], [221, 223, ["scene", "parsing"], "Task"], [234, 237, ["scene", "parsing", "task"], "Task"], [248, 250, ["LMO", "dataset"], "Material"], [258, 261, ["PASCAL", "context", "datasets"], "Material"], [282, 285, ["ADE20", "K", "dataset"], "Material"], [337, 340, ["scene", "parsing", "frameworks"], "Method"], [345, 348, ["fully", "convolutional", "network"], "Method"], [349, 350, ["FCN"], "Method"], [354, 357, ["convolutional", "neural", "network"], "Method"], [358, 359, ["CNN"], "Method"], [363, 366, ["dynamic", "object", "understanding"], "Task"], [440, 443, ["accurate", "scene", "perception"], "Task"], [445, 447, ["knowledge", "graph"], "Method"], [463, 466, ["FCN", "based", "models"], "Method"], [480, 483, ["complex", "scene", "understanding"], "Task"], [494, 497, ["spatial", "pyramid", "pooling"], "Method"], [508, 511, ["overall", "scene", "interpretation"], "Task"], [512, 516, ["Spatial", "pyramid", "pooling", "network"], "Method"], [534, 538, ["pyramid", "scene", "parsing", "network"], "Method"], [539, 540, ["PSPNet"], "Method"], [546, 548, ["dilated", "FCN"], "Method"], [549, 551, ["pixel", "prediction"], "Task"], [563, 567, ["global", "pyramid", "pooling", "one"], "Method"], [577, 578, ["prediction"], "Task"], [585, 587, ["optimization", "strategy"], "Method"], [588, 591, ["deeply", "supervised", "loss"], "Task"], [640, 645, ["ImageNet", "scene", "parsing", "challenge", "2016"], "Task"], [672, 673, ["PSPNet"], "Method"], [678, 683, ["pixel", "-", "level", "prediction", "tasks"], "Task"], [694, 696, ["optical", "flow"], "Task"], [697, 699, ["depth", "estimation"], "Task"], [717, 721, ["pyramid", "scene", "parsing", "network"], "Method"], [729, 734, ["FCN", "based", "pixel", "prediction", "framework"], "Method"], [739, 741, ["optimization", "strategy"], "Method"], [742, 744, ["deep", "ResNet"], "Task"], [746, 749, ["deeply", "supervised", "loss"], "Task"], [759, 765, ["-", "the", "-", "art", "scene", "parsing"], "Task"], [766, 768, ["semantic", "segmentation"], "Task"], [789, 791, ["scene", "parsing"], "Task"], [799, 802, ["deep", "neural", "networks"], "Method"], [803, 808, ["pixel", "-", "level", "prediction", "tasks"], "Task"], [809, 811, ["scene", "parsing"], "Task"], [812, 814, ["semantic", "segmentation"], "Task"], [821, 825, ["fully", "-", "connected", "layer"], "Method"], [826, 827, ["classification"], "Method"], [829, 831, ["convolution", "layer"], "Method"], [838, 840, ["neural", "networks"], "Method"], [844, 846, ["dilated", "convolution"], "Method"], [850, 856, ["coarse", "-", "to", "-", "fine", "structure"], "Method"], [857, 859, ["deconvolution", "network"], "Method"], [869, 873, ["FCN", "and", "dilated", "network"], "Method"], [886, 891, ["multi", "-", "scale", "feature", "ensembling"], "Method"], [894, 896, ["deep", "networks"], "Method"], [926, 928, ["structure", "prediction"], "Task"], [933, 938, ["conditional", "random", "field", "(", "CRF"], "Method"], [940, 942, ["post", "processing"], "Task"], [945, 946, ["segmentation"], "Task"], [953, 959, ["end", "-", "to", "-", "end", "modeling"], "Method"], [967, 969, ["localization", "ability"], "Task"], [970, 972, ["scene", "parsing"], "Task"], [1004, 1007, ["diverse", "scene", "understanding"], "Task"], [1019, 1022, ["deep", "neural", "networks"], "Method"], [1028, 1031, ["object", "detection", "frameworks"], "Method"], [1035, 1038, ["global", "average", "pooling"], "Method"], [1039, 1040, ["FCN"], "Method"], [1042, 1044, ["semantic", "segmentation"], "Task"], [1053, 1055, ["global", "descriptors"], "Method"], [1070, 1072, ["global", "pooling"], "Method"], [1085, 1090, ["region", "-", "based", "context", "aggregation"], "Method"], [1092, 1096, ["pyramid", "scene", "parsing", "network"], "Method"], [1099, 1103, ["Pyramid", "Scene", "Parsing", "Network"], "Method"], [1111, 1114, ["representative", "failure", "cases"], "Task"], [1116, 1118, ["FCN", "methods"], "Method"], [1119, 1121, ["scene", "parsing"], "Task"], [1127, 1130, ["pyramid", "pooling", "module"], "Method"], [1138, 1142, ["pyramid", "scene", "parsing", "network"], "Method"], [1143, 1144, ["PSPNet"], "Method"], [1159, 1166, ["open", "-", "vocabulary", "object", "and", "stuff", "identification"], "Task"], [1167, 1170, ["complex", "scene", "parsing"], "Task"], [1177, 1180, ["ADE20", "K", "dataset"], "Material"], [1232, 1233, ["prediction"], "Task"], [1236, 1238, ["FCN", "baseline"], "Method"], [1247, 1251, ["complex", "-", "scene", "parsing"], "Task"], [1264, 1267, ["complex", "scene", "understanding"], "Task"], [1310, 1311, ["FCN"], "Method"], [1368, 1371, ["ADE20", "K", "dataset"], "Material"], [1375, 1376, ["classification"], "Task"], [1414, 1416, ["pixel", "error"], "Metric"], [1431, 1432, ["FCN"], "Method"], [1527, 1528, ["FCN"], "Method"], [1531, 1533, ["discontinuous", "prediction"], "Task"], [1620, 1622, ["deep", "network"], "Method"], [1637, 1639, ["scene", "parsing"], "Task"], [1642, 1645, ["Pyramid", "Pooling", "Module"], "Method"], [1656, 1659, ["pyramid", "pooling", "module"], "Method"], [1673, 1676, ["deep", "neural", "network"], "Method"], [1698, 1699, ["ResNet"], "Method"], [1718, 1719, ["CNN"], "Method"], [1754, 1757, ["global", "prior", "representation"], "Method"], [1758, 1761, ["Global", "average", "pooling"], "Method"], [1777, 1780, ["image", "classification", "tasks"], "Task"], [1788, 1790, ["semantic", "segmentation"], "Task"], [1799, 1801, ["ADE20", "K"], "Material"], [1888, 1892, ["scene", "/", "image", "classification"], "Task"], [1902, 1904, ["pyramid", "pooling"], "Method"], [1914, 1917, ["fully", "connected", "layer"], "Method"], [1918, 1919, ["classification"], "Task"], [1933, 1934, ["CNN"], "Method"], [1935, 1937, ["image", "classification"], "Task"], [1973, 1976, ["pyramid", "pooling", "module"], "Method"], [1977, 1981, ["global", "scene", "prior", "construction"], "Task"], [1992, 1995, ["deep", "neural", "network"], "Method"], [2011, 2014, ["pyramid", "pooling", "module"], "Method"], [2029, 2031, ["global", "pooling"], "Method"], [2053, 2055, ["pooled", "representation"], "Method"], [2066, 2069, ["pyramid", "pooling", "module"], "Method"], [2087, 2089, ["convolution", "layer"], "Method"], [2098, 2100, ["context", "representation"], "Method"], [2135, 2137, ["bilinear", "interpolation"], "Method"], [2184, 2187, ["pyramid", "pooling", "layer"], "Method"], [2197, 2202, ["varying", "-", "size", "pooling", "kernels"], "Method"], [2209, 2213, ["multi", "-", "stage", "kernels"], "Method"], [2222, 2225, ["pyramid", "pooling", "module"], "Method"], [2267, 2269, ["Network", "Architecture"], "Method"], [2271, 2274, ["pyramid", "pooling", "module"], "Method"], [2278, 2282, ["pyramid", "scene", "parsing", "network"], "Method"], [2283, 2284, ["PSPNet"], "Method"], [2311, 2314, ["pretrained", "ResNet", "model"], "Method"], [2316, 2319, ["dilated", "network", "strategy"], "Method"], [2327, 2330, ["feature", "map", "size"], "Metric"], [2357, 2360, ["pyramid", "pooling", "module"], "Method"], [2372, 2376, ["4", "-", "level", "pyramid"], "Method"], [2378, 2380, ["pooling", "kernels"], "Method"], [2426, 2428, ["convolution", "layer"], "Method"], [2444, 2445, ["PSPNet"], "Method"], [2448, 2451, ["global", "contextual", "prior"], "Method"], [2452, 2457, ["pixel", "-", "level", "scene", "parsing"], "Task"], [2459, 2462, ["pyramid", "pooling", "module"], "Method"], [2471, 2473, ["global", "pooling"], "Method"], [2477, 2479, ["computational", "cost"], "Metric"], [2481, 2482, ["PSPNet"], "Method"], [2491, 2494, ["dilated", "FCN", "network"], "Method"], [2496, 2502, ["end", "-", "to", "-", "end", "learning"], "Task"], [2504, 2508, ["global", "pyramid", "pooling", "module"], "Method"], [2520, 2522, ["Deep", "Supervision"], "Task"], [2523, 2527, ["ResNet", "-", "Based", "FCN"], "Method"], [2527, 2530, ["Deep", "pretrained", "networks"], "Method"], [2545, 2547, ["optimization", "difficulty"], "Metric"], [2551, 2553, ["image", "classification"], "Task"], [2554, 2555, ["ResNet"], "Method"], [2568, 2570, ["deep", "ResNet"], "Method"], [2603, 2604, ["optimization"], "Task"], [2606, 2608, ["deep", "network"], "Method"], [2623, 2627, ["deeply", "supervised", "ResNet101", "model"], "Method"], [2648, 2649, ["classifier"], "Method"], [2651, 2652, ["classifier"], "Method"], [2668, 2670, ["relay", "backpropagation"], "Method"], [2699, 2701, ["learning", "process"], "Task"], [2704, 2707, ["master", "branch", "loss"], "Method"], [2741, 2742, ["prediction"], "Task"], [2746, 2750, ["deeply", "supervised", "training", "strategy"], "Method"], [2751, 2755, ["ResNet", "-", "based", "FCN"], "Method"], [2769, 2771, ["ResNet", "model"], "Method"], [2779, 2781, ["learning", "strategy"], "Method"], [2801, 2803, ["scene", "parsing"], "Task"], [2820, 2825, ["ImageNet", "scene", "parsing", "challenge", "2016"], "Task"], [2826, 2831, ["PASCAL", "VOC", "2012", "semantic", "segmentation"], "Task"], [2845, 2848, ["deep", "learning", "system"], "Method"], [2862, 2865, ["public", "platform", "Caffe"], "Method"], [2873, 2878, ["poly", "”", "learning", "rate", "policy"], "Method"], [2880, 2882, ["learning", "rate"], "Metric"], [2892, 2894, ["learning", "rate"], "Metric"], [2909, 2911, ["iteration", "number"], "Metric"], [2919, 2921, ["ImageNet", "experiment"], "Material"], [2925, 2927, ["PASCAL", "VOC"], "Material"], [2931, 2932, ["Cityscapes"], "Material"], [2946, 2948, ["data", "augmentation"], "Task"], [2981, 2982, ["ImageNet"], "Material"], [2983, 2985, ["PASCAL", "VOC"], "Material"], [2988, 2991, ["data", "augmentation", "scheme"], "Method"], [3000, 3003, ["dilated", "convolution", "following"], "Method"], [3029, 3032, ["batch", "normalization", "layer"], "Method"], [3055, 3056, ["training"], "Task"], [3063, 3064, ["Caffe"], "Method"], [3067, 3068, ["branch"], "Method"], [3072, 3074, ["batch", "normalization"], "Task"], [3082, 3083, ["OpenMPI"], "Method"], [3086, 3088, ["auxiliary", "loss"], "Task"], [3100, 3105, ["ImageNet", "Scene", "Parsing", "Challenge", "2016"], "Task"], [3109, 3111, ["Evaluation", "Metrics"], "Metric"], [3112, 3115, ["ADE20", "K", "dataset"], "Material"], [3118, 3123, ["ImageNet", "scene", "parsing", "challenge", "2016"], "Task"], [3129, 3131, ["ADE20", "K"], "Material"], [3171, 3172, ["validation"], "Task"], [3173, 3174, ["testing"], "Task"], [3202, 3206, ["pixel", "-", "wise", "accuracy"], "Metric"], [3214, 3220, ["class", "-", "wise", "intersection", "over", "union"], "Metric"], [3221, 3224, ["Mean", "IoU", ")"], "Metric"], [3229, 3231, ["Ablation", "Study"], "Task"], [3232, 3233, ["PSPNet"], "Method"], [3235, 3236, ["PSPNet"], "Method"], [3248, 3251, ["max", "and", "average"], "Method"], [3252, 3253, ["pooling"], "Method"], [3267, 3269, ["dimension", "reduction"], "Method"], [3271, 3273, ["pooling", "operation"], "Method"], [3275, 3276, ["concatenation"], "Method"], [3285, 3287, ["average", "pooling"], "Method"], [3290, 3292, ["max", "pooling"], "Method"], [3296, 3297, ["Pooling"], "Method"], [3298, 3300, ["pyramid", "parsing"], "Method"], [3303, 3305, ["global", "pooling"], "Method"], [3307, 3309, ["dimension", "reduction"], "Method"], [3319, 3320, ["PSPNet"], "Method"], [3332, 3334, ["Mean", "IoU"], "Metric"], [3335, 3337, ["Pixel", "Acc"], "Metric"], [3343, 3346, ["global", "average", "pooling"], "Metric"], [3365, 3366, ["PSPNet"], "Method"], [3375, 3377, ["absolute", "improvement"], "Metric"], [3384, 3386, ["relative", "difference"], "Metric"], [3389, 3391, ["Ablation", "Study"], "Task"], [3392, 3394, ["Auxiliary", "Loss"], "Task"], [3401, 3403, ["learning", "process"], "Task"], [3406, 3407, ["learning"], "Task"], [3437, 3441, ["ResNet50", "-", "based", "FCN"], "Method"], [3442, 3444, ["dilated", "network"], "Method"], [3453, 3454, ["optimization"], "Task"], [3482, 3484, ["Mean", "IoU"], "Metric"], [3485, 3487, ["Pixel", "Acc"], "Metric"], [3508, 3510, ["Ablation", "Study"], "Task"], [3513, 3515, ["trained", "Model"], "Method"], [3515, 3518, ["Deeper", "neural", "networks"], "Method"], [3528, 3532, ["large", "scale", "data", "classification"], "Task"], [3536, 3537, ["PSPNet"], "Method"], [3548, 3549, ["ResNet"], "Method"], [3592, 3594, ["Mean", "IoU"], "Metric"], [3595, 3597, ["Pixel", "Acc"], "Metric"], [3617, 3618, ["PSPNet"], "Method"], [3623, 3626, ["depth", "ResNet", "models"], "Method"], [3651, 3653, ["ADE20", "K"], "Material"], [3675, 3684, ["ResNet269", "+", "DA", "+", "AL", "+", "PSP", "+", "MS"], "Method"], [3686, 3690, ["multi", "-", "scale", "testing"], "Task"], [3696, 3700, ["ResNet50", "with", "dilated", "network"], "Method"], [3703, 3705, ["MeanIoU", "34.28"], "Metric"], [3706, 3708, ["Pixel", "Acc"], "Metric"], [3722, 3723, ["ResNet"], "Method"], [3736, 3738, ["data", "augmentation"], "Method"], [3773, 3774, ["PSPNet"], "Method"], [3808, 3810, ["absolute", "improvement"], "Metric"], [3820, 3821, ["relativity"], "Metric"], [3826, 3827, ["ResNet269"], "Method"], [3840, 3845, ["multi", "-", "scale", "testing", "scheme"], "Method"], [3871, 3876, ["ImageNet", "scene", "parsing", "challenge", "2016"], "Task"], [3906, 3907, ["score"], "Metric"], [3918, 3923, ["multi", "-", "model", "ensemble", "submissions"], "Method"], [3962, 3963, ["PSPNet"], "Method"], [3968, 3969, ["FCN"], "Method"], [3978, 3979, ["parsing"], "Task"], [3984, 3985, ["ADE20K."], "Material"], [4000, 4003, ["PASCAL", "VOC", "2012"], "Material"], [4004, 4005, ["PSPNet"], "Method"], [4009, 4011, ["semantic", "segmentation"], "Task"], [4018, 4023, ["PASCAL", "VOC", "2012", "segmentation", "dataset"], "Material"], [4063, 4064, ["validation"], "Task"], [4065, 4066, ["testing"], "Task"], [4078, 4079, ["PSPNet"], "Method"], [4099, 4102, ["pre", "-", "training"], "Method"], [4113, 4116, ["MS", "-", "COCO"], "Material"], [4128, 4131, ["ResNet", "based", "frameworks"], "Method"], [4145, 4146, ["ResNet101"], "Method"], [4148, 4151, ["post", "-", "processing"], "Method"], [4152, 4153, ["CRF"], "Method"], [4156, 4157, ["PSPNet"], "Method"], [4177, 4178, ["PSPNet"], "Method"], [4188, 4191, ["VOC", "2012", "data"], "Material"], [4196, 4197, ["accuracy"], "Metric"], [4202, 4203, ["accuracy"], "Metric"], [4209, 4210, ["PSPNet"], "Method"], [4224, 4225, ["accuracy"], "Metric"], [4235, 4236, ["accuracy"], "Metric"], [4240, 4241, ["PSPNet"], "Method"], [4244, 4247, ["VOC", "2012", "data"], "Material"], [4266, 4269, ["based", "classification", "model"], "Method"], [4277, 4278, ["ResNet"], "Method"], [4310, 4311, ["FCRNs"], "Method"], [4312, 4313, ["LRR"], "Method"], [4315, 4316, ["DeepLab"], "Method"], [4331, 4334, ["post", "-", "processing"], "Method"], [4337, 4338, ["CRF"], "Method"], [4376, 4377, ["PSPNet"], "Method"], [4396, 4397, ["PSPNet"], "Method"], [4417, 4418, ["PSPNet"], "Method"], [4441, 4442, ["PSPNet"], "Method"], [4456, 4457, ["Cityscapes"], "Material"], [4457, 4458, ["Cityscapes"], "Material"], [4464, 4468, ["semantic", "urban", "scene", "understanding"], "Task"], [4511, 4512, ["validation"], "Task"], [4513, 4514, ["testing"], "Task"], [4585, 4586, ["ResNet101"], "Method"], [4588, 4589, ["DeepLab"], "Method"], [4610, 4611, ["PSPNet"], "Method"], [4631, 4632, ["accuracy"], "Metric"], [4669, 4673, ["pyramid", "scene", "parsing", "network"], "Method"], [4674, 4677, ["complex", "scene", "understanding"], "Task"], [4679, 4683, ["global", "pyramid", "pooling", "feature"], "Method"], [4694, 4697, ["supervised", "optimization", "strategy"], "Method"], [4698, 4703, ["ResNet", "-", "based", "FCN", "network"], "Method"], [4720, 4722, ["scene", "parsing"], "Task"], [4723, 4725, ["semantic", "segmentation"], "Task"], [4750, 4752, ["classification", "models"], "Method"], [253, 255, ["PASCAL", "VOC"], "Material"], [255, 257, ["semantic", "segmentation"], "Task"], [652, 654, ["PASCAL", "VOC"], "Material"], [655, 657, ["semantic", "segmentation"], "Task"], [666, 667, ["Cityscapes"], "Material"], [688, 689, ["CNN"], "Method"], [792, 794, ["semantic", "segmentation"], "Task"], [1062, 1064, ["ADE20", "K"], "Material"], [2804, 2806, ["semantic", "segmentation"], "Task"], [2836, 2837, ["Cityscapes"], "Material"], [3207, 3209, ["Pixel", "Acc"], "Metric"], [3589, 3590, ["score"], "Metric"], [3893, 3894, ["score"], "Metric"], [3925, 3926, ["score"], "Metric"], [4103, 4106, ["MS", "-", "COCO"], "Material"], [4135, 4137, ["semantic", "segmentation"], "Task"], [4215, 4218, ["MS", "-", "COCO"], "Material"], [4253, 4256, ["MS", "-", "COCO"], "Material"]]} {"doc_id": "325af39d281d5903a269c01fab8f53d7400a4c49", "text": ["document", ":", "ArtTrack", ":", "Articulated", "Multi", "-", "person", "Tracking", "in", "the", "Wild", "In", "this", "paper", "we", "propose", "an", "approach", "for", "articulated", "tracking", "of", "multiple", "people", "in", "unconstrained", "videos", ".", "Our", "starting", "point", "is", "a", "model", "that", "resembles", "existing", "architectures", "for", "single", "-", "frame", "pose", "estimation", "but", "is", "substantially", "faster", ".", "We", "achieve", "this", "in", "two", "ways", ":", "(", "1", ")", "by", "simplifying", "and", "sparsifying", "the", "body", "-", "part", "relationship", "graph", "and", "leveraging", "recent", "methods", "for", "faster", "inference", ",", "and", "(", "2", ")", "by", "offloading", "a", "substantial", "share", "of", "computation", "onto", "a", "feed", "-", "forward", "convolutional", "architecture", "that", "is", "able", "to", "detect", "and", "associate", "body", "joints", "of", "the", "same", "person", "even", "in", "clutter", ".", "We", "use", "this", "model", "to", "generate", "proposals", "for", "body", "joint", "locations", "and", "formulate", "articulated", "tracking", "as", "spatio", "-", "temporal", "grouping", "of", "such", "proposals", ".", "This", "allows", "to", "jointly", "solve", "the", "association", "problem", "for", "all", "people", "in", "the", "scene", "by", "propagating", "evidence", "from", "strong", "detections", "through", "time", "and", "enforcing", "constraints", "that", "each", "proposal", "can", "be", "assigned", "to", "one", "person", "only", ".", "We", "report", "results", "on", "a", "public", "“", "MPII", "Human", "Pose", "”", "benchmark", "and", "on", "a", "new", "“", "MPII", "Video", "Pose", "”", "dataset", "of", "image", "sequences", "with", "multiple", "people", ".", "We", "demonstrate", "that", "our", "model", "achieves", "state", "-", "of", "-", "the", "-", "art", "results", "while", "using", "only", "a", "fraction", "of", "time", "and", "is", "able", "to", "leverage", "temporal", "information", "to", "improve", "state", "-", "of", "-", "the", "-", "art", "for", "crowded", "scenes", ".", "section", ":", "Introduction", "This", "paper", "addresses", "the", "task", "of", "articulated", "human", "pose", "tracking", "in", "monocular", "video", ".", "We", "focus", "on", "scenes", "of", "realistic", "complexity", "that", "often", "include", "fast", "motions", ",", "large", "variability", "in", "appearance", "and", "clothing", ",", "and", "person", "-", "person", "occlusions", ".", "A", "successful", "approach", "must", "thus", "identify", "the", "number", "of", "people", "in", "each", "video", "frame", ",", "determine", "locations", "of", "the", "joints", "of", "each", "person", "and", "associate", "the", "joints", "over", "time", ".", "One", "of", "the", "key", "challenges", "in", "such", "scenes", "is", "that", "people", "might", "overlap", "and", "only", "a", "subset", "of", "joints", "of", "the", "person", "might", "be", "visible", "in", "each", "frame", "either", "due", "to", "person", "-", "person", "occlusion", "or", "truncation", "by", "image", "boundaries", "(", ".", "Fig", ".", "[", "reference", "]", ")", ".", "Arguably", ",", "resolving", "such", "cases", "correctly", "requires", "reasoning", "beyond", "purely", "geometric", "information", "on", "the", "arrangement", "of", "body", "joints", "in", "the", "image", ",", "and", "requires", "incorporation", "of", "a", "variety", "of", "image", "cues", "and", "joint", "modeling", "of", "several", "persons", ".", "The", "design", "of", "our", "model", "is", "motivated", "by", "two", "factors", ".", "We", "would", "like", "to", "leverage", "bottom", "-", "up", "end", "-", "to", "-", "end", "learning", "to", "directly", "capture", "image", "information", ".", "At", "the", "same", "time", "we", "aim", "to", "address", "a", "complex", "multi", "-", "person", "articulated", "tracking", "problem", "that", "does", "not", "naturally", "lend", "itself", "to", "an", "end", "-", "to", "-", "end", "prediction", "task", "and", "for", "which", "training", "data", "is", "not", "available", "in", "the", "amounts", "usually", "required", "for", "end", "-", "to", "-", "end", "learning", ".", "To", "leverage", "the", "available", "image", "information", "we", "learn", "a", "model", "for", "associating", "a", "body", "joint", "to", "a", "specific", "person", "in", "an", "end", "-", "to", "-", "end", "fashion", "relying", "on", "a", "convolutional", "network", ".", "We", "then", "incorporate", "these", "part", "-", "to", "-", "person", "association", "responses", "into", "a", "framework", "for", "jointly", "reasoning", "about", "assignment", "of", "body", "joints", "within", "the", "image", "and", "over", "time", ".", "To", "that", "end", "we", "use", "the", "graph", "partitioning", "formulation", "that", "has", "been", "used", "for", "people", "tracking", "and", "pose", "estimation", "in", "the", "past", ",", "but", "has", "not", "been", "shown", "to", "enable", "articulated", "people", "tracking", ".", "To", "facilitate", "efficient", "inference", "in", "video", "we", "resort", "to", "fast", "inference", "methods", "based", "on", "local", "combinatorial", "optimization", "and", "aim", "for", "a", "sparse", "model", "that", "keeps", "the", "number", "of", "connections", "between", "variables", "to", "a", "minimum", ".", "As", "we", "demonstrate", ",", "in", "combination", "with", "feed", "-", "forward", "reasoning", "for", "joint", "-", "to", "-", "person", "association", "this", "allows", "us", "to", "achieve", "substantial", "speed", "-", "ups", "compared", "to", "state", "-", "of", "-", "the", "-", "art", "while", "maintaining", "the", "same", "level", "of", "accuracy", ".", "The", "main", "contribution", "of", "this", "work", "is", "a", "new", "articulated", "tracking", "model", "that", "operates", "by", "bottom", "-", "up", "assembly", "of", "part", "detections", "within", "each", "frame", "and", "over", "time", ".", "In", "contrast", "to", "this", "model", "is", "suitable", "for", "scenes", "with", "an", "unknown", "number", "of", "subjects", "and", "reasons", "jointly", "across", "multiple", "people", "incorporating", "inter", "-", "person", "exclusion", "constraints", "and", "propagating", "strong", "observations", "to", "neighboring", "frames", ".", "Our", "second", "contribution", "is", "a", "formulation", "for", "single", "-", "frame", "pose", "estimation", "that", "relies", "on", "a", "sparse", "graph", "between", "body", "parts", "and", "a", "mechanism", "for", "generating", "body", "-", "part", "proposals", "conditioned", "on", "a", "person", "’s", "location", ".", "This", "is", "in", "contrast", "to", "state", "-", "of", "-", "the", "-", "art", "approaches", "that", "perform", "expensive", "inference", "in", "a", "full", "graph", "and", "rely", "on", "generic", "bottom", "-", "up", "proposals", ".", "We", "demonstrate", "that", "a", "sparse", "model", "with", "a", "few", "spatial", "edges", "performs", "competitively", "with", "a", "fully", "-", "connected", "model", "while", "being", "much", "more", "efficient", ".", "Notably", ",", "a", "simple", "model", "that", "operates", "in", "top", "-", "down", "/", "bottom", "-", "up", "fashion", "exceeds", "the", "performance", "of", "a", "fully", "-", "connected", "model", "while", "being", "x", "faster", "at", "inference", "time", "(", "cf", ".", "Tab", ".", "[", "reference", "]", ")", ".", "This", "is", "due", "to", "offloading", "of", "a", "large", "share", "of", "the", "reasoning", "about", "body", "-", "part", "association", "onto", "a", "feed", "-", "forward", "convolutional", "architecture", ".", "Finally", ",", "we", "contribute", "a", "new", "challenging", "dataset", "for", "evaluation", "of", "articulated", "body", "joint", "tracking", "in", "crowded", "realistic", "environments", "with", "multiple", "overlapping", "people", ".", "Related", "work", ".", "Convolutional", "networks", "have", "emerged", "as", "an", "effective", "approach", "to", "localizing", "body", "joints", "of", "people", "in", "images", "and", "have", "also", "been", "extended", "for", "joint", "estimation", "of", "body", "configurations", "over", "time", ",", "and", "3D", "pose", "estimation", "in", "outdoor", "environments", "in", "multi", "-", "camera", "setting", ".", "Current", "approaches", "are", "increasingly", "effective", "for", "estimating", "body", "configurations", "of", "single", "people", "achieving", "high", "accuracies", "on", "this", "task", ",", "but", "are", "still", "failing", "on", "fast", "moving", "and", "articulated", "limbs", ".", "More", "complex", "recent", "models", "jointly", "reason", "about", "entire", "scenes", ",", "but", "are", "too", "complex", "and", "inefficient", "to", "directly", "generalize", "to", "image", "sequences", ".", "Recent", "feed", "-", "forward", "models", "are", "able", "to", "jointly", "infer", "body", "joints", "of", "the", "same", "person", "and", "even", "operate", "over", "time", "but", "consider", "isolated", "persons", "only", "and", "do", "not", "generalize", "to", "the", "case", "of", "multiple", "overlapping", "people", ".", "Similarly", ",", "consider", "a", "simplified", "task", "of", "tracking", "upper", "body", "poses", "of", "isolated", "upright", "individuals", ".", "We", "build", "on", "recent", "CNN", "detectors", "that", "are", "effective", "in", "localizing", "body", "joints", "in", "cluttered", "scenes", "and", "explore", "different", "mechanisms", "for", "assembling", "the", "joints", "into", "multiple", "person", "configurations", ".", "To", "that", "end", "we", "rely", "on", "a", "graph", "partitioning", "approach", "closely", "related", "to", ".", "In", "contrast", "to", "who", "focus", "on", "pedestrian", "tracking", ",", "and", "who", "perform", "single", "frame", "multi", "-", "person", "pose", "estimation", ",", "we", "solve", "a", "more", "complex", "problem", "of", "articulated", "multi", "-", "person", "pose", "tracking", ".", "Earlier", "approaches", "to", "articulated", "pose", "tracking", "in", "monocular", "videos", "rely", "on", "hand", "-", "crafted", "image", "representations", "and", "focus", "on", "simplified", "tasks", ",", "such", "as", "tracking", "upper", "body", "poses", "of", "frontal", "isolated", "people", ",", "or", "tracking", "walking", "pedestrians", "with", "little", "degree", "of", "articulation", ".", "In", "contrast", ",", "we", "address", "a", "harder", "problem", "of", "multi", "-", "person", "articulated", "pose", "tracking", "and", "do", "not", "make", "assumptions", "about", "the", "type", "of", "body", "motions", "or", "activities", "of", "people", ".", "Our", "approach", "is", "closely", "related", "to", "who", "propose", "a", "similar", "formulation", "based", "on", "graph", "partitioning", ".", "Our", "approach", "differs", "from", "primarily", "in", "the", "type", "of", "body", "-", "part", "proposals", "and", "the", "structure", "of", "the", "spatio", "-", "temporal", "graph", ".", "In", "our", "approach", "we", "introduce", "a", "person", "-", "conditioned", "model", "that", "is", "trained", "to", "associate", "body", "parts", "of", "a", "specific", "person", "already", "at", "the", "detection", "stage", ".", "This", "is", "in", "contrast", "to", "the", "approach", "of", "that", "relies", "on", "the", "generic", "body", "-", "part", "detectors", ".", "Overview", ".", "Our", "model", "consists", "of", "the", "two", "components", ":", "(", "1", ")", "a", "convolutional", "network", "for", "generating", "body", "part", "proposals", "and", "(", "2", ")", "an", "approach", "to", "group", "the", "proposals", "into", "spatio", "-", "temporal", "clusters", ".", "In", "Sec", ".", "[", "reference", "]", "we", "introduce", "a", "general", "formulation", "for", "multi", "-", "target", "tracking", "that", "follows", "and", "allows", "us", "to", "define", "pose", "estimation", "and", "articulated", "tracking", "in", "a", "unified", "framework", ".", "We", "then", "describe", "the", "details", "of", "our", "articulated", "tracking", "approach", "in", "Sec", ".", "[", "reference", "]", ",", "and", "introduce", "two", "variants", "of", "our", "formulation", ":", "bottom", "-", "up", "(", "BU", ")", "and", "top", "-", "down", "/", "bottom", "-", "up", "(", "TD", "/", "BU", ")", ".", "We", "present", "experimental", "results", "in", "Sec", ".", "[", "reference", "]", ".", "section", ":", "Tracking", "by", "Spatio", "-", "temporal", "Grouping", "Our", "body", "part", "detector", "generates", "a", "set", "of", "proposals", "for", "each", "frame", "of", "the", "video", ".", "Each", "proposal", "is", "given", "by", ",", "where", "denotes", "the", "index", "of", "the", "video", "frame", ",", "is", "the", "spatial", "location", "of", "the", "proposal", "in", "image", "coordinates", ",", "is", "the", "probability", "of", "correct", "detection", ",", "and", "is", "the", "type", "of", "the", "body", "joint", "(", "ankle", "or", "shoulder", ")", ".", "Let", "be", "a", "graph", "whose", "nodes", "are", "the", "joint", "detections", "in", "a", "video", "and", "whose", "edges", "connect", "pairs", "of", "detections", "that", "hypothetically", "correspond", "to", "the", "same", "target", ".", "The", "output", "of", "the", "tracking", "algorithm", "is", "a", "subgraph", "of", ",", "where", "is", "a", "subset", "of", "nodes", "after", "filtering", "redundant", "and", "erroneous", "detections", "and", "are", "edges", "linking", "nodes", "corresponding", "to", "the", "same", "target", ".", "We", "specify", "via", "binary", "variables", "and", "that", "define", "subsets", "of", "edges", "and", "nodes", "included", "in", ".", "In", "particular", "each", "track", "will", "correspond", "to", "a", "connected", "component", "in", ".", "As", "a", "general", "way", "to", "introduce", "constraints", "on", "edge", "configurations", "that", "correspond", "to", "a", "valid", "tracking", "solution", "we", "introduce", "a", "set", "and", "define", "a", "combination", "of", "edge", "and", "node", "indicator", "variables", "to", "be", "feasible", "if", "and", "only", "if", ".", "An", "example", "of", "a", "constraint", "encoded", "through", "is", "that", "endpoint", "nodes", "of", "an", "edge", "included", "by", "must", "also", "be", "included", "by", ".", "Note", "that", "the", "variables", "and", "are", "coupled", "though", ".", "Moreover", ",", "assuming", "that", "we", "are", "free", "to", "set", "components", "of", "and", "independently", "to", "maximize", "the", "tracking", "objective", ".", "Given", "image", "observations", "we", "compute", "a", "set", "of", "features", "for", "each", "node", "and", "edge", "in", "the", "graph", ".", "We", "denote", "such", "node", "and", "edge", "features", "as", "and", "respectively", ".", "Assuming", "independence", "of", "the", "feature", "vectors", "the", "conditional", "probability", "of", "indicator", "functions", "of", "nodes", "and", "of", "edges", "given", "features", "and", "and", "given", "a", "feasible", "set", "is", "given", "by", "where", "assigns", "a", "constant", "non", "-", "zero", "probability", "to", "every", "feasible", "solution", "and", "is", "equal", "to", "zero", "otherwise", ".", "Minimizing", "the", "negative", "log", "-", "likelihood", "of", "Eq", ".", "[", "reference", "]", "is", "equivalent", "to", "solving", "the", "following", "integer", "-", "linear", "program", ":", "where", "is", "the", "cost", "of", "retaining", "as", "part", "of", "the", "solution", ",", "and", "is", "the", "cost", "of", "assigning", "the", "detections", "linked", "by", "an", "edge", "to", "the", "same", "track", ".", "We", "define", "the", "set", "of", "constraints", "as", "in", ":", "Jointly", "with", "the", "objective", "in", "Eq", ".", "[", "reference", "]", "the", "constraints", "(", "[", "reference", "]", ")", "-", "(", "[", "reference", "]", ")", "define", "an", "instance", "of", "the", "minimum", "cost", "subgraph", "multicut", "problem", ".", "The", "constraints", "(", "[", "reference", "]", ")", "and", "(", "[", "reference", "]", ")", "ensure", "that", "assignment", "of", "node", "and", "edge", "variables", "is", "consistent", ".", "The", "constraint", "(", "[", "reference", "]", ")", "ensures", "that", "for", "every", "two", "nodes", "either", "all", "or", "none", "of", "the", "paths", "between", "these", "nodes", "in", "graph", "are", "contained", "in", "one", "of", "the", "connected", "components", "of", "subgraph", ".", "This", "constraint", "is", "necessary", "to", "unambigously", "assign", "person", "identity", "to", "a", "body", "part", "proposal", "based", "on", "its", "membership", "in", "a", "specific", "connnected", "component", "of", ".", "section", ":", "Articulated", "Multi", "-", "person", "Tracking", "In", "Sec", ".", "[", "reference", "]", "we", "introduced", "a", "general", "framework", "for", "multi", "-", "object", "tracking", "by", "solving", "an", "instance", "of", "the", "subgraph", "multicut", "problem", ".", "The", "subgraph", "multicut", "problem", "is", "NP", "-", "hard", ",", "but", "recent", "work", "has", "shown", "that", "efficient", "approximate", "inference", "is", "possible", "with", "local", "search", "methods", ".", "The", "framework", "allows", "for", "a", "variety", "of", "graphs", "and", "connectivity", "patterns", ".", "Simpler", "connectivity", "allows", "for", "faster", "and", "more", "efficient", "processing", "at", "the", "cost", "of", "ignoring", "some", "of", "the", "potentially", "informative", "dependencies", "between", "model", "variables", ".", "Our", "goal", "is", "to", "design", "a", "model", "that", "is", "efficient", ",", "with", "as", "few", "edges", "as", "possible", ",", "yet", "effective", "in", "crowded", "scenes", ",", "and", "that", "allows", "us", "to", "model", "temporal", "continuity", "and", "inter", "-", "person", "exclusion", ".", "Our", "articulated", "tracking", "approach", "proceeds", "by", "constructing", "a", "graph", "that", "couples", "body", "part", "proposals", "within", "the", "same", "frame", "and", "across", "neighboring", "frames", ".", "In", "general", "the", "graph", "will", "have", "three", "types", "of", "edges", ":", "(", "1", ")", "cross", "-", "type", "edges", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "a", ")", "and", "Fig", ".", "[", "reference", "]", "(", "b", ")", "that", "connect", "two", "parts", "of", "different", "types", ",", "(", "2", ")", "same", "-", "type", "edges", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "b", ")", "that", "connect", "two", "nodes", "of", "the", "same", "type", "in", "the", "same", "image", ",", "and", "(", "3", ")", "temporal", "edges", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "c", ")", "that", "connect", "nodes", "in", "the", "neighboring", "frames", ".", "We", "now", "define", "two", "variants", "of", "our", "model", "that", "we", "denote", "as", "Bottom", "-", "Up", "(", "BU", ")", "and", "Top", "-", "Down", "/", "Bottom", "-", "Up", "(", "TD", "/", "BU", ")", ".", "In", "the", "BU", "model", "the", "body", "part", "proposals", "are", "generated", "with", "our", "publicly", "available", "convolutional", "part", "detector", ".", "In", "the", "TD", "/", "BU", "model", "we", "substitute", "these", "generic", "part", "detectors", "with", "a", "new", "convolutional", "body", "-", "part", "detector", "that", "is", "trained", "to", "output", "consistent", "body", "configurations", "conditioned", "on", "the", "person", "location", ".", "This", "alows", "to", "further", "reduce", "the", "complexity", "of", "the", "model", "graph", "since", "the", "task", "of", "associating", "body", "parts", "is", "addressed", "within", "the", "proposal", "mechanism", ".", "As", "we", "show", "in", "Sec", ".", "[", "reference", "]", "this", "leads", "to", "considerable", "gains", "in", "performance", "and", "allows", "for", "faster", "inference", ".", "Note", "that", "the", "BU", "and", "TD", "/", "BU", "models", "have", "identical", "same", "-", "type", "and", "temporal", "pairwise", "terms", ",", "but", "differ", "in", "the", "form", "of", "cross", "-", "type", "pairwise", "terms", ",", "and", "the", "connectivity", "of", "the", "nodes", "in", ".", "For", "both", "models", "we", "rely", "on", "the", "solver", "from", "for", "inference", ".", "subsection", ":", "Bottom", "-", "Up", "Model", "(", "BU", ")", ".", "For", "each", "body", "part", "proposal", "the", "detector", "outputs", "image", "location", ",", "probability", "of", "detection", ",", "and", "a", "label", "that", "indicates", "the", "type", "of", "the", "detected", "part", "(", "shoulder", "or", "ankle", ")", ".", "We", "directly", "use", "the", "probability", "of", "detection", "to", "derive", "the", "unary", "costs", "in", "Eq", ".", "[", "reference", "]", "as", ".", "Image", "features", "in", "this", "case", "correspond", "to", "the", "image", "representation", "generated", "by", "the", "convolutional", "network", ".", "We", "consider", "two", "connectivity", "patterns", "for", "nodes", "in", "the", "graph", ".", "We", "either", "define", "edges", "for", "every", "pair", "of", "proposals", "which", "results", "in", "a", "fully", "connected", "graph", "in", "each", "image", ".", "Alternatively", "we", "obtain", "a", "sparse", "version", "of", "the", "model", "by", "defining", "edges", "for", "a", "subset", "of", "part", "types", "only", "as", "is", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "a", ")", ".", "The", "rationale", "behind", "the", "sparse", "version", "is", "to", "obtain", "a", "simpler", "and", "faster", "version", "of", "the", "model", "by", "omitting", "edges", "between", "parts", "that", "carry", "little", "information", "about", "each", "other", "’s", "image", "location", "(", "left", "ankle", "and", "right", "arm", ")", ".", "Edge", "costs", ".", "In", "our", "Bottom", "-", "Up", "model", "the", "cost", "of", "the", "edges", "connecting", "two", "body", "part", "detections", "and", "is", "defined", "as", "a", "function", "of", "the", "detection", "types", "and", ".", "Following", "we", "thus", "train", "for", "each", "pair", "of", "part", "types", "a", "regression", "function", "that", "predicts", "relative", "image", "location", "of", "the", "parts", "in", "the", "pair", ".", "The", "cost", "is", "given", "by", "the", "output", "of", "the", "logistic", "regression", "given", "the", "features", "computed", "from", "offset", "and", "angle", "of", "the", "predicted", "and", "actual", "location", "of", "the", "other", "joint", "in", "the", "pair", ".", "We", "refer", "to", "for", "more", "details", "on", "these", "pairwise", "terms", ".", "Note", "that", "our", "model", "generalizes", "in", "that", "the", "edge", "cost", "depends", "on", "the", "type", "of", "nodes", "linked", "by", "the", "edge", ".", "It", "also", "generalizes", "by", "allowing", "to", "be", "sparse", ".", "This", "is", "achieved", "by", "reformulating", "the", "model", "with", "a", "more", "general", "type", "of", "cycle", "constraint", "(", "[", "reference", "]", ")", ",", "in", "contrast", "to", "simple", "triangle", "inequalities", "used", "in", ".", "subsection", ":", "Top", "-", "Down", "/", "Bottom", "-", "up", "Model", "(", "TD", "/", "BU", ")", "We", "now", "introduce", "a", "version", "of", "our", "model", "that", "operates", "by", "first", "generating", "body", "part", "proposals", "conditioned", "on", "the", "locations", "of", "people", "in", "the", "image", "and", "then", "performing", "joint", "reasoning", "to", "group", "these", "proposals", "into", "spatio", "-", "temporal", "clusters", "corresponding", "to", "different", "people", ".", "We", "follow", "the", "intuition", "that", "it", "is", "considerably", "easier", "to", "identify", "and", "detect", "individual", "people", "(", "e.g.", "by", "detecting", "their", "heads", ")", "compared", "to", "correctly", "associating", "body", "parts", "such", "as", "ankles", "and", "wrists", "to", "each", "person", ".", "We", "select", "person", "’s", "head", "as", "a", "root", "part", "that", "is", "responsible", "for", "representing", "the", "person", "location", ",", "and", "delegate", "the", "task", "of", "identifying", "body", "parts", "of", "the", "person", "corresponding", "to", "a", "head", "location", "to", "a", "convolutional", "network", ".", "The", "structure", "of", "TD", "/", "BU", "model", "is", "illustrated", "in", "Fig", ".", "[", "reference", "]", "(", "b", ")", "for", "the", "simplified", "case", "of", "two", "distinct", "head", "detections", ".", "Let", "us", "denote", "the", "set", "of", "all", "root", "part", "detections", "as", ".", "For", "each", "pair", "of", "the", "root", "nodes", "we", "explicitly", "set", "the", "corresponding", "edge", "indicator", "variables", ".", "This", "implements", "a", "“", "must", "-", "not", "-", "link", "”", "constraint", "between", "these", "nodes", ",", "and", "in", "combination", "with", "the", "cycle", "inequality", "(", "[", "reference", "]", ")", "implies", "that", "each", "proposal", "can", "be", "connected", "to", "one", "of", "the", "“", "person", "nodes", "”", "only", ".", "The", "cost", "for", "an", "edge", "connecting", "detection", "proposal", "and", "a", "“", "person", "node", "”", "is", "based", "on", "the", "conditional", "distribution", "generated", "by", "the", "convolutional", "network", ".", "The", "output", "of", "such", "network", "is", "a", "set", "of", "conditional", "distributions", ",", "one", "for", "each", "node", "type", ".", "We", "augment", "the", "graph", "with", "attractive", "/", "repulsive", "and", "temporal", "terms", "as", "described", "in", "Sec", ".", "[", "reference", "]", "and", "Sec", ".", "[", "reference", "]", "and", "set", "the", "unary", "costs", "for", "all", "indicator", "variables", "to", "a", "constant", ".", "Any", "proposal", "not", "connected", "to", "any", "of", "the", "root", "nodes", "is", "excluded", "from", "the", "final", "solution", ".", "We", "use", "the", "solver", "from", "for", "consistency", ",", "but", "a", "simpler", "KL", "-", "based", "solver", "as", "in", "could", "be", "used", "as", "well", "since", "the", "TD", "/", "BU", "model", "effectively", "ignores", "the", "unary", "variables", ".", "The", "processing", "stages", "of", "TD", "/", "BU", "model", "are", "shown", "in", "Fig", ".", "[", "reference", "]", ".", "Note", "that", "the", "body", "-", "part", "heatmaps", "change", "depending", "on", "the", "person", "-", "identity", "signal", "provided", "by", "the", "person", "’s", "neck", ",", "and", "that", "the", "bottom", "-", "up", "step", "was", "able", "to", "correct", "the", "predictions", "on", "the", "forearms", "of", "the", "front", "person", ".", "Implementation", "details", ".", "For", "head", "detection", ",", "we", "use", "a", "version", "of", "our", "model", "that", "contains", "the", "two", "head", "parts", "(", "neck", "and", "head", "top", ")", ".", "This", "makes", "our", "TD", "/", "BU", "model", "related", "to", "the", "hierarchical", "model", "defined", "in", "that", "also", "uses", "easier", "-", "to", "-", "detect", "parts", "to", "guide", "the", "rest", "of", "the", "inference", "process", ".", "However", "here", "we", "replace", "all", "the", "stages", "in", "the", "hierarchical", "inference", "except", "the", "first", "one", "with", "a", "convolutional", "network", ".", "The", "structure", "of", "the", "convolutional", "network", "used", "to", "generate", "person", "-", "conditioned", "proposals", "is", "shown", "on", "Fig", ".", "[", "reference", "]", ".", "The", "network", "uses", "the", "ResNet", "-", "101", "from", "that", "we", "modify", "to", "bring", "the", "stride", "of", "the", "network", "down", "to", "8", "pixels", ".", "The", "network", "generates", "predictions", "for", "all", "body", "parts", "after", "the", "conv4_4", "block", ".", "We", "use", "the", "cross", "-", "entropy", "binary", "classification", "loss", "at", "this", "stage", "to", "predict", "the", "part", "heatmaps", ".", "At", "each", "training", "iteration", "we", "forward", "pass", "an", "image", "with", "multiple", "people", "potentially", "in", "close", "proximity", "to", "each", "other", ".", "We", "select", "a", "single", "person", "from", "the", "image", "and", "condition", "the", "network", "on", "the", "person", "’s", "neck", "location", "by", "zeroing", "out", "the", "heatmap", "of", "the", "neck", "joint", "outside", "the", "ground", "-", "truth", "region", ".", "We", "then", "pass", "the", "neck", "heatmap", "through", "a", "convolutional", "layer", "to", "match", "the", "dimensionality", "of", "the", "feature", "channels", "and", "add", "them", "to", "the", "main", "stream", "of", "the", "ResNet", ".", "We", "finally", "add", "a", "joint", "prediction", "layer", "at", "the", "end", "of", "the", "network", "with", "a", "loss", "that", "considers", "predictions", "to", "be", "correct", "only", "if", "they", "correspond", "to", "the", "body", "joints", "of", "the", "selected", "person", ".", "Spatial", "propagation", "(", "SP", ")", ".", "In", "our", "network", "the", "person", "identity", "signal", "is", "provided", "by", "the", "location", "of", "the", "head", ".", "In", "principle", "the", "receptive", "field", "size", "of", "the", "network", "is", "large", "enough", "to", "propagate", "this", "signal", "to", "all", "body", "parts", ".", "However", "we", "found", "that", "it", "is", "useful", "to", "introduce", "an", "additional", "mechanism", "to", "propagate", "the", "person", "identity", "signal", ".", "To", "that", "end", "we", "inject", "intermediate", "supervision", "layers", "for", "individual", "body", "parts", "arranged", "in", "the", "order", "of", "kinematic", "proximity", "to", "the", "root", "joint", "(", "Fig", ".", "[", "reference", "]", ")", ".", "We", "place", "prediction", "layers", "for", "shoulders", "at", "conv4_8", ",", "for", "elbows", "and", "hips", "at", "conv4_14", "and", "for", "knees", "at", "conv4_18", ".", "We", "empirically", "found", "that", "such", "an", "explicit", "form", "of", "spatial", "propagation", "significantly", "improves", "performance", "on", "joints", "such", "as", "ankles", ",", "that", "are", "typically", "far", "from", "the", "head", "in", "the", "image", "space", "(", "see", "Tab", ".", "[", "reference", "]", "for", "details", ")", ".", "Training", ".", "We", "use", "Caffe", "’s", "ResNet", "implementation", "and", "initialize", "from", "the", "ImageNet", "-", "pre", "-", "trained", "models", ".", "Networks", "are", "trained", "on", "the", "MPII", "Human", "Pose", "dataset", "with", "SGD", "for", "1", "M", "iterations", "with", "stepwise", "learning", "rate", "(", "lr=0.002", "for", "400k", ",", "lr=0.0002", "for", "300k", "and", "lr=0.0001", "for", "300k", ")", ".", "subsection", ":", "Attractive", "/", "Repulsive", "Edges", "Attractive", "/", "repulsive", "edges", "are", "defined", "between", "two", "proposals", "of", "the", "same", "type", "within", "the", "same", "image", ".", "The", "costs", "of", "these", "edges", "is", "inversely", "-", "proportional", "to", "distance", ".", "The", "decision", "to", "group", "two", "nodes", "is", "made", "based", "on", "the", "evidence", "from", "the", "entire", "image", ",", "which", "is", "in", "contrast", "to", "typical", "non", "-", "maximum", "suppression", "based", "on", "the", "state", "of", "just", "two", "detections", ".", "Inversely", ",", "these", "edges", "prevent", "grouping", "of", "multiple", "distant", "hypothesis", "of", "the", "same", "type", ",", "prevent", "merging", "two", "heads", "of", "different", "people", ".", "subsection", ":", "Temporal", "Model", "Regardless", "of", "the", "type", "of", "within", "frame", "model", "(", "BU", "or", "TD", "/", "BU", ")", "we", "rely", "on", "the", "same", "type", "of", "temporal", "edges", "that", "connect", "nodes", "of", "the", "same", "type", "in", "adjacent", "frames", ".", "We", "derive", "the", "costs", "for", "such", "temporal", "edges", "via", "logistic", "regression", ".", "Given", "the", "feature", "vector", "the", "probability", "that", "the", "two", "proposals", "and", "in", "adjacent", "frames", "correspond", "to", "the", "same", "body", "part", "is", "given", "by", ":", ",", "where", ",", "and", "-", ",", "is", "Euclidean", "distance", "between", "the", "SIFT", "descriptors", "computed", "at", "and", ",", "and", "and", "measure", "the", "agreement", "with", "the", "dense", "motion", "field", "computed", "with", "the", "DeepMatching", "approach", "of", ".", "For", "SIFT", "features", "we", "specify", "the", "location", "of", "the", "detection", "proposal", ",", "but", "rely", "on", "SIFT", "to", "identify", "the", "local", "orientation", ".", "In", "cases", "with", "multiple", "local", "maxima", "in", "orientation", "estimation", "we", "compute", "SIFT", "descriptor", "for", "each", "orientation", "and", "set", "to", "the", "minimal", "distance", "among", "all", "pairs", "of", "descriptors", ".", "We", "found", "that", "this", "makes", "the", "SIFT", "distance", "more", "robust", "in", "the", "presence", "of", "rotations", "of", "the", "body", "limbs", ".", "We", "define", "the", "features", "and", "as", "in", ".", "Let", "be", "an", "squared", "image", "region", "centered", "on", "the", "part", "proposal", ".", "We", "define", "as", "a", "ratio", "of", "the", "number", "of", "point", "correspondences", "between", "the", "regions", "and", "and", "the", "total", "number", "of", "point", "correspondences", "in", "either", "of", "them", ".", "Specifically", ",", "let", "be", "a", "set", "of", "point", "correspondences", "between", "the", "two", "images", "computed", "with", "DeepMatching", ",", "where", "and", "and", "denote", "the", "corresponding", "points", "in", "the", "first", "and", "second", "image", "respectively", ".", "Using", "this", "notation", "we", "define", ":", "The", "rationale", "behind", "computing", "by", "aggregating", "across", "multiple", "correspondences", "is", "to", "make", "the", "feature", "robust", "to", "outliers", "and", "to", "inaccuracies", "in", "body", "part", "detection", ".", "is", "defined", "analogously", ",", "but", "using", "the", "DeepMatching", "correspondences", "obtained", "by", "inverting", "the", "order", "of", "images", ".", "Discussion", ".", "As", "we", "demonstrate", "in", "Sec", ".", "[", "reference", "]", ",", "we", "found", "the", "set", "of", "features", "described", "above", "to", "be", "complementary", "to", "each", "other", ".", "Euclidean", "distance", "between", "proposals", "is", "informative", "for", "finding", "correspondences", "for", "slow", "motions", ",", "but", "fails", "for", "faster", "motions", "and", "in", "the", "presence", "of", "multiple", "people", ".", "DeepMatching", "is", "usually", "effective", "in", "finding", "corresponding", "regions", "between", "the", "two", "images", ",", "but", "occasionally", "fails", "in", "the", "case", "of", "sudden", "background", "changes", "due", "to", "fast", "motion", "or", "large", "changes", "in", "body", "limb", "orientation", ".", "In", "these", "cases", "SIFT", "is", "often", "still", "able", "to", "provide", "a", "meaningful", "measure", "of", "similarity", "due", "to", "its", "rotation", "invariance", ".", "section", ":", "Experiments", "subsection", ":", "Datasets", "and", "evaluation", "measure", "Single", "frame", ".", "We", "evaluate", "our", "single", "frame", "models", "on", "the", "MPII", "Multi", "-", "Person", "dataset", ".", "We", "report", "all", "intermediate", "results", "on", "a", "validation", "set", "of", "images", "sampled", "uniformly", "at", "random", "(", "MPII", "Multi", "-", "Person", "Val", ")", ",", "while", "major", "results", "and", "comparison", "to", "the", "state", "of", "the", "art", "are", "reported", "on", "the", "test", "set", ".", "Video", ".", "In", "order", "to", "evaluate", "video", "-", "based", "models", "we", "introduce", "a", "novel", "“", "MPII", "Video", "Pose", "”", "dataset", ".", "To", "this", "end", "we", "manually", "selected", "challenging", "keyframes", "from", "MPII", "Multi", "-", "Person", "dataset", ".", "Selected", "keyframes", "represent", "crowded", "scenes", "with", "highly", "articulated", "people", "engaging", "in", "various", "dynamic", "activities", ".", "In", "addition", "to", "each", "keyframe", ",", "we", "include", "+", "/-", "neighboring", "frames", "from", "the", "corresponding", "publicly", "available", "video", "sequences", ",", "and", "annotate", "every", "second", "frame", ".", "Each", "body", "pose", "was", "annotated", "following", "the", "standard", "annotation", "procedure", ",", "while", "maintaining", "person", "identity", "throughout", "the", "sequence", ".", "In", "contrast", "to", "MPII", "Multi", "-", "Person", "where", "some", "frames", "may", "contain", "non", "-", "annotated", "people", ",", "we", "annotate", "all", "people", "participating", "in", "the", "activity", "captured", "in", "the", "video", ",", "and", "add", "ignore", "regions", "for", "areas", "that", "contain", "dense", "crowds", "(", "e.g.", "static", "spectators", "in", "the", "dancing", "sequences", ")", ".", "In", "total", ",", "our", "dataset", "consists", "of", "sequences", "with", "over", "annotated", "poses", ".", "Evaluation", "details", ".", "The", "average", "precision", "(", "AP", ")", "measure", "is", "used", "for", "evaluation", "of", "pose", "estimation", "accuracy", ".", "For", "each", "algorithm", "we", "also", "report", "run", "time", "of", "the", "proposal", "generation", "and", "of", "the", "graph", "partitioning", "stages", ".", "All", "time", "measurements", "were", "conducted", "on", "a", "single", "core", "Intel", "Xeon", "GHz", ".", "Finally", "we", "also", "evaluate", "tracking", "perfomance", "using", "standard", "MOTA", "metric", ".", "Evaluation", "on", "our", "“", "MPII", "Video", "Pose", "”", "dataset", "is", "performed", "on", "the", "full", "frames", "using", "the", "publicly", "available", "evaluation", "kit", "of", ".", "On", "MPII", "Multi", "-", "Person", "we", "follow", "the", "official", "evaluation", "protocol", "and", "evaluate", "on", "groups", "using", "the", "provided", "rough", "group", "location", "and", "scale", ".", "subsection", ":", "Single", "-", "frame", "models", "We", "compare", "the", "performance", "of", "different", "variants", "of", "our", "Bottom", "-", "Up", "(", "BU", ")", "and", "Top", "-", "Down", "/", "Bottom", "-", "Up", "(", "TD", "/", "BU", ")", "models", "introduced", "in", "Sec", ".", "[", "reference", "]", "and", "Sec", ".", "[", "reference", "]", ".", "For", "BU", "we", "consider", "a", "model", "that", "(", "1", ")", "uses", "a", "fully", "-", "connected", "graph", "with", "up", "to", "detection", "proposals", "and", "jointly", "performs", "partitioning", "and", "body", "-", "part", "labeling", "similar", "to", "(", "BU", "-", "full", ",", "label", ")", ";", "(", "2", ")", "is", "same", "as", "(", "1", ")", ",", "but", "labeling", "of", "detection", "proposals", "is", "done", "based", "on", "detection", "score", "(", "BU", "-", "full", ")", ";", "(", "3", ")", "is", "same", "as", "(", "2", ")", ",", "but", "uses", "a", "sparsely", "-", "connected", "graph", "(", "BU", "-", "sparse", ")", ".", "The", "results", "are", "shown", "in", "Tab", ".", "[", "reference", "]", ".", "BU", "-", "full", ",", "label", "achieves", "%", "AP", "with", "a", "median", "inference", "run", "-", "time", "of", "s", "/", "f", ".", "BU", "-", "full", "achieves", "run", "-", "time", "reduction", "(", "vs.", "s", "/", "f", ")", ":", "pre", "-", "labeling", "detection", "candidates", "based", "on", "detection", "score", "significantly", "reduces", "the", "number", "of", "variables", "in", "the", "problem", "graph", ".", "Interestingly", ",", "pre", "-", "labeling", "also", "improves", "the", "performance", "(", "vs.", "%", "AP", ")", ":", "some", "of", "the", "low", "-", "scoring", "detections", "may", "complicate", "the", "search", "for", "an", "optimal", "labeling", ".", "BU", "-", "sparse", "further", "reduces", "run", "-", "time", "(", "vs.", "s", "/", "f", ")", ",", "as", "it", "reduces", "the", "complexity", "of", "the", "initial", "problem", "by", "sparsifying", "the", "graph", ",", "at", "a", "price", "of", "a", "drop", "in", "performance", "(", "vs.", "%", "AP", ")", ".", "In", "Tab", ".", "[", "reference", "]", "we", "compare", "the", "variants", "of", "the", "TD", "/", "BU", "model", ".", "Our", "TD", "approach", "achieves", "%", "AP", ",", "performing", "on", "par", "with", "a", "more", "complex", "BU", "-", "full", ".", "Explicit", "spatial", "propagation", "(", "TD", "+", "SP", ")", "further", "improves", "the", "results", "(", "vs.", "%", "AP", ")", ".", "The", "largest", "improvement", "is", "observed", "for", "ankles", ":", "progressive", "prediction", "that", "conditions", "on", "the", "close", "-", "by", "parts", "in", "the", "tree", "hierarchy", "reduces", "the", "distance", "between", "the", "conditioning", "signal", "and", "the", "location", "of", "the", "predicted", "body", "part", "and", "simplifies", "the", "prediction", "task", ".", "Performing", "inference", "(", "TD", "/", "BU", "+", "SP", ")", "improves", "the", "performance", "to", "%", "AP", ",", "due", "to", "more", "optimal", "assignment", "of", "part", "detection", "candidates", "to", "corresponding", "persons", ".", "Graph", "simplification", "in", "TD", "/", "BU", "allows", "to", "further", "reduce", "the", "inference", "time", "for", "graph", "partitioning", "(", "vs.", "for", "BU", "-", "sparse", ")", ".", "Single", "Frame", "Tracking", "Single", "Frame", "Tracking", "Comparison", "to", "the", "State", "of", "the", "Art", ".", "We", "compare", "the", "proposed", "single", "-", "frame", "approaches", "to", "the", "state", "of", "the", "art", "on", "MPII", "Multi", "-", "Person", "Test", "and", "WAF", "datasets", ".", "Comparison", "on", "MPII", "is", "shown", "in", "Tab", ".", "[", "reference", "]", ".", "Both", "BU", "-", "full", "and", "TD", "/", "BU", "improve", "over", "the", "best", "published", "result", "of", "DeeperCut", ",", "achieving", "and", "%", "AP", "respectively", "vs.", "%", "AP", "by", "DeeperCut", ".", "For", "the", "TD", "/", "BU", "the", "improvements", "on", "articulated", "parts", "(", "elbows", ",", "wrists", ",", "ankles", ",", "knees", ")", "are", "particularly", "pronounced", ".", "We", "argue", "that", "this", "is", "due", "to", "using", "the", "network", "that", "is", "directly", "trained", "to", "disambiguate", "body", "parts", "of", "different", "people", ",", "instead", "of", "using", "explicit", "geometric", "pairwise", "terms", "that", "only", "serve", "as", "a", "proxy", "to", "person", "’s", "identity", ".", "Overall", ",", "the", "performance", "of", "our", "best", "TD", "/", "BU", "method", "is", "noticeably", "higher", "(", "vs.", "%", "AP", ")", ".", "Remarkably", ",", "its", "run", "-", "time", "of", "graph", "partitioning", "stage", "is", "orders", "of", "magnitude", "faster", "compared", "to", "DeeperCut", ".", "This", "speed", "-", "up", "is", "due", "to", "two", "factors", ".", "First", ",", "TD", "/", "BU", "relies", "on", "a", "faster", "solver", "that", "tackles", "the", "graph", "-", "partitioning", "problem", "via", "local", "search", ",", "in", "contrast", "to", "the", "exact", "solver", "used", "in", ".", "Second", ",", "in", "the", "case", "of", "TD", "/", "BU", "model", "the", "graph", "is", "sparse", "and", "a", "large", "portion", "of", "the", "computation", "is", "performed", "by", "the", "feed", "-", "forward", "CNN", "introduced", "in", "Sec", ".", "[", "reference", "]", ".", "On", "WAF", "dataset", "TD", "/", "BU", "substantially", "improves", "over", "the", "best", "published", "result", "(", "vs.", "%", "AP", "by", ")", ".", "We", "refer", "to", "supplemental", "material", "for", "details", ".", "subsection", ":", "Multi", "-", "frame", "models", "Comparison", "of", "video", "-", "based", "models", ".", "Performance", "of", "the", "proposed", "video", "-", "based", "models", "is", "compared", "in", "Tab", ".", "[", "reference", "]", ".", "Video", "-", "based", "models", "outperform", "single", "-", "frame", "models", "in", "each", "case", ".", "BU", "-", "full", "+", "temporal", "slightly", "outperforms", "BU", "-", "full", ",", "where", "improvements", "are", "noticeable", "for", "ankle", ",", "knee", "and", "head", ".", "BU", "-", "sparse", "+", "temporal", "noticeably", "improves", "over", "BU", "-", "sparse", "(", "vs.", "%", "AP", ")", ".", "We", "observe", "significant", "improvements", "on", "the", "most", "difficult", "parts", "such", "as", "ankles", "(", "%", "AP", ")", "and", "wrists", "(", "%", "AP", ")", ".", "Interestingly", ",", "BU", "-", "sparse", "+", "temporal", "outperforms", ":", "longer", "-", "range", "connections", "such", "as", ",", ",", "head", "to", "ankle", ",", "may", "introduce", "additional", "confusion", "when", "information", "is", "propagated", "over", "time", ".", "Finally", ",", "TD", "/", "BU", "+", "temporal", "improves", "over", "TD", "/", "BU", "(", "%", "AP", ")", ".", "Similarly", "to", "BU", "-", "sparse", "+", "temporal", ",", "improvement", "is", "most", "prominent", "on", "ankles", "(", "%", "AP", ")", "and", "wrists", "(", "%", "AP", ")", ".", "Note", "that", "even", "the", "single", "-", "frame", "TD", "/", "BU", "outperforms", "the", "best", "temporal", "BU", "model", ".", "We", "show", "examples", "of", "articulated", "tracking", "on", "“", "MPII", "Video", "Pose", "”", "in", "Fig", ".", "[", "reference", "]", ".", "Temporal", "reasoning", "helps", "in", "cases", "when", "image", "information", "is", "ambiguous", "due", "to", "close", "proximity", "of", "multiple", "people", ".", "For", "example", "the", "video", "-", "based", "approach", "succeeds", "in", "correctly", "localizing", "legs", "of", "the", "person", "in", "Fig", ".", "[", "reference", "]", "(", "d", ")", "and", "(", "h", ")", ".", "Temporal", "features", ".", "We", "perform", "an", "ablative", "experiment", "on", "the", "“", "MPII", "Video", "Pose", "”", "dataset", "to", "evaluate", "the", "individual", "contribution", "of", "the", "temporal", "features", "introduced", "in", "Sec", ".", "[", "reference", "]", ".", "The", "Euclidean", "distance", "alone", "achieves", "AP", ",", "adding", "DeepMatching", "features", "improves", "the", "resuls", "to", "AP", ",", "whereas", "the", "combination", "of", "all", "features", "achieves", "the", "best", "result", "of", "AP", "(", "details", "in", "supplemental", "material", ")", ".", "Tracking", "evaluation", ".", "In", "Tab", ".", "[", "reference", "]", "we", "present", "results", "of", "the", "evaluation", "of", "multi", "-", "person", "articulated", "body", "tracking", ".", "We", "treat", "each", "body", "joint", "of", "each", "person", "as", "a", "tracking", "target", "and", "measure", "tracking", "performance", "using", "a", "standard", "multiple", "object", "tracking", "accuracy", "(", "MOTA", ")", "metric", "that", "incorporates", "identity", "switches", ",", "false", "positives", "and", "false", "negatives", ".", "We", "experimentally", "compare", "to", "a", "baseline", "model", "that", "first", "tracks", "people", "across", "frames", "and", "then", "performs", "per", "-", "frame", "pose", "estimation", ".", "To", "track", "a", "person", "we", "use", "a", "reduced", "version", "of", "our", "algorithm", "that", "operates", "on", "the", "two", "head", "joints", "only", ".", "This", "allows", "to", "achieve", "near", "perfect", "person", "tracking", "results", "in", "most", "cases", ".", "Our", "tracker", "still", "fails", "when", "the", "person", "head", "is", "occluded", "for", "multiple", "frames", "as", "it", "does", "not", "incorporate", "long", "-", "range", "connectivity", "between", "target", "hypothesis", ".", "We", "leave", "handling", "of", "long", "-", "term", "occlusions", "for", "the", "future", "work", ".", "For", "full", "-", "body", "tracking", "we", "use", "the", "same", "inital", "head", "tracks", "and", "add", "them", "to", "the", "set", "of", "body", "part", "proposals", ",", "while", "also", "adding", "must", "-", "link", "and", "must", "-", "cut", "constraints", "for", "the", "temporal", "edges", "corresponding", "to", "the", "head", "parts", "detections", ".", "The", "rest", "of", "the", "graph", "remains", "unchanged", "so", "that", "at", "inference", "time", "the", "body", "parts", "can", "be", "freely", "assigned", "to", "different", "person", "tracks", ".", "For", "the", "BU", "-", "sparse", "the", "full", "body", "tracking", "improves", "performance", "by", "and", "MOTA", "on", "wrists", "and", "ankles", ",", "and", "by", "and", "MOTA", "on", "elbows", "and", "knees", "respectively", ".", "TD", "/", "BU", "benefits", "from", "adding", "temporal", "connections", "between", "body", "parts", "as", "well", ",", "but", "to", "a", "lesser", "extent", "than", "BU", "-", "sparse", ".", "The", "most", "significant", "improvement", "is", "for", "ankles", "(", "MOTA", ")", ".", "BU", "-", "sparse", "also", "achieves", "the", "best", "overall", "score", "of", "compared", "to", "by", "TD", "/", "BU", ".", "This", "is", "surprising", "since", "TD", "/", "BU", "outperformed", "BU", "-", "sparse", "on", "the", "pose", "estimation", "task", "(", "see", "Tab", ".", "[", "reference", "]", "and", "[", "reference", "]", ")", ".", "We", "hypothesize", "that", "limited", "improvement", "of", "TD", "/", "BU", "could", "be", "due", "to", "balancing", "issues", "between", "the", "temporal", "and", "spatial", "pairwise", "terms", "that", "are", "estimated", "independently", "of", "each", "other", ".", "section", ":", "Conclusion", "In", "this", "paper", "we", "introduced", "an", "efficient", "and", "effective", "approach", "to", "articulated", "body", "tracking", "in", "monocular", "video", ".", "Our", "approach", "defines", "a", "model", "that", "jointly", "groups", "body", "part", "proposals", "within", "each", "video", "frame", "and", "across", "time", ".", "Grouping", "is", "formulated", "as", "a", "graph", "partitioning", "problem", "that", "lends", "itself", "to", "efficient", "inference", "with", "recent", "local", "search", "techniques", ".", "Our", "approach", "improves", "over", "state", "-", "of", "-", "the", "-", "art", "while", "being", "substantially", "faster", "compared", "to", "other", "related", "work", ".", "Acknowledgements", ".", "This", "work", "has", "been", "supported", "by", "the", "Max", "Planck", "Center", "for", "Visual", "Computing", "and", "Communication", ".", "The", "authors", "thank", "Varvara", "Obolonchykova", "and", "Bahar", "Tarakameh", "for", "their", "help", "in", "creating", "the", "video", "dataset", ".", "section", ":", "Additional", "Results", "on", "the", "MPII", "Multi", "-", "Person", "Dataset", "We", "perform", "qualitative", "comparison", "of", "the", "proposed", "single", "-", "frame", "based", "TD", "/", "BU", "and", "BU", "-", "full", "methods", "on", "challenging", "scenes", "containing", "highly", "articulated", "and", "strongly", "overlapping", "individuals", ".", "Results", "are", "shown", "in", "Fig", ".", "[", "reference", "]", "and", "Figure", "[", "reference", "]", ".", "The", "BU", "-", "full", "works", "well", "when", "persons", "are", "sufficiently", "separated", "(", "images", "11", "and", "12", ")", ".", "However", ",", "it", "fails", "on", "images", "where", "people", "significantly", "overlap", "(", "images", "1", "-", "3", ",", "5", "-", "10", ")", "or", "exhibit", "high", "degree", "of", "articulation", "(", "image", "4", ")", ".", "This", "is", "due", "to", "the", "fact", "that", "geometric", "image", "-", "conditioned", "pairwise", "may", "get", "confused", "in", "the", "presence", "of", "multiple", "overlapping", "individuals", "and", "thus", "mislead", "post", "-", "CNN", "bottom", "-", "up", "assembling", "of", "body", "poses", ".", "In", "contrast", ",", "TD", "/", "BU", "performs", "explicit", "modeling", "of", "person", "identity", "via", "top", "-", "dop", "bottom", "-", "up", "reasoning", "while", "offloading", "the", "larger", "share", "of", "the", "reasoning", "about", "body", "-", "part", "association", "onto", "feed", "-", "forward", "convolutional", "architecture", ",", "and", "thus", "is", "able", "to", "resolve", "such", "challenging", "cases", ".", "Interestingly", ",", "TD", "/", "BU", "is", "able", "to", "correctly", "predict", "lower", "limbs", "of", "people", "in", "the", "back", "through", "partial", "occlusion", "(", "image", "3", ",", "5", ",", "7", ",", "10", ")", ".", "TD", "/", "BU", "model", "occasionally", "incorrectly", "assembles", "body", "parts", "in", "kinematically", "implausible", "manner", "(", "image", "12", ")", ",", "as", "it", "does", "not", "explicitly", "model", "geometric", "body", "part", "relations", ".", "Finally", ",", "both", "models", "fail", "in", "presense", "of", "high", "variations", "in", "scale", "(", "image", "13", ")", ".", "We", "envision", "that", "reasoning", "over", "multiple", "scales", "is", "likely", "to", "improve", "the", "results", ".", "BU", "TD", "/", "BU", "BU", "TD", "/", "BU", "BU", "TD", "/", "BU", "BU", "TD", "/", "BU", "section", ":", "Results", "on", "the", "We", "Are", "Family", "dataset", "We", "compare", "our", "proposed", "TD", "/", "BU", "model", "to", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "the", "“", "We", "Are", "Family", "”", "(", "WAF", ")", "dataset", "and", "present", "results", "in", "Table", "[", "reference", "]", ".", "We", "use", "evaluation", "protocol", "from", "and", "report", "the", "AP", "evaluation", "measure", ".", "TD", "/", "BU", "model", "outperforms", "the", "best", "published", "results", "across", "all", "body", "parts", "(", "vs", "%", "AP", ")", "as", "well", "improves", "on", "articulated", "parts", "such", "as", "wrists", "(", "%", "AP", ")", "and", "elbows", "(", "%", "AP", ")", ".", "We", "attribute", "that", "to", "the", "ability", "of", "top", "-", "down", "model", "to", "better", "learn", "part", "associations", "compared", "to", "explicit", "modeling", "geometric", "pairwise", "relations", "as", "in", ".", "section", ":", "Evaluation", "of", "temporal", "features", ".", "We", "evaluate", "the", "importance", "of", "combining", "temporal", "features", "introduced", "in", "Sec", ".", "3.4", "of", "the", "paper", "on", "our", "Multi", "-", "Person", "Video", "dataset", ".", "To", "that", "end", ",", "we", "consider", "BU", "-", "sparse", "+", "temporal", "model", "and", "compare", "results", "to", "BU", "-", "sparse", "in", "Tab", ".", "[", "reference", "]", ".", "Single", "-", "frame", "BU", "-", "sparse", "achieves", "%", "AP", ".", "It", "can", "be", "seen", "that", "using", "geometry", "based", "det", "-", "distance", "features", "slightly", "improves", "the", "results", "to", "%", "AP", ",", "as", "it", "enables", "the", "propagation", "of", "information", "from", "neighboring", "frames", ".", "Using", "deepmatch", "features", "slightly", "improves", "the", "performance", "further", "as", "it", "helps", "to", "link", "the", "same", "body", "part", "of", "the", "same", "person", "over", "time", "based", "on", "the", "body", "part", "appearance", ".", "It", "is", "especially", "helpful", "in", "the", "case", "of", "fast", "motion", "where", "det", "-", "distance", "may", "fail", ".", "The", "combination", "of", "both", "geometry", "and", "appearance", "based", "features", "further", "improves", "the", "performance", "to", "%", ",", "which", "shows", "their", "complementarity", ".", "Finally", ",", "adding", "the", "sift", "-", "distance", "feature", "improves", "the", "results", "to", "%", ",", "since", "it", "copes", "better", "with", "the", "sudden", "changes", "in", "background", "and", "body", "part", "orientations", ".", "Overall", ",", "using", "a", "combination", "of", "temporal", "features", "in", "BU", "-", "sparse", "+", "temporal", "results", "in", "a", "%", "AP", "improvement", "over", "the", "single", "-", "frame", "BU", "-", "sparse", ".", "This", "demonstrates", "the", "advantages", "of", "the", "proposed", "approach", "to", "improve", "pose", "estimation", "performance", "using", "temporal", "information", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["ArtTrack"], "Method"], [4, 9, ["Articulated", "Multi", "-", "person", "Tracking"], "Method"], [20, 28, ["articulated", "tracking", "of", "multiple", "people", "in", "unconstrained", "videos"], "Task"], [40, 45, ["single", "-", "frame", "pose", "estimation"], "Task"], [65, 70, ["body", "-", "part", "relationship", "graph"], "Method"], [75, 77, ["faster", "inference"], "Task"], [91, 96, ["feed", "-", "forward", "convolutional", "architecture"], "Method"], [126, 128, ["articulated", "tracking"], "Task"], [129, 133, ["spatio", "-", "temporal", "grouping"], "Task"], [143, 145, ["association", "problem"], "Task"], [240, 242, ["crowded", "scenes"], "Task"], [252, 256, ["articulated", "human", "pose", "tracking"], "Task"], [397, 399, ["joint", "modeling"], "Method"], [422, 428, ["end", "-", "to", "-", "end", "learning"], "Method"], [444, 450, ["multi", "-", "person", "articulated", "tracking", "problem"], "Task"], [458, 465, ["end", "-", "to", "-", "end", "prediction", "task"], "Task"], [483, 485, ["end", "learning"], "Method"], [516, 518, ["convolutional", "network"], "Method"], [534, 541, ["jointly", "reasoning", "about", "assignment", "of", "body", "joints"], "Task"], [554, 557, ["graph", "partitioning", "formulation"], "Method"], [562, 567, ["people", "tracking", "and", "pose", "estimation"], "Task"], [578, 581, ["articulated", "people", "tracking"], "Task"], [585, 586, ["inference"], "Task"], [587, 588, ["video"], "Task"], [592, 594, ["inference", "methods"], "Method"], [596, 599, ["local", "combinatorial", "optimization"], "Method"], [603, 605, ["sparse", "model"], "Method"], [624, 628, ["feed", "-", "forward", "reasoning"], "Method"], [629, 635, ["joint", "-", "to", "-", "person", "association"], "Task"], [659, 660, ["accuracy"], "Metric"], [670, 673, ["articulated", "tracking", "model"], "Method"], [681, 683, ["part", "detections"], "Method"], [732, 737, ["single", "-", "frame", "pose", "estimation"], "Task"], [741, 743, ["sparse", "graph"], "Method"], [751, 755, ["body", "-", "part", "proposals"], "Task"], [777, 779, ["expensive", "inference"], "Task"], [786, 791, ["generic", "bottom", "-", "up", "proposals"], "Method"], [796, 798, ["sparse", "model"], "Method"], [807, 811, ["fully", "-", "connected", "model"], "Method"], [825, 833, ["top", "-", "down", "/", "bottom", "-", "up", "fashion"], "Method"], [838, 842, ["fully", "-", "connected", "model"], "Method"], [872, 876, ["body", "-", "part", "association"], "Task"], [878, 883, ["feed", "-", "forward", "convolutional", "architecture"], "Method"], [895, 899, ["articulated", "body", "joint", "tracking"], "Task"], [911, 913, ["Convolutional", "networks"], "Method"], [920, 927, ["localizing", "body", "joints", "of", "people", "in", "images"], "Task"], [933, 940, ["joint", "estimation", "of", "body", "configurations", "over", "time"], "Task"], [942, 945, ["3D", "pose", "estimation"], "Task"], [946, 948, ["outdoor", "environments"], "Task"], [949, 953, ["multi", "-", "camera", "setting"], "Task"], [960, 966, ["estimating", "body", "configurations", "of", "single", "people"], "Task"], [968, 969, ["accuracies"], "Metric"], [1008, 1012, ["feed", "-", "forward", "models"], "Method"], [1052, 1060, ["tracking", "upper", "body", "poses", "of", "isolated", "upright", "individuals"], "Task"], [1065, 1067, ["CNN", "detectors"], "Method"], [1071, 1074, ["localizing", "body", "joints"], "Task"], [1097, 1100, ["graph", "partitioning", "approach"], "Method"], [1110, 1112, ["pedestrian", "tracking"], "Task"], [1116, 1123, ["single", "frame", "multi", "-", "person", "pose", "estimation"], "Task"], [1131, 1137, ["articulated", "multi", "-", "person", "pose", "tracking"], "Task"], [1141, 1144, ["articulated", "pose", "tracking"], "Task"], [1149, 1154, ["hand", "-", "crafted", "image", "representations"], "Method"], [1157, 1159, ["simplified", "tasks"], "Task"], [1162, 1170, ["tracking", "upper", "body", "poses", "of", "frontal", "isolated", "people"], "Task"], [1172, 1175, ["tracking", "walking", "pedestrians"], "Task"], [1190, 1196, ["multi", "-", "person", "articulated", "pose", "tracking"], "Task"], [1225, 1227, ["graph", "partitioning"], "Method"], [1237, 1241, ["body", "-", "part", "proposals"], "Task"], [1257, 1261, ["person", "-", "conditioned", "model"], "Method"], [1275, 1277, ["detection", "stage"], "Task"], [1290, 1295, ["generic", "body", "-", "part", "detectors"], "Method"], [1310, 1312, ["convolutional", "network"], "Method"], [1314, 1317, ["body", "part", "proposals"], "Task"], [1345, 1349, ["multi", "-", "target", "tracking"], "Task"], [1356, 1358, ["pose", "estimation"], "Task"], [1359, 1361, ["articulated", "tracking"], "Task"], [1373, 1376, ["articulated", "tracking", "approach"], "Method"], [1424, 1425, ["Tracking"], "Task"], [1426, 1430, ["Spatio", "-", "temporal", "Grouping"], "Method"], [1431, 1434, ["body", "part", "detector"], "Method"], [1477, 1478, ["detection"], "Task"], [1525, 1527, ["tracking", "algorithm"], "Method"], [1598, 1600, ["tracking", "solution"], "Task"], [1669, 1671, ["tracking", "objective"], "Metric"], [1766, 1770, ["integer", "-", "linear", "program"], "Task"], [1837, 1842, ["minimum", "cost", "subgraph", "multicut", "problem"], "Task"], [1930, 1935, ["Articulated", "Multi", "-", "person", "Tracking"], "Method"], [1947, 1951, ["multi", "-", "object", "tracking"], "Task"], [1957, 1960, ["subgraph", "multicut", "problem"], "Task"], [1962, 1965, ["subgraph", "multicut", "problem"], "Task"], [1977, 1979, ["approximate", "inference"], "Task"], [1982, 1985, ["local", "search", "methods"], "Method"], [1999, 2000, ["connectivity"], "Method"], [2006, 2007, ["processing"], "Task"], [2061, 2064, ["articulated", "tracking", "approach"], "Method"], [2068, 2069, ["graph"], "Method"], [2198, 2199, ["BU"], "Method"], [2216, 2218, ["BU", "model"], "Method"], [2228, 2231, ["convolutional", "part", "detector"], "Method"], [2241, 2244, ["generic", "part", "detectors"], "Method"], [2247, 2252, ["convolutional", "body", "-", "part", "detector"], "Method"], [2272, 2273, ["complexity"], "Metric"], [2275, 2277, ["model", "graph"], "Method"], [2281, 2284, ["associating", "body", "parts"], "Task"], [2288, 2290, ["proposal", "mechanism"], "Method"], [2311, 2312, ["inference"], "Task"], [2359, 2360, ["solver"], "Method"], [2362, 2363, ["inference"], "Task"], [2366, 2370, ["Bottom", "-", "Up", "Model"], "Method"], [2371, 2372, ["BU"], "Method"], [2387, 2388, ["detection"], "Task"], [2434, 2436, ["image", "representation"], "Method"], [2439, 2441, ["convolutional", "network"], "Method"], [2477, 2479, ["sparse", "version"], "Method"], [2509, 2511, ["sparse", "version"], "Method"], [2545, 2547, ["Edge", "costs"], "Metric"], [2550, 2554, ["Bottom", "-", "Up", "model"], "Method"], [2587, 2589, ["regression", "function"], "Method"], [2610, 2612, ["logistic", "regression"], "Method"], [2653, 2655, ["edge", "cost"], "Metric"], [2707, 2715, ["Top", "-", "Down", "/", "Bottom", "-", "up", "Model"], "Method"], [2748, 2750, ["joint", "reasoning"], "Method"], [2824, 2827, ["identifying", "body", "parts"], "Task"], [2837, 2839, ["convolutional", "network"], "Method"], [2963, 2965, ["convolutional", "network"], "Method"], [3045, 3046, ["consistency"], "Task"], [3050, 3054, ["KL", "-", "based", "solver"], "Method"], [3074, 3076, ["processing", "stages"], "Method"], [3137, 3139, ["head", "detection"], "Task"], [3170, 3172, ["hierarchical", "model"], "Method"], [3189, 3191, ["inference", "process"], "Task"], [3201, 3203, ["hierarchical", "inference"], "Task"], [3209, 3211, ["convolutional", "network"], "Method"], [3216, 3218, ["convolutional", "network"], "Method"], [3273, 3279, ["cross", "-", "entropy", "binary", "classification", "loss"], "Metric"], [3350, 3352, ["convolutional", "layer"], "Method"], [3369, 3370, ["ResNet"], "Method"], [3375, 3378, ["joint", "prediction", "layer"], "Method"], [3406, 3408, ["Spatial", "propagation"], "Method"], [3409, 3410, ["SP"], "Method"], [3501, 3503, ["prediction", "layers"], "Method"], [3529, 3531, ["spatial", "propagation"], "Method"], [3566, 3570, ["Caffe", "’s", "ResNet", "implementation"], "Method"], [3574, 3580, ["ImageNet", "-", "pre", "-", "trained", "models"], "Method"], [3586, 3590, ["MPII", "Human", "Pose", "dataset"], "Material"], [3591, 3592, ["SGD"], "Method"], [3597, 3600, ["stepwise", "learning", "rate"], "Metric"], [3601, 3602, ["lr=0.002"], "Method"], [3605, 3606, ["lr=0.0002"], "Method"], [3673, 3677, ["non", "-", "maximum", "suppression"], "Method"], [3711, 3713, ["Temporal", "Model"], "Method"], [3718, 3721, ["within", "frame", "model"], "Method"], [3757, 3759, ["logistic", "regression"], "Method"], [3814, 3816, ["DeepMatching", "approach"], "Method"], [3833, 3834, ["SIFT"], "Method"], [3847, 3849, ["orientation", "estimation"], "Task"], [3851, 3853, ["SIFT", "descriptor"], "Method"], [3950, 3951, ["DeepMatching"], "Method"], [3994, 3997, ["body", "part", "detection"], "Task"], [4049, 4051, ["finding", "correspondences"], "Task"], [4068, 4069, ["DeepMatching"], "Method"], [4074, 4076, ["corresponding", "regions"], "Task"], [4106, 4107, ["SIFT"], "Method"], [4117, 4118, ["similarity"], "Metric"], [4131, 4133, ["evaluation", "measure"], "Metric"], [4139, 4142, ["single", "frame", "models"], "Method"], [4197, 4201, ["video", "-", "based", "models"], "Method"], [4276, 4278, ["annotation", "procedure"], "Method"], [4350, 4352, ["Evaluation", "details"], "Metric"], [4363, 4368, ["evaluation", "of", "pose", "estimation", "accuracy"], "Task"], [4375, 4377, ["run", "time"], "Metric"], [4379, 4381, ["proposal", "generation"], "Task"], [4384, 4387, ["graph", "partitioning", "stages"], "Method"], [4405, 4407, ["tracking", "perfomance"], "Metric"], [4409, 4411, ["MOTA", "metric"], "Metric"], [4461, 4465, ["Single", "-", "frame", "models"], "Method"], [4509, 4510, ["BU"], "Method"], [4520, 4524, ["fully", "-", "connected", "graph"], "Method"], [4532, 4538, ["partitioning", "and", "body", "-", "part", "labeling"], "Task"], [4559, 4563, ["labeling", "of", "detection", "proposals"], "Task"], [4567, 4569, ["detection", "score"], "Metric"], [4570, 4571, ["BU"], "Method"], [4616, 4617, ["AP"], "Metric"], [4619, 4624, ["median", "inference", "run", "-", "time"], "Metric"], [4633, 4637, ["run", "-", "time", "reduction"], "Metric"], [4639, 4642, ["s", "/", "f"], "Metric"], [4646, 4649, ["labeling", "detection", "candidates"], "Task"], [4651, 4653, ["detection", "score"], "Metric"], [4666, 4669, ["pre", "-", "labeling"], "Method"], [4676, 4677, ["AP"], "Metric"], [4700, 4703, ["run", "-", "time"], "Metric"], [4707, 4708, ["f"], "Metric"], [4714, 4715, ["complexity"], "Metric"], [4735, 4736, ["AP"], "Metric"], [4760, 4761, ["AP"], "Metric"], [4773, 4776, ["Explicit", "spatial", "propagation"], "Method"], [4799, 4801, ["progressive", "prediction"], "Task"], [4831, 4833, ["prediction", "task"], "Task"], [4835, 4836, ["inference"], "Method"], [4848, 4849, ["AP"], "Metric"], [4863, 4865, ["Graph", "simplification"], "Method"], [4874, 4876, ["inference", "time"], "Metric"], [4877, 4879, ["graph", "partitioning"], "Task"], [4887, 4893, ["Single", "Frame", "Tracking", "Single", "Frame", "Tracking"], "Task"], [4905, 4909, ["single", "-", "frame", "approaches"], "Method"], [4916, 4921, ["MPII", "Multi", "-", "Person", "Test"], "Material"], [4922, 4924, ["WAF", "datasets"], "Material"], [4927, 4928, ["MPII"], "Material"], [4957, 4958, ["AP"], "Metric"], [5045, 5046, ["AP"], "Metric"], [5051, 5054, ["run", "-", "time"], "Metric"], [5055, 5058, ["graph", "partitioning", "stage"], "Method"], [5085, 5087, ["faster", "solver"], "Method"], [5090, 5094, ["graph", "-", "partitioning", "problem"], "Task"], [5095, 5097, ["local", "search"], "Task"], [5102, 5104, ["exact", "solver"], "Method"], [5132, 5136, ["feed", "-", "forward", "CNN"], "Method"], [5145, 5147, ["WAF", "dataset"], "Material"], [5149, 5150, ["BU"], "Method"], [5174, 5178, ["Multi", "-", "frame", "models"], "Method"], [5180, 5184, ["video", "-", "based", "models"], "Method"], [5189, 5193, ["video", "-", "based", "models"], "Method"], [5202, 5206, ["Video", "-", "based", "models"], "Method"], [5207, 5211, ["single", "-", "frame", "models"], "Method"], [5251, 5252, ["AP"], "Metric"], [5323, 5324, ["AP"], "Metric"], [5372, 5374, ["articulated", "tracking"], "Task"], [5387, 5389, ["Temporal", "reasoning"], "Task"], [5408, 5412, ["video", "-", "based", "approach"], "Method"], [5472, 5473, ["AP"], "Metric"], [5494, 5495, ["AP"], "Metric"], [5502, 5504, ["Tracking", "evaluation"], "Task"], [5518, 5524, ["multi", "-", "person", "articulated", "body", "tracking"], "Task"], [5539, 5540, ["tracking"], "Task"], [5544, 5548, ["multiple", "object", "tracking", "accuracy"], "Metric"], [5549, 5552, ["MOTA", ")", "metric"], "Metric"], [5557, 5559, ["false", "positives"], "Metric"], [5560, 5562, ["false", "negatives"], "Metric"], [5568, 5570, ["baseline", "model"], "Method"], [5579, 5584, ["per", "-", "frame", "pose", "estimation"], "Task"], [5612, 5614, ["person", "tracking"], "Task"], [5649, 5653, ["long", "-", "term", "occlusions"], "Task"], [5659, 5663, ["full", "-", "body", "tracking"], "Task"], [5733, 5736, ["full", "body", "tracking"], "Task"], [5740, 5741, ["MOTA"], "Method"], [5749, 5750, ["MOTA"], "Method"], [5788, 5789, ["MOTA"], "Task"], [5821, 5824, ["pose", "estimation", "task"], "Task"], [5881, 5884, ["articulated", "body", "tracking"], "Task"], [5907, 5908, ["Grouping"], "Task"], [5912, 5915, ["graph", "partitioning", "problem"], "Task"], [5920, 5921, ["inference"], "Task"], [5923, 5926, ["local", "search", "techniques"], "Method"], [5961, 5965, ["Visual", "Computing", "and", "Communication"], "Task"], [6001, 6006, ["single", "-", "frame", "based", "TD"], "Method"], [6007, 6008, ["BU"], "Method"], [6009, 6013, ["BU", "-", "full", "methods"], "Method"], [6115, 6123, ["CNN", "bottom", "-", "up", "assembling", "of", "body", "poses"], "Task"], [6127, 6128, ["TD"], "Method"], [6129, 6130, ["BU"], "Method"], [6131, 6136, ["explicit", "modeling", "of", "person", "identity"], "Method"], [6137, 6144, ["top", "-", "dop", "bottom", "-", "up", "reasoning"], "Method"], [6158, 6163, ["feed", "-", "forward", "convolutional", "architecture"], "Method"], [6176, 6177, ["TD"], "Method"], [6178, 6179, ["BU"], "Method"], [6205, 6206, ["TD"], "Method"], [6207, 6208, ["BU"], "Method"], [6288, 6290, ["Family", "dataset"], "Material"], [6294, 6295, ["TD"], "Method"], [6296, 6298, ["BU", "model"], "Method"], [6340, 6341, ["TD"], "Method"], [6342, 6344, ["BU", "model"], "Method"], [6385, 6389, ["top", "-", "down", "model"], "Method"], [6406, 6410, ["Evaluation", "of", "temporal", "features"], "Task"], [6429, 6434, ["Multi", "-", "Person", "Video", "dataset"], "Material"], [6469, 6470, ["AP"], "Metric"], [6489, 6490, ["AP"], "Metric"], [6540, 6542, ["fast", "motion"], "Task"], [6574, 6578, ["sift", "-", "distance", "feature"], "Method"], [6617, 6618, ["AP"], "Metric"], [6638, 6640, ["pose", "estimation"], "Task"], [180, 181, ["MPII"], "Material"], [190, 191, ["MPII"], "Material"], [1395, 1396, ["BU"], "Method"], [1406, 1407, ["TD"], "Method"], [1408, 1409, ["BU"], "Method"], [2209, 2210, ["TD"], "Method"], [2211, 2212, ["BU"], "Method"], [2234, 2235, ["TD"], "Method"], [2236, 2237, ["BU"], "Method"], [2316, 2317, ["BU"], "Method"], [2318, 2319, ["TD"], "Method"], [2320, 2321, ["BU"], "Method"], [2716, 2717, ["TD"], "Method"], [2718, 2719, ["BU"], "Method"], [2843, 2844, ["TD"], "Method"], [2845, 2846, ["BU"], "Method"], [3063, 3064, ["TD"], "Method"], [3065, 3066, ["BU"], "Method"], [3077, 3078, ["TD"], "Method"], [3079, 3080, ["BU"], "Method"], [3163, 3164, ["TD"], "Method"], [3165, 3166, ["BU"], "Method"], [3722, 3723, ["BU"], "Method"], [3724, 3725, ["TD"], "Method"], [3726, 3727, ["BU"], "Method"], [4144, 4145, ["MPII"], "Material"], [4166, 4167, ["MPII"], "Material"], [4206, 4207, ["MPII"], "Material"], [4221, 4222, ["MPII"], "Material"], [4290, 4291, ["MPII"], "Material"], [4357, 4358, ["AP"], "Metric"], [4416, 4417, ["MPII"], "Material"], [4436, 4437, ["MPII"], "Material"], [4478, 4479, ["BU"], "Method"], [4489, 4490, ["TD"], "Method"], [4491, 4492, ["BU"], "Method"], [4541, 4542, ["BU"], "Method"], [4593, 4594, ["BU"], "Method"], [4609, 4610, ["BU"], "Method"], [4629, 4630, ["BU"], "Method"], [4695, 4696, ["BU"], "Method"], [4750, 4751, ["TD"], "Method"], [4752, 4753, ["BU"], "Method"], [4756, 4757, ["TD"], "Method"], [4769, 4770, ["BU"], "Method"], [4777, 4778, ["TD"], "Method"], [4788, 4789, ["AP"], "Metric"], [4837, 4838, ["TD"], "Method"], [4839, 4840, ["BU"], "Method"], [4866, 4867, ["TD"], "Method"], [4868, 4869, ["BU"], "Method"], [4882, 4883, ["BU"], "Method"], [4938, 4939, ["BU"], "Method"], [4942, 4943, ["TD"], "Method"], [4944, 4945, ["BU"], "Method"], [4961, 4962, ["AP"], "Metric"], [4967, 4968, ["TD"], "Method"], [4969, 4970, ["BU"], "Method"], [5035, 5036, ["TD"], "Method"], [5037, 5038, ["BU"], "Method"], [5079, 5080, ["TD"], "Method"], [5081, 5082, ["BU"], "Method"], [5113, 5114, ["TD"], "Method"], [5115, 5116, ["BU"], "Method"], [5147, 5148, ["TD"], "Method"], [5160, 5161, ["AP"], "Metric"], [5215, 5216, ["BU"], "Method"], [5222, 5223, ["BU"], "Method"], [5237, 5238, ["BU"], "Method"], [5245, 5246, ["BU"], "Method"], [5268, 5269, ["AP"], "Metric"], [5274, 5275, ["AP"], "Metric"], [5279, 5280, ["BU"], "Method"], [5311, 5312, ["TD"], "Method"], [5313, 5314, ["BU"], "Method"], [5318, 5319, ["TD"], "Method"], [5320, 5321, ["BU"], "Method"], [5328, 5329, ["BU"], "Method"], [5342, 5343, ["AP"], "Metric"], [5348, 5349, ["AP"], "Metric"], [5358, 5359, ["TD"], "Method"], [5360, 5361, ["BU"], "Method"], [5365, 5366, ["BU"], "Method"], [5376, 5377, ["MPII"], "Material"], [5445, 5446, ["MPII"], "Material"], [5481, 5482, ["AP"], "Metric"], [5729, 5730, ["BU"], "Method"], [5756, 5757, ["TD"], "Method"], [5758, 5759, ["BU"], "Method"], [5776, 5777, ["BU"], "Method"], [5791, 5792, ["BU"], "Method"], [5804, 5805, ["TD"], "Method"], [5806, 5807, ["BU"], "Method"], [5812, 5813, ["TD"], "Method"], [5814, 5815, ["BU"], "Method"], [5816, 5817, ["BU"], "Method"], [5843, 5844, ["TD"], "Method"], [5845, 5846, ["BU"], "Method"], [5989, 5990, ["MPII"], "Material"], [6040, 6041, ["BU"], "Method"], [6265, 6266, ["BU"], "Method"], [6266, 6267, ["TD"], "Method"], [6268, 6269, ["BU"], "Method"], [6269, 6270, ["BU"], "Method"], [6270, 6271, ["TD"], "Method"], [6272, 6273, ["BU"], "Method"], [6273, 6274, ["BU"], "Method"], [6274, 6275, ["TD"], "Method"], [6276, 6277, ["BU"], "Method"], [6277, 6278, ["BU"], "Method"], [6278, 6279, ["TD"], "Method"], [6280, 6281, ["BU"], "Method"], [6336, 6337, ["AP"], "Metric"], [6356, 6357, ["AP"], "Metric"], [6369, 6370, ["AP"], "Metric"], [6375, 6376, ["AP"], "Metric"], [6441, 6442, ["BU"], "Method"], [6451, 6452, ["BU"], "Method"], [6464, 6465, ["BU"], "Method"], [6608, 6609, ["BU"], "Method"], [6624, 6625, ["BU"], "Method"]]} {"doc_id": "33261d252218007147a71e40f8367ed152fa2fe0", "text": ["Question", "Answering", "with", "Subgraph", "Embeddings", "section", ":", "Abstract", ".", "This", "paper", "presents", "a", "system", "which", "learns", "to", "answer", "questions", "on", "a", "broad", "range", "of", "topics", "from", "a", "knowledge", "base", "using", "few", "handcrafted", "features", ".", "Our", "model", "learns", "low", "-", "dimensional", "embeddings", "of", "words", "and", "knowledge", "base", "constituents", ";", "these", "representations", "are", "used", "to", "score", "natural", "language", "questions", "against", "candidate", "answers", ".", "Training", "our", "system", "using", "pairs", "of", "questions", "and", "structured", "representations", "of", "their", "answers", ",", "and", "pairs", "of", "question", "paraphrases", ",", "yields", "competitive", "results", "on", "a", "recent", "benchmark", "of", "the", "literature", ".", "section", ":", "Introduction", "Teaching", "machines", "how", "to", "automatically", "answer", "questions", "asked", "in", "natural", "language", "on", "any", "topic", "or", "in", "any", "domain", "has", "always", "been", "a", "long", "standing", "goal", "in", "Artificial", "Intelligence", ".", "With", "the", "rise", "of", "large", "scale", "structured", "knowledge", "bases", "(", "KBs", ")", ",", "this", "problem", ",", "known", "as", "open", "-", "domain", "question", "answering", "(", "or", "open", "QA", ")", ",", "boils", "down", "to", "being", "able", "to", "query", "efficiently", "such", "databases", "with", "natural", "language", ".", "These", "KBs", ",", "such", "as", "Freebase", "[", "reference", "]", "encompass", "huge", "ever", "growing", "amounts", "of", "information", "and", "ease", "open", "QA", "by", "organizing", "a", "great", "variety", "of", "answers", "in", "a", "structured", "format", ".", "However", ",", "the", "scale", "and", "the", "difficulty", "for", "machines", "to", "interpret", "natural", "language", "still", "makes", "this", "task", "a", "challenging", "problem", ".", "The", "state", "-", "of", "-", "the", "-", "art", "techniques", "in", "open", "QA", "can", "be", "classified", "into", "two", "main", "classes", ",", "namely", ",", "information", "retrieval", "based", "and", "semantic", "parsing", "based", ".", "Information", "retrieval", "systems", "first", "retrieve", "a", "broad", "set", "of", "candidate", "answers", "by", "querying", "the", "search", "API", "of", "KBs", "with", "a", "transformation", "of", "the", "question", "into", "a", "valid", "query", "and", "then", "use", "fine", "-", "grained", "detection", "heuristics", "to", "identify", "the", "exact", "answer", "[", "reference", "][", "reference", "][", "reference", "]", ".", "On", "the", "other", "hand", ",", "semantic", "parsing", "methods", "focus", "on", "the", "correct", "interpretation", "of", "the", "meaning", "of", "a", "question", "by", "a", "semantic", "parsing", "system", ".", "A", "correct", "interpretation", "converts", "a", "question", "into", "the", "exact", "database", "query", "that", "returns", "the", "correct", "answer", ".", "Interestingly", ",", "recent", "works", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "have", "shown", "that", "such", "systems", "can", "be", "efficiently", "trained", "under", "indirect", "and", "imperfect", "supervision", "and", "hence", "scale", "to", "large", "-", "scale", "regimes", ",", "while", "bypassing", "most", "of", "the", "annotation", "costs", ".", "Yet", ",", "even", "if", "both", "kinds", "of", "system", "have", "shown", "the", "ability", "to", "handle", "largescale", "KBs", ",", "they", "still", "require", "experts", "to", "hand", "-", "craft", "lexicons", ",", "grammars", ",", "and", "KB", "schema", "to", "be", "effective", ".", "This", "non", "-", "negligible", "human", "intervention", "might", "not", "be", "generic", "enough", "to", "conveniently", "scale", "up", "to", "new", "databases", "with", "other", "schema", ",", "broader", "vocabularies", "or", "languages", "other", "than", "English", ".", "In", "contrast", ",", "[", "reference", "]", "proposed", "a", "framework", "for", "open", "QA", "requiring", "almost", "no", "human", "annotation", ".", "Despite", "being", "an", "interesting", "approach", ",", "this", "method", "is", "outperformed", "by", "other", "competing", "methods", ".", "[", "reference", "]", "introduced", "an", "embedding", "model", ",", "which", "learns", "low", "-", "dimensional", "vector", "representations", "of", "words", "and", "symbols", "(", "such", "as", "KBs", "constituents", ")", "and", "can", "be", "trained", "with", "even", "less", "supervision", "than", "the", "system", "of", "[", "reference", "]", "while", "being", "able", "to", "achieve", "better", "prediction", "performance", ".", "However", ",", "this", "approach", "is", "only", "compared", "with", "[", "reference", "]", "which", "operates", "in", "a", "simplified", "setting", "and", "has", "not", "been", "applied", "in", "more", "realistic", "conditions", "nor", "evaluated", "against", "the", "best", "performing", "methods", ".", "In", "this", "paper", ",", "we", "improve", "the", "model", "of", "[", "reference", "]", "by", "providing", "the", "ability", "to", "answer", "more", "complicated", "questions", ".", "sThe", "main", "contributions", "of", "the", "paper", "are", ":", "(", "1", ")", "a", "more", "sophisticated", "inference", "procedure", "that", "is", "both", "efficient", "and", "can", "consider", "longer", "paths", "(", "[", "reference", "]", "considered", "only", "answers", "directly", "connected", "to", "the", "question", "in", "the", "graph", ")", ";", "and", "(", "2", ")", "a", "richer", "representation", "of", "the", "answers", "which", "encodes", "the", "question", "-", "answer", "path", "and", "surrounding", "subgraph", "of", "the", "KB", ".", "Our", "approach", "is", "competitive", "with", "the", "current", "state", "-", "of", "-", "the", "-", "art", "on", "the", "recent", "benchmark", "WebQuestions", "[", "reference", "]", "without", "using", "any", "lexicon", ",", "rules", "or", "additional", "system", "for", "partof", "-", "speech", "tagging", ",", "syntactic", "or", "dependency", "parsing", "during", "training", "as", "most", "other", "systems", "do", ".", "section", ":", "Task", "Definition", "Our", "main", "motivation", "is", "to", "provide", "a", "system", "for", "open", "QA", "able", "to", "be", "trained", "as", "long", "as", "it", "has", "access", "to", ":", "(", "1", ")", "a", "training", "set", "of", "questions", "paired", "with", "answers", "and", "(", "2", ")", "a", "KB", "providing", "a", "structure", "among", "answers", ".", "We", "suppose", "that", "all", "potential", "answers", "are", "entities", "in", "the", "KB", "and", "that", "questions", "are", "sequences", "of", "words", "that", "include", "one", "identified", "KB", "entity", ".", "When", "this", "entity", "is", "not", "given", ",", "plain", "string", "matching", "is", "used", "to", "perform", "entity", "resolution", ".", "Smarter", "methods", "could", "be", "used", "but", "this", "is", "not", "our", "focus", ".", "We", "use", "WebQuestions", "[", "1", "]", "as", "our", "evaluation", "bemchmark", ".", "Since", "it", "contains", "few", "training", "samples", ",", "it", "is", "impossible", "to", "learn", "on", "it", "alone", ",", "and", "this", "section", "describes", "the", "various", "data", "sources", "that", "were", "used", "for", "training", ".", "These", "are", "similar", "to", "those", "used", "in", "[", "reference", "]", ".", "WebQuestions", "This", "dataset", "is", "built", "using", "Freebase", "as", "the", "KB", "and", "contains", "5", ",", "810", "question", "-", "answer", "pairs", ".", "It", "was", "created", "by", "crawling", "questions", "through", "the", "Google", "Suggest", "API", ",", "and", "then", "obtaining", "answers", "using", "Amazon", "Mechanical", "Turk", ".", "We", "used", "the", "original", "split", "(", "3", ",", "778", "examples", "for", "training", "and", "2", ",", "032", "for", "testing", ")", ",", "and", "isolated", "1k", "questions", "from", "the", "training", "set", "for", "validation", ".", "WebQuestions", "is", "built", "on", "Freebase", "since", "all", "answers", "are", "defined", "as", "Freebase", "entities", ".", "In", "each", "question", ",", "we", "identified", "one", "Freebase", "entity", "using", "string", "matching", "between", "words", "of", "the", "question", "and", "entity", "names", "in", "Freebase", ".", "When", "the", "same", "string", "matches", "multiple", "entities", ",", "only", "the", "entity", "appearing", "in", "most", "triples", ",", "i.e.", "the", "most", "popular", "in", "Freebase", ",", "was", "kept", ".", "Example", "questions", "(", "answers", ")", "in", "the", "dataset", "include", "\"", "Where", "did", "Edgar", "Allan", "Poe", "died", "?", "\"", "(", "baltimore", ")", "or", "\"", "What", "degrees", "did", "Barack", "Obama", "get", "?", "\"", "(", "bachelor", "of", "arts", ",", "juris", "doctor", ")", ".", "Freebase", "Freebase", "[", "reference", "]", "is", "a", "huge", "and", "freely", "available", "database", "of", "general", "facts", ";", "data", "is", "organized", "as", "triplets", "(", "subject", ",", "type1.type2.predicate", ",", "object", ")", ",", "where", "two", "entities", "subject", "and", "object", "(", "identified", "by", "mids", ")", "are", "connected", "by", "the", "relation", "type", "type1.type2.predicate", ".", "We", "used", "a", "subset", ",", "created", "by", "only", "keeping", "triples", "where", "one", "of", "the", "entities", "was", "appearing", "in", "either", "the", "WebQuestions", "training", "/", "validation", "set", "or", "in", "ClueWeb", "extractions", ".", "We", "also", "removed", "all", "entities", "appearing", "less", "than", "5", "times", "and", "finally", "obtained", "a", "Freebase", "set", "containing", "14", "M", "triples", "made", "of", "2.2", "M", "entities", "and", "7k", "relation", "types", ".", "section", ":", "WebQuestions", "1", "Since", "the", "format", "of", "triples", "does", "not", "correspond", "to", "any", "structure", "one", "could", "find", "in", "language", ",", "we", "decided", "to", "transform", "them", "into", "automatically", "generated", "questions", ".", "Hence", ",", "all", "triples", "were", "converted", "into", "questions", "\"", "What", "is", "the", "predicate", "of", "the", "type2", "subject", "?", "\"", "(", "using", "the", "mid", "of", "the", "subject", ")", "with", "the", "answer", "being", "object", ".", "An", "example", "is", "\"", "What", "is", "the", "nationality", "of", "the", "person", "barack", "obama", "?", "\"", "(", "united", "states", ")", ".", "More", "examples", "and", "details", "are", "given", "in", "a", "longer", "version", "of", "this", "paper", "[", "reference", "]", ".", "section", ":", "ClueWeb", "Extractions", "Freebase", "data", "allows", "to", "train", "our", "model", "on", "14", "M", "questions", "but", "these", "have", "a", "fixed", "lexicon", "and", "vocabulary", ",", "which", "is", "not", "realistic", ".", "Following", "[", "reference", "]", ",", "we", "also", "created", "questions", "using", "ClueWeb", "extractions", "provided", "by", "[", "reference", "]", ".", "Using", "string", "matching", ",", "we", "ended", "up", "with", "2", "M", "extractions", "structured", "as", "(", "subject", ",", "\"", "text", "string", "\"", ",", "object", ")", "with", "both", "subject", "and", "object", "linked", "to", "Freebase", ".", "We", "also", "converted", "these", "triples", "into", "questions", "by", "using", "simple", "patterns", "and", "Freebase", "types", ".", "An", "example", "of", "generated", "question", "is", "\"", "Where", "barack", "obama", "was", "allegedly", "bear", "in", "?", "\"", "(", "hawaii", ")", ".", "Paraphrases", "The", "automatically", "generated", "questions", "that", "are", "useful", "to", "connect", "Freebase", "triples", "and", "natural", "language", ",", "do", "not", "provide", "a", "satisfactory", "modeling", "of", "natural", "language", "because", "of", "their", "semi", "-", "automatic", "wording", "and", "rigid", "syntax", ".", "To", "overcome", "this", "issue", ",", "we", "follow", "[", "reference", "]", "Table", "2", ".", "Examples", "of", "questions", ",", "answer", "paths", "and", "paraphrases", "used", "in", "this", "paper", ".", "as", "rephrasings", "of", "each", "other", ":", "[", "reference", "]", "harvested", "a", "set", "of", "2", "M", "distinct", "questions", "from", "WikiAnswers", ",", "which", "were", "grouped", "into", "350k", "paraphrase", "clusters", ".", "section", ":", "Embedding", "Questions", "and", "Answers", "Inspired", "by", "[", "reference", "]", ",", "our", "model", "works", "by", "learning", "low", "-", "dimensional", "vector", "embeddings", "of", "words", "appearing", "in", "questions", "and", "of", "entities", "and", "relation", "types", "of", "Freebase", ",", "so", "that", "representations", "of", "questions", "and", "of", "their", "corresponding", "answers", "are", "close", "to", "each", "other", "in", "the", "joint", "embedding", "space", ".", "Let", "q", "denote", "a", "question", "and", "a", "a", "candidate", "answer", ".", "Learning", "embeddings", "is", "achieved", "by", "learning", "a", "scoring", "function", "S", "(", "q", ",", "a", ")", ",", "so", "that", "S", "generates", "a", "high", "score", "if", "a", "is", "the", "correct", "answer", "to", "the", "question", "q", ",", "and", "a", "low", "score", "otherwise", ".", "Note", "that", "both", "q", "and", "a", "are", "represented", "as", "a", "combination", "of", "the", "embeddings", "of", "their", "individual", "words", "and", "/", "or", "symbols", ";", "hence", ",", "learning", "S", "essentially", "involves", "learning", "these", "embeddings", ".", "In", "our", "model", ",", "the", "form", "of", "the", "scoring", "function", "is", ":", "Let", "W", "be", "a", "matrix", "of", "R", "k×N", ",", "where", "k", "is", "the", "dimension", "of", "the", "embedding", "space", "which", "is", "fixed", "a", "-", "priori", ",", "and", "N", "is", "the", "dictionary", "of", "embeddings", "to", "be", "learned", ".", "Let", "N", "W", "denote", "the", "total", "number", "of", "words", "and", "N", "S", "the", "total", "number", "of", "entities", "and", "relation", "types", ".", "With", "N", "=", "N", "W", "+", "N", "S", ",", "the", "i", "-", "th", "column", "of", "W", "is", "the", "embedding", "of", "the", "i", "-", "th", "element", "(", "word", ",", "entity", "or", "relation", "type", ")", "in", "the", "dictionary", ".", "The", "function", "f", "(", ".", ")", ",", "which", "maps", "the", "questions", "into", "the", "embedding", "space", "R", "k", "is", "defined", "as", "f", "(", "q", ")", "=", "Wφ", "(", "q", ")", ",", "where", "φ", "(", "q", ")", "∈", "N", "N", ",", "is", "a", "sparse", "vector", "indicating", "the", "number", "of", "times", "each", "word", "appears", "in", "the", "question", "q", "(", "usually", "0", "or", "1", ")", ".", "Likewise", "the", "function", "g", "(", ".", ")", "which", "maps", "the", "answer", "into", "the", "same", "embedding", "space", "R", "k", "as", "the", "questions", ",", "is", "given", "by", "g", "(", "a", ")", "=", "Wψ", "(", "a", ")", ".", "Here", "ψ", "(", "a", ")", "∈", "N", "N", "is", "a", "sparse", "vector", "representation", "of", "the", "answer", "a", ",", "which", "we", "now", "detail", ".", "section", ":", "Embedding", "model", "Freebase", "subgraph", "section", ":", "Binary", "encoding", "of", "the", "subgraph", "ψ", "(", "a", ")", "section", ":", "Embedding", "of", "the", "subgraph", "g", "(", "a", ")", "Binary", "encoding", "of", "the", "ques0on", "Φ", "(", "q", ")", "section", ":", "Embedding", "of", "the", "ques0on", "f", "(", "q", ")", "Ques0on", "q", "Subgraph", "of", "a", "candidate", "answer", "a", "(", "here", "K.", "Preston", ")", "section", ":", "Score", "S", "(", "q", ",", "a", ")", "How", "the", "candidate", "answer", "fits", "the", "ques0on", "section", ":", "Dot", "product", "Embedding", "matrix", "W", "Fig", ".", "1", ".", "Illustration", "of", "the", "subgraph", "embedding", "model", "scoring", "a", "candidate", "answer", ":", "(", "i", ")", "locate", "entity", "in", "the", "question", ";", "(", "ii", ")", "compute", "path", "from", "entity", "to", "answer", ";", "(", "iii", ")", "represent", "answer", "as", "path", "plus", "all", "connected", "entities", "to", "the", "answer", "(", "the", "subgraph", ")", ";", "(", "iv", ")", "embed", "both", "the", "question", "and", "the", "answer", "subgraph", "separately", "using", "the", "learnt", "embedding", "vectors", ",", "and", "score", "the", "match", "via", "their", "dot", "product", ".", "section", ":", "Representing", "Candidate", "Answers", "We", "now", "describe", "possible", "feature", "representations", "for", "a", "single", "candidate", "answer", ".", "(", "When", "there", "are", "multiple", "correct", "answers", ",", "we", "average", "these", "representations", ",", "see", "Section", "3.4", ".", ")", "We", "consider", "three", "different", "types", "of", "representation", ",", "corresponding", "to", "different", "subgraphs", "of", "Freebase", "around", "it", ".", "(", "i", ")", "Single", "Entity", ".", "The", "answer", "is", "represented", "as", "a", "single", "entity", "from", "Freebase", ":", "ψ", "(", "a", ")", "is", "a", "1", "-", "of", "-", "N", "S", "coded", "vector", "with", "1", "corresponding", "to", "the", "entity", "of", "the", "answer", ",", "and", "0", "elsewhere", ".", "(", "ii", ")", "Path", "Representation", ".", "The", "answer", "is", "represented", "as", "a", "path", "from", "the", "entity", "mentioned", "in", "the", "question", "to", "the", "answer", "entity", ".", "In", "our", "experiments", ",", "we", "considered", "1", "-", "or", "2", "-", "hops", "paths", "(", "i.e.", "with", "either", "1", "or", "2", "edges", "to", "traverse", ")", ":", "(", "barack", "obama", ",", "people.person.place", "of", "birth", ",", "honolulu", ")", "is", "a", "1", "-", "hop", "path", "and", "(", "barack", "obama", ",", "people.person.place", "of", "birth", ",", "location", ".", "location.containedby", ",", "hawaii", ")", "a", "2", "-", "hops", "path", ".", "This", "results", "in", "a", "ψ", "(", "a", ")", "which", "is", "a", "3", "-", "of", "-", "N", "S", "or", "4", "-", "of", "-", "N", "S", "coded", "vector", ",", "expressing", "the", "start", "and", "end", "entities", "of", "the", "path", "and", "the", "relation", "types", "(", "but", "not", "entities", ")", "in", "-", "between", ".", "(", "iii", ")", "Subgraph", "Representation", ".", "We", "encode", "both", "the", "path", "representation", "from", "(", "ii", ")", ",", "and", "the", "entire", "subgraph", "of", "entities", "connected", "to", "the", "candidate", "answer", "entity", ".", "That", "is", ",", "for", "each", "entity", "connected", "to", "the", "answer", "we", "include", "both", "the", "relation", "type", "and", "the", "entity", "itself", "in", "the", "representation", "ψ", "(", "a", ")", ".", "In", "order", "to", "represent", "the", "answer", "path", "differently", "to", "the", "surrounding", "subgraph", "(", "so", "the", "model", "can", "differentiate", "them", ")", ",", "we", "double", "the", "dictionary", "size", "for", "entities", ",", "and", "use", "one", "embedding", "representation", "if", "they", "are", "in", "the", "path", "and", "another", "if", "they", "are", "in", "the", "subgraph", ".", "Thus", "we", "now", "learn", "a", "parameter", "matrix", "R", "k×N", "where", "N", "=", "N", "W", "+", "2N", "S", "(", "N", "S", "is", "the", "total", "number", "of", "entities", "and", "relation", "types", ")", ".", "If", "there", "are", "C", "connected", "entities", "with", "D", "relation", "types", "to", "the", "candidate", "answer", ",", "its", "representation", "is", "a", "3", "+", "C", "+", "D", "or", "4", "+", "C", "+", "D", "-", "of", "-", "N", "S", "coded", "vector", ",", "depending", "on", "the", "path", "length", ".", "Our", "hypothesis", "is", "that", "including", "more", "information", "about", "the", "answer", "in", "its", "representation", "will", "lead", "to", "improved", "results", ".", "While", "it", "is", "possible", "that", "all", "required", "information", "could", "be", "encoded", "in", "the", "k", "dimensional", "embedding", "of", "the", "single", "entity", "(", "i", ")", ",", "it", "is", "unclear", "what", "dimension", "k", "should", "be", "to", "make", "this", "possible", ".", "For", "example", "the", "embedding", "of", "a", "country", "entity", "encoding", "all", "of", "its", "citizens", "seems", "unrealistic", ".", "Similarly", ",", "only", "having", "access", "to", "the", "path", "ignores", "all", "the", "other", "information", "we", "have", "about", "the", "answer", "entity", ",", "unless", "it", "is", "encoded", "in", "the", "embeddings", "of", "either", "the", "entity", "of", "the", "question", ",", "the", "answer", "or", "the", "relations", "linking", "them", ",", "which", "might", "be", "quite", "complicated", "as", "well", ".", "We", "thus", "adopt", "the", "subgraph", "approach", ".", "Figure", "1", "illustrates", "our", "model", ".", "section", ":", "Training", "and", "Loss", "Function", "As", "in", "[", "reference", "]", ",", "we", "train", "our", "model", "using", "a", "margin", "-", "based", "ranking", "loss", "function", ".", "Let", "D", "=", "{", "(", "q", "i", ",", "a", "i", ")", ":", "i", "=", "1", ",", ".", ".", ".", ",", "|D|", "}", "be", "the", "training", "set", "of", "questions", "q", "i", "paired", "with", "their", "correct", "answer", "a", "i", ".", "The", "loss", "function", "we", "minimize", "is", "where", "m", "is", "the", "margin", "(", "fixed", "to", "0.1", ")", ".", "Minimizing", "Eq", ".", "(", "2", ")", "learns", "the", "embedding", "matrix", "W", "so", "that", "the", "score", "of", "a", "question", "paired", "with", "a", "correct", "answer", "is", "greater", "than", "with", "any", "incorrect", "answerā", "by", "at", "least", "m.ā", "is", "sampled", "from", "a", "set", "of", "incorrect", "candidatesĀ.", "This", "is", "achieved", "by", "sampling", "50", "%", "of", "the", "time", "from", "the", "set", "of", "entities", "connected", "to", "the", "entity", "of", "the", "question", "(", "i.e.", "other", "candidate", "paths", ")", ",", "and", "by", "replacing", "the", "answer", "entity", "by", "a", "random", "one", "otherwise", ".", "Optimization", "is", "accomplished", "using", "stochastic", "gradient", "descent", ",", "multi", "-", "threaded", "with", "Hogwild", "!", "[", "reference", "]", ",", "with", "the", "constraint", "that", "the", "columns", "w", "i", "of", "W", "remain", "within", "the", "unit", "-", "ball", ",", "i.e.", ",", "∀", "i", ",", "||w", "i", "||", "2", "≤", "1", ".", "section", ":", "Multitask", "Training", "of", "Embeddings", "Since", "a", "large", "number", "of", "questions", "in", "our", "training", "datasets", "are", "synthetically", "generated", ",", "they", "do", "not", "adequately", "cover", "the", "range", "of", "syntax", "used", "in", "natural", "language", ".", "Hence", ",", "we", "also", "multi", "-", "task", "the", "training", "of", "our", "model", "with", "the", "task", "of", "paraphrase", "prediction", ".", "We", "do", "so", "by", "alternating", "the", "training", "of", "S", "with", "that", "of", "a", "scoring", "function", "S", "prp", "(", "q", "1", ",", "q", "2", ")", "=", "f", "(", "q", "1", ")", "f", "(", "q", "2", ")", ",", "which", "uses", "the", "same", "embedding", "matrix", "W", "and", "makes", "the", "embeddings", "of", "a", "pair", "of", "questions", "(", "q", "1", ",", "q", "2", ")", "similar", "to", "each", "other", "if", "they", "are", "paraphrases", "(", "i.e.", "if", "they", "belong", "to", "the", "same", "paraphrase", "cluster", ")", ",", "and", "make", "them", "different", "otherwise", ".", "Training", "S", "prp", "is", "similar", "to", "that", "of", "S", "except", "that", "negative", "samples", "are", "obtained", "by", "sampling", "a", "question", "from", "another", "paraphrase", "cluster", ".", "We", "also", "multitask", "the", "training", "of", "the", "embeddings", "with", "the", "mapping", "of", "the", "mids", "of", "Freebase", "entities", "to", "the", "actual", "words", "of", "their", "names", ",", "so", "that", "the", "model", "learns", "that", "the", "embedding", "of", "the", "mid", "of", "an", "entity", "should", "be", "similar", "to", "the", "embedding", "of", "the", "word", "(", "s", ")", "that", "compose", "its", "name", "(", "s", ")", ".", "section", ":", "Inference", "Once", "W", "is", "trained", ",", "at", "test", "time", ",", "for", "a", "given", "question", "q", "the", "model", "predicts", "the", "answer", "with", ":", "â", "=", "argmax", "a", "∈A", "(", "q", ")", "S", "(", "q", ",", "a", ")", "where", "A", "(", "q", ")", "is", "the", "candidate", "answer", "set", ".", "This", "candidate", "set", "could", "be", "the", "whole", "KB", "but", "this", "has", "both", "speed", "and", "potentially", "precision", "issues", ".", "Instead", ",", "we", "create", "a", "candidate", "set", "A", "(", "q", ")", "for", "each", "question", ".", "We", "recall", "that", "each", "question", "contains", "one", "identified", "Freebase", "entity", ".", "A", "(", "q", ")", "is", "first", "populated", "with", "all", "triples", "from", "Freebase", "involving", "this", "entity", ".", "This", "allows", "to", "answer", "simple", "factual", "questions", "whose", "answers", "are", "directly", "connected", "to", "them", "(", "i.e.", "1", "-", "hop", "paths", ")", ".", "This", "strategy", "is", "denoted", "C", "1", ".", "Since", "a", "system", "able", "to", "answer", "only", "such", "questions", "would", "be", "limited", ",", "we", "supplement", "A", "(", "q", ")", "with", "examples", "situated", "in", "the", "KB", "graph", "at", "2", "-", "hops", "from", "the", "entity", "of", "the", "question", ".", "We", "do", "not", "add", "all", "such", "quadruplets", "since", "this", "would", "lead", "to", "very", "large", "candidate", "sets", ".", "Instead", ",", "we", "consider", "the", "following", "general", "approach", ":", "given", "that", "we", "are", "predicting", "a", "path", ",", "we", "can", "predict", "its", "elements", "in", "turn", "using", "a", "beam", "search", ",", "and", "hence", "avoid", "scoring", "all", "candidates", ".", "Specifically", ",", "our", "model", "first", "ranks", "relation", "types", "using", "Eq", ".", "(", "1", ")", ",", "i.e.", "selects", "which", "relation", "types", "are", "the", "most", "likely", "to", "be", "expressed", "in", "q.", "We", "keep", "the", "top", "10", "types", "(", "10", "was", "selected", "on", "the", "validation", "set", ")", "and", "only", "add", "2", "-", "hops", "candidates", "to", "A", "(", "q", ")", "when", "these", "relations", "appear", "in", "their", "path", ".", "Scores", "of", "1", "-", "hop", "triples", "are", "weighted", "by", "1.5", "since", "they", "have", "one", "less", "element", "than", "2", "-", "hops", "quadruplets", ".", "This", "strategy", ",", "denoted", "C", "2", ",", "is", "used", "by", "default", ".", "A", "prediction", "a", "can", "commonly", "actually", "be", "a", "set", "of", "candidate", "answers", ",", "not", "just", "one", "answer", ",", "for", "example", "for", "questions", "like", "\"", "Who", "are", "David", "Beckham", "'s", "children", "?", "\"", ".", "This", "is", "achieved", "by", "considering", "a", "prediction", "to", "be", "all", "the", "entities", "that", "lie", "on", "the", "same", "1", "-", "hop", "or", "2", "-", "hops", "path", "from", "the", "entity", "found", "in", "the", "question", ".", "Hence", ",", "all", "answers", "to", "the", "above", "question", "are", "connected", "to", "david", "beckham", "via", "the", "same", "path", "(", "david", "beckham", ",", "people.person.children", ",", "*", ")", ".", "The", "feature", "representation", "of", "the", "prediction", "is", "then", "the", "average", "over", "each", "candidate", "entity", "'s", "features", "(", "see", "Section", "3.1", ")", ",", "i.e.", "ψ", "all", "(", "a", ")", "=", "1", "|a", "|", "a", "j", ":", "a", "ψ", "(", "a", "j", ")", "where", "a", "j", "are", "the", "individual", "entities", "in", "the", "overall", "prediction", "a", ".", "In", "the", "results", ",", "we", "compare", "to", "a", "baseline", "method", "that", "can", "only", "predict", "single", "candidates", ",", "which", "understandly", "performs", "poorly", ".", "section", ":", "Experiments", "We", "compare", "our", "system", "in", "terms", "of", "F1", "score", "as", "computed", "by", "the", "official", "evaluation", "script", "2", "(", "F1", "(", "Berant", ")", ")", "but", "also", "with", "a", "slightly", "different", "F1", "definition", ",", "termed", "F1", "(", "Yao", ")", "which", "was", "used", "in", "[", "reference", "]", "(", "the", "difference", "being", "the", "way", "that", "questions", "with", "no", "answers", "are", "dealt", "with", ")", ",", "and", "precision", "@", "1", "(", "p@1", ")", "of", "the", "first", "candidate", "entity", "(", "even", "when", "there", "are", "a", "set", "of", "correct", "answers", ")", ",", "comparing", "to", "recently", "published", "systems", ".", "3", "The", "upper", "part", "of", "Table", "3", "indicates", "that", "our", "approach", "outperforms", "[", "reference", "]", ",", "[", "reference", "]", "and", "[", "reference", "]", ",", "and", "performs", "similarly", "as", "[", "reference", "]", ".", "The", "lower", "part", "of", "Table", "3", "compares", "various", "versions", "of", "our", "model", ".", "Our", "default", "approach", "uses", "the", "Subgraph", "representation", "for", "answers", "and", "C", "2", "as", "the", "candidate", "answers", "set", ".", "Replacing", "C", "2", "by", "C", "1", "induces", "a", "large", "drop", "in", "performance", "because", "many", "questions", "do", "not", "have", "answers", "thatare", "directly", "connected", "to", "their", "inluded", "entity", "(", "not", "in", "C", "1", ")", ".", "However", ",", "using", "all", "2", "-", "hops", "connections", "as", "a", "candidate", "set", "is", "also", "detrimental", ",", "because", "the", "larger", "number", "of", "candidates", "confuses", "(", "and", "slows", "a", "lot", ")", "our", "ranking", "based", "inference", ".", "Our", "results", "also", "verify", "our", "hypothesis", "of", "Section", "3.1", ",", "that", "a", "richer", "representation", "for", "answers", "(", "using", "the", "local", "subgraph", ")", "can", "store", "more", "pertinent", "information", ".", "Finally", ",", "we", "demonstrate", "that", "we", "greatly", "improve", "upon", "the", "model", "of", "[", "reference", "]", ",", "which", "actually", "corresponds", "to", "a", "setting", "with", "the", "Path", "representation", "and", "C", "1", "as", "candidate", "set", ".", "We", "also", "considered", "an", "ensemble", "of", "our", "approach", "and", "that", "of", "[", "reference", "]", ".", "As", "we", "only", "had", "access", "to", "their", "test", "predictions", "we", "used", "the", "following", "combination", "method", ".", "Our", "approach", "gives", "a", "score", "S", "(", "q", ",", "a", ")", "for", "the", "answer", "it", "predicts", ".", "We", "chose", "a", "threshold", "such", "that", "our", "approach", "predicts", "50", "%", "of", "the", "time", "(", "when", "S", "(", "q", ",", "a", ")", "is", "above", "its", "value", ")", ",", "and", "the", "other", "50", "%", "of", "the", "time", "we", "use", "the", "prediction", "of", "[", "reference", "]", "instead", ".", "We", "aimed", "for", "a", "50", "/", "50", "ratio", "because", "both", "methods", "perform", "similarly", ".", "The", "ensemble", "improves", "the", "state", "-", "of", "-", "the", "-", "art", ",", "and", "indicates", "that", "our", "models", "are", "significantly", "different", "in", "their", "design", ".", "section", ":", "Conclusion", "This", "paper", "presented", "an", "embedding", "model", "that", "learns", "to", "perform", "open", "QA", "using", "training", "data", "made", "of", "questions", "paired", "with", "their", "answers", "and", "of", "a", "KB", "to", "provide", "a", "structure", "among", "answers", ",", "and", "can", "achieve", "promising", "performance", "on", "the", "competitive", "benchmark", "WebQuestions", ".", "section", ":"], "entity": [[0, 2, ["Question", "Answering"], "Task"], [3, 5, ["Subgraph", "Embeddings"], "Method"], [95, 97, ["Teaching", "machines"], "Task"], [121, 123, ["Artificial", "Intelligence"], "Task"], [142, 147, ["open", "-", "domain", "question", "answering"], "Task"], [149, 151, ["open", "QA"], "Task"], [172, 173, ["Freebase"], "Material"], [185, 187, ["open", "QA"], "Task"], [207, 208, ["machines"], "Task"], [230, 232, ["open", "QA"], "Task"], [242, 244, ["information", "retrieval"], "Task"], [246, 248, ["semantic", "parsing"], "Task"], [250, 253, ["Information", "retrieval", "systems"], "Task"], [284, 286, ["detection", "heuristics"], "Method"], [304, 307, ["semantic", "parsing", "methods"], "Method"], [320, 323, ["semantic", "parsing", "system"], "Method"], [382, 384, ["annotation", "costs"], "Metric"], [415, 417, ["KB", "schema"], "Method"], [461, 463, ["open", "QA"], "Task"], [466, 468, ["human", "annotation"], "Task"], [489, 491, ["embedding", "model"], "Method"], [530, 531, ["prediction"], "Task"], [603, 605, ["inference", "procedure"], "Method"], [673, 674, ["WebQuestions"], "Material"], [687, 691, ["partof", "-", "speech", "tagging"], "Task"], [692, 696, ["syntactic", "or", "dependency", "parsing"], "Task"], [717, 719, ["open", "QA"], "Task"], [747, 748, ["KB"], "Method"], [786, 789, ["plain", "string", "matching"], "Method"], [793, 795, ["entity", "resolution"], "Task"], [796, 798, ["Smarter", "methods"], "Method"], [810, 811, ["WebQuestions"], "Material"], [860, 861, ["WebQuestions"], "Material"], [866, 867, ["Freebase"], "Material"], [888, 891, ["Google", "Suggest", "API"], "Material"], [897, 900, ["Amazon", "Mechanical", "Turk"], "Material"], [930, 931, ["validation"], "Task"], [932, 933, ["WebQuestions"], "Material"], [936, 937, ["Freebase"], "Material"], [967, 968, ["Freebase"], "Material"], [990, 991, ["Freebase"], "Material"], [1110, 1112, ["ClueWeb", "extractions"], "Task"], [1246, 1248, ["ClueWeb", "Extractions"], "Task"], [1292, 1294, ["string", "matching"], "Method"], [1321, 1322, ["Freebase"], "Material"], [1438, 1439, ["WikiAnswers"], "Material"], [1465, 1470, ["low", "-", "dimensional", "vector", "embeddings"], "Method"], [1516, 1518, ["Learning", "embeddings"], "Task"], [1523, 1528, ["scoring", "function", "S", "(", "q"], "Method"], [1581, 1583, ["learning", "S"], "Task"], [1597, 1599, ["scoring", "function"], "Metric"], [1802, 1809, ["sparse", "vector", "representation", "of", "the", "answer", "a"], "Method"], [1817, 1819, ["Embedding", "model"], "Method"], [1823, 1831, ["Binary", "encoding", "of", "the", "subgraph", "ψ", "(", "a"], "Method"], [1834, 1841, ["Embedding", "of", "the", "subgraph", "g", "(", "a"], "Task"], [1842, 1844, ["Binary", "encoding"], "Method"], [1892, 1894, ["Dot", "product"], "Method"], [1904, 1907, ["subgraph", "embedding", "model"], "Method"], [1974, 1976, ["dot", "product"], "Method"], [1986, 1988, ["feature", "representations"], "Method"], [2044, 2045, ["Freebase"], "Material"], [2077, 2079, ["Path", "Representation"], "Method"], [2165, 2169, ["ψ", "(", "a", ")"], "Method"], [2213, 2215, ["Subgraph", "Representation"], "Method"], [2220, 2222, ["path", "representation"], "Method"], [2300, 2302, ["embedding", "representation"], "Method"], [2424, 2427, ["k", "dimensional", "embedding"], "Method"], [2519, 2521, ["subgraph", "approach"], "Method"], [2532, 2534, ["Loss", "Function"], "Metric"], [2546, 2552, ["margin", "-", "based", "ranking", "loss", "function"], "Method"], [2608, 2610, ["Minimizing", "Eq"], "Task"], [2691, 2692, ["Optimization"], "Task"], [2695, 2698, ["stochastic", "gradient", "descent"], "Method"], [2703, 2704, ["Hogwild"], "Method"], [2740, 2744, ["Multitask", "Training", "of", "Embeddings"], "Task"], [2788, 2790, ["paraphrase", "prediction"], "Task"], [2799, 2800, ["S"], "Method"], [2804, 2806, ["scoring", "function"], "Method"], [2961, 2962, ["Inference"], "Task"], [3023, 3025, ["precision", "issues"], "Metric"], [3063, 3064, ["Freebase"], "Material"], [3177, 3179, ["beam", "search"], "Method"], [3378, 3380, ["feature", "representation"], "Method"], [3382, 3383, ["prediction"], "Task"], [3463, 3465, ["F1", "score"], "Metric"], [3469, 3472, ["official", "evaluation", "script"], "Metric"], [3474, 3475, ["F1"], "Metric"], [3489, 3490, ["F1"], "Metric"], [3517, 3518, ["precision"], "Metric"], [3592, 3594, ["default", "approach"], "Method"], [3596, 3598, ["Subgraph", "representation"], "Method"], [3672, 3675, ["ranking", "based", "inference"], "Method"], [3728, 3730, ["Path", "representation"], "Method"], [3765, 3767, ["combination", "method"], "Method"], [3876, 3878, ["embedding", "model"], "Method"], [3882, 3884, ["open", "QA"], "Task"], [3914, 3915, ["WebQuestions"], "Material"], [943, 944, ["Freebase"], "Material"], [953, 954, ["Freebase"], "Material"], [1035, 1036, ["Freebase"], "Material"], [1036, 1037, ["Freebase"], "Material"], [1103, 1104, ["WebQuestions"], "Material"], [1127, 1128, ["Freebase"], "Material"], [1145, 1146, ["WebQuestions"], "Material"], [1248, 1249, ["Freebase"], "Material"], [1335, 1336, ["Freebase"], "Material"], [1368, 1369, ["Freebase"], "Material"], [1482, 1483, ["Freebase"], "Material"], [1819, 1820, ["Freebase"], "Material"], [2025, 2026, ["Freebase"], "Material"], [2915, 2916, ["Freebase"], "Material"], [3049, 3050, ["Freebase"], "Material"], [3485, 3486, ["F1"], "Metric"]]} {"doc_id": "3448e6a5039417dc1ae890efeca3bef5390ace7c", "text": ["document", ":", "[", "Combinatorial", "features", "are", "essential", "for", "the", "success", "of", "many", "commercial", "models", ".", "Manually", "crafting", "these", "features", "usually", "comes", "with", "high", "cost", "due", "to", "the", "variety", ",", "volume", "and", "velocity", "of", "raw", "data", "in", "web", "-", "scale", "systems", ".", "Factorization", "based", "models", ",", "which", "measure", "interactions", "in", "terms", "of", "vector", "product", ",", "can", "learn", "patterns", "of", "combinatorial", "features", "automatically", "and", "generalize", "to", "unseen", "features", "as", "well", ".", "With", "the", "great", "success", "of", "deep", "neural", "networks", "(", "DNNs", ")", "in", "various", "fields", ",", "recently", "researchers", "have", "proposed", "several", "DNN", "-", "based", "factorization", "model", "to", "learn", "both", "low", "-", "and", "high", "-", "order", "feature", "interactions", ".", "Despite", "the", "powerful", "ability", "of", "learning", "an", "arbitrary", "function", "from", "data", ",", "plain", "DNNs", "generate", "feature", "interactions", "implicitly", "and", "at", "the", "bit", "-", "wise", "level", ".", "In", "this", "paper", ",", "we", "propose", "a", "novel", "Compressed", "Interaction", "Network", "(", "CIN", ")", ",", "which", "aims", "to", "generate", "feature", "interactions", "in", "an", "explicit", "fashion", "and", "at", "the", "vector", "-", "wise", "level", ".", "We", "show", "that", "the", "CIN", "share", "some", "functionalities", "with", "convolutional", "neural", "networks", "(", "CNNs", ")", "and", "recurrent", "neural", "networks", "(", "RNNs", ")", ".", "We", "further", "combine", "a", "CIN", "and", "a", "classical", "DNN", "into", "one", "unified", "model", ",", "and", "named", "this", "new", "model", "eXtreme", "Deep", "Factorization", "Machine", "(", "xDeepFM", ")", ".", "On", "one", "hand", ",", "the", "xDeepFM", "is", "able", "to", "learn", "certain", "bounded", "-", "degree", "feature", "interactions", "explicitly", ";", "on", "the", "other", "hand", ",", "it", "can", "learn", "arbitrary", "low", "-", "and", "high", "-", "order", "feature", "interactions", "implicitly", ".", "We", "conduct", "comprehensive", "experiments", "on", "three", "real", "-", "world", "datasets", ".", "Our", "results", "demonstrate", "that", "xDeepFM", "outperforms", "state", "-", "of", "-", "the", "-", "art", "models", ".", "We", "have", "released", "the", "source", "code", "of", "xDeepFM", "at", ".", "Combining", "Explicit", "and", "Implicit", "Feature", "Interactions", "for", "Recommender", "Systems", "]", "xDeepFM", ":", "Combining", "Explicit", "and", "Implicit", "Feature", "Interactions", "for", "Recommender", "Systems", "J.", "Lian", "]", "Jianxun", "Lian", "X.", "Zhou", "]", "Xiaohuan", "Zhou", "F.", "Zhang", "]", "Fuzheng", "Zhang", "Z.", "Chen", "]", "Zhongxia", "Chen", "X.", "Xie", "]", "Xing", "Xie", "G.", "Sun", "]", "Guangzhong", "Sun", "¡", "ccs2012", "¿", "¡", "concept", "¿", "¡", "concept_id¿10002951.10003260.10003261.10003271¡", "/", "concept_id", "¿", "¡", "concept_desc¿Information", "systems", "Personalization¡", "/", "concept_desc", "¿", "¡", "concept_significance¿500¡", "/", "concept_significance", "¿", "¡", "/", "concept", "¿", "¡", "concept", "¿", "¡", "concept_id¿10010147.10010257.10010293.10010294¡", "/", "concept_id", "¿", "¡", "concept_desc¿Computing", "methodologies", "Neural", "networks¡", "/", "concept_desc", "¿", "¡", "concept_significance¿500¡", "/", "concept_significance", "¿", "¡", "/", "concept", "¿", "¡", "concept", "¿", "¡", "concept_id¿10010147.10010257.10010293.10010309¡", "/", "concept_id", "¿", "¡", "concept_desc¿Computing", "methodologies", "Factorization", "methods¡", "/", "concept_desc", "¿", "¡", "concept_significance¿500¡", "/", "concept_significance", "¿", "¡", "/", "concept", "¿", "¡", "/", "ccs2012", "¿", "[", "500", "]", "Information", "systems", "Personalization", "[", "500", "]", "Computing", "methodologies", "Neural", "networks", "[", "500", "]", "Computing", "methodologies", "Factorization", "methods", "2018", "2018", "acmcopyright", "[", "KDD", "’", "18", "]", "The", "24th", "ACM", "SIGKDD", "International", "Conference", "on", "Knowledge", "Discovery", "&", "Data", "MiningAugust", "19–23", ",", "2018London", ",", "United", "Kingdom", "KDD", "’", "18", ":", "The", "24th", "ACM", "SIGKDD", "International", "Conference", "on", "Knowledge", "Discovery", "&", "Data", "Mining", ",", "August", "19–23", ",", "2018", ",", "London", ",", "United", "Kingdom", "15.00", "10.1145", "/", "3219819.3220023", "978", "-", "1", "-", "4503", "-", "5552", "-", "0", "/", "18", "/", "08", "section", ":", "introduction", "Features", "play", "a", "central", "role", "in", "the", "success", "of", "many", "predictive", "systems", ".", "Because", "using", "raw", "features", "can", "rarely", "lead", "to", "optimal", "results", ",", "data", "scientists", "usually", "spend", "a", "lot", "of", "work", "on", "the", "transformation", "of", "raw", "features", "in", "order", "to", "generate", "best", "predictive", "systems", "or", "to", "win", "data", "mining", "games", ".", "One", "major", "type", "of", "feature", "transformation", "is", "the", "cross", "-", "product", "transformation", "over", "categorical", "features", ".", "These", "features", "are", "called", "cross", "features", "or", "multi", "-", "way", "features", ",", "they", "measure", "the", "interactions", "of", "multiple", "raw", "features", ".", "For", "instance", ",", "a", "3", "-", "way", "feature", "AND", "(", "user_organization", "=", "msra", ",", "item_category", "=", "deeplearning", ",", "time", "=", "monday", ")", "has", "value", "1", "if", "the", "user", "works", "at", "Microsoft", "Research", "Asia", "and", "is", "shown", "a", "technical", "article", "about", "deep", "learning", "on", "a", "Monday", ".", "There", "are", "three", "major", "downsides", "for", "traditional", "cross", "feature", "engineering", ".", "First", ",", "obtaining", "high", "-", "quality", "features", "comes", "with", "a", "high", "cost", ".", "Because", "right", "features", "are", "usually", "task", "-", "specific", ",", "data", "scientists", "need", "spend", "a", "lot", "of", "time", "exploring", "the", "potential", "patterns", "from", "the", "product", "data", "before", "they", "become", "domain", "experts", "and", "extract", "meaningful", "cross", "features", ".", "Second", ",", "in", "large", "-", "scale", "predictive", "systems", "such", "as", "web", "-", "scale", "recommender", "systems", ",", "the", "huge", "number", "of", "raw", "features", "makes", "it", "infeasible", "to", "extract", "all", "cross", "features", "manually", ".", "Third", ",", "hand", "-", "crafted", "cross", "features", "do", "not", "generalize", "to", "unseen", "interactions", "in", "the", "training", "data", ".", "Therefore", ",", "learning", "to", "interact", "features", "without", "manual", "engineering", "is", "a", "meaningful", "task", ".", "Factorization", "Machines", "(", "FM", ")", "embed", "each", "feature", "to", "a", "latent", "factor", "vector", ",", "and", "pairwise", "feature", "interactions", "are", "modeled", "as", "the", "inner", "product", "of", "latent", "vectors", ":", ".", "In", "this", "paper", "we", "use", "the", "term", "bit", "to", "denote", "a", "element", "(", "such", "as", ")", "in", "latent", "vectors", ".", "The", "classical", "FM", "can", "be", "extended", "to", "arbitrary", "higher", "-", "order", "feature", "interactions", ",", "but", "one", "major", "downside", "is", "that", ",", "proposes", "to", "model", "all", "feature", "interactions", ",", "including", "both", "useful", "and", "useless", "combinations", ".", "As", "revealed", "in", ",", "the", "interactions", "with", "useless", "features", "may", "introduce", "noises", "and", "degrade", "the", "performance", ".", "In", "recent", "years", ",", "deep", "neural", "networks", "(", "DNNs", ")", "have", "become", "successful", "in", "computer", "vision", ",", "speech", "recognition", ",", "and", "natural", "language", "processing", "with", "their", "great", "power", "of", "feature", "representation", "learning", ".", "It", "is", "promising", "to", "exploit", "DNNs", "to", "learn", "sophisticated", "and", "selective", "feature", "interactions", ".", "proposes", "a", "Factorisation", "-", "machine", "supported", "Neural", "Network", "(", "FNN", ")", "to", "learn", "high", "-", "order", "feature", "interactions", ".", "It", "uses", "the", "pre", "-", "trained", "factorization", "machines", "for", "field", "embedding", "before", "applying", "DNN", ".", "further", "proposes", "a", "Product", "-", "based", "Neural", "Network", "(", "PNN", ")", ",", "which", "introduces", "a", "product", "layer", "between", "embedding", "layer", "and", "DNN", "layer", ",", "and", "does", "not", "rely", "on", "pre", "-", "trained", "FM", ".", "The", "major", "downside", "of", "FNN", "and", "PNN", "is", "that", "they", "focus", "more", "on", "high", "-", "order", "feature", "interactions", "while", "capture", "little", "low", "-", "order", "interactions", ".", "The", "Wide", "&", "Deep", "and", "DeepFM", "models", "overcome", "this", "problem", "by", "introducing", "hybrid", "architectures", ",", "which", "contain", "a", "shallow", "component", "and", "a", "deep", "component", "with", "the", "purpose", "of", "learning", "both", "memorization", "and", "generalization", ".", "Therefore", "they", "can", "jointly", "learn", "low", "-", "order", "and", "high", "-", "order", "feature", "interactions", ".", "All", "the", "abovementioned", "models", "leverage", "DNNs", "for", "learning", "high", "-", "order", "feature", "interactions", ".", "However", ",", "DNNs", "model", "high", "-", "order", "feature", "interactions", "in", "an", "implicit", "fashion", ".", "The", "final", "function", "learned", "by", "DNNs", "can", "be", "arbitrary", ",", "and", "there", "is", "no", "theoretical", "conclusion", "on", "what", "the", "maximum", "degree", "of", "feature", "interactions", "is", ".", "In", "addition", ",", "DNNs", "model", "feature", "interactions", "at", "the", "bit", "-", "wise", "level", ",", "which", "is", "different", "from", "the", "traditional", "FM", "framework", "which", "models", "feature", "interactions", "at", "the", "vector", "-", "wise", "level", ".", "Thus", ",", "in", "the", "field", "of", "recommender", "systems", ",", "whether", "DNNs", "are", "indeed", "the", "most", "effective", "model", "in", "representing", "high", "-", "order", "feature", "interactions", "remains", "an", "open", "question", ".", "In", "this", "paper", ",", "we", "propose", "a", "neural", "network", "-", "based", "model", "to", "learn", "feature", "interactions", "in", "an", "explicit", ",", "vector", "-", "wise", "fashion", ".", "Our", "approach", "is", "based", "on", "the", "Deep", "&", "Cross", "Network", "(", "DCN", ")", ",", "which", "aims", "to", "efficiently", "capture", "feature", "interactions", "of", "bounded", "degrees", ".", "However", ",", "we", "will", "argue", "in", "Section", "[", "reference", "]", "that", "DCN", "will", "lead", "to", "a", "special", "format", "of", "interactions", ".", "We", "thus", "design", "a", "novel", "compressed", "interaction", "network", "(", "CIN", ")", "to", "replace", "the", "cross", "network", "in", "the", "DCN", ".", "CIN", "learns", "feature", "interactions", "explicitly", ",", "and", "the", "degree", "of", "interactions", "grows", "with", "the", "depth", "of", "the", "network", ".", "Following", "the", "spirit", "of", "the", "Wide", "&", "Deep", "and", "DeepFM", "models", ",", "we", "combine", "the", "explicit", "high", "-", "order", "interaction", "module", "with", "implicit", "interaction", "module", "and", "traditional", "FM", "module", ",", "and", "name", "the", "joint", "model", "eXtreme", "Deep", "Factorization", "Machine", "(", "xDeepFM", ")", ".", "The", "new", "model", "requires", "no", "manual", "feature", "engineering", "and", "release", "data", "scientists", "from", "tedious", "feature", "searching", "work", ".", "To", "summarize", ",", "we", "make", "the", "following", "contributions", ":", "We", "propose", "a", "novel", "model", ",", "named", "eXtreme", "Deep", "Factorization", "Machine", "(", "xDeepFM", ")", ",", "that", "jointly", "learns", "explicit", "and", "implicit", "high", "-", "order", "feature", "interactions", "effectively", "and", "requires", "no", "manual", "feature", "engineering", ".", "We", "design", "a", "compressed", "interaction", "network", "(", "CIN", ")", "in", "xDeepFM", "that", "learns", "high", "-", "order", "feature", "interactions", "explicitly", ".", "We", "show", "that", "the", "degree", "of", "feature", "interactions", "increases", "at", "each", "layer", ",", "and", "features", "interact", "at", "the", "vector", "-", "wise", "level", "rather", "than", "the", "bit", "-", "wise", "level", ".", "We", "conduct", "extensive", "experiments", "on", "three", "real", "-", "world", "dataset", ",", "and", "the", "results", "demonstrate", "that", "our", "xDeepFM", "outperforms", "several", "state", "-", "of", "-", "the", "-", "art", "models", "significantly", ".", "The", "rest", "of", "this", "paper", "is", "organized", "as", "follows", ".", "Section", "[", "reference", "]", "provides", "some", "preliminary", "knowledge", "which", "is", "necessary", "for", "understanding", "deep", "learning", "-", "based", "recommender", "systems", ".", "Section", "[", "reference", "]", "introduces", "our", "proposed", "CIN", "and", "xDeepFM", "model", "in", "detail", ".", "We", "will", "present", "experimental", "explorations", "on", "multiple", "datasets", "in", "Section", "[", "reference", "]", ".", "Related", "works", "are", "discussed", "in", "Section", "[", "reference", "]", ".", "Section", "[", "reference", "]", "concludes", "this", "paper", ".", "section", ":", "Preliminaries", "subsection", ":", "Embedding", "Layer", "In", "computer", "vision", "or", "natural", "language", "understanding", ",", "the", "input", "data", "are", "usually", "images", "or", "textual", "signals", ",", "which", "are", "known", "to", "be", "spatially", "and", "/", "or", "temporally", "correlated", ",", "so", "DNNs", "can", "be", "applied", "directly", "on", "the", "raw", "feature", "with", "dense", "structures", ".", "However", ",", "in", "web", "-", "scale", "recommender", "systems", ",", "the", "input", "features", "are", "sparse", ",", "of", "huge", "dimension", ",", "and", "present", "no", "clear", "spatial", "or", "temporal", "correlation", ".", "Therefore", ",", "multi", "-", "field", "categorical", "form", "is", "widely", "used", "by", "related", "works", ".", "For", "example", ",", "one", "input", "instance", "[", "user_id", "=", "s02", ",", "gender", "=", "male", ",", "organization", "=", "msra", ",", "interests", "=", "comedy", "&", "rock", "]", "is", "normally", "transformed", "into", "a", "high", "-", "dimensional", "sparse", "features", "via", "field", "-", "aware", "one", "-", "hot", "encoding", ":", "An", "embedding", "layer", "is", "applied", "upon", "the", "raw", "feature", "input", "to", "compress", "it", "to", "a", "low", "dimensional", ",", "dense", "real", "-", "value", "vector", ".", "If", "the", "field", "is", "univalent", ",", "the", "feature", "embedding", "is", "used", "as", "the", "field", "embedding", ".", "Take", "the", "above", "instance", "as", "an", "example", ",", "the", "embedding", "of", "feature", "male", "is", "taken", "as", "the", "embedding", "of", "field", "gender", ".", "If", "the", "field", "is", "multivalent", ",", "the", "sum", "of", "feature", "embedding", "is", "used", "as", "the", "field", "embedding", ".", "The", "embedding", "layer", "is", "illustrated", "in", "Figure", "[", "reference", "]", ".", "The", "result", "of", "embedding", "layer", "is", "a", "wide", "concatenated", "vector", ":", "where", "denotes", "the", "number", "of", "fields", ",", "and", "denotes", "the", "embedding", "of", "one", "field", ".", "Although", "the", "feature", "lengths", "of", "instances", "can", "be", "various", ",", "their", "embeddings", "are", "of", "the", "same", "length", ",", "where", "is", "the", "dimension", "of", "field", "embedding", ".", "subsection", ":", "Implicit", "High", "-", "order", "Interactions", "FNN", ",", "Deep", "Crossing", ",", "and", "the", "deep", "part", "in", "Wide", "&", "Deep", "exploit", "a", "feed", "-", "forward", "neural", "network", "on", "the", "field", "embedding", "vector", "to", "learn", "high", "-", "order", "feature", "interactions", ".", "The", "forward", "process", "is", ":", "where", "is", "the", "layer", "depth", ",", "is", "an", "activation", "function", ",", "and", "is", "the", "output", "of", "the", "-", "th", "layer", ".", "The", "visual", "structure", "is", "very", "similar", "to", "what", "is", "shown", "in", "Figure", "[", "reference", "]", ",", "except", "that", "they", "do", "not", "include", "the", "FM", "or", "Product", "layer", ".", "This", "architecture", "models", "the", "interaction", "in", "a", "bit", "-", "wise", "fashion", ".", "That", "is", "to", "say", ",", "even", "the", "elements", "within", "the", "same", "field", "embedding", "vector", "will", "influence", "each", "other", ".", "PNN", "and", "DeepFM", "modify", "the", "above", "architecture", "slightly", ".", "Besides", "applying", "DNNs", "on", "the", "embedding", "vector", ",", "they", "add", "a", "two", "-", "way", "interaction", "layer", "in", "the", "architecture", ".", "Therefore", ",", "both", "bit", "-", "wise", "and", "vector", "-", "wise", "interaction", "is", "included", "in", "their", "model", ".", "The", "major", "difference", "between", "PNN", "and", "DeepFM", ",", "is", "that", "PNN", "connects", "the", "outputs", "of", "product", "layer", "to", "the", "DNNs", ",", "whereas", "DeepFM", "connects", "the", "FM", "layer", "directly", "to", "the", "output", "unit", "(", "refer", "to", "Figure", "[", "reference", "]", ")", ".", "subsection", ":", "Explicit", "High", "-", "order", "Interactions", "proposes", "the", "Cross", "Network", "(", "CrossNet", ")", "whose", "architecture", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "It", "aims", "to", "explicitly", "model", "the", "high", "-", "order", "feature", "interactions", ".", "Unlike", "the", "classical", "fully", "-", "connected", "feed", "-", "forward", "network", ",", "the", "hidden", "layers", "are", "calculated", "by", "the", "following", "cross", "operation", ":", "where", "are", "weights", ",", "bias", "and", "output", "of", "the", "-", "th", "layer", ",", "respectively", ".", "We", "argue", "that", "the", "CrossNet", "learns", "a", "special", "type", "of", "high", "-", "order", "feature", "interactions", ",", "where", "each", "hidden", "layer", "in", "the", "CrossNet", "is", "a", "scalar", "multiple", "of", ".", "Consider", "a", "-", "layer", "cross", "network", "with", "the", "(", "i", "+", "1", ")-", "th", "layer", "defined", "as", ".", "Then", ",", "the", "output", "of", "the", "cross", "network", "is", "a", "scalar", "multiple", "of", ".", "When", "=", "1", ",", "according", "to", "the", "associative", "law", "and", "distributive", "law", "for", "matrix", "multiplication", ",", "we", "have", ":", "where", "the", "scalar", "is", "actually", "a", "linear", "regression", "of", ".", "Thus", ",", "is", "a", "scalar", "multiple", "of", ".", "Suppose", "the", "scalar", "multiple", "statement", "holds", "for", "=", ".", "For", "=", ",", "we", "have", ":", "where", ",", "is", "a", "scalar", ".", "Thus", "is", "still", "a", "scalar", "multiple", "of", ".", "By", "induction", "hypothesis", ",", "the", "output", "of", "cross", "network", "is", "a", "scalar", "multiple", "of", ".", "Note", "that", "the", "scalar", "multiple", "does", "not", "mean", "is", "linear", "with", ".", "The", "coefficient", "is", "sensitive", "with", ".", "The", "CrossNet", "can", "learn", "feature", "interactions", "very", "efficiently", "(", "the", "complexity", "is", "negligible", "compared", "with", "a", "DNN", "model", ")", ",", "however", "the", "downsides", "are", ":", "(", "1", ")", "the", "output", "of", "CrossNet", "is", "limited", "in", "a", "special", "form", ",", "with", "each", "hidden", "layer", "is", "a", "scalar", "multiple", "of", ";", "(", "2", ")", "interactions", "come", "in", "a", "bit", "-", "wise", "fashion", ".", "section", ":", "Our", "proposed", "model", ".33", ".32", ".32", "subsection", ":", "Compressed", "Interaction", "Network", "We", "design", "a", "new", "cross", "network", ",", "named", "Compressed", "Interaction", "Network", "(", "CIN", ")", ",", "with", "the", "following", "considerations", ":", "(", "1", ")", "interactions", "are", "applied", "at", "vector", "-", "wise", "level", ",", "not", "at", "bit", "-", "wise", "level", ";", "(", "2", ")", "high", "-", "order", "feature", "interactions", "is", "measured", "explicitly", ";", "(", "3", ")", "the", "complexity", "of", "network", "will", "not", "grow", "exponentially", "with", "the", "degree", "of", "interactions", ".", "Since", "an", "embedding", "vector", "is", "regarded", "as", "a", "unit", "for", "vector", "-", "wise", "interactions", ",", "hereafter", "we", "formulate", "the", "output", "of", "field", "embedding", "as", "a", "matrix", ",", "where", "the", "-", "th", "row", "in", "is", "the", "embedding", "vector", "of", "the", "-", "th", "field", ":", ",", "and", "is", "the", "dimension", "of", "the", "field", "embedding", ".", "The", "output", "of", "the", "-", "th", "layer", "in", "CIN", "is", "also", "a", "matrix", ",", "where", "denotes", "the", "number", "of", "(", "embedding", ")", "feature", "vectors", "in", "the", "-", "th", "layer", "and", "we", "let", ".", "For", "each", "layer", ",", "are", "calculated", "via", ":", "where", ",", "is", "the", "parameter", "matrix", "for", "the", "-", "th", "feature", "vector", ",", "and", "denotes", "the", "Hadamard", "product", ",", "for", "example", ",", ".", "Note", "that", "is", "derived", "via", "the", "interactions", "between", "and", ",", "thus", "feature", "interactions", "are", "measured", "explicitly", "and", "the", "degree", "of", "interactions", "increases", "with", "the", "layer", "depth", ".", "The", "structure", "of", "CIN", "is", "very", "similar", "to", "the", "Recurrent", "Neural", "Network", "(", "RNN", ")", ",", "where", "the", "outputs", "of", "the", "next", "hidden", "layer", "are", "dependent", "on", "the", "last", "hidden", "layer", "and", "an", "additional", "input", ".", "We", "hold", "the", "structure", "of", "embedding", "vectors", "at", "all", "layers", ",", "thus", "the", "interactions", "are", "applied", "at", "the", "vector", "-", "wise", "level", ".", "It", "is", "interesting", "to", "point", "out", "that", "Equation", "[", "reference", "]", "has", "strong", "connections", "with", "the", "well", "-", "known", "Convolutional", "Neural", "Networks", "(", "CNNs", ")", "in", "computer", "vision", ".", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "we", "introduce", "an", "intermediate", "tensor", ",", "which", "is", "the", "outer", "products", "(", "along", "each", "embedding", "dimension", ")", "of", "hidden", "layer", "and", "original", "feature", "matrix", ".", "Then", "can", "be", "regarded", "as", "a", "special", "type", "of", "image", "and", "is", "a", "filter", ".", "We", "slide", "the", "filter", "across", "along", "the", "embedding", "dimension", "(", "D", ")", "as", "shown", "in", "Figure", "[", "reference", "]", ",", "and", "get", "an", "hidden", "vector", ",", "which", "is", "usually", "called", "a", "feature", "map", "in", "computer", "vision", ".", "Therefore", ",", "is", "a", "collection", "of", "different", "feature", "maps", ".", "The", "term", "“", "compressed", "”", "in", "the", "name", "of", "CIN", "indicates", "that", "the", "-", "th", "hidden", "layer", "compress", "the", "potential", "space", "of", "vectors", "down", "to", "vectors", ".", "Figure", "[", "reference", "]", "provides", "an", "overview", "of", "the", "architecture", "of", "CIN", ".", "Let", "T", "denotes", "the", "depth", "of", "the", "network", ".", "Every", "hidden", "layer", "has", "a", "connection", "with", "output", "units", ".", "We", "first", "apply", "sum", "pooling", "on", "each", "feature", "map", "of", "the", "hidden", "layer", ":", "for", ".", "Thus", ",", "we", "have", "a", "pooling", "vector", "with", "length", "for", "the", "-", "th", "hidden", "layer", ".", "All", "pooling", "vectors", "from", "hidden", "layers", "are", "concatenated", "before", "connected", "to", "output", "units", ":", ".", "If", "we", "use", "CIN", "directly", "for", "binary", "classification", ",", "the", "output", "unit", "is", "a", "sigmoid", "node", "on", ":", "where", "are", "the", "regression", "parameters", ".", "subsection", ":", "CIN", "Analysis", "We", "analyze", "the", "proposed", "CIN", "to", "study", "the", "model", "complexity", "and", "the", "potential", "effectiveness", ".", "subsubsection", ":", "Space", "Complexity", "The", "-", "th", "feature", "map", "at", "the", "-", "th", "layer", "contains", "parameters", ",", "which", "is", "exactly", "the", "size", "of", ".", "Thus", ",", "there", "are", "parameters", "at", "the", "-", "th", "layer", ".", "Considering", "the", "last", "regression", "layer", "for", "the", "output", "unit", ",", "which", "has", "parameters", ",", "the", "total", "number", "of", "parameters", "for", "CIN", "is", ".", "Note", "that", "CIN", "is", "independent", "of", "the", "embedding", "dimension", ".", "In", "contrast", ",", "a", "plain", "-", "layers", "DNN", "contains", "parameters", ",", "and", "the", "number", "of", "parameters", "will", "increase", "with", "the", "embedding", "dimension", ".", "Usually", "and", "will", "not", "be", "very", "large", ",", "so", "the", "scale", "of", "is", "acceptable", ".", "When", "necessary", ",", "we", "can", "exploit", "a", "-", "order", "decomposition", "and", "replace", "with", "two", "smaller", "matrices", "and", ":", "where", "and", ".", "Hereafter", "we", "assume", "that", "each", "hidden", "layer", "has", "the", "same", "number", "(", "which", "is", ")", "of", "feature", "maps", "for", "simplicity", ".", "Through", "the", "-", "order", "decomposition", ",", "the", "space", "complexity", "of", "CIN", "is", "reduced", "from", "to", ".", "In", "contrast", ",", "the", "space", "complexity", "of", "the", "plain", "DNN", "is", ",", "which", "is", "sensitive", "to", "the", "dimension", "(", "D", ")", "of", "field", "embedding", ".", "subsubsection", ":", "Time", "Complexity", "The", "cost", "of", "computing", "tensor", "(", "as", "shown", "in", "Figure", "[", "reference", "]", ")", "is", "time", ".", "Because", "we", "have", "feature", "maps", "in", "one", "hidden", "layer", ",", "computing", "a", "-", "layers", "CIN", "takes", "time", ".", "A", "-", "layers", "plain", "DNN", ",", "by", "contrast", ",", "takes", "time", ".", "Therefore", ",", "the", "major", "downside", "of", "CIN", "lies", "in", "the", "time", "complexity", ".", "subsubsection", ":", "Polynomial", "Approximation", "Next", "we", "examine", "the", "high", "-", "order", "interaction", "properties", "of", "CIN", ".", "For", "simplicity", ",", "we", "assume", "that", "numbers", "of", "feature", "maps", "at", "hidden", "layers", "are", "all", "equal", "to", "the", "number", "of", "fields", ".", "Let", "denote", "the", "set", "of", "positive", "integers", "that", "are", "less", "than", "or", "equal", "to", ".", "The", "-", "th", "feature", "map", "at", "the", "first", "layer", ",", "denoted", "as", ",", "is", "calculated", "via", ":", "Therefore", ",", "each", "feature", "map", "at", "the", "first", "layer", "models", "pair", "-", "wise", "interactions", "with", "coefficients", ".", "Similarly", ",", "the", "-", "th", "feature", "map", "at", "the", "second", "layer", "is", ":", "Note", "that", "all", "calculations", "related", "to", "the", "subscript", "and", "is", "already", "finished", "at", "the", "previous", "hidden", "layer", ".", "We", "expand", "the", "factors", "in", "Equation", "[", "reference", "]", "just", "for", "clarity", ".", "We", "can", "observe", "that", "each", "feature", "map", "at", "the", "second", "layer", "models", "3", "-", "way", "interactions", "with", "new", "parameters", ".", "A", "classical", "-", "order", "polynomial", "has", "coefficients", ".", "We", "show", "that", "CIN", "approximate", "this", "class", "of", "polynomial", "with", "only", "parameters", "in", "terms", "of", "a", "chain", "of", "feature", "maps", ".", "By", "induction", "hypothesis", ",", "we", "can", "prove", "that", "the", "-", "th", "feature", "map", "at", "the", "-", "th", "layer", "is", ":", "For", "better", "illustration", ",", "here", "we", "borrow", "the", "notations", "from", ".", "Let", "denote", "a", "multi", "-", "index", ",", "and", ".", "We", "omit", "the", "original", "superscript", "from", ",", "and", "use", "to", "denote", "it", "since", "we", "only", "we", "the", "feature", "maps", "from", "the", "-", "th", "layer", "(", "which", "is", "exactly", "the", "field", "embeddings", ")", "for", "the", "final", "expanded", "expression", "(", "refer", "to", "Eq", ".", "[", "reference", "]", ")", ".", "Now", "a", "superscript", "is", "used", "to", "denote", "the", "vector", "operation", ",", "such", "as", ".", "Let", "denote", "a", "multi", "-", "vector", "polynomial", "of", "degree", ":", "Each", "vector", "polylnomial", "in", "this", "class", "has", "coefficients", ".", "Then", ",", "our", "CIN", "approaches", "the", "coefficient", "with", ":", "where", ",", "is", "a", "multi", "-", "index", ",", "and", "is", "the", "set", "of", "all", "the", "permutations", "of", "the", "indices", ".", "subsection", ":", "Combination", "with", "Implicit", "Networks", "As", "discussed", "in", "Section", "[", "reference", "]", ",", "plain", "DNNs", "learn", "implicit", "high", "-", "order", "feature", "interactions", ".", "Since", "CIN", "and", "plain", "DNNs", "can", "complement", "each", "other", ",", "an", "intuitive", "way", "to", "make", "the", "model", "stronger", "is", "to", "combine", "these", "two", "structures", ".", "The", "resulting", "model", "is", "very", "similar", "to", "the", "Wide", "&", "Deep", "or", "DeepFM", "model", ".", "The", "architecture", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "We", "name", "the", "new", "model", "eXtreme", "Deep", "Factorization", "Machine", "(", "xDeepFM", ")", ",", "considering", "that", "on", "one", "hand", ",", "it", "includes", "both", "low", "-", "order", "and", "high", "-", "order", "feature", "interactions", ";", "on", "the", "other", "hand", ",", "it", "includes", "both", "implicit", "feature", "interactions", "and", "explicit", "feature", "interactions", ".", "Its", "resulting", "output", "unit", "becomes", ":", "where", "is", "the", "sigmoid", "function", ",", "is", "the", "raw", "features", ".", "are", "the", "outputs", "of", "the", "plain", "DNN", "and", "CIN", ",", "respectively", ".", "and", "are", "learnable", "parameters", ".", "For", "binary", "classifications", ",", "the", "loss", "function", "is", "the", "log", "loss", ":", "where", "is", "the", "total", "number", "of", "training", "instances", ".", "The", "optimization", "process", "is", "to", "minimize", "the", "following", "objective", "function", ":", "where", "denotes", "the", "regularization", "term", "and", "denotes", "the", "set", "of", "parameters", ",", "including", "these", "in", "linear", "part", ",", "CIN", "part", ",", "and", "DNN", "part", ".", "subsubsection", ":", "Relationship", "with", "FM", "and", "DeepFM", "Suppose", "all", "fields", "are", "univalent", ".", "It", "’s", "not", "hard", "to", "observe", "from", "Figure", "[", "reference", "]", "that", ",", "when", "the", "depth", "and", "feature", "maps", "of", "the", "CIN", "part", "are", "both", "set", "to", "1", ",", "xDeepFM", "is", "a", "generalization", "of", "DeepFM", "by", "learning", "the", "linear", "regression", "weights", "for", "the", "FM", "layer", "(", "note", "that", "in", "DeepFM", ",", "units", "of", "FM", "layer", "are", "directly", "linked", "to", "the", "output", "unit", "without", "any", "coefficients", ")", ".", "When", "we", "further", "remove", "the", "DNN", "part", ",", "and", "at", "the", "same", "time", "use", "a", "constant", "sum", "filter", "(", "which", "simply", "takes", "the", "sum", "of", "inputs", "without", "any", "parameter", "learning", ")", "for", "the", "feature", "map", ",", "then", "xDeepFM", "is", "downgraded", "to", "the", "traditional", "FM", "model", ".", "section", ":", "Experiments", "In", "this", "section", ",", "we", "conduct", "extensive", "experiments", "to", "answer", "the", "following", "questions", ":", "(", "Q1", ")", "How", "does", "our", "proposed", "CIN", "perform", "in", "high", "-", "order", "feature", "interactions", "learning", "?", "(", "Q2", ")", "Is", "it", "necessary", "to", "combine", "explicit", "and", "implicit", "high", "-", "order", "feature", "interactions", "for", "recommender", "systems", "?", "(", "Q3", ")", "How", "does", "the", "settings", "of", "networks", "influence", "the", "performance", "of", "xDeepFM", "?", "We", "will", "answer", "these", "questions", "after", "presenting", "some", "fundamental", "experimental", "settings", ".", "subsection", ":", "Experiment", "Setup", "subsubsection", ":", "Datasets", ".", "We", "evaluate", "our", "proposed", "models", "on", "the", "following", "three", "datasets", ":", "1", ".", "Criteo", "Dataset", ".", "It", "is", "a", "famous", "industry", "benchmarking", "dataset", "for", "developing", "models", "predicting", "ad", "click", "-", "through", "rate", ",", "and", "is", "publicly", "accessible", ".", "Given", "a", "user", "and", "the", "page", "he", "is", "visiting", ",", "the", "goal", "is", "to", "predict", "the", "probability", "that", "he", "will", "clik", "on", "a", "given", "ad", ".", "2", ".", "Dianping", "Dataset", ".", "Dianping.com", "is", "the", "largest", "consumer", "review", "site", "in", "China", ".", "It", "provides", "diverse", "functions", "such", "as", "reviews", ",", "check", "-", "ins", ",", "and", "shops", "’", "meta", "information", "(", "including", "geographical", "messages", "and", "shop", "attributes", ")", ".", "We", "collect", "6", "months", "’", "users", "check", "-", "in", "activities", "for", "restaurant", "recommendation", "experiments", ".", "Given", "a", "user", "’s", "profile", ",", "a", "restaurant", "’s", "attributes", "and", "the", "user", "’s", "last", "three", "visited", "POIs", "(", "point", "of", "interest", ")", ",", "we", "want", "to", "predict", "the", "probability", "that", "he", "will", "visit", "the", "restaurant", ".", "For", "each", "restaurant", "in", "a", "user", "’s", "check", "-", "in", "instance", ",", "we", "sample", "four", "restaurants", "which", "are", "within", "3", "kilometers", "as", "negative", "instances", "by", "POI", "popularity", ".", "3", ".", "Bing", "News", "Dataset", ".", "Bing", "News", "is", "part", "of", "Microsoft", "’s", "Bing", "search", "engine", ".", "In", "order", "to", "evaluate", "the", "performance", "of", "our", "model", "in", "a", "real", "commercial", "dataset", ",", "we", "collect", "five", "consecutive", "days", "’", "impression", "logs", "on", "news", "reading", "service", ".", "We", "use", "the", "first", "three", "days", "’", "data", "for", "training", "and", "validation", ",", "and", "the", "next", "two", "days", "for", "testing", ".", "For", "the", "Criteo", "dataset", "and", "the", "Dianping", "dataset", ",", "we", "randomly", "split", "instances", "by", "8:1:1", "for", "training", ",", "validation", "and", "test", ".", "The", "characteristics", "of", "the", "three", "datasets", "are", "summarized", "in", "Table", "[", "reference", "]", ".", "subsubsection", ":", "Evaluation", "Metrics", ".", "We", "use", "two", "metrics", "for", "model", "evaluation", ":", "AUC", "(", "Area", "Under", "the", "ROC", "curve", ")", "and", "Logloss", "(", "cross", "entropy", ")", ".", "These", "two", "metrics", "evaluate", "the", "performance", "from", "two", "different", "angels", ":", "AUC", "measures", "the", "probability", "that", "a", "positive", "instance", "will", "be", "ranked", "higher", "than", "a", "randomly", "chosen", "negative", "one", ".", "It", "only", "takes", "into", "account", "the", "order", "of", "predicted", "instances", "and", "is", "insensitive", "to", "class", "imbalance", "problem", ".", "Logloss", ",", "in", "contrast", ",", "measures", "the", "distance", "between", "the", "predicted", "score", "and", "the", "true", "label", "for", "each", "instance", ".", "Sometimes", "we", "rely", "more", "on", "Logloss", "because", "we", "need", "to", "use", "the", "predicted", "probability", "to", "estimate", "the", "benefit", "of", "a", "ranking", "strategy", "(", "which", "is", "usually", "adjusted", "as", "CTR", "bid", ")", ".", "subsubsection", ":", "Baselines", ".", "We", "compare", "our", "xDeepFM", "with", "LR", "(", "logistic", "regression", ")", ",", "FM", ",", "DNN", "(", "plain", "deep", "neural", "network", ")", ",", "PNN", "(", "choose", "the", "better", "one", "from", "iPNN", "and", "oPNN", ")", ",", "Wide", "&", "Deep", ",", "DCN", "(", "Deep", "&", "Cross", "Network", ")", "and", "DeepFM", ".", "As", "introduced", "and", "discussed", "in", "Section", "[", "reference", "]", ",", "these", "models", "are", "highly", "related", "to", "our", "xDeepFM", "and", "some", "of", "them", "are", "state", "-", "of", "-", "the", "-", "art", "models", "for", "recommender", "systems", ".", "Note", "that", "the", "focus", "of", "this", "paper", "is", "to", "learn", "feature", "interactions", "automatically", ",", "so", "we", "do", "not", "include", "any", "hand", "-", "crafted", "cross", "features", ".", "subsubsection", ":", "Reproducibility", "We", "implement", "our", "method", "using", "Tensorflow", ".", "Hyper", "-", "parameters", "of", "each", "model", "are", "tuned", "by", "grid", "-", "searching", "on", "the", "validation", "set", ",", "and", "the", "best", "settings", "for", "each", "model", "will", "be", "shown", "in", "corresponding", "sections", ".", "Learning", "rate", "is", "set", "to", "0.001", ".", "For", "optimization", "method", ",", "we", "use", "the", "Adam", "with", "a", "mini", "-", "batch", "size", "of", "4096", ".", "We", "use", "a", "L2", "regularization", "with", "for", "DNN", ",", "DCN", ",", "Wide", "&", "Deep", ",", "DeepFM", "and", "xDeepFM", ",", "and", "use", "dropout", "0.5", "for", "PNN", ".", "The", "default", "setting", "for", "number", "of", "neurons", "per", "layer", "is", ":", "(", "1", ")", "400", "for", "DNN", "layers", ";", "(", "2", ")", "200", "for", "CIN", "layers", "on", "Criteo", "dataset", ",", "and", "100", "for", "CIN", "layers", "on", "Dianping", "and", "Bing", "News", "datasets", ".", "Since", "we", "focus", "on", "neural", "networks", "structures", "in", "this", "paper", ",", "we", "make", "the", "dimension", "of", "field", "embedding", "for", "all", "models", "be", "a", "fixed", "value", "of", "10", ".", "We", "conduct", "experiments", "of", "different", "settings", "in", "parallel", "with", "5", "Tesla", "K80", "GPUs", ".", "The", "source", "code", "is", "available", "at", ".", "subsection", ":", "Performance", "Comparison", "among", "Individual", "Neural", "Components", "(", "Q1", ")", "We", "want", "to", "know", "how", "CIN", "performs", "individually", ".", "Note", "that", "FM", "measures", "2", "-", "order", "feature", "interactions", "explicitly", ",", "DNN", "model", "high", "-", "order", "feature", "interactions", "implicitly", ",", "CrossNet", "tries", "to", "model", "high", "-", "order", "feature", "interactions", "with", "a", "small", "number", "of", "parameters", "(", "which", "is", "proven", "not", "effective", "in", "Section", "[", "reference", "]", ")", ",", "and", "CIN", "models", "high", "-", "order", "feature", "interactions", "explicitly", ".", "There", "is", "no", "theoretic", "guarantee", "of", "the", "superiority", "of", "one", "individual", "model", "over", "the", "others", ",", "due", "to", "that", "it", "really", "depends", "on", "the", "dataset", ".", "For", "example", ",", "if", "the", "practical", "dataset", "does", "not", "require", "high", "-", "order", "feature", "interactions", ",", "FM", "may", "be", "the", "best", "individual", "model", ".", "Thus", "we", "do", "not", "have", "any", "expectation", "for", "which", "model", "will", "perform", "the", "best", "in", "this", "experiment", ".", "Table", "[", "reference", "]", "shows", "the", "results", "of", "individual", "models", "on", "the", "three", "practical", "datasets", ".", "Surprisingly", ",", "our", "CIN", "outperform", "the", "other", "models", "consistently", ".", "On", "one", "hand", ",", "the", "results", "indicate", "that", "for", "practical", "datasets", ",", "higher", "-", "order", "interactions", "over", "sparse", "features", "are", "necessary", ",", "and", "this", "can", "be", "verified", "through", "the", "fact", "that", "DNN", ",", "CrossNet", "and", "CIN", "outperform", "FM", "significantly", "on", "all", "the", "three", "datasets", ".", "On", "the", "other", "hand", ",", "CIN", "is", "the", "best", "individual", "model", ",", "which", "demonstrates", "the", "effectiveness", "of", "CIN", "on", "modeling", "explicit", "high", "-", "order", "feature", "interactions", ".", "Note", "that", "a", "-", "layer", "CIN", "can", "model", "-", "degree", "feature", "interactions", ".", "It", "is", "also", "interesting", "to", "see", "that", "it", "take", "5", "layers", "for", "CIN", "to", "yield", "the", "best", "result", "ON", "the", "Bing", "News", "dataset", ".", "subsection", ":", "Performance", "of", "Integrated", "Models", "(", "Q2", ")", "xDeepFM", "integrates", "CIN", "and", "DNN", "into", "an", "end", "-", "to", "-", "end", "model", ".", "While", "CIN", "and", "DNN", "covers", "two", "distinct", "properties", "in", "learning", "feature", "interactions", ",", "we", "are", "interested", "to", "know", "whether", "it", "is", "indeed", "necessary", "and", "effective", "to", "combine", "them", "together", "for", "jointly", "explicit", "and", "implicit", "learning", ".", "Here", "we", "compare", "several", "strong", "baselines", "which", "are", "not", "limited", "to", "individual", "models", ",", "and", "the", "results", "are", "shown", "in", "Table", "[", "reference", "]", ".", "We", "observe", "that", "LR", "is", "far", "worse", "than", "all", "the", "rest", "models", ",", "which", "demonstrates", "that", "factorization", "-", "based", "models", "are", "essential", "for", "measuring", "sparse", "features", ".", "Wide", "&", "Deep", ",", "DCN", ",", "DeepFM", "and", "xDeepFM", "are", "significantly", "better", "than", "DNN", ",", "which", "directly", "reflects", "that", ",", "despite", "their", "simplicity", ",", "incorporating", "hybrid", "components", "are", "important", "for", "boosting", "the", "accuracy", "of", "predictive", "systems", ".", "Our", "proposed", "xDeepFM", "achieves", "the", "best", "performance", "on", "all", "datasets", ",", "which", "demonstrates", "that", "combining", "explicit", "and", "implicit", "high", "-", "order", "feature", "interaction", "is", "necessary", ",", "and", "xDeepFM", "is", "effective", "in", "learning", "this", "class", "of", "combination", ".", "Another", "interesting", "observation", "is", "that", ",", "all", "the", "neural", "-", "based", "models", "do", "not", "require", "a", "very", "deep", "network", "structure", "for", "the", "best", "performance", ".", "Typical", "settings", "for", "the", "depth", "hyper", "-", "parameter", "are", "2", "and", "3", ",", "and", "the", "best", "depth", "setting", "for", "xDeepFM", "is", "3", ",", "which", "indicates", "that", "the", "interactions", "we", "learned", "are", "at", "most", "4", "-", "order", ".", "subsection", ":", "Hyper", "-", "Parameter", "Study", "(", "Q3", ")", "We", "study", "the", "impact", "of", "hyper", "-", "parameters", "on", "xDeepFM", "in", "this", "section", ",", "including", "(", "1", ")", "the", "number", "of", "hidden", "layers", ";", "(", "2", ")", "the", "number", "of", "neurons", "per", "layer", ";", "and", "(", "3", ")", "activation", "functions", ".", "We", "conduct", "experiments", "via", "holding", "the", "best", "settings", "for", "the", "DNN", "part", "while", "varying", "the", "settings", "for", "the", "CIN", "part", ".", ".32", ".32", ".32", ".32", ".32", ".32", "Depth", "of", "Network", ".", "Figure", "[", "reference", "]", "and", "[", "reference", "]", "demonstrate", "the", "impact", "of", "number", "of", "hidden", "layers", ".", "We", "can", "observe", "that", "the", "performance", "of", "xDeepFM", "increases", "with", "the", "depth", "of", "network", "at", "the", "beginning", ".", "However", ",", "model", "performance", "degrades", "when", "the", "depth", "of", "network", "is", "set", "greater", "than", "3", ".", "It", "is", "caused", "by", "overfitting", "evidenced", "by", "that", "we", "notice", "that", "the", "loss", "of", "training", "data", "still", "keeps", "decreasing", "when", "we", "add", "more", "hidden", "layers", ".", "Number", "of", "Neurons", "per", "Layer", ".", "Adding", "the", "number", "of", "neurons", "per", "layer", "indicates", "increasing", "the", "number", "of", "feature", "maps", "in", "CIN", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "and", "[", "reference", "]", ",", "model", "performance", "on", "Bing", "News", "dataset", "increases", "steadily", "when", "we", "increase", "the", "number", "of", "neurons", "from", "to", ",", "while", "on", "Dianping", "dataset", ",", "is", "a", "more", "suitable", "setting", "for", "the", "number", "of", "neurons", "per", "layer", ".", "In", "this", "experiment", "we", "fix", "the", "depth", "of", "network", "at", "3", ".", "Activation", "Function", ".", "Note", "that", "we", "exploit", "the", "identity", "as", "activation", "function", "on", "neurons", "of", "CIN", ",", "as", "shown", "in", "Eq", ".", "[", "reference", "]", ".", "A", "common", "practice", "in", "deep", "learning", "literature", "is", "to", "employ", "non", "-", "linear", "activation", "functions", "on", "hidden", "neurons", ".", "We", "thus", "compare", "the", "results", "of", "different", "activation", "functions", "on", "CIN", "(", "for", "neurons", "in", "DNN", ",", "we", "keep", "the", "activation", "function", "with", "relu", ")", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "and", "[", "reference", "]", ",", "identify", "function", "is", "indeed", "the", "most", "suitable", "one", "for", "neurons", "in", "CIN", ".", "section", ":", "related", "work", "subsection", ":", "Classical", "Recommender", "Systems", "subsubsection", ":", "Non", "-", "factorization", "Models", "For", "web", "-", "scale", "recommender", "systems", "(", "RSs", ")", ",", "the", "input", "features", "are", "usually", "sparse", ",", "categorical", "-", "continuous", "-", "mixed", ",", "and", "high", "-", "dimensional", ".", "Linear", "models", ",", "such", "as", "logistic", "regression", "with", "FTRL", ",", "are", "widely", "adopted", "as", "they", "are", "easy", "to", "manage", ",", "maintain", ",", "and", "deploy", ".", "Because", "linear", "models", "lack", "the", "ability", "of", "learning", "feature", "interactions", ",", "data", "scientists", "have", "to", "spend", "a", "lot", "of", "work", "on", "engineering", "cross", "features", "in", "order", "to", "achieve", "better", "performance", ".", "Considering", "that", "some", "hidden", "features", "are", "hard", "to", "design", "manually", ",", "some", "researchers", "exploit", "boosting", "decision", "trees", "to", "help", "build", "feature", "transformations", ".", "subsubsection", ":", "Factorization", "Models", "A", "major", "downside", "of", "the", "aforementioned", "models", "is", "that", "they", "can", "not", "generalize", "to", "unseen", "feature", "interactions", "in", "the", "training", "set", ".", "Factorization", "Machines", "overcome", "this", "problem", "via", "embedding", "each", "feature", "into", "a", "low", "dimension", "latent", "vector", ".", "Matrix", "factorization", "(", "MF", ")", ",", "which", "only", "considers", "IDs", "as", "features", ",", "can", "be", "regarded", "as", "a", "special", "kind", "of", "FM", ".", "Recommendations", "are", "made", "via", "the", "product", "of", "two", "latent", "vectors", ",", "thus", "it", "does", "not", "require", "the", "co", "-", "occurrence", "of", "user", "and", "item", "in", "the", "training", "set", ".", "MF", "is", "the", "most", "popular", "model", "-", "based", "collaborative", "filtering", "method", "in", "the", "RS", "literature", ".", "extend", "MF", "to", "leveraging", "side", "information", ",", "in", "which", "both", "a", "linear", "model", "and", "a", "MF", "model", "are", "included", ".", "On", "the", "other", "hand", ",", "for", "many", "recommender", "systems", ",", "only", "implicit", "feedback", "datasets", "such", "as", "users", "’", "watching", "history", "and", "browsing", "activities", "are", "available", ".", "Thus", "researchers", "extend", "the", "factorization", "models", "to", "a", "Bayesian", "Personalized", "Ranking", "(", "BPR", ")", "framework", "for", "implicit", "feedback", ".", "subsection", ":", "Recommender", "Systems", "with", "Deep", "Learning", "Deep", "learning", "techniques", "have", "achieved", "great", "success", "in", "computer", "vision", ",", "speech", "recognition", "and", "natural", "language", "understanding", ".", "As", "a", "result", ",", "an", "increasing", "number", "of", "researchers", "are", "interested", "in", "employing", "DNNs", "for", "recommender", "systems", ".", "subsubsection", ":", "Deep", "Learning", "for", "High", "-", "Order", "Interactions", "To", "avoid", "manually", "building", "up", "high", "-", "order", "cross", "features", ",", "researchers", "apply", "DNNs", "on", "field", "embedding", ",", "thus", "patterns", "from", "categorical", "feature", "interactions", "can", "be", "learned", "automatically", ".", "Representative", "models", "include", "FNN", ",", "PNN", ",", "DeepCross", ",", "NFM", ",", "DCN", ",", "Wide", "&", "Deep", ",", "and", "DeepFM", ".", "These", "models", "are", "highly", "related", "to", "our", "proposed", "xDeepFM", ".", "Since", "we", "have", "reviewed", "them", "in", "Section", "[", "reference", "]", "and", "Section", "[", "reference", "]", ",", "we", "do", "not", "further", "discuss", "them", "in", "detail", "in", "this", "section", ".", "We", "have", "demonstrated", "that", "our", "proposed", "xDeepFM", "has", "two", "special", "properties", "in", "comparison", "with", "these", "models", ":", "(", "1", ")", "xDeepFM", "learns", "high", "-", "order", "feature", "interactions", "in", "both", "explicit", "and", "implicit", "fashions", ";", "(", "2", ")", "xDeepFM", "learns", "feature", "interactions", "at", "the", "vector", "-", "wise", "level", "rather", "than", "at", "the", "bit", "-", "wise", "level", ".", "subsubsection", ":", "Deep", "Learning", "for", "Elaborate", "Representation", "Learning", "We", "include", "some", "other", "deep", "learning", "-", "based", "RSs", "in", "this", "section", "due", "to", "that", "they", "are", "less", "focused", "on", "learning", "feature", "interactions", ".", "Some", "early", "work", "employs", "deep", "learning", "mainly", "to", "model", "auxiliary", "information", ",", "such", "as", "visual", "data", "and", "audio", "data", ".", "Recently", ",", "deep", "neural", "networks", "are", "used", "to", "model", "the", "collaborative", "filtering", "(", "CF", ")", "in", "RSs", ".", "proposes", "a", "Neural", "Collaborative", "Filtering", "(", "NCF", ")", "so", "that", "the", "inner", "product", "in", "MF", "can", "be", "replaced", "with", "an", "arbitrary", "function", "via", "a", "neural", "architecture", ".", "model", "CF", "base", "on", "the", "autoencoder", "paradigm", ",", "and", "they", "have", "empirically", "demonstrated", "that", "autoencoder", "-", "based", "CF", "outperforms", "several", "classical", "MF", "models", ".", "Autoencoders", "can", "be", "further", "employed", "for", "jointly", "modeling", "CF", "and", "side", "information", "with", "the", "purpose", "of", "generating", "better", "latent", "factors", ".", "employ", "neural", "networks", "to", "jointly", "train", "multiple", "domains", "’", "latent", "factors", ".", "proposes", "the", "Attentive", "Collaborative", "Filtering", "(", "ACF", ")", "to", "learn", "more", "elaborate", "preference", "at", "both", "item", "-", "level", "and", "component", "-", "level", ".", "shows", "tha", "traditional", "RSs", "can", "not", "capture", "interest", "diversity", "and", "local", "activation", "effectively", ",", "so", "they", "introduce", "a", "Deep", "Interest", "Network", "(", "DIN", ")", "to", "represent", "users", "’", "diverse", "interests", "with", "an", "attentive", "activation", "mechanism", ".", "section", ":", "Conclusions", "In", "this", "paper", ",", "we", "propose", "a", "novel", "network", "named", "Compressed", "Interaction", "Network", "(", "CIN", ")", ",", "which", "aims", "to", "learn", "high", "-", "order", "feature", "interactions", "explicitly", ".", "CIN", "has", "two", "special", "virtues", ":", "(", "1", ")", "it", "can", "learn", "certain", "bounded", "-", "degree", "feature", "interactions", "effectively", ";", "(", "2", ")", "it", "learns", "feature", "interactions", "at", "a", "vector", "-", "wise", "level", ".", "Following", "the", "spirit", "of", "several", "popular", "models", ",", "we", "incorporate", "a", "CIN", "and", "a", "DNN", "in", "an", "end", "-", "to", "-", "end", "framework", ",", "and", "named", "the", "resulting", "model", "eXtreme", "Deep", "Factorization", "Machine", "(", "xDeepFM", ")", ".", "Thus", "xDeepFM", "can", "automatically", "learn", "high", "-", "order", "feature", "interactions", "in", "both", "explicit", "and", "implicit", "fashions", ",", "which", "is", "of", "great", "significance", "to", "reducing", "manual", "feature", "engineering", "work", ".", "We", "conduct", "comprehensive", "experiments", "and", "the", "results", "demonstrate", "that", "our", "xDeepFM", "outperforms", "state", "-", "of", "-", "the", "-", "art", "models", "consistently", "on", "three", "real", "-", "world", "datasets", ".", "There", "are", "two", "directions", "for", "future", "work", ".", "First", ",", "currently", "we", "simply", "employ", "a", "sum", "pooling", "for", "embedding", "multivalent", "fields", ".", "We", "can", "explore", "the", "usage", "of", "the", "DIN", "mechanism", "to", "capture", "the", "related", "activation", "according", "to", "the", "candidate", "item", ".", "Second", ",", "as", "discussed", "in", "section", "[", "reference", "]", ",", "the", "time", "complexity", "of", "the", "CIN", "module", "is", "high", ".", "We", "are", "interested", "in", "developing", "a", "distributed", "version", "of", "xDeepFM", "which", "can", "be", "trained", "efficiently", "on", "a", "GPU", "cluster", ".", "section", ":", "Acknowledgements", "The", "authors", "would", "like", "to", "thank", "the", "anonymous", "reviewers", "for", "their", "insightful", "reviews", ",", "which", "are", "very", "helpful", "on", "the", "revision", "of", "this", "paper", ".", "This", "work", "is", "supported", "in", "part", "by", "Youth", "Innovation", "Promotion", "Association", "of", "CAS", ".", "bibliography", ":", "References"], "entity": [[23, 24, ["cost"], "Metric"], [36, 40, ["web", "-", "scale", "systems"], "Task"], [41, 44, ["Factorization", "based", "models"], "Method"], [74, 77, ["deep", "neural", "networks"], "Method"], [78, 79, ["DNNs"], "Method"], [119, 120, ["DNNs"], "Method"], [140, 143, ["Compressed", "Interaction", "Network"], "Method"], [144, 145, ["CIN"], "Method"], [169, 170, ["CIN"], "Method"], [174, 177, ["convolutional", "neural", "networks"], "Method"], [178, 179, ["CNNs"], "Method"], [181, 184, ["recurrent", "neural", "networks"], "Method"], [185, 186, ["RNNs"], "Method"], [192, 193, ["CIN"], "Method"], [196, 197, ["DNN"], "Method"], [199, 201, ["unified", "model"], "Method"], [208, 211, ["Deep", "Factorization", "Machine"], "Method"], [212, 213, ["xDeepFM"], "Method"], [220, 221, ["xDeepFM"], "Method"], [267, 268, ["xDeepFM"], "Method"], [285, 286, ["xDeepFM"], "Method"], [295, 297, ["Recommender", "Systems"], "Task"], [298, 299, ["xDeepFM"], "Method"], [307, 309, ["Recommender", "Systems"], "Task"], [431, 433, ["Neural", "networks"], "Method"], [438, 440, ["Factorization", "methods"], "Method"], [444, 445, ["KDD"], "Method"], [450, 454, ["ACM", "SIGKDD", "International", "Conference"], "Task"], [455, 457, ["Knowledge", "Discovery"], "Task"], [458, 460, ["Data", "MiningAugust"], "Task"], [472, 476, ["ACM", "SIGKDD", "International", "Conference"], "Task"], [477, 482, ["Knowledge", "Discovery", "&", "Data", "Mining"], "Task"], [522, 524, ["predictive", "systems"], "Method"], [546, 550, ["transformation", "of", "raw", "features"], "Task"], [555, 557, ["predictive", "systems"], "Method"], [560, 563, ["data", "mining", "games"], "Task"], [568, 570, ["feature", "transformation"], "Method"], [572, 576, ["cross", "-", "product", "transformation"], "Method"], [641, 643, ["deep", "learning"], "Method"], [654, 657, ["cross", "feature", "engineering"], "Method"], [680, 682, ["data", "scientists"], "Task"], [710, 715, ["large", "-", "scale", "predictive", "systems"], "Task"], [717, 722, ["web", "-", "scale", "recommender", "systems"], "Task"], [764, 766, ["manual", "engineering"], "Task"], [771, 773, ["Factorization", "Machines"], "Method"], [774, 775, ["FM"], "Method"], [822, 823, ["FM"], "Method"], [876, 879, ["deep", "neural", "networks"], "Method"], [880, 881, ["DNNs"], "Method"], [886, 888, ["computer", "vision"], "Task"], [889, 891, ["speech", "recognition"], "Task"], [893, 896, ["natural", "language", "processing"], "Task"], [901, 904, ["feature", "representation", "learning"], "Method"], [910, 911, ["DNNs"], "Method"], [921, 927, ["Factorisation", "-", "machine", "supported", "Neural", "Network"], "Method"], [928, 929, ["FNN"], "Method"], [944, 946, ["factorization", "machines"], "Method"], [947, 949, ["field", "embedding"], "Task"], [951, 952, ["DNN"], "Method"], [956, 961, ["Product", "-", "based", "Neural", "Network"], "Method"], [962, 963, ["PNN"], "Method"], [968, 970, ["product", "layer"], "Method"], [971, 973, ["embedding", "layer"], "Method"], [985, 986, ["FM"], "Method"], [991, 992, ["FNN"], "Method"], [993, 994, ["PNN"], "Method"], [1014, 1017, ["Wide", "&", "Deep"], "Method"], [1018, 1019, ["DeepFM"], "Method"], [1025, 1027, ["hybrid", "architectures"], "Method"], [1031, 1033, ["shallow", "component"], "Method"], [1035, 1037, ["deep", "component"], "Method"], [1043, 1046, ["memorization", "and", "generalization"], "Task"], [1067, 1068, ["DNNs"], "Method"], [1069, 1075, ["learning", "high", "-", "order", "feature", "interactions"], "Task"], [1078, 1079, ["DNNs"], "Method"], [1095, 1096, ["DNNs"], "Method"], [1119, 1120, ["DNNs"], "Method"], [1155, 1157, ["recommender", "systems"], "Task"], [1159, 1160, ["DNNs"], "Method"], [1167, 1173, ["representing", "high", "-", "order", "feature", "interactions"], "Task"], [1185, 1190, ["neural", "network", "-", "based", "model"], "Method"], [1209, 1213, ["Deep", "&", "Cross", "Network"], "Method"], [1214, 1215, ["DCN"], "Method"], [1239, 1240, ["DCN"], "Method"], [1254, 1257, ["compressed", "interaction", "network"], "Method"], [1258, 1259, ["CIN"], "Method"], [1263, 1265, ["cross", "network"], "Method"], [1267, 1268, ["DCN"], "Method"], [1269, 1270, ["CIN"], "Method"], [1303, 1309, ["explicit", "high", "-", "order", "interaction", "module"], "Method"], [1310, 1313, ["implicit", "interaction", "module"], "Method"], [1321, 1323, ["joint", "model"], "Method"], [1324, 1327, ["Deep", "Factorization", "Machine"], "Method"], [1328, 1329, ["xDeepFM"], "Method"], [1336, 1339, ["manual", "feature", "engineering"], "Task"], [1341, 1343, ["data", "scientists"], "Task"], [1345, 1347, ["feature", "searching"], "Task"], [1370, 1371, ["xDeepFM"], "Method"], [1388, 1391, ["manual", "feature", "engineering"], "Method"], [1395, 1398, ["compressed", "interaction", "network"], "Method"], [1399, 1400, ["CIN"], "Method"], [1402, 1403, ["xDeepFM"], "Method"], [1459, 1460, ["xDeepFM"], "Method"], [1495, 1501, ["deep", "learning", "-", "based", "recommender", "systems"], "Method"], [1509, 1510, ["CIN"], "Method"], [1511, 1513, ["xDeepFM", "model"], "Method"], [1553, 1555, ["Embedding", "Layer"], "Method"], [1556, 1558, ["computer", "vision"], "Task"], [1559, 1562, ["natural", "language", "understanding"], "Task"], [1586, 1587, ["DNNs"], "Method"], [1602, 1607, ["web", "-", "scale", "recommender", "systems"], "Task"], [1677, 1684, ["field", "-", "aware", "one", "-", "hot", "encoding"], "Method"], [1686, 1688, ["embedding", "layer"], "Method"], [1716, 1718, ["feature", "embedding"], "Method"], [1722, 1724, ["field", "embedding"], "Method"], [1754, 1758, ["sum", "of", "feature", "embedding"], "Method"], [1762, 1764, ["field", "embedding"], "Method"], [1766, 1768, ["embedding", "layer"], "Method"], [1779, 1781, ["embedding", "layer"], "Method"], [1825, 1827, ["field", "embedding"], "Method"], [1830, 1835, ["Implicit", "High", "-", "order", "Interactions"], "Task"], [1835, 1836, ["FNN"], "Method"], [1837, 1839, ["Deep", "Crossing"], "Method"], [1842, 1844, ["deep", "part"], "Method"], [1845, 1848, ["Wide", "&", "Deep"], "Method"], [1850, 1855, ["feed", "-", "forward", "neural", "network"], "Method"], [1869, 1871, ["forward", "process"], "Method"], [1953, 1954, ["PNN"], "Method"], [1955, 1956, ["DeepFM"], "Method"], [1964, 1965, ["DNNs"], "Method"], [1975, 1978, ["way", "interaction", "layer"], "Method"], [2003, 2004, ["PNN"], "Method"], [2005, 2006, ["DeepFM"], "Method"], [2009, 2010, ["PNN"], "Method"], [2018, 2019, ["DNNs"], "Method"], [2021, 2022, ["DeepFM"], "Method"], [2049, 2051, ["Cross", "Network"], "Method"], [2052, 2053, ["CrossNet"], "Method"], [2079, 2086, ["fully", "-", "connected", "feed", "-", "forward", "network"], "Method"], [2095, 2097, ["cross", "operation"], "Method"], [2117, 2118, ["CrossNet"], "Method"], [2135, 2136, ["CrossNet"], "Method"], [2144, 2148, ["-", "layer", "cross", "network"], "Method"], [2166, 2168, ["cross", "network"], "Method"], [2184, 2186, ["distributive", "law"], "Method"], [2187, 2189, ["matrix", "multiplication"], "Task"], [2199, 2201, ["linear", "regression"], "Method"], [2247, 2249, ["cross", "network"], "Method"], [2274, 2275, ["CrossNet"], "Method"], [2283, 2284, ["complexity"], "Metric"], [2304, 2305, ["CrossNet"], "Method"], [2344, 2347, ["Compressed", "Interaction", "Network"], "Method"], [2351, 2353, ["cross", "network"], "Method"], [2355, 2358, ["Compressed", "Interaction", "Network"], "Method"], [2359, 2360, ["CIN"], "Method"], [2402, 2403, ["complexity"], "Metric"], [2436, 2438, ["field", "embedding"], "Method"], [2465, 2467, ["field", "embedding"], "Method"], [2472, 2475, ["-", "th", "layer"], "Method"], [2476, 2477, ["CIN"], "Method"], [2562, 2563, ["CIN"], "Method"], [2568, 2571, ["Recurrent", "Neural", "Network"], "Method"], [2572, 2573, ["RNN"], "Method"], [2637, 2640, ["Convolutional", "Neural", "Networks"], "Method"], [2641, 2642, ["CNNs"], "Method"], [2644, 2646, ["computer", "vision"], "Task"], [2693, 2694, ["filter"], "Method"], [2726, 2728, ["feature", "map"], "Task"], [2729, 2731, ["computer", "vision"], "Task"], [2757, 2759, ["hidden", "layer"], "Method"], [2780, 2781, ["CIN"], "Method"], [2792, 2794, ["hidden", "layer"], "Method"], [2804, 2806, ["sum", "pooling"], "Method"], [2851, 2852, ["CIN"], "Method"], [2854, 2856, ["binary", "classification"], "Task"], [2880, 2881, ["CIN"], "Method"], [2884, 2886, ["model", "complexity"], "Metric"], [2893, 2895, ["Space", "Complexity"], "Metric"], [2929, 2931, ["regression", "layer"], "Method"], [2946, 2947, ["CIN"], "Method"], [2951, 2952, ["CIN"], "Method"], [3004, 3007, ["-", "order", "decomposition"], "Method"], [3041, 3044, ["-", "order", "decomposition"], "Method"], [3059, 3061, ["space", "complexity"], "Metric"], [3063, 3065, ["plain", "DNN"], "Method"], [3077, 3079, ["field", "embedding"], "Method"], [3080, 3081, ["subsubsection"], "Method"], [3082, 3084, ["Time", "Complexity"], "Metric"], [3123, 3124, ["DNN"], "Method"], [3137, 3138, ["CIN"], "Method"], [3141, 3143, ["time", "complexity"], "Metric"], [3144, 3145, ["subsubsection"], "Method"], [3146, 3148, ["Polynomial", "Approximation"], "Method"], [3158, 3159, ["CIN"], "Method"], [3306, 3307, ["CIN"], "Method"], [3319, 3323, ["chain", "of", "feature", "maps"], "Method"], [3447, 3448, ["CIN"], "Method"], [3477, 3479, ["Implicit", "Networks"], "Method"], [3488, 3489, ["DNNs"], "Method"], [3498, 3499, ["CIN"], "Method"], [3500, 3502, ["plain", "DNNs"], "Method"], [3530, 3533, ["Wide", "&", "Deep"], "Method"], [3534, 3536, ["DeepFM", "model"], "Method"], [3553, 3556, ["Deep", "Factorization", "Machine"], "Method"], [3557, 3558, ["xDeepFM"], "Method"], [3617, 3619, ["plain", "DNN"], "Method"], [3620, 3621, ["CIN"], "Method"], [3630, 3632, ["binary", "classifications"], "Task"], [3634, 3636, ["loss", "function"], "Metric"], [3638, 3640, ["log", "loss"], "Metric"], [3651, 3653, ["optimization", "process"], "Task"], [3658, 3660, ["objective", "function"], "Metric"], [3690, 3691, ["FM"], "Method"], [3692, 3693, ["DeepFM"], "Method"], [3728, 3729, ["xDeepFM"], "Method"], [3737, 3740, ["linear", "regression", "weights"], "Method"], [3748, 3749, ["DeepFM"], "Method"], [3781, 3784, ["constant", "sum", "filter"], "Method"], [3794, 3796, ["parameter", "learning"], "Method"], [3803, 3804, ["xDeepFM"], "Method"], [3836, 3837, ["CIN"], "Method"], [3839, 3845, ["high", "-", "order", "feature", "interactions", "learning"], "Task"], [3863, 3865, ["recommender", "systems"], "Task"], [3879, 3880, ["xDeepFM"], "Method"], [3914, 3916, ["Criteo", "Dataset"], "Material"], [3927, 3933, ["predicting", "ad", "click", "-", "through", "rate"], "Task"], [3967, 3969, ["Dianping", "Dataset"], "Material"], [3970, 3971, ["Dianping.com"], "Method"], [4017, 4019, ["restaurant", "recommendation"], "Task"], [4088, 4091, ["Bing", "News", "Dataset"], "Material"], [4092, 4094, ["Bing", "News"], "Material"], [4142, 4143, ["validation"], "Task"], [4154, 4156, ["Criteo", "dataset"], "Material"], [4170, 4171, ["validation"], "Task"], [4190, 4192, ["Evaluation", "Metrics"], "Metric"], [4198, 4200, ["model", "evaluation"], "Metric"], [4201, 4202, ["AUC"], "Metric"], [4203, 4208, ["Area", "Under", "the", "ROC", "curve"], "Metric"], [4210, 4211, ["Logloss"], "Metric"], [4212, 4214, ["cross", "entropy"], "Metric"], [4227, 4228, ["AUC"], "Metric"], [4260, 4263, ["class", "imbalance", "problem"], "Task"], [4264, 4265, ["Logloss"], "Metric"], [4289, 4290, ["Logloss"], "Metric"], [4304, 4306, ["ranking", "strategy"], "Method"], [4312, 4313, ["CTR"], "Task"], [4323, 4324, ["xDeepFM"], "Method"], [4325, 4326, ["LR"], "Method"], [4327, 4329, ["logistic", "regression"], "Method"], [4331, 4332, ["FM"], "Method"], [4333, 4334, ["DNN"], "Method"], [4335, 4339, ["plain", "deep", "neural", "network"], "Method"], [4341, 4342, ["PNN"], "Method"], [4348, 4349, ["iPNN"], "Method"], [4350, 4351, ["oPNN"], "Method"], [4353, 4356, ["Wide", "&", "Deep"], "Method"], [4357, 4358, ["DCN"], "Method"], [4359, 4363, ["Deep", "&", "Cross", "Network"], "Method"], [4365, 4366, ["DeepFM"], "Method"], [4384, 4385, ["xDeepFM"], "Method"], [4399, 4401, ["recommender", "systems"], "Task"], [4436, 4437, ["Tensorflow"], "Method"], [4447, 4450, ["grid", "-", "searching"], "Method"], [4469, 4471, ["Learning", "rate"], "Metric"], [4477, 4479, ["optimization", "method"], "Method"], [4483, 4484, ["Adam"], "Method"], [4486, 4490, ["mini", "-", "batch", "size"], "Method"], [4496, 4498, ["L2", "regularization"], "Method"], [4500, 4501, ["DNN"], "Method"], [4502, 4503, ["DCN"], "Method"], [4504, 4507, ["Wide", "&", "Deep"], "Method"], [4508, 4509, ["DeepFM"], "Method"], [4510, 4511, ["xDeepFM"], "Method"], [4514, 4516, ["dropout", "0.5"], "Method"], [4517, 4518, ["PNN"], "Method"], [4546, 4548, ["Criteo", "dataset"], "Material"], [4555, 4556, ["Dianping"], "Material"], [4557, 4560, ["Bing", "News", "datasets"], "Material"], [4565, 4568, ["neural", "networks", "structures"], "Task"], [4577, 4579, ["field", "embedding"], "Method"], [4599, 4602, ["Tesla", "K80", "GPUs"], "Method"], [4616, 4618, ["Neural", "Components"], "Method"], [4626, 4627, ["CIN"], "Method"], [4632, 4633, ["FM"], "Method"], [4641, 4642, ["DNN"], "Method"], [4650, 4651, ["CrossNet"], "Method"], [4679, 4680, ["CIN"], "Method"], [4730, 4731, ["FM"], "Method"], [4775, 4776, ["CIN"], "Method"], [4813, 4814, ["DNN"], "Method"], [4815, 4816, ["CrossNet"], "Method"], [4817, 4818, ["CIN"], "Method"], [4819, 4820, ["FM"], "Method"], [4832, 4833, ["CIN"], "Method"], [4844, 4845, ["CIN"], "Method"], [4847, 4853, ["explicit", "high", "-", "order", "feature", "interactions"], "Task"], [4879, 4880, ["CIN"], "Method"], [4887, 4890, ["Bing", "News", "dataset"], "Material"], [4895, 4897, ["Integrated", "Models"], "Method"], [4900, 4901, ["xDeepFM"], "Method"], [4902, 4903, ["CIN"], "Method"], [4904, 4905, ["DNN"], "Method"], [4907, 4913, ["end", "-", "to", "-", "end", "model"], "Method"], [4915, 4916, ["CIN"], "Method"], [4917, 4918, ["DNN"], "Method"], [4923, 4926, ["learning", "feature", "interactions"], "Task"], [4944, 4949, ["jointly", "explicit", "and", "implicit", "learning"], "Task"], [4978, 4979, ["LR"], "Method"], [4991, 4995, ["factorization", "-", "based", "models"], "Method"], [4998, 5001, ["measuring", "sparse", "features"], "Task"], [5002, 5005, ["Wide", "&", "Deep"], "Method"], [5006, 5007, ["DCN"], "Method"], [5008, 5009, ["DeepFM"], "Method"], [5010, 5011, ["xDeepFM"], "Method"], [5015, 5016, ["DNN"], "Method"], [5027, 5029, ["hybrid", "components"], "Method"], [5034, 5035, ["accuracy"], "Metric"], [5036, 5038, ["predictive", "systems"], "Method"], [5041, 5042, ["xDeepFM"], "Method"], [5066, 5067, ["xDeepFM"], "Method"], [5074, 5075, ["combination"], "Task"], [5084, 5088, ["neural", "-", "based", "models"], "Method"], [5093, 5096, ["deep", "network", "structure"], "Method"], [5120, 5121, ["xDeepFM"], "Method"], [5156, 5157, ["xDeepFM"], "Method"], [5243, 5244, ["xDeepFM"], "Method"], [5334, 5337, ["Bing", "News", "dataset"], "Material"], [5379, 5381, ["Activation", "Function"], "Method"], [5409, 5412, ["deep", "learning", "literature"], "Task"], [5415, 5420, ["non", "-", "linear", "activation", "functions"], "Method"], [5431, 5433, ["activation", "functions"], "Method"], [5434, 5435, ["CIN"], "Method"], [5439, 5440, ["DNN"], "Method"], [5447, 5448, ["relu"], "Method"], [5462, 5464, ["identify", "function"], "Method"], [5486, 5490, ["Non", "-", "factorization", "Models"], "Method"], [5491, 5496, ["web", "-", "scale", "recommender", "systems"], "Task"], [5497, 5498, ["RSs"], "Task"], [5518, 5520, ["Linear", "models"], "Method"], [5523, 5525, ["logistic", "regression"], "Method"], [5526, 5527, ["FTRL"], "Method"], [5544, 5546, ["linear", "models"], "Method"], [5588, 5591, ["boosting", "decision", "trees"], "Method"], [5599, 5601, ["Factorization", "Models"], "Method"], [5623, 5625, ["Factorization", "Machines"], "Method"], [5639, 5641, ["Matrix", "factorization"], "Method"], [5642, 5643, ["MF"], "Method"], [5660, 5661, ["FM"], "Method"], [5662, 5663, ["Recommendations"], "Task"], [5691, 5692, ["MF"], "Method"], [5708, 5709, ["MF"], "Method"], [5718, 5720, ["linear", "model"], "Method"], [5734, 5736, ["recommender", "systems"], "Task"], [5757, 5759, ["factorization", "models"], "Method"], [5761, 5764, ["Bayesian", "Personalized", "Ranking"], "Method"], [5765, 5766, ["BPR"], "Method"], [5769, 5771, ["implicit", "feedback"], "Task"], [5774, 5776, ["Recommender", "Systems"], "Task"], [5777, 5779, ["Deep", "Learning"], "Method"], [5779, 5782, ["Deep", "learning", "techniques"], "Method"], [5787, 5789, ["computer", "vision"], "Task"], [5790, 5792, ["speech", "recognition"], "Task"], [5793, 5796, ["natural", "language", "understanding"], "Task"], [5810, 5811, ["DNNs"], "Method"], [5812, 5814, ["recommender", "systems"], "Task"], [5817, 5819, ["Deep", "Learning"], "Method"], [5820, 5824, ["High", "-", "Order", "Interactions"], "Task"], [5837, 5838, ["DNNs"], "Method"], [5839, 5841, ["field", "embedding"], "Method"], [5856, 5857, ["FNN"], "Method"], [5858, 5859, ["PNN"], "Method"], [5860, 5861, ["DeepCross"], "Method"], [5862, 5863, ["NFM"], "Method"], [5864, 5865, ["DCN"], "Method"], [5866, 5869, ["Wide", "&", "Deep"], "Method"], [5871, 5872, ["DeepFM"], "Method"], [5881, 5882, ["xDeepFM"], "Method"], [5917, 5918, ["xDeepFM"], "Method"], [5931, 5932, ["xDeepFM"], "Method"], [5948, 5949, ["xDeepFM"], "Method"], [5967, 5968, ["subsubsection"], "Method"], [5969, 5971, ["Deep", "Learning"], "Method"], [5972, 5975, ["Elaborate", "Representation", "Learning"], "Task"], [5979, 5984, ["deep", "learning", "-", "based", "RSs"], "Method"], [6003, 6005, ["deep", "learning"], "Method"], [6021, 6024, ["deep", "neural", "networks"], "Method"], [6029, 6031, ["collaborative", "filtering"], "Method"], [6032, 6033, ["CF"], "Method"], [6039, 6042, ["Neural", "Collaborative", "Filtering"], "Method"], [6043, 6045, ["NCF", ")"], "Method"], [6051, 6052, ["MF"], "Method"], [6061, 6063, ["neural", "architecture"], "Method"], [6069, 6071, ["autoencoder", "paradigm"], "Method"], [6088, 6089, ["Autoencoders"], "Method"], [6094, 6097, ["jointly", "modeling", "CF"], "Method"], [6098, 6100, ["side", "information"], "Method"], [6110, 6112, ["neural", "networks"], "Method"], [6123, 6126, ["Attentive", "Collaborative", "Filtering"], "Method"], [6127, 6128, ["ACF"], "Method"], [6147, 6148, ["RSs"], "Method"], [6162, 6165, ["Deep", "Interest", "Network"], "Method"], [6166, 6167, ["DIN"], "Method"], [6176, 6179, ["attentive", "activation", "mechanism"], "Method"], [6193, 6196, ["Compressed", "Interaction", "Network"], "Method"], [6197, 6198, ["CIN"], "Method"], [6211, 6212, ["CIN"], "Method"], [6256, 6257, ["CIN"], "Method"], [6259, 6260, ["DNN"], "Method"], [6262, 6268, ["end", "-", "to", "-", "end", "framework"], "Method"], [6275, 6278, ["Deep", "Factorization", "Machine"], "Method"], [6279, 6280, ["xDeepFM"], "Method"], [6283, 6284, ["xDeepFM"], "Method"], [6306, 6310, ["manual", "feature", "engineering", "work"], "Task"], [6321, 6322, ["xDeepFM"], "Method"], [6354, 6356, ["sum", "pooling"], "Method"], [6357, 6360, ["embedding", "multivalent", "fields"], "Task"], [6368, 6370, ["DIN", "mechanism"], "Method"], [6392, 6394, ["time", "complexity"], "Metric"], [6407, 6409, ["distributed", "version"], "Method"], [6410, 6411, ["xDeepFM"], "Method"], [6418, 6420, ["GPU", "cluster"], "Method"], [89, 90, ["DNN"], "Method"], [974, 975, ["DNN"], "Method"], [1136, 1137, ["FM"], "Method"], [1293, 1296, ["Wide", "&", "Deep"], "Method"], [1297, 1298, ["DeepFM"], "Method"], [1315, 1316, ["FM"], "Method"], [1366, 1369, ["Deep", "Factorization", "Machine"], "Method"], [1917, 1918, ["FM"], "Method"], [2024, 2025, ["FM"], "Method"], [2289, 2290, ["DNN"], "Method"], [2751, 2752, ["CIN"], "Method"], [2874, 2875, ["CIN"], "Method"], [2966, 2967, ["DNN"], "Method"], [3049, 3050, ["CIN"], "Method"], [3115, 3116, ["CIN"], "Method"], [3679, 3680, ["CIN"], "Method"], [3683, 3684, ["DNN"], "Method"], [3720, 3721, ["CIN"], "Method"], [3733, 3734, ["DeepFM"], "Method"], [3742, 3743, ["FM"], "Method"], [3752, 3753, ["FM"], "Method"], [3771, 3772, ["DNN"], "Method"], [3809, 3810, ["FM"], "Method"], [4158, 4159, ["Dianping"], "Material"], [4535, 4536, ["DNN"], "Method"], [4543, 4544, ["CIN"], "Method"], [4552, 4553, ["CIN"], "Method"], [4859, 4860, ["CIN"], "Method"], [5198, 5199, ["DNN"], "Method"], [5206, 5207, ["CIN"], "Method"], [5317, 5318, ["CIN"], "Method"], [5351, 5352, ["Dianping"], "Material"], [5394, 5395, ["CIN"], "Method"], [5473, 5474, ["CIN"], "Method"], [5482, 5484, ["Recommender", "Systems"], "Task"], [5699, 5701, ["collaborative", "filtering"], "Method"], [5711, 5713, ["side", "information"], "Method"], [5722, 5723, ["MF"], "Method"], [6065, 6066, ["CF"], "Method"], [6081, 6082, ["CF"], "Method"], [6085, 6086, ["MF"], "Method"], [6396, 6397, ["CIN"], "Method"]]} {"doc_id": "35734e8724559fb0d494e5cba6a28ad7a3d5dd4d", "text": ["document", ":", "Explaining", "and", "Harnessing", "Adversarial", "Examples", "Several", "machine", "learning", "models", ",", "including", "neural", "networks", ",", "consistently", "misclassify", "adversarial", "examples", "—inputs", "formed", "by", "applying", "small", "but", "intentionally", "worst", "-", "case", "perturbations", "to", "examples", "from", "the", "dataset", ",", "such", "that", "the", "perturbed", "input", "results", "in", "the", "model", "outputting", "an", "incorrect", "answer", "with", "high", "confidence", ".", "Early", "attempts", "at", "explaining", "this", "phenomenon", "focused", "on", "nonlinearity", "and", "overfitting", ".", "We", "argue", "instead", "that", "the", "primary", "cause", "of", "neural", "networks", "’", "vulnerability", "to", "adversarial", "perturbation", "is", "their", "linear", "nature", ".", "This", "explanation", "is", "supported", "by", "new", "quantitative", "results", "while", "giving", "the", "first", "explanation", "of", "the", "most", "intriguing", "fact", "about", "them", ":", "their", "generalization", "across", "architectures", "and", "training", "sets", ".", "Moreover", ",", "this", "view", "yields", "a", "simple", "and", "fast", "method", "of", "generating", "adversarial", "examples", ".", "Using", "this", "approach", "to", "provide", "examples", "for", "adversarial", "training", ",", "we", "reduce", "the", "test", "set", "error", "of", "a", "maxout", "network", "on", "the", "MNIST", "dataset", ".", "*", "section", ":", "0pt1pt1pt", "section", ":", "Introduction", "Szegedy", "-", "ICLR2014", "made", "an", "intriguing", "discovery", ":", "several", "machine", "learning", "models", ",", "including", "state", "-", "of", "-", "the", "-", "art", "neural", "networks", ",", "are", "vulnerable", "to", "adversarial", "examples", ".", "That", "is", ",", "these", "machine", "learning", "models", "misclassify", "examples", "that", "are", "only", "slightly", "different", "from", "correctly", "classified", "examples", "drawn", "from", "the", "data", "distribution", ".", "In", "many", "cases", ",", "a", "wide", "variety", "of", "models", "with", "different", "architectures", "trained", "on", "different", "subsets", "of", "the", "training", "data", "misclassify", "the", "same", "adversarial", "example", ".", "This", "suggests", "that", "adversarial", "examples", "expose", "fundamental", "blind", "spots", "in", "our", "training", "algorithms", ".", "The", "cause", "of", "these", "adversarial", "examples", "was", "a", "mystery", ",", "and", "speculative", "explanations", "have", "suggested", "it", "is", "due", "to", "extreme", "nonlinearity", "of", "deep", "neural", "networks", ",", "perhaps", "combined", "with", "insufficient", "model", "averaging", "and", "insufficient", "regularization", "of", "the", "purely", "supervised", "learning", "problem", ".", "We", "show", "that", "these", "speculative", "hypotheses", "are", "unnecessary", ".", "Linear", "behavior", "in", "high", "-", "dimensional", "spaces", "is", "sufficient", "to", "cause", "adversarial", "examples", ".", "This", "view", "enables", "us", "to", "design", "a", "fast", "method", "of", "generating", "adversarial", "examples", "that", "makes", "adversarial", "training", "practical", ".", "We", "show", "that", "adversarial", "training", "can", "provide", "an", "additional", "regularization", "benefit", "beyond", "that", "provided", "by", "using", "dropout", "dropout", "alone", ".", "Generic", "regularization", "strategies", "such", "as", "dropout", ",", "pretraining", ",", "and", "model", "averaging", "do", "not", "confer", "a", "significant", "reduction", "in", "a", "model", "’s", "vulnerability", "to", "adversarial", "examples", ",", "but", "changing", "to", "nonlinear", "model", "families", "such", "as", "RBF", "networks", "can", "do", "so", ".", "Our", "explanation", "suggests", "a", "fundamental", "tension", "between", "designing", "models", "that", "are", "easy", "to", "train", "due", "to", "their", "linearity", "and", "designing", "models", "that", "use", "nonlinear", "effects", "to", "resist", "adversarial", "perturbation", ".", "In", "the", "long", "run", ",", "it", "may", "be", "possible", "to", "escape", "this", "tradeoff", "by", "designing", "more", "powerful", "optimization", "methods", "that", "can", "succesfully", "train", "more", "nonlinear", "models", ".", "section", ":", "Related", "work", "Szegedy", "-", "ICLR2014", "demonstrated", "a", "variety", "of", "intriguing", "properties", "of", "neural", "networks", "and", "related", "models", ".", "Those", "most", "relevant", "to", "this", "paper", "include", ":", "Box", "-", "constrained", "L", "-", "BFGS", "can", "reliably", "find", "adversarial", "examples", ".", "On", "some", "datasets", ",", "such", "as", "ImageNet", "ImageNet", ",", "the", "adversarial", "examples", "were", "so", "close", "to", "the", "original", "examples", "that", "the", "differences", "were", "indistinguishable", "to", "the", "human", "eye", ".", "The", "same", "adversarial", "example", "is", "often", "misclassified", "by", "a", "variety", "of", "classifiers", "with", "different", "architectures", "or", "trained", "on", "different", "subsets", "of", "the", "training", "data", ".", "Shallow", "softmax", "regression", "models", "are", "also", "vulnerable", "to", "adversarial", "examples", ".", "Training", "on", "adversarial", "examples", "can", "regularize", "the", "model", "—", "however", ",", "this", "was", "not", "practical", "at", "the", "time", "due", "to", "the", "need", "for", "expensive", "constrained", "optimization", "in", "the", "inner", "loop", ".", "These", "results", "suggest", "that", "classifiers", "based", "on", "modern", "machine", "learning", "techniques", ",", "even", "those", "that", "obtain", "excellent", "performance", "on", "the", "test", "set", ",", "are", "not", "learning", "the", "true", "underlying", "concepts", "that", "determine", "the", "correct", "output", "label", ".", "Instead", ",", "these", "algorithms", "have", "built", "a", "Potemkin", "village", "that", "works", "well", "on", "naturally", "occuring", "data", ",", "but", "is", "exposed", "as", "a", "fake", "when", "one", "visits", "points", "in", "space", "that", "do", "not", "have", "high", "probability", "in", "the", "data", "distribution", ".", "This", "is", "particularly", "disappointing", "because", "a", "popular", "approach", "in", "computer", "vision", "is", "to", "use", "convolutional", "network", "features", "as", "a", "space", "where", "Euclidean", "distance", "approximates", "perceptual", "distance", ".", "This", "resemblance", "is", "clearly", "flawed", "if", "images", "that", "have", "an", "immeasurably", "small", "perceptual", "distance", "correspond", "to", "completely", "different", "classes", "in", "the", "network", "’s", "representation", ".", "These", "results", "have", "often", "been", "interpreted", "as", "being", "a", "flaw", "in", "deep", "networks", "in", "particular", ",", "even", "though", "linear", "classifiers", "have", "the", "same", "problem", ".", "We", "regard", "the", "knowledge", "of", "this", "flaw", "as", "an", "opportunity", "to", "fix", "it", ".", "Indeed", ",", "Luca", "and", "causal", "have", "already", "begun", "the", "first", "steps", "toward", "designing", "models", "that", "resist", "adversarial", "perturbation", ",", "though", "no", "model", "has", "yet", "succesfully", "done", "so", "while", "maintaining", "state", "of", "the", "art", "accuracy", "on", "clean", "inputs", ".", "section", ":", "The", "linear", "explanation", "of", "adversarial", "examples", "We", "start", "with", "explaining", "the", "existence", "of", "adversarial", "examples", "for", "linear", "models", ".", "In", "many", "problems", ",", "the", "precision", "of", "an", "individual", "input", "feature", "is", "limited", ".", "For", "example", ",", "digital", "images", "often", "use", "only", "8", "bits", "per", "pixel", "so", "they", "discard", "all", "information", "below", "of", "the", "dynamic", "range", ".", "Because", "the", "precision", "of", "the", "features", "is", "limited", ",", "it", "is", "not", "rational", "for", "the", "classifier", "to", "respond", "differently", "to", "an", "input", "than", "to", "an", "adversarial", "input", "if", "every", "element", "of", "the", "perturbation", "is", "smaller", "than", "the", "precision", "of", "the", "features", ".", "Formally", ",", "for", "problems", "with", "well", "-", "separated", "classes", ",", "we", "expect", "the", "classifier", "to", "assign", "the", "same", "class", "to", "and", "so", "long", "as", ",", "where", "is", "small", "enough", "to", "be", "discarded", "by", "the", "sensor", "or", "data", "storage", "apparatus", "associated", "with", "our", "problem", ".", "Consider", "the", "dot", "product", "between", "a", "weight", "vector", "and", "an", "adversarial", "example", ":", "The", "adversarial", "perturbation", "causes", "the", "activation", "to", "grow", "by", ".", "We", "can", "maximize", "this", "increase", "subject", "to", "the", "max", "norm", "constraint", "on", "by", "assigning", ".", "If", "has", "dimensions", "and", "the", "average", "magnitude", "of", "an", "element", "of", "the", "weight", "vector", "is", ",", "then", "the", "activation", "will", "grow", "by", ".", "Since", "does", "not", "grow", "with", "the", "dimensionality", "of", "the", "problem", "but", "the", "change", "in", "activation", "caused", "by", "perturbation", "by", "can", "grow", "linearly", "with", ",", "then", "for", "high", "dimensional", "problems", ",", "we", "can", "make", "many", "infinitesimal", "changes", "to", "the", "input", "that", "add", "up", "to", "one", "large", "change", "to", "the", "output", ".", "We", "can", "think", "of", "this", "as", "a", "sort", "of", "“", "accidental", "steganography", ",", "”", "where", "a", "linear", "model", "is", "forced", "to", "attend", "exclusively", "to", "the", "signal", "that", "aligns", "most", "closely", "with", "its", "weights", ",", "even", "if", "multiple", "signals", "are", "present", "and", "other", "signals", "have", "much", "greater", "amplitude", ".", "This", "explanation", "shows", "that", "a", "simple", "linear", "model", "can", "have", "adversarial", "examples", "if", "its", "input", "has", "sufficient", "dimensionality", ".", "Previous", "explanations", "for", "adversarial", "examples", "invoked", "hypothesized", "properties", "of", "neural", "networks", ",", "such", "as", "their", "supposed", "highly", "non", "-", "linear", "nature", ".", "Our", "hypothesis", "based", "on", "linearity", "is", "simpler", ",", "and", "can", "also", "explain", "why", "softmax", "regression", "is", "vulnerable", "to", "adversarial", "examples", ".", "section", ":", "Linear", "perturbation", "of", "non", "-", "linear", "models", "The", "linear", "view", "of", "adversarial", "examples", "suggests", "a", "fast", "way", "of", "generating", "them", ".", "We", "hypothesize", "that", "neural", "networks", "are", "too", "linear", "to", "resist", "linear", "adversarial", "perturbation", ".", "LSTMs", "lstm", ",", "ReLUs", "Jarrett", "-", "ICCV2009", ",", "Glorot", "+", "al", "-", "AI", "-", "2011", ",", "and", "maxout", "networks", "Goodfellow", "-", "et", "-", "al", "-", "ICML2013", "are", "all", "intentionally", "designed", "to", "behave", "in", "very", "linear", "ways", ",", "so", "that", "they", "are", "easier", "to", "optimize", ".", "More", "nonlinear", "models", "such", "as", "sigmoid", "networks", "are", "carefully", "tuned", "to", "spend", "most", "of", "their", "time", "in", "the", "non", "-", "saturating", ",", "more", "linear", "regime", "for", "the", "same", "reason", ".", "This", "linear", "behavior", "suggests", "that", "cheap", ",", "analytical", "perturbations", "of", "a", "linear", "model", "should", "also", "damage", "neural", "networks", ".", "Let", "be", "the", "parameters", "of", "a", "model", ",", "the", "input", "to", "the", "model", ",", "the", "targets", "associated", "with", "(", "for", "machine", "learning", "tasks", "that", "have", "targets", ")", "and", "be", "the", "cost", "used", "to", "train", "the", "neural", "network", ".", "We", "can", "linearize", "the", "cost", "function", "around", "the", "current", "value", "of", ",", "obtaining", "an", "optimal", "max", "-", "norm", "constrained", "pertubation", "of", "We", "refer", "to", "this", "as", "the", "“", "fast", "gradient", "sign", "method", "”", "of", "generating", "adversarial", "examples", ".", "Note", "that", "the", "required", "gradient", "can", "be", "computed", "efficiently", "using", "backpropagation", ".", "We", "find", "that", "this", "method", "reliably", "causes", "a", "wide", "variety", "of", "models", "to", "misclassify", "their", "input", ".", "See", "Fig", ".", "[", "reference", "]", "for", "a", "demonstration", "on", "ImageNet", ".", "We", "find", "that", "using", ",", "we", "cause", "a", "shallow", "softmax", "classifier", "to", "have", "an", "error", "rate", "of", "99.9", "%", "with", "an", "average", "confidence", "of", "79.3", "%", "on", "the", "MNIST", "LeCun", "+", "98", "test", "set", ".", "In", "the", "same", "setting", ",", "a", "maxout", "network", "misclassifies", "89.4", "%", "of", "our", "adversarial", "examples", "with", "an", "average", "confidence", "of", "97.6", "%", ".", "Similarly", ",", "using", ",", "we", "obtain", "an", "error", "rate", "of", "87.15", "%", "and", "an", "average", "probability", "of", "96.6", "%", "assigned", "to", "the", "incorrect", "labels", "when", "using", "a", "convolutional", "maxout", "network", "on", "a", "preprocessed", "version", "of", "the", "CIFAR", "-", "10", "KrizhevskyHinton2009", "test", "set", ".", "Other", "simple", "methods", "of", "generating", "adversarial", "examples", "are", "possible", ".", "For", "example", ",", "we", "also", "found", "that", "rotating", "by", "a", "small", "angle", "in", "the", "direction", "of", "the", "gradient", "reliably", "produces", "adversarial", "examples", ".", "The", "fact", "that", "these", "simple", ",", "cheap", "algorithms", "are", "able", "to", "generate", "misclassified", "examples", "serves", "as", "evidence", "in", "favor", "of", "our", "interpretation", "of", "adversarial", "examples", "as", "a", "result", "of", "linearity", ".", "The", "algorithms", "are", "also", "useful", "as", "a", "way", "of", "speeding", "up", "adversarial", "training", "or", "even", "just", "analysis", "of", "trained", "networks", ".", "section", ":", "Adversarial", "training", "of", "linear", "models", "versus", "weight", "decay", "Perhaps", "the", "simplest", "possible", "model", "we", "can", "consider", "is", "logistic", "regression", ".", "In", "this", "case", ",", "the", "fast", "gradient", "sign", "method", "is", "exact", ".", "We", "can", "use", "this", "case", "to", "gain", "some", "intuition", "for", "how", "adversarial", "examples", "are", "generated", "in", "a", "simple", "setting", ".", "See", "Fig", ".", "[", "reference", "]", "for", "instructive", "images", ".", "If", "we", "train", "a", "single", "model", "to", "recognize", "labels", "with", "where", "is", "the", "logistic", "sigmoid", "function", ",", "then", "training", "consists", "of", "gradient", "descent", "on", "where", "is", "the", "softplus", "function", ".", "We", "can", "derive", "a", "simple", "analytical", "form", "for", "training", "on", "the", "worst", "-", "case", "adversarial", "perturbation", "of", "rather", "than", "itself", ",", "based", "on", "gradient", "sign", "perturbation", ".", "Note", "that", "the", "sign", "of", "the", "gradient", "is", "just", ",", "and", "that", ".", "The", "adversarial", "version", "of", "logistic", "regression", "is", "therefore", "to", "minimize", "This", "is", "somewhat", "similar", "to", "regularization", ".", "However", ",", "there", "are", "some", "important", "differences", ".", "Most", "significantly", ",", "the", "penalty", "is", "subtracted", "off", "the", "model", "’s", "activation", "during", "training", ",", "rather", "than", "added", "to", "the", "training", "cost", ".", "This", "means", "that", "the", "penalty", "can", "eventually", "start", "to", "disappear", "if", "the", "model", "learns", "to", "make", "confident", "enough", "predictions", "that", "saturates", ".", "This", "is", "not", "guaranteed", "to", "happen", "—", "in", "the", "underfitting", "regime", ",", "adversarial", "training", "will", "simply", "worsen", "underfitting", ".", "We", "can", "thus", "view", "weight", "decay", "as", "being", "more", "“", "worst", "case", "”", "than", "adversarial", "training", ",", "because", "it", "fails", "to", "deactivate", "in", "the", "case", "of", "good", "margin", ".", "If", "we", "move", "beyond", "logistic", "regression", "to", "multiclass", "softmax", "regression", ",", "weight", "decay", "becomes", "even", "more", "pessimistic", ",", "because", "it", "treats", "each", "of", "the", "softmax", "’s", "outputs", "as", "independently", "perturbable", ",", "when", "in", "fact", "it", "is", "usually", "not", "possible", "to", "find", "a", "single", "that", "aligns", "with", "all", "of", "the", "class", "’s", "weight", "vectors", ".", "Weight", "decay", "overestimates", "the", "damage", "achievable", "with", "perturbation", "even", "more", "in", "the", "case", "of", "a", "deep", "network", "with", "multiple", "hidden", "units", ".", "Because", "weight", "decay", "overestimates", "the", "amount", "of", "damage", "an", "adversary", "can", "do", ",", "it", "is", "necessary", "to", "use", "a", "smaller", "weight", "decay", "coefficient", "than", "the", "associated", "with", "the", "precision", "of", "our", "features", ".", "When", "training", "maxout", "networks", "on", "MNIST", ",", "we", "obtained", "good", "results", "using", "adversarial", "training", "with", ".", "When", "applying", "weight", "decay", "to", "the", "first", "layer", ",", "we", "found", "that", "even", "a", "coefficient", "of", ".0025", "was", "too", "large", ",", "and", "caused", "the", "model", "to", "get", "stuck", "with", "over", "5", "%", "error", "on", "the", "training", "set", ".", "Smaller", "weight", "decay", "coefficients", "permitted", "succesful", "training", "but", "conferred", "no", "regularization", "benefit", ".", "section", ":", "Adversarial", "training", "of", "deep", "networks", "The", "criticism", "of", "deep", "networks", "as", "vulnerable", "to", "adversarial", "examples", "is", "somewhat", "misguided", ",", "because", "unlike", "shallow", "linear", "models", ",", "deep", "networks", "are", "at", "least", "able", "to", "represent", "functions", "that", "resist", "adversarial", "perturbation", ".", "The", "universal", "approximator", "theorem", "Hornik89", "guarantees", "that", "a", "neural", "network", "with", "at", "least", "one", "hidden", "layer", "can", "represent", "any", "function", "to", "an", "arbitary", "degree", "of", "accuracy", "so", "long", "as", "its", "hidden", "layer", "is", "permitted", "to", "have", "enough", "units", ".", "Shallow", "linear", "models", "are", "not", "able", "to", "become", "constant", "near", "training", "points", "while", "also", "assigning", "different", "outputs", "to", "different", "training", "points", ".", "Of", "course", ",", "the", "universal", "approximator", "theorem", "does", "not", "say", "anything", "about", "whether", "a", "training", "algorithm", "will", "be", "able", "to", "discover", "a", "function", "with", "all", "of", "the", "desired", "properties", ".", "Obviously", ",", "standard", "supervised", "training", "does", "not", "specify", "that", "the", "chosen", "function", "be", "resistant", "to", "adversarial", "examples", ".", "This", "must", "be", "encoded", "in", "the", "training", "procedure", "somehow", ".", "Szegedy", "-", "ICLR2014", "showed", "that", "by", "training", "on", "a", "mixture", "of", "adversarial", "and", "clean", "examples", ",", "a", "neural", "network", "could", "be", "regularized", "somewhat", ".", "Training", "on", "adversarial", "examples", "is", "somewhat", "different", "from", "other", "data", "augmentation", "schemes", ";", "usually", ",", "one", "augments", "the", "data", "with", "transformations", "such", "as", "translations", "that", "are", "expected", "to", "actually", "occur", "in", "the", "test", "set", ".", "This", "form", "of", "data", "augmentation", "instead", "uses", "inputs", "that", "are", "unlikely", "to", "occur", "naturally", "but", "that", "expose", "flaws", "in", "the", "ways", "that", "the", "model", "conceptualizes", "its", "decision", "function", ".", "At", "the", "time", ",", "this", "procedure", "was", "never", "demonstrated", "to", "improve", "beyond", "dropout", "on", "a", "state", "of", "the", "art", "benchmark", ".", "However", ",", "this", "was", "partially", "because", "it", "was", "difficult", "to", "experiment", "extensively", "with", "expensive", "adversarial", "examples", "based", "on", "L", "-", "BFGS", ".", "We", "found", "that", "training", "with", "an", "adversarial", "objective", "function", "based", "on", "the", "fast", "gradient", "sign", "method", "was", "an", "effective", "regularizer", ":", "In", "all", "of", "our", "experiments", ",", "we", "used", ".", "Other", "values", "may", "work", "better", ";", "our", "initial", "guess", "of", "this", "hyperparameter", "worked", "well", "enough", "that", "we", "did", "not", "feel", "the", "need", "to", "explore", "more", ".", "This", "approach", "means", "that", "we", "continually", "update", "our", "supply", "of", "adversarial", "examples", ",", "to", "make", "them", "resist", "the", "current", "version", "of", "the", "model", ".", "Using", "this", "approach", "to", "train", "a", "maxout", "network", "that", "was", "also", "regularized", "with", "dropout", ",", "we", "were", "able", "to", "reduce", "the", "error", "rate", "from", "0.94", "%", "without", "adversarial", "training", "to", "0.84", "%", "with", "adversarial", "training", ".", "We", "observed", "that", "we", "were", "not", "reaching", "zero", "error", "rate", "on", "adversarial", "examples", "on", "the", "training", "set", ".", "We", "fixed", "this", "problem", "by", "making", "two", "changes", ".", "First", ",", "we", "made", "the", "model", "larger", ",", "using", "1600", "units", "per", "layer", "rather", "than", "the", "240", "used", "by", "the", "original", "maxout", "network", "for", "this", "problem", ".", "Without", "adversarial", "training", ",", "this", "causes", "the", "model", "to", "overfit", "slightly", ",", "and", "get", "an", "error", "rate", "of", "1.14", "%", "on", "the", "test", "set", ".", "With", "adversarial", "training", ",", "we", "found", "that", "the", "validation", "set", "error", "leveled", "off", "over", "time", ",", "and", "made", "very", "slow", "progress", ".", "The", "original", "maxout", "result", "uses", "early", "stopping", ",", "and", "terminates", "learning", "after", "the", "validation", "set", "error", "rate", "has", "not", "decreased", "for", "100", "epochs", ".", "We", "found", "that", "while", "the", "validation", "set", "error", "was", "very", "flat", ",", "the", "adversarial", "validation", "set", "error", "was", "not", ".", "We", "therefore", "used", "early", "stopping", "on", "the", "adversarial", "validation", "set", "error", ".", "Using", "this", "criterion", "to", "choose", "the", "number", "of", "epochs", "to", "train", "for", ",", "we", "then", "retrained", "on", "all", "60", ",", "000", "examples", ".", "Five", "different", "training", "runs", "using", "different", "seeds", "for", "the", "random", "number", "generators", "used", "to", "select", "minibatches", "of", "training", "examples", ",", "initialize", "model", "weights", ",", "and", "generate", "dropout", "masks", "result", "in", "four", "trials", "that", "each", "had", "an", "error", "rate", "of", "0.77", "%", "on", "the", "test", "set", "and", "one", "trial", "that", "had", "an", "error", "rate", "of", "0.83", "%", ".", "The", "average", "of", "0.782", "%", "is", "the", "best", "result", "reported", "on", "the", "permutation", "invariant", "version", "of", "MNIST", ",", "though", "statistically", "indistinguishable", "from", "the", "result", "obtained", "by", "fine", "-", "tuning", "DBMs", "with", "dropout", "dropout", "at", "0.79", "%", ".", "The", "model", "also", "became", "somewhat", "resistant", "to", "adversarial", "examples", ".", "Recall", "that", "without", "adversarial", "training", ",", "this", "same", "kind", "of", "model", "had", "an", "error", "rate", "of", "89.4", "%", "on", "adversarial", "examples", "based", "on", "the", "fast", "gradient", "sign", "method", ".", "With", "adversarial", "training", ",", "the", "error", "rate", "fell", "to", "17.9", "%", ".", "Adversarial", "examples", "are", "transferable", "between", "the", "two", "models", "but", "with", "the", "adversarially", "trained", "model", "showing", "greater", "robustness", ".", "Adversarial", "examples", "generated", "via", "the", "original", "model", "yield", "an", "error", "rate", "of", "19.6", "%", "on", "the", "adversarially", "trained", "model", ",", "while", "adversarial", "examples", "generated", "via", "the", "new", "model", "yield", "an", "error", "rate", "of", "40.9", "%", "on", "the", "original", "model", ".", "When", "the", "adversarially", "trained", "model", "does", "misclassify", "an", "adversarial", "example", ",", "its", "predictions", "are", "unfortunately", "still", "highly", "confident", ".", "The", "average", "confidence", "on", "a", "misclassified", "example", "was", "81.4", "%", ".", "We", "also", "found", "that", "the", "weights", "of", "the", "learned", "model", "changed", "significantly", ",", "with", "the", "weights", "of", "the", "adversarially", "trained", "model", "being", "significantly", "more", "localized", "and", "interpretable", "(", "see", "Fig", ".", "[", "reference", "]", ")", ".", "The", "adversarial", "training", "procedure", "can", "be", "seen", "as", "minimizing", "the", "worst", "case", "error", "when", "the", "data", "is", "perturbed", "by", "an", "adversary", ".", "That", "can", "be", "interpreted", "as", "learning", "to", "play", "an", "adversarial", "game", ",", "or", "as", "minimizing", "an", "upper", "bound", "on", "the", "expected", "cost", "over", "noisy", "samples", "with", "noise", "from", "added", "to", "the", "inputs", ".", "Adversarial", "training", "can", "also", "be", "seen", "as", "a", "form", "of", "active", "learning", ",", "where", "the", "model", "is", "able", "to", "request", "labels", "on", "new", "points", ".", "In", "this", "case", "the", "human", "labeler", "is", "replaced", "with", "a", "heuristic", "labeler", "that", "copies", "labels", "from", "nearby", "points", ".", "We", "could", "also", "regularize", "the", "model", "to", "be", "insensitive", "to", "changes", "in", "its", "features", "that", "are", "smaller", "than", "the", "precision", "simply", "by", "training", "on", "all", "points", "within", "the", "max", "norm", "box", ",", "or", "sampling", "many", "points", "within", "this", "box", ".", "This", "corresponds", "to", "adding", "noise", "with", "max", "norm", "during", "training", ".", "However", ",", "noise", "with", "zero", "mean", "and", "zero", "covariance", "is", "very", "inefficient", "at", "preventing", "adversarial", "examples", ".", "The", "expected", "dot", "product", "between", "any", "reference", "vector", "and", "such", "a", "noise", "vector", "is", "zero", ".", "This", "means", "that", "in", "many", "cases", "the", "noise", "will", "have", "essentially", "no", "effect", "rather", "than", "yielding", "a", "more", "difficult", "input", ".", "In", "fact", ",", "in", "many", "cases", "the", "noise", "will", "actualy", "result", "in", "a", "lower", "objective", "function", "value", ".", "We", "can", "think", "of", "adversarial", "training", "as", "doing", "hard", "example", "mining", "among", "the", "set", "of", "noisy", "inputs", ",", "in", "order", "to", "train", "more", "efficiently", "by", "considering", "only", "those", "noisy", "points", "that", "strongly", "resist", "classification", ".", "As", "control", "experiments", ",", "we", "trained", "training", "a", "maxout", "network", "with", "noise", "based", "on", "randomly", "adding", "to", "each", "pixel", ",", "or", "adding", "noise", "in", "to", "each", "pixel", ".", "These", "obtained", "an", "error", "rate", "of", "86.2", "%", "with", "confidence", "97.3", "%", "and", "an", "error", "rate", "of", "90.4", "%", "with", "a", "confidence", "of", "97.8", "%", "respectively", "on", "fast", "gradient", "sign", "adversarial", "examples", ".", "Because", "the", "derivative", "of", "the", "sign", "function", "is", "zero", "or", "undefined", "everywhere", ",", "gradient", "descent", "on", "the", "adversarial", "objective", "function", "based", "on", "the", "fast", "gradient", "sign", "method", "does", "not", "allow", "the", "model", "to", "anticipate", "how", "the", "adversary", "will", "react", "to", "changes", "in", "the", "parameters", ".", "If", "we", "instead", "adversarial", "examples", "based", "on", "small", "rotations", "or", "addition", "of", "the", "scaled", "gradient", ",", "then", "the", "perturbation", "process", "is", "itself", "differentiable", "and", "the", "learning", "can", "take", "the", "reaction", "of", "the", "adversary", "into", "account", ".", "However", ",", "we", "did", "not", "find", "nearly", "as", "powerful", "of", "a", "regularizing", "result", "from", "this", "process", ",", "perhaps", "because", "these", "kinds", "of", "adversarial", "examples", "are", "not", "as", "difficult", "to", "solve", ".", "One", "natural", "question", "is", "whether", "it", "is", "better", "to", "perturb", "the", "input", "or", "the", "hidden", "layers", "or", "both", ".", "Here", "the", "results", "are", "inconsistent", ".", "Szegedy", "-", "ICLR2014", "reported", "that", "adversarial", "perturbations", "yield", "the", "best", "regularization", "when", "applied", "to", "the", "hidden", "layers", ".", "That", "result", "was", "obtained", "on", "a", "sigmoidal", "network", ".", "In", "our", "experiments", "with", "the", "fast", "gradient", "sign", "method", ",", "we", "find", "that", "networks", "with", "hidden", "units", "whose", "activations", "are", "unbounded", "simply", "respond", "by", "making", "their", "hidden", "unit", "activations", "very", "large", ",", "so", "it", "is", "usually", "better", "to", "just", "perturb", "the", "original", "input", ".", "On", "saturating", "models", "such", "as", "the", "Rust", "model", "we", "found", "that", "perturbation", "of", "the", "input", "performed", "comparably", "to", "perturbation", "of", "the", "hidden", "layers", ".", "Perturbations", "based", "on", "rotating", "the", "hidden", "layers", "solve", "the", "problem", "of", "unbounded", "activations", "growing", "to", "make", "additive", "perturbations", "smaller", "by", "comparison", ".", "We", "were", "able", "to", "succesfully", "train", "maxout", "networks", "with", "rotational", "perturbations", "of", "the", "hidden", "layers", ".", "However", ",", "this", "did", "not", "yield", "nearly", "as", "strong", "of", "a", "regularizing", "effect", "as", "additive", "perturbation", "of", "the", "input", "layer", ".", "Our", "view", "of", "adversarial", "training", "is", "that", "it", "is", "only", "clearly", "useful", "when", "the", "model", "has", "the", "capacity", "to", "learn", "to", "resist", "adversarial", "examples", ".", "This", "is", "only", "clearly", "the", "case", "when", "a", "universal", "approximator", "theorem", "applies", ".", "Because", "the", "last", "layer", "of", "a", "neural", "network", ",", "the", "linear", "-", "sigmoid", "or", "linear", "-", "softmax", "layer", ",", "is", "not", "a", "universal", "approximator", "of", "functions", "of", "the", "final", "hidden", "layer", ",", "this", "suggests", "that", "one", "is", "likely", "to", "encounter", "problems", "with", "underfitting", "when", "applying", "adversarial", "perturbations", "to", "the", "final", "hidden", "layer", ".", "We", "indeed", "found", "this", "effect", ".", "Our", "best", "results", "with", "training", "using", "perturbations", "of", "hidden", "layers", "never", "involved", "perturbations", "of", "the", "final", "hidden", "layer", ".", "section", ":", "Different", "kinds", "of", "model", "capacity", "One", "reason", "that", "the", "existence", "of", "adversarial", "examples", "can", "seem", "counter", "-", "intuitive", "is", "that", "most", "of", "us", "have", "poor", "intuitions", "for", "high", "dimensional", "spaces", ".", "We", "live", "in", "three", "dimensions", ",", "so", "we", "are", "not", "used", "to", "small", "effects", "in", "hundreds", "of", "dimensions", "adding", "up", "to", "create", "a", "large", "effect", ".", "There", "is", "another", "way", "that", "our", "intuitions", "serve", "us", "poorly", ".", "Many", "people", "think", "of", "models", "with", "low", "capacity", "as", "being", "unable", "to", "make", "many", "different", "confident", "predictions", ".", "This", "is", "not", "correct", ".", "Some", "models", "with", "low", "capacity", "do", "exhibit", "this", "behavior", ".", "For", "example", "shallow", "RBF", "networks", "with", "are", "only", "able", "to", "confidently", "predict", "that", "the", "positive", "class", "is", "present", "in", "the", "vicinity", "of", ".", "Elsewhere", ",", "they", "default", "to", "predicting", "the", "class", "is", "absent", ",", "or", "have", "low", "-", "confidence", "predictions", ".", "RBF", "networks", "are", "naturally", "immune", "to", "adversarial", "examples", ",", "in", "the", "sense", "that", "they", "have", "low", "confidence", "when", "they", "are", "fooled", ".", "A", "shallow", "RBF", "network", "with", "no", "hidden", "layers", "gets", "an", "error", "rate", "of", "55.4", "%", "on", "MNIST", "using", "adversarial", "examples", "generated", "with", "the", "fast", "gradient", "sign", "method", "and", ".", "However", ",", "its", "confidence", "on", "mistaken", "examples", "is", "only", ".", "Its", "average", "confidence", "on", "clean", "test", "examples", "is", "%", ".", "We", "ca", "n’t", "expect", "a", "model", "with", "such", "low", "capacity", "to", "get", "the", "right", "answer", "at", "all", "points", "of", "space", ",", "but", "it", "does", "correctly", "respond", "by", "reducing", "its", "confidence", "considerably", "on", "points", "it", "does", "not", "“", "understand", ".", "”", "RBF", "units", "are", "unfortunately", "not", "invariant", "to", "any", "significant", "transformations", "so", "they", "can", "not", "generalize", "very", "well", ".", "We", "can", "view", "linear", "units", "and", "RBF", "units", "as", "different", "points", "on", "a", "precision", "-", "recall", "tradeoff", "curve", ".", "Linear", "units", "achieve", "high", "recall", "by", "responding", "to", "every", "input", "in", "a", "certain", "direction", ",", "but", "may", "have", "low", "precision", "due", "to", "responding", "too", "strongly", "in", "unfamiliar", "situations", ".", "RBF", "units", "achieve", "high", "precision", "by", "responding", "only", "to", "a", "specific", "point", "in", "space", ",", "but", "in", "doing", "so", "sacrifice", "recall", ".", "Motivated", "by", "this", "idea", ",", "we", "decided", "to", "explore", "a", "variety", "of", "models", "involving", "quadratic", "units", ",", "including", "deep", "RBF", "networks", ".", "We", "found", "this", "to", "be", "a", "difficult", "task", "—", "very", "model", "with", "sufficient", "quadratic", "inhibition", "to", "resist", "adversarial", "perturbation", "obtained", "high", "training", "set", "error", "when", "trained", "with", "SGD", ".", "section", ":", "Why", "do", "adversarial", "examples", "generalize", "?", "An", "intriguing", "aspect", "of", "adversarial", "examples", "is", "that", "an", "example", "generated", "for", "one", "model", "is", "often", "misclassified", "by", "other", "models", ",", "even", "when", "they", "have", "different", "architecures", "or", "were", "trained", "on", "disjoint", "training", "sets", ".", "Moreover", ",", "when", "these", "different", "models", "misclassify", "an", "adversarial", "example", ",", "they", "often", "agree", "with", "each", "other", "on", "its", "class", ".", "Explanations", "based", "on", "extreme", "non", "-", "linearity", "and", "overfitting", "can", "not", "readily", "account", "for", "this", "behavior", "—", "why", "should", "multiple", "extremely", "non", "-", "linear", "model", "with", "excess", "capacity", "consistently", "label", "out", "-", "of", "-", "distribution", "points", "in", "the", "same", "way", "?", "This", "behavior", "is", "especially", "surprising", "from", "the", "view", "of", "the", "hypothesis", "that", "adversarial", "examples", "finely", "tile", "space", "like", "the", "rational", "numbers", "among", "the", "reals", ",", "because", "in", "this", "view", "adversarial", "examples", "are", "common", "but", "occur", "only", "at", "very", "precise", "locations", ".", "Under", "the", "linear", "view", ",", "adversarial", "examples", "occur", "in", "broad", "subspaces", ".", "The", "direction", "need", "only", "have", "positive", "dot", "product", "with", "the", "gradient", "of", "the", "cost", "function", ",", "and", "need", "only", "be", "large", "enough", ".", "Fig", ".", "[", "reference", "]", "demonstrates", "this", "phenomenon", ".", "By", "tracing", "out", "different", "values", "of", "we", "see", "that", "adversarial", "examples", "occur", "in", "contiguous", "regions", "of", "the", "1", "-", "D", "subspace", "defined", "by", "the", "fast", "gradient", "sign", "method", ",", "not", "in", "fine", "pockets", ".", "This", "explains", "why", "adversarial", "examples", "are", "abundant", "and", "why", "an", "example", "misclassified", "by", "one", "classifier", "has", "a", "fairly", "high", "prior", "probability", "of", "being", "misclassified", "by", "another", "classifier", ".", "To", "explain", "why", "mutiple", "classifiers", "assign", "the", "same", "class", "to", "adversarial", "examples", ",", "we", "hypothesize", "that", "neural", "networks", "trained", "with", "current", "methodologies", "all", "resemble", "the", "linear", "classifier", "learned", "on", "the", "same", "training", "set", ".", "This", "reference", "classifier", "is", "able", "to", "learn", "approximately", "the", "same", "classification", "weights", "when", "trained", "on", "different", "subsets", "of", "the", "training", "set", ",", "simply", "because", "machine", "learning", "algorithms", "are", "able", "to", "generalize", ".", "The", "stability", "of", "the", "underlying", "classification", "weights", "in", "turn", "results", "in", "the", "stability", "of", "adversarial", "examples", ".", "To", "test", "this", "hypothesis", ",", "we", "generated", "adversarial", "examples", "on", "a", "deep", "maxout", "network", "and", "classified", "these", "examples", "using", "a", "shallow", "softmax", "network", "and", "a", "shallow", "RBF", "network", ".", "On", "examples", "that", "were", "misclassified", "by", "the", "maxout", "network", ",", "the", "RBF", "network", "predicted", "the", "maxout", "network", "’s", "class", "assignment", "only", "16.0", "%", "of", "the", "time", ",", "while", "the", "softmax", "classifier", "predict", "the", "maxout", "network", "’s", "class", "correctly", "54.6", "%", "of", "the", "time", ".", "These", "numbers", "are", "largely", "driven", "by", "the", "differing", "error", "rate", "of", "the", "different", "models", "though", ".", "If", "we", "exclude", "our", "attention", "to", "cases", "where", "both", "models", "being", "compared", "make", "a", "mistake", ",", "then", "softmax", "regression", "predict", "’s", "maxout", "’s", "class", "84.6", "%", "of", "the", "time", ",", "while", "the", "RBF", "network", "is", "able", "to", "predict", "maxout", "’s", "class", "only", "54.3", "%", "of", "the", "time", ".", "For", "comparison", ",", "the", "RBF", "network", "can", "predict", "softmax", "regression", "’s", "class", "53.6", "%", "of", "the", "time", ",", "so", "it", "does", "have", "a", "strong", "linear", "component", "to", "its", "own", "behavior", ".", "Our", "hypothesis", "does", "not", "explain", "all", "of", "the", "maxout", "network", "’s", "mistakes", "or", "all", "of", "the", "mistakes", "that", "generalize", "across", "models", ",", "but", "clearly", "a", "significant", "proportion", "of", "them", "are", "consistent", "with", "linear", "behavior", "being", "a", "major", "cause", "of", "cross", "-", "model", "generalization", ".", "section", ":", "Alternative", "hypotheses", "We", "now", "consider", "and", "refute", "some", "alternative", "hypotheses", "for", "the", "existence", "of", "adversarial", "examples", ".", "First", ",", "one", "hypothesis", "is", "that", "generative", "training", "could", "provide", "more", "constraint", "on", "the", "training", "process", ",", "or", "cause", "the", "model", "to", "learn", "what", "to", "distinguish", "“", "real", "”", "from", "“", "fake", "”", "data", "and", "be", "confident", "only", "on", "“", "real", "”", "data", ".", "The", "MP", "-", "DBM", "mpdbm", "provides", "a", "good", "model", "to", "test", "this", "hypothesis", ".", "Its", "inference", "procedure", "gets", "good", "classification", "accuracy", "(", "an", "0.88", "%", "error", "rate", ")", "on", "MNIST", ".", "This", "inference", "procedure", "is", "differentiable", ".", "Other", "generative", "models", "either", "have", "non", "-", "differentiable", "inference", "procedures", ",", "making", "it", "harder", "to", "compute", "adversarial", "examples", ",", "or", "require", "an", "additional", "non", "-", "generative", "discriminator", "model", "to", "get", "good", "classification", "accuracy", "on", "MNIST", ".", "In", "the", "case", "of", "the", "MP", "-", "DBM", ",", "we", "can", "be", "sure", "that", "the", "generative", "model", "itself", "is", "responding", "to", "adversarial", "examples", ",", "rather", "than", "the", "non", "-", "generative", "classifier", "model", "on", "top", ".", "We", "find", "that", "the", "model", "is", "vulnerable", "to", "adversarial", "examples", ".", "With", "an", "of", "0.25", ",", "we", "find", "an", "error", "rate", "of", "97.5", "%", "on", "adversarial", "examples", "generated", "from", "the", "MNIST", "test", "set", ".", "It", "remains", "possible", "that", "some", "other", "form", "of", "generative", "training", "could", "confer", "resistance", ",", "but", "clearly", "the", "mere", "fact", "of", "being", "generative", "is", "not", "alone", "sufficient", ".", "Another", "hypothesis", "about", "why", "adversarial", "examples", "exist", "is", "that", "individual", "models", "have", "strange", "quirks", "but", "averaging", "over", "many", "models", "can", "cause", "adversarial", "examples", "to", "wash", "out", ".", "To", "test", "this", "hypothesis", ",", "we", "trained", "an", "ensemble", "of", "twelve", "maxout", "networks", "on", "MNIST", ".", "Each", "network", "was", "trained", "using", "a", "different", "seed", "for", "the", "random", "number", "generator", "used", "to", "initialize", "the", "weights", ",", "generate", "dropout", "masks", ",", "and", "select", "minibatches", "of", "data", "for", "stochastic", "gradient", "descent", ".", "The", "ensemble", "gets", "an", "error", "rate", "of", "91.1", "%", "on", "adversarial", "examples", "designed", "to", "perturb", "the", "entire", "ensemble", "with", ".", "If", "we", "instead", "use", "adversarial", "examples", "designed", "to", "perturb", "only", "one", "member", "of", "the", "ensemble", ",", "the", "error", "rate", "falls", "to", "87.9", "%", ".", "Ensembling", "provides", "only", "limited", "resistance", "to", "adversarial", "perturbation", ".", "section", ":", "Summary", "and", "discussion", "As", "a", "summary", ",", "this", "paper", "has", "made", "the", "following", "observations", ":", "Adversarial", "examples", "can", "be", "explained", "as", "a", "property", "of", "high", "-", "dimensional", "dot", "products", ".", "They", "are", "a", "result", "of", "models", "being", "too", "linear", ",", "rather", "than", "too", "nonlinear", ".", "The", "generalization", "of", "adversarial", "examples", "across", "different", "models", "can", "be", "explained", "as", "a", "result", "of", "adversarial", "perturbations", "being", "highly", "aligned", "with", "the", "weight", "vectors", "of", "a", "model", ",", "and", "different", "models", "learning", "similar", "functions", "when", "trained", "to", "perform", "the", "same", "task", ".", "The", "direction", "of", "perturbation", ",", "rather", "than", "the", "specific", "point", "in", "space", ",", "matters", "most", ".", "Space", "is", "not", "full", "of", "pockets", "of", "adversarial", "examples", "that", "finely", "tile", "the", "reals", "like", "the", "rational", "numbers", ".", "Because", "it", "is", "the", "direction", "that", "matters", "most", ",", "adversarial", "perturbations", "generalize", "across", "different", "clean", "examples", ".", "We", "have", "introduced", "a", "family", "of", "fast", "methods", "for", "generating", "adversarial", "examples", ".", "We", "have", "demonstrated", "that", "adversarial", "training", "can", "result", "in", "regularization", ";", "even", "further", "regularization", "than", "dropout", ".", "We", "have", "run", "control", "experiments", "that", "failed", "to", "reproduce", "this", "effect", "with", "simpler", "but", "less", "efficient", "regularizers", "including", "weight", "decay", "and", "adding", "noise", ".", "Models", "that", "are", "easy", "to", "optimize", "are", "easy", "to", "perturb", ".", "Linear", "models", "lack", "the", "capacity", "to", "resist", "adversarial", "perturbation", ";", "only", "structures", "with", "a", "hidden", "layer", "(", "where", "the", "universal", "approximator", "theorem", "applies", ")", "should", "be", "trained", "to", "resist", "adversarial", "perturbation", ".", "RBF", "networks", "are", "resistant", "to", "adversarial", "examples", ".", "Models", "trained", "to", "model", "the", "input", "distribution", "are", "not", "resistant", "to", "adversarial", "examples", ".", "Ensembles", "are", "not", "resistant", "to", "adversarial", "examples", ".", "Some", "further", "observations", "concerning", "rubbish", "class", "examples", "are", "presented", "in", "the", "appendix", ":", "Rubbish", "class", "examples", "are", "ubiquitous", "and", "easily", "generated", ".", "Shallow", "linear", "models", "are", "not", "resistant", "to", "rubbish", "class", "examples", ".", "RBF", "networks", "are", "resistant", "to", "rubbish", "class", "examples", ".", "Gradient", "-", "based", "optimization", "is", "the", "workhorse", "of", "modern", "AI", ".", "Using", "a", "network", "that", "has", "been", "designed", "to", "be", "sufficiently", "linear", "–", "whether", "it", "is", "a", "ReLU", "or", "maxout", "network", ",", "an", "LSTM", ",", "or", "a", "sigmoid", "network", "that", "has", "been", "carefully", "configured", "not", "to", "saturate", "too", "much–", "we", "are", "able", "to", "fit", "most", "problems", "we", "care", "about", ",", "at", "least", "on", "the", "training", "set", ".", "The", "existence", "of", "adversarial", "examples", "suggests", "that", "being", "able", "to", "explain", "the", "training", "data", "or", "even", "being", "able", "to", "correctly", "label", "the", "test", "data", "does", "not", "imply", "that", "our", "models", "truly", "understand", "the", "tasks", "we", "have", "asked", "them", "to", "perform", ".", "Instead", ",", "their", "linear", "responses", "are", "overly", "confident", "at", "points", "that", "do", "not", "occur", "in", "the", "data", "distribution", ",", "and", "these", "confident", "predictions", "are", "often", "highly", "incorrect", ".", "This", "work", "has", "shown", "we", "can", "partially", "correct", "for", "this", "problem", "by", "explicitly", "identifying", "problematic", "points", "and", "correcting", "the", "model", "at", "each", "of", "these", "points", ".", "However", ",", "one", "may", "also", "conclude", "that", "the", "model", "families", "we", "use", "are", "intrinsically", "flawed", ".", "Ease", "of", "optimization", "has", "come", "at", "the", "cost", "of", "models", "that", "are", "easily", "misled", ".", "This", "motivates", "the", "development", "of", "optimization", "procedures", "that", "are", "able", "to", "train", "models", "whose", "behavior", "is", "more", "locally", "stable", ".", "subsubsection", ":", "Acknowledgments", "We", "would", "like", "to", "thank", "Geoffrey", "Hinton", "and", "Ilya", "Sutskever", "for", "helpful", "discussions", ".", "We", "would", "also", "like", "to", "thank", "Jeff", "Dean", ",", "Greg", "Corrado", ",", "and", "Oriol", "Vinyals", "for", "their", "feedback", "on", "drafts", "of", "this", "article", ".", "We", "would", "like", "to", "thank", "the", "developers", "of", "Theano", "bergstra", "+", "al:2010", "-", "scipy", ",", "Bastien", "-", "Theano", "-", "2012", ",", "Pylearn2", "pylearn2_arxiv_2013", ",", "and", "DistBelief", "distbelief", ".", "plus", "0.3ex", "bibliography", ":", "References", "appendix", ":", "Rubbish", "class", "examples", "A", "concept", "related", "to", "adversarial", "examples", "is", "the", "concept", "of", "examples", "drawn", "from", "a", "“", "rubbish", "class", ".", "”", "These", "examples", "are", "degenerate", "inputs", "that", "a", "human", "would", "classify", "as", "not", "belonging", "to", "any", "of", "the", "categories", "in", "the", "training", "set", ".", "If", "we", "call", "these", "classes", "in", "the", "training", "set", "“", "the", "positive", "classes", ",", "”", "then", "we", "want", "to", "be", "careful", "to", "avoid", "false", "positives", "on", "rubbish", "inputs", "–", "i.e", ".", ",", "we", "do", "not", "want", "to", "classify", "a", "degenerate", "input", "as", "being", "something", "real", ".", "In", "the", "case", "of", "separate", "binary", "classifiers", "for", "each", "class", ",", "we", "want", "all", "classes", "output", "near", "zero", "probability", "of", "the", "class", "being", "present", ",", "and", "in", "the", "case", "of", "a", "multinoulli", "distribution", "over", "only", "the", "positive", "classes", ",", "we", "would", "prefer", "that", "the", "classifier", "output", "a", "high", "-", "entropy", "(", "nearly", "uniform", ")", "distribution", "over", "the", "classes", ".", "The", "traditional", "approach", "to", "reducing", "vulnerability", "to", "rubbish", "inputs", "is", "to", "introduce", "an", "extra", ",", "constant", "output", "to", "the", "model", "representing", "the", "rubbish", "class", "LeCun", "+", "98", ".", "fool", "recently", "re", "-", "popularized", "the", "concept", "of", "the", "rubbish", "class", "in", "the", "context", "of", "computer", "vision", "under", "the", "name", "fooling", "images", ".", "As", "with", "adversarial", "examples", ",", "there", "has", "been", "a", "misconception", "that", "rubbish", "class", "false", "positives", "are", "hard", "to", "find", ",", "and", "that", "they", "are", "primarily", "a", "problem", "faced", "by", "deep", "networks", ".", "Our", "explanation", "of", "adversarial", "examples", "as", "the", "result", "of", "linearity", "and", "high", "dimensional", "spaces", "also", "applies", "to", "analyzing", "the", "behavior", "of", "the", "model", "on", "rubbish", "class", "examples", ".", "Linear", "models", "produce", "more", "extreme", "predictions", "at", "points", "that", "are", "far", "from", "the", "training", "data", "than", "at", "points", "that", "are", "near", "the", "training", "data", ".", "In", "order", "to", "find", "high", "confidence", "rubbish", "false", "positives", "for", "such", "a", "model", ",", "we", "need", "only", "generate", "a", "point", "that", "is", "far", "from", "the", "data", ",", "with", "larger", "norms", "yielding", "more", "confidence", ".", "RBF", "networks", ",", "which", "are", "not", "able", "to", "confidently", "predict", "the", "presence", "of", "any", "class", "far", "from", "the", "training", "data", ",", "are", "not", "fooled", "by", "this", "phenomenon", ".", "We", "generated", "10", ",", "000", "samples", "from", "and", "fed", "them", "into", "various", "classifiers", "on", "the", "MNIST", "dataset", ".", "In", "this", "context", ",", "we", "consider", "assigning", "a", "probability", "greater", "than", "0.5", "to", "any", "class", "to", "be", "an", "error", ".", "A", "naively", "trained", "maxout", "network", "with", "a", "softmax", "layer", "on", "top", "had", "an", "error", "rate", "of", "98.35", "%", "on", "Gaussian", "rubbish", "examples", "with", "an", "average", "confidence", "of", "92.8", "%", "on", "mistakes", ".", "Changing", "the", "top", "layer", "to", "independent", "sigmoids", "dropped", "the", "error", "rate", "to", "68", "%", "with", "an", "average", "confidence", "on", "mistakes", "of", "87.9", "%", ".", "On", "CIFAR", "-", "10", ",", "using", "1", ",", "000", "samples", "from", ",", "a", "convolutional", "maxout", "net", "obtains", "an", "error", "rate", "of", "93.4", "%", ",", "with", "an", "average", "confidence", "of", "84.4", "%", ".", "These", "experiments", "suggest", "that", "the", "optimization", "algorithms", "employed", "by", "fool", "are", "overkill", "(", "or", "perhaps", "only", "needed", "on", "ImageNet", ")", ",", "and", "that", "the", "rich", "geometric", "structure", "in", "their", "fooling", "images", "are", "due", "to", "the", "priors", "encoded", "in", "their", "search", "procedures", ",", "rather", "than", "those", "structures", "being", "uniquely", "able", "to", "cause", "false", "positives", ".", "Though", "fool", "focused", "their", "attention", "on", "deep", "networks", ",", "shallow", "linear", "models", "have", "the", "same", "problem", ".", "A", "softmax", "regression", "model", "has", "an", "error", "rate", "of", "59.8", "%", "on", "the", "rubbish", "examples", ",", "with", "an", "average", "confidence", "on", "mistakes", "of", "70.8", "%", ".", "If", "we", "use", "instead", "an", "RBF", "network", ",", "which", "does", "not", "behave", "like", "a", "linear", "function", ",", "we", "find", "an", "error", "rate", "of", "0", "%", ".", "Note", "that", "when", "the", "error", "rate", "is", "zero", "the", "average", "confidence", "on", "a", "mistake", "is", "undefined", ".", "fool", "focused", "on", "the", "problem", "of", "generating", "fooling", "images", "for", "a", "specific", "class", ",", "which", "is", "a", "harder", "problem", "than", "simply", "finding", "points", "that", "the", "network", "confidently", "classifies", "as", "belonging", "to", "any", "one", "class", "despite", "being", "defective", ".", "The", "above", "methods", "on", "MNIST", "and", "CIFAR", "-", "10", "tend", "to", "have", "a", "very", "skewed", "distribution", "over", "classes", ".", "On", "MNIST", ",", "45.3", "%", "of", "a", "naively", "trained", "maxout", "network", "’s", "false", "positives", "were", "classified", "as", "5s", ",", "and", "none", "were", "classified", "as", "8s", ".", "Likewise", ",", "on", "CIFAR", "-", "10", ",", "49.7", "%", "of", "the", "convolutional", "network", "’s", "false", "positives", "were", "classified", "as", "frogs", ",", "and", "none", "were", "classified", "as", "airplanes", ",", "automobiles", ",", "horses", ",", "ships", ",", "or", "trucks", ".", "To", "solve", "the", "problem", "introduced", "by", "fool", "of", "generating", "a", "fooling", "image", "for", "a", "particular", "class", ",", "we", "propose", "adding", "to", "a", "Gaussian", "sample", "as", "a", "fast", "method", "of", "generating", "a", "fooling", "image", "classified", "as", "class", ".", "If", "we", "repeat", "this", "sampling", "process", "until", "it", "succeeds", ",", "we", "a", "randomized", "algorithm", "with", "variable", "runtime", ".", "On", "CIFAR", "-", "10", ",", "we", "found", "that", "one", "sampling", "step", "had", "a", "100", "%", "success", "rate", "for", "frogs", "and", "trucks", ",", "and", "the", "hardest", "class", "was", "airplanes", ",", "with", "a", "success", "rate", "of", "24.7", "%", "per", "sampling", "step", ".", "Averaged", "over", "all", "ten", "classes", ",", "the", "method", "has", "an", "average", "per", "-", "step", "success", "rate", "of", "75.3", "%", ".", "We", "can", "thus", "generate", "any", "desired", "class", "with", "a", "handful", "of", "samples", "and", "no", "special", "priors", ",", "rather", "than", "tens", "of", "thousands", "of", "generations", "of", "evolution", ".", "To", "confirm", "that", "the", "resulting", "examples", "are", "indeed", "fooling", "images", ",", "and", "not", "images", "of", "real", "classes", "rendered", "by", "the", "gradient", "sign", "method", ",", "see", "Fig", ".", "[", "reference", "]", ".", "The", "success", "rate", "of", "this", "method", "in", "terms", "of", "generating", "members", "of", "class", "may", "degrade", "for", "datasets", "with", "more", "classes", ",", "since", "the", "risk", "of", "inadvertently", "increasing", "the", "activation", "of", "a", "different", "class", "increases", "in", "that", "case", ".", "We", "found", "that", "we", "were", "able", "to", "train", "a", "maxout", "network", "to", "have", "a", "zero", "percent", "error", "rate", "on", "Gaussian", "rubbish", "examples", "(", "it", "was", "still", "vulnerable", "to", "rubbish", "examples", "generated", "by", "applying", "a", "fast", "gradient", "sign", "step", "to", "a", "Gaussian", "sample", ")", "with", "no", "negative", "impact", "on", "its", "ability", "to", "classify", "clean", "examples", ".", "Unfortunately", ",", "unlike", "training", "on", "adversarial", "examples", ",", "this", "did", "not", "result", "in", "any", "significant", "reduction", "of", "the", "model", "’s", "test", "set", "error", "rate", ".", "In", "conclusion", ",", "it", "appears", "that", "a", "randomly", "selected", "input", "to", "deep", "or", "shallow", "models", "built", "from", "linear", "parts", "is", "overwhelmingly", "likely", "to", "be", "processed", "incorrectly", ",", "and", "that", "these", "models", "only", "behave", "reasonably", "on", "a", "very", "thin", "manifold", "encompassing", "the", "training", "data", "."], "entity": [[8, 11, ["machine", "learning", "models"], "Method"], [13, 15, ["neural", "networks"], "Method"], [74, 76, ["neural", "networks"], "Method"], [79, 81, ["adversarial", "perturbation"], "Task"], [110, 111, ["architectures"], "Method"], [126, 129, ["generating", "adversarial", "examples"], "Task"], [137, 139, ["adversarial", "training"], "Method"], [143, 146, ["test", "set", "error"], "Metric"], [148, 150, ["maxout", "network"], "Method"], [152, 154, ["MNIST", "dataset"], "Material"], [171, 174, ["machine", "learning", "models"], "Method"], [183, 185, ["neural", "networks"], "Method"], [196, 199, ["machine", "learning", "models"], "Method"], [253, 255, ["training", "algorithms"], "Method"], [278, 281, ["deep", "neural", "networks"], "Method"], [286, 288, ["model", "averaging"], "Method"], [290, 291, ["regularization"], "Method"], [294, 297, ["supervised", "learning", "problem"], "Task"], [331, 334, ["generating", "adversarial", "examples"], "Task"], [336, 338, ["adversarial", "training"], "Method"], [343, 345, ["adversarial", "training"], "Method"], [356, 358, ["dropout", "dropout"], "Method"], [360, 363, ["Generic", "regularization", "strategies"], "Method"], [365, 366, ["dropout"], "Method"], [367, 368, ["pretraining"], "Method"], [370, 372, ["model", "averaging"], "Method"], [390, 393, ["nonlinear", "model", "families"], "Method"], [395, 397, ["RBF", "networks"], "Method"], [448, 450, ["optimization", "methods"], "Method"], [455, 457, ["nonlinear", "models"], "Method"], [464, 465, ["ICLR2014"], "Method"], [472, 474, ["neural", "networks"], "Method"], [486, 492, ["Box", "-", "constrained", "L", "-", "BFGS"], "Method"], [504, 506, ["ImageNet", "ImageNet"], "Material"], [538, 539, ["classifiers"], "Method"], [541, 542, ["architectures"], "Method"], [552, 556, ["Shallow", "softmax", "regression", "models"], "Method"], [587, 589, ["constrained", "optimization"], "Task"], [598, 599, ["classifiers"], "Method"], [602, 605, ["machine", "learning", "techniques"], "Method"], [638, 640, ["Potemkin", "village"], "Method"], [680, 682, ["computer", "vision"], "Task"], [685, 688, ["convolutional", "network", "features"], "Method"], [719, 722, ["network", "’s", "representation"], "Method"], [734, 736, ["deep", "networks"], "Task"], [741, 743, ["linear", "classifiers"], "Method"], [764, 765, ["Luca"], "Method"], [795, 796, ["accuracy"], "Metric"], [803, 808, ["linear", "explanation", "of", "adversarial", "examples"], "Method"], [818, 820, ["linear", "models"], "Method"], [873, 874, ["classifier"], "Method"], [913, 914, ["classifier"], "Method"], [1031, 1034, ["high", "dimensional", "problems"], "Task"], [1065, 1067, ["accidental", "steganography"], "Task"], [1071, 1073, ["linear", "model"], "Method"], [1109, 1111, ["linear", "model"], "Method"], [1125, 1127, ["adversarial", "examples"], "Task"], [1131, 1133, ["neural", "networks"], "Method"], [1157, 1159, ["softmax", "regression"], "Method"], [1167, 1174, ["Linear", "perturbation", "of", "non", "-", "linear", "models"], "Task"], [1175, 1180, ["linear", "view", "of", "adversarial", "examples"], "Method"], [1191, 1193, ["neural", "networks"], "Method"], [1198, 1201, ["linear", "adversarial", "perturbation"], "Method"], [1202, 1204, ["LSTMs", "lstm"], "Method"], [1219, 1221, ["maxout", "networks"], "Method"], [1227, 1228, ["ICML2013"], "Method"], [1248, 1250, ["nonlinear", "models"], "Method"], [1252, 1254, ["sigmoid", "networks"], "Method"], [1284, 1286, ["analytical", "perturbations"], "Method"], [1288, 1290, ["linear", "model"], "Method"], [1293, 1295, ["neural", "networks"], "Method"], [1316, 1319, ["machine", "learning", "tasks"], "Task"], [1331, 1333, ["neural", "network"], "Method"], [1349, 1354, ["max", "-", "norm", "constrained", "pertubation"], "Method"], [1362, 1366, ["fast", "gradient", "sign", "method"], "Method"], [1368, 1371, ["generating", "adversarial", "examples"], "Task"], [1382, 1383, ["backpropagation"], "Method"], [1411, 1412, ["ImageNet"], "Material"], [1421, 1424, ["shallow", "softmax", "classifier"], "Method"], [1434, 1436, ["average", "confidence"], "Metric"], [1454, 1456, ["maxout", "network"], "Method"], [1465, 1467, ["average", "confidence"], "Metric"], [1485, 1487, ["average", "probability"], "Metric"], [1498, 1501, ["convolutional", "maxout", "network"], "Method"], [1518, 1521, ["generating", "adversarial", "examples"], "Task"], [1553, 1555, ["cheap", "algorithms"], "Method"], [1589, 1591, ["adversarial", "training"], "Method"], [1594, 1598, ["analysis", "of", "trained", "networks"], "Task"], [1601, 1606, ["Adversarial", "training", "of", "linear", "models"], "Method"], [1607, 1609, ["weight", "decay"], "Method"], [1618, 1620, ["logistic", "regression"], "Method"], [1626, 1630, ["fast", "gradient", "sign", "method"], "Method"], [1684, 1686, ["gradient", "descent"], "Method"], [1698, 1700, ["analytical", "form"], "Method"], [1701, 1702, ["training"], "Task"], [1716, 1719, ["gradient", "sign", "perturbation"], "Method"], [1734, 1736, ["adversarial", "version"], "Method"], [1737, 1739, ["logistic", "regression"], "Method"], [1748, 1749, ["regularization"], "Method"], [1778, 1780, ["training", "cost"], "Metric"], [1812, 1814, ["underfitting", "regime"], "Task"], [1815, 1817, ["adversarial", "training"], "Method"], [1826, 1828, ["weight", "decay"], "Method"], [1836, 1838, ["adversarial", "training"], "Method"], [1855, 1857, ["logistic", "regression"], "Method"], [1858, 1861, ["multiclass", "softmax", "regression"], "Method"], [1862, 1864, ["weight", "decay"], "Method"], [1905, 1907, ["Weight", "decay"], "Method"], [1920, 1922, ["deep", "network"], "Method"], [1928, 1930, ["weight", "decay"], "Method"], [1955, 1956, ["precision"], "Metric"], [1962, 1964, ["maxout", "networks"], "Method"], [1965, 1966, ["MNIST"], "Material"], [1972, 1974, ["adversarial", "training"], "Method"], [1978, 1980, ["weight", "decay"], "Method"], [2008, 2009, ["error"], "Metric"], [2029, 2034, ["Adversarial", "training", "of", "deep", "networks"], "Task"], [2037, 2039, ["deep", "networks"], "Method"], [2050, 2053, ["shallow", "linear", "models"], "Method"], [2054, 2056, ["deep", "networks"], "Method"], [2069, 2073, ["universal", "approximator", "theorem", "Hornik89"], "Method"], [2076, 2078, ["neural", "network"], "Method"], [2093, 2094, ["accuracy"], "Metric"], [2107, 2110, ["Shallow", "linear", "models"], "Method"], [2133, 2136, ["universal", "approximator", "theorem"], "Method"], [2143, 2145, ["training", "algorithm"], "Method"], [2162, 2164, ["supervised", "training"], "Method"], [2183, 2185, ["training", "procedure"], "Method"], [2189, 2190, ["ICLR2014"], "Method"], [2204, 2206, ["neural", "network"], "Method"], [2211, 2212, ["Training"], "Task"], [2220, 2223, ["data", "augmentation", "schemes"], "Method"], [2249, 2251, ["data", "augmentation"], "Task"], [2287, 2288, ["dropout"], "Method"], [2314, 2317, ["L", "-", "BFGS"], "Method"], [2330, 2334, ["fast", "gradient", "sign", "method"], "Method"], [2337, 2338, ["regularizer"], "Method"], [2404, 2406, ["maxout", "network"], "Method"], [2411, 2412, ["dropout"], "Method"], [2425, 2427, ["adversarial", "training"], "Method"], [2431, 2433, ["adversarial", "training"], "Method"], [2482, 2484, ["maxout", "network"], "Method"], [2489, 2491, ["adversarial", "training"], "Method"], [2514, 2516, ["adversarial", "training"], "Method"], [2537, 2538, ["maxout"], "Method"], [2545, 2546, ["learning"], "Task"], [2572, 2576, ["adversarial", "validation", "set", "error"], "Metric"], [2582, 2584, ["early", "stopping"], "Method"], [2586, 2590, ["adversarial", "validation", "set", "error"], "Metric"], [2623, 2626, ["random", "number", "generators"], "Method"], [2683, 2686, ["permutation", "invariant", "version"], "Method"], [2687, 2688, ["MNIST"], "Material"], [2697, 2701, ["fine", "-", "tuning", "DBMs"], "Method"], [2702, 2704, ["dropout", "dropout"], "Method"], [2721, 2723, ["adversarial", "training"], "Method"], [2742, 2746, ["fast", "gradient", "sign", "method"], "Method"], [2748, 2750, ["adversarial", "training"], "Method"], [2770, 2773, ["adversarially", "trained", "model"], "Method"], [2793, 2796, ["adversarially", "trained", "model"], "Method"], [2819, 2822, ["adversarially", "trained", "model"], "Method"], [2865, 2868, ["adversarially", "trained", "model"], "Method"], [2914, 2916, ["adversarial", "game"], "Method"], [2925, 2927, ["expected", "cost"], "Metric"], [2938, 2940, ["Adversarial", "training"], "Method"], [2948, 2950, ["active", "learning"], "Method"], [2967, 2969, ["human", "labeler"], "Task"], [2973, 2975, ["heuristic", "labeler"], "Method"], [3028, 3030, ["max", "norm"], "Method"], [3101, 3104, ["objective", "function", "value"], "Metric"], [3109, 3111, ["adversarial", "training"], "Method"], [3113, 3116, ["hard", "example", "mining"], "Task"], [3138, 3139, ["classification"], "Task"], [3148, 3150, ["maxout", "network"], "Method"], [3177, 3178, ["confidence"], "Metric"], [3189, 3190, ["confidence"], "Metric"], [3195, 3200, ["fast", "gradient", "sign", "adversarial", "examples"], "Task"], [3214, 3216, ["gradient", "descent"], "Method"], [3224, 3228, ["fast", "gradient", "sign", "method"], "Method"], [3264, 3266, ["perturbation", "process"], "Method"], [3271, 3272, ["learning"], "Method"], [3340, 3341, ["ICLR2014"], "Method"], [3362, 3364, ["sigmoidal", "network"], "Method"], [3370, 3374, ["fast", "gradient", "sign", "method"], "Method"], [3410, 3412, ["saturating", "models"], "Method"], [3415, 3417, ["Rust", "model"], "Method"], [3461, 3463, ["maxout", "networks"], "Method"], [3489, 3491, ["input", "layer"], "Method"], [3495, 3497, ["adversarial", "training"], "Method"], [3525, 3528, ["universal", "approximator", "theorem"], "Method"], [3536, 3538, ["neural", "network"], "Method"], [3540, 3548, ["linear", "-", "sigmoid", "or", "linear", "-", "softmax", "layer"], "Method"], [3552, 3556, ["universal", "approximator", "of", "functions"], "Method"], [3713, 3716, ["shallow", "RBF", "networks"], "Method"], [3752, 3754, ["RBF", "networks"], "Method"], [3775, 3778, ["shallow", "RBF", "network"], "Method"], [3790, 3791, ["MNIST"], "Material"], [3797, 3801, ["fast", "gradient", "sign", "method"], "Method"], [3815, 3816, ["confidence"], "Metric"], [3863, 3865, ["RBF", "units"], "Method"], [3884, 3886, ["linear", "units"], "Method"], [3887, 3889, ["RBF", "units"], "Method"], [3894, 3899, ["precision", "-", "recall", "tradeoff", "curve"], "Metric"], [3900, 3902, ["Linear", "units"], "Method"], [3904, 3905, ["recall"], "Metric"], [3919, 3920, ["precision"], "Metric"], [3929, 3931, ["RBF", "units"], "Method"], [3933, 3934, ["precision"], "Metric"], [3949, 3950, ["recall"], "Metric"], [3965, 3967, ["quadratic", "units"], "Method"], [3969, 3972, ["deep", "RBF", "networks"], "Method"], [4000, 4001, ["SGD"], "Method"], [4074, 4075, ["overfitting"], "Method"], [4087, 4091, ["non", "-", "linear", "model"], "Method"], [4216, 4220, ["fast", "gradient", "sign", "method"], "Method"], [4240, 4241, ["classifier"], "Method"], [4252, 4253, ["classifier"], "Method"], [4257, 4259, ["mutiple", "classifiers"], "Method"], [4270, 4272, ["neural", "networks"], "Method"], [4279, 4281, ["linear", "classifier"], "Method"], [4289, 4291, ["reference", "classifier"], "Method"], [4312, 4315, ["machine", "learning", "algorithms"], "Method"], [4348, 4351, ["deep", "maxout", "network"], "Method"], [4357, 4360, ["shallow", "softmax", "network"], "Method"], [4362, 4365, ["shallow", "RBF", "network"], "Method"], [4373, 4375, ["maxout", "network"], "Method"], [4377, 4379, ["RBF", "network"], "Method"], [4381, 4383, ["maxout", "network"], "Method"], [4384, 4386, ["class", "assignment"], "Method"], [4395, 4397, ["softmax", "classifier"], "Method"], [4399, 4401, ["maxout", "network"], "Method"], [4443, 4445, ["softmax", "regression"], "Method"], [4458, 4460, ["RBF", "network"], "Method"], [4478, 4480, ["RBF", "network"], "Method"], [4482, 4484, ["softmax", "regression"], "Method"], [4498, 4500, ["linear", "component"], "Method"], [4513, 4515, ["maxout", "network"], "Method"], [4544, 4548, ["cross", "-", "model", "generalization"], "Method"], [4563, 4567, ["existence", "of", "adversarial", "examples"], "Task"], [4574, 4576, ["generative", "training"], "Method"], [4613, 4617, ["MP", "-", "DBM", "mpdbm"], "Method"], [4627, 4629, ["inference", "procedure"], "Method"], [4641, 4642, ["MNIST"], "Material"], [4644, 4646, ["inference", "procedure"], "Method"], [4650, 4652, ["generative", "models"], "Method"], [4656, 4659, ["differentiable", "inference", "procedures"], "Method"], [4672, 4677, ["non", "-", "generative", "discriminator", "model"], "Method"], [4683, 4684, ["MNIST"], "Material"], [4690, 4693, ["MP", "-", "DBM"], "Method"], [4700, 4702, ["generative", "model"], "Method"], [4712, 4717, ["non", "-", "generative", "classifier", "model"], "Method"], [4762, 4764, ["generative", "training"], "Method"], [4816, 4821, ["ensemble", "of", "twelve", "maxout", "networks"], "Method"], [4822, 4823, ["MNIST"], "Material"], [4834, 4837, ["random", "number", "generator"], "Method"], [4853, 4856, ["stochastic", "gradient", "descent"], "Method"], [4901, 4902, ["Ensembling"], "Method"], [5060, 5063, ["generating", "adversarial", "examples"], "Task"], [5068, 5070, ["adversarial", "training"], "Method"], [5073, 5074, ["regularization"], "Task"], [5077, 5078, ["regularization"], "Task"], [5079, 5080, ["dropout"], "Method"], [5097, 5098, ["regularizers"], "Method"], [5099, 5101, ["weight", "decay"], "Method"], [5116, 5118, ["Linear", "models"], "Method"], [5135, 5138, ["universal", "approximator", "theorem"], "Method"], [5148, 5150, ["RBF", "networks"], "Method"], [5170, 5171, ["Ensembles"], "Method"], [5200, 5203, ["Shallow", "linear", "models"], "Method"], [5211, 5213, ["RBF", "networks"], "Method"], [5220, 5224, ["Gradient", "-", "based", "optimization"], "Method"], [5229, 5230, ["AI"], "Task"], [5247, 5251, ["ReLU", "or", "maxout", "network"], "Method"], [5253, 5254, ["LSTM"], "Method"], [5257, 5259, ["sigmoid", "network"], "Method"], [5400, 5401, ["optimization"], "Task"], [5418, 5420, ["optimization", "procedures"], "Method"], [5495, 5497, ["Pylearn2", "pylearn2_arxiv_2013"], "Method"], [5499, 5501, ["DistBelief", "distbelief"], "Method"], [5605, 5607, ["binary", "classifiers"], "Method"], [5644, 5645, ["classifier"], "Method"], [5696, 5698, ["rubbish", "class"], "Method"], [5702, 5704, ["computer", "vision"], "Task"], [5739, 5741, ["deep", "networks"], "Method"], [5770, 5772, ["Linear", "models"], "Method"], [5829, 5831, ["RBF", "networks"], "Method"], [5869, 5870, ["classifiers"], "Method"], [5872, 5874, ["MNIST", "dataset"], "Material"], [5893, 5894, ["error"], "Metric"], [5898, 5900, ["maxout", "network"], "Method"], [5902, 5904, ["softmax", "layer"], "Method"], [5919, 5921, ["average", "confidence"], "Metric"], [5933, 5934, ["sigmoids"], "Method"], [5944, 5945, ["confidence"], "Metric"], [5946, 5947, ["mistakes"], "Metric"], [5952, 5955, ["CIFAR", "-", "10"], "Material"], [5964, 5967, ["convolutional", "maxout", "net"], "Method"], [5977, 5979, ["average", "confidence"], "Metric"], [5988, 5990, ["optimization", "algorithms"], "Method"], [5992, 5993, ["fool"], "Method"], [6001, 6002, ["ImageNet"], "Material"], [6022, 6024, ["search", "procedures"], "Method"], [6043, 6045, ["deep", "networks"], "Method"], [6046, 6049, ["shallow", "linear", "models"], "Method"], [6055, 6058, ["softmax", "regression", "model"], "Method"], [6072, 6074, ["average", "confidence"], "Metric"], [6085, 6087, ["RBF", "network"], "Method"], [6115, 6117, ["average", "confidence"], "Metric"], [6123, 6124, ["fool"], "Method"], [6129, 6132, ["generating", "fooling", "images"], "Task"], [6165, 6166, ["MNIST"], "Material"], [6167, 6170, ["CIFAR", "-", "10"], "Material"], [6181, 6182, ["MNIST"], "Material"], [6189, 6191, ["maxout", "network"], "Method"], [6192, 6194, ["false", "positives"], "Metric"], [6209, 6212, ["CIFAR", "-", "10"], "Material"], [6217, 6219, ["convolutional", "network"], "Method"], [6220, 6222, ["false", "positives"], "Metric"], [6253, 6255, ["fooling", "image"], "Task"], [6265, 6267, ["Gaussian", "sample"], "Method"], [6274, 6276, ["fooling", "image"], "Task"], [6292, 6294, ["randomized", "algorithm"], "Method"], [6299, 6302, ["CIFAR", "-", "10"], "Material"], [6313, 6315, ["success", "rate"], "Metric"], [6329, 6331, ["success", "rate"], "Metric"], [6349, 6354, ["per", "-", "step", "success", "rate"], "Metric"], [6405, 6408, ["gradient", "sign", "method"], "Method"], [6417, 6419, ["success", "rate"], "Metric"], [6425, 6429, ["generating", "members", "of", "class"], "Task"], [6463, 6465, ["maxout", "network"], "Method"], [6545, 6549, ["deep", "or", "shallow", "models"], "Method"], [6551, 6553, ["linear", "parts"], "Method"], [1214, 1215, ["AI"], "Task"], [1427, 1428, ["error"], "Metric"], [1441, 1442, ["MNIST"], "Material"], [1478, 1479, ["error"], "Metric"], [1507, 1510, ["CIFAR", "-", "10"], "Material"], [2419, 2420, ["error"], "Metric"], [2442, 2443, ["error"], "Metric"], [2503, 2504, ["error"], "Metric"], [2523, 2524, ["error"], "Metric"], [2550, 2551, ["error"], "Metric"], [2566, 2567, ["error"], "Metric"], [2650, 2651, ["error"], "Metric"], [2665, 2666, ["error"], "Metric"], [2731, 2732, ["error"], "Metric"], [2752, 2753, ["error"], "Metric"], [2786, 2787, ["error"], "Metric"], [2807, 2808, ["error"], "Metric"], [2884, 2886, ["adversarial", "training"], "Method"], [2895, 2896, ["error"], "Metric"], [3171, 3172, ["error"], "Metric"], [3182, 3183, ["error"], "Metric"], [3784, 3785, ["error"], "Metric"], [3996, 3997, ["error"], "Metric"], [4298, 4299, ["classification"], "Task"], [4325, 4326, ["classification"], "Task"], [4418, 4419, ["error"], "Metric"], [4631, 4632, ["classification"], "Task"], [4637, 4638, ["error"], "Metric"], [4680, 4681, ["classification"], "Task"], [4739, 4740, ["error"], "Metric"], [4750, 4751, ["MNIST"], "Material"], [4861, 4862, ["error"], "Metric"], [4894, 4895, ["error"], "Metric"], [5908, 5909, ["error"], "Metric"], [5936, 5937, ["error"], "Metric"], [5969, 5970, ["error"], "Metric"], [6060, 6061, ["error"], "Metric"], [6100, 6101, ["error"], "Metric"], [6110, 6111, ["error"], "Metric"], [6470, 6471, ["error"], "Metric"], [6529, 6532, ["test", "set", "error"], "Metric"]]} {"doc_id": "364c1a3df58d87cb40ab33fdf3831cf2862f3570", "text": ["document", ":", "aNMM", ":", "Ranking", "Short", "Answer", "Texts", "with", "Attention", "-", "Based", "Neural", "Matching", "Model", "As", "an", "alternative", "to", "question", "answering", "methods", "based", "on", "feature", "engineering", ",", "deep", "learning", "approaches", "such", "as", "convolutional", "neural", "networks", "(", "CNNs", ")", "and", "Long", "Short", "-", "Term", "Memory", "Models", "(", "LSTMs", ")", "have", "recently", "been", "proposed", "for", "semantic", "matching", "of", "questions", "and", "answers", ".", "To", "achieve", "good", "results", ",", "however", ",", "these", "models", "have", "been", "combined", "with", "additional", "features", "such", "as", "word", "overlap", "or", "BM25", "scores", ".", "Without", "this", "combination", ",", "these", "models", "perform", "significantly", "worse", "than", "methods", "based", "on", "linguistic", "feature", "engineering", ".", "In", "this", "paper", ",", "we", "propose", "an", "attention", "based", "neural", "matching", "model", "for", "ranking", "short", "answer", "text", ".", "We", "adopt", "value", "-", "shared", "weighting", "scheme", "instead", "of", "position", "-", "shared", "weighting", "scheme", "for", "combining", "different", "matching", "signals", "and", "incorporate", "question", "term", "importance", "learning", "using", "question", "attention", "network", ".", "Using", "the", "popular", "benchmark", "TREC", "QA", "data", ",", "we", "show", "that", "the", "relatively", "simple", "aNMM", "model", "can", "significantly", "outperform", "other", "neural", "network", "models", "that", "have", "been", "used", "for", "the", "question", "answering", "task", ",", "and", "is", "competitive", "with", "models", "that", "are", "combined", "with", "additional", "features", ".", "When", "aNMM", "is", "combined", "with", "additional", "features", ",", "it", "outperforms", "all", "baselines", ".", "2016", ",", "October", "24", "-", "28", ",", "2016", ",", "Indianapolis", ",", "IN", ",", "USA", "¡", "ccs2012", "¿", "¡", "concept", "¿", "¡", "concept_id¿10002951.10003317.10003338¡", "/", "concept_id", "¿", "¡", "concept_desc¿Information", "systems", "Retrieval", "models", "and", "ranking¡", "/", "concept_desc", "¿", "¡", "concept_significance¿500¡", "/", "concept_significance", "¿", "¡", "/", "concept", "¿", "¡", "concept", "¿", "¡", "concept_id¿10002951.10003317.10003347.10003348¡", "/", "concept_id", "¿", "¡", "concept_desc¿Information", "systems", "Question", "answering¡", "/", "concept_desc", "¿", "¡", "concept_significance¿500¡", "/", "concept_significance", "¿", "¡", "/", "concept", "¿", "¡", "/", "ccs2012", "¿", "[", "500", "]", "Information", "systems", "Retrieval", "models", "and", "ranking", "systems", "Question", "answering", "section", ":", "Introduction", "Question", "answering", "(", "QA", ")", ",", "which", "returns", "exact", "answers", "as", "either", "short", "facts", "or", "long", "passages", "to", "natural", "language", "questions", "issued", "by", "users", ",", "is", "a", "challenging", "task", "and", "plays", "a", "central", "role", "in", "the", "next", "generation", "of", "advanced", "web", "search", ".", "Many", "of", "current", "QA", "systems", "use", "a", "learning", "to", "rank", "approach", "that", "encodes", "question", "/", "answer", "pairs", "with", "complex", "linguistic", "features", "including", "lexical", ",", "syntactic", "and", "semantic", "features", ".", "For", "instance", ",", "Surdeanu", "et", "al", ".", "investigated", "a", "wide", "range", "of", "feature", "types", "including", "similarity", "features", ",", "translation", "features", ",", "density", "/", "frequency", "features", "and", "web", "correlation", "features", "for", "learning", "to", "rank", "answers", "and", "show", "improvements", "in", "accuracy", ".", "However", ",", "such", "methods", "rely", "on", "manual", "feature", "engineering", ",", "which", "is", "often", "time", "-", "consuming", "and", "requires", "domain", "dependent", "expertise", "and", "experience", ".", "Moreover", ",", "they", "may", "need", "additional", "NLP", "parsers", "or", "external", "knowledge", "sources", "that", "may", "not", "be", "available", "for", "some", "languages", ".", "Recently", ",", "researchers", "have", "been", "studying", "deep", "learning", "approaches", "to", "automatically", "learn", "semantic", "match", "between", "questions", "and", "answers", ".", "Such", "methods", "are", "built", "on", "the", "top", "of", "neural", "network", "models", "such", "as", "convolutional", "neural", "networks", "(", "CNNs", ")", "and", "Long", "Short", "-", "Term", "Memory", "Models", "(", "LSTMs", ")", ".", "The", "proposed", "models", "have", "the", "benefit", "of", "not", "requiring", "hand", "-", "crafted", "linguistic", "features", "and", "external", "resources", ".", "Some", "of", "them", "achieve", "state", "-", "of", "-", "the", "art", "performance", "for", "the", "answer", "sentence", "selection", "task", "benchmarked", "by", "the", "TREC", "QA", "track", ".", "However", ",", "the", "weakness", "of", "the", "existing", "studies", "is", "that", "the", "proposed", "deep", "models", ",", "either", "based", "on", "CNNs", "or", "LSTMs", ",", "need", "to", "be", "combined", "with", "additional", "features", "such", "as", "word", "overlap", "features", "and", "BM25", "to", "perform", "well", ".", "Without", "combining", "these", "additional", "features", ",", "their", "performance", "is", "significantly", "worse", "than", "the", "results", "obtained", "by", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "based", "on", "linguistic", "feature", "engineering", ".", "This", "led", "us", "to", "propose", "the", "following", "research", "questions", ":", "RQ1", "Without", "combining", "additional", "features", ",", "could", "we", "build", "deep", "learning", "models", "that", "can", "achieve", "comparable", "or", "even", "better", "performance", "than", "methods", "using", "feature", "engineering", "?", "RQ2", "By", "combining", "additional", "features", ",", "could", "our", "model", "outperform", "state", "-", "of", "-", "the", "-", "art", "models", "for", "question", "answering", "?", "To", "address", "these", "research", "questions", ",", "we", "analyze", "the", "existing", "current", "deep", "learning", "architectures", "for", "answer", "ranking", "and", "make", "the", "following", "two", "key", "observations", ":", "Architectures", "not", "specifically", "designed", "for", "question", "/", "answer", "matching", ":", "Some", "methods", "employ", "CNNs", "for", "question", "/", "answer", "matching", ".", "However", ",", "CNNs", "are", "originally", "designed", "for", "computer", "vision", "(", "CV", ")", ",", "which", "uses", "position", "-", "shared", "weights", "with", "local", "perceptive", "filters", ",", "to", "learn", "spatial", "regularities", "in", "many", "CV", "tasks", ".", "However", ",", "such", "spatial", "regularities", "may", "not", "exist", "in", "semantic", "matching", "between", "questions", "and", "answers", ",", "since", "important", "similarity", "signals", "between", "question", "and", "answer", "terms", "could", "appear", "in", "any", "position", "due", "to", "the", "complex", "linguistic", "property", "of", "natural", "languages", ".", "Meanwhile", ",", "models", "based", "on", "LSTMs", "view", "the", "question", "/", "answer", "matching", "problem", "in", "a", "sequential", "way", ".", "Without", "direct", "interactions", "between", "question", "and", "answer", "terms", ",", "the", "model", "may", "not", "be", "able", "to", "capture", "sufficiently", "detailed", "matching", "signals", "between", "them", ".", "Lack", "of", "modeling", "question", "focus", ":", "Understanding", "the", "focus", "of", "questions", ",", "e.g.", ",", "important", "terms", "in", "a", "question", ",", "is", "helpful", "for", "ranking", "the", "answers", "correctly", ".", "For", "example", ",", "given", "a", "question", "like", "“", "Where", "was", "the", "first", "burger", "king", "restaurant", "opened", "?", "”", ",", "it", "is", "critical", "for", "the", "answer", "to", "talk", "about", "“", "burger", "”", ",", "“", "king", "”", ",", "“", "open", "”", ",", "etc", ".", "Most", "existing", "text", "matching", "models", "do", "not", "explicitly", "model", "question", "focus", ".", "For", "example", ",", "models", "based", "on", "CNNs", "treat", "all", "the", "question", "terms", "as", "equally", "important", "when", "matching", "to", "answer", "terms", ".", "Models", "based", "on", "LSTMs", "usually", "model", "question", "terms", "closer", "to", "the", "end", "to", "be", "more", "important", ".", "To", "handle", "these", "issues", "in", "the", "existing", "deep", "learning", "architectures", "for", "ranking", "answers", ",", "we", "propose", "an", "attention", "based", "neural", "matching", "model", "(", "aNMM", ")", ".", "The", "novel", "properties", "of", "the", "proposed", "model", "and", "our", "contributions", "can", "be", "summarized", "as", "follows", ":", "Deep", "neural", "network", "with", "value", "-", "shared", "weights", ":", "We", "introduce", "a", "novel", "value", "-", "shared", "weighting", "scheme", "in", "deep", "neural", "networks", "as", "a", "counterpart", "of", "the", "position", "-", "shared", "weighting", "scheme", "in", "CNNs", ",", "based", "on", "the", "idea", "that", "semantic", "matching", "between", "a", "question", "and", "answer", "is", "mainly", "about", "the", "(", "semantic", "similarity", ")", "value", "regularities", "rather", "than", "spatial", "regularities", ".", "Incorporate", "attention", "scheme", "over", "question", "terms", ":", "We", "incorporate", "the", "attention", "scheme", "over", "the", "question", "terms", "using", "a", "gating", "function", ",", "so", "that", "we", "can", "explicitly", "discriminate", "the", "question", "term", "importance", ".", "Extensive", "experimental", "evaluation", "and", "promising", "results", ".", "We", "perform", "a", "thorough", "experimental", "study", "based", "on", "the", "TREC", "QA", "dataset", "from", "TREC", "QA", "tracks", "8", "-", "13", ",", "which", "appears", "to", "be", "one", "of", "the", "most", "widely", "used", "benchmarks", "for", "answer", "re", "-", "ranking", ".", "Unlike", "previous", "methods", "using", "CNNs", "and", "LSTMs", ",", "which", "showed", "inferior", "results", "without", "combining", "additional", "features", ",", "our", "model", "can", "achieve", "better", "performance", "than", "a", "state", "-", "of", "-", "art", "method", "using", "linguistic", "feature", "engineering", "and", "comparable", "performance", "with", "previous", "deep", "learning", "models", "with", "combined", "additional", "features", ".", "If", "we", "combine", "our", "model", "with", "a", "simple", "additional", "feature", "like", "QL", ",", "our", "method", "can", "achieve", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "among", "current", "existing", "methods", "for", "ranking", "answers", "under", "multiple", "metrics", ".", "Roadmap", ".", "The", "rest", "of", "our", "paper", "is", "organized", "as", "follows", ".", "We", "will", "review", "related", "work", "in", "Section", "[", "reference", "]", ".", "In", "Section", "[", "reference", "]", ",", "we", "will", "present", "the", "proposed", "aNMM", "model", "with", "two", "components", ":", "value", "-", "shared", "weights", "and", "question", "attention", "network", "with", "gating", "functions", ".", "Two", "different", "architectures", "will", "be", "presented", "and", "analyzed", ".", "Section", "[", "reference", "]", "is", "a", "systematic", "experimental", "analysis", "using", "the", "TREC", "QA", "benchmark", "dataset", ".", "Finally", ",", "we", "conclude", "our", "paper", "and", "discuss", "future", "work", "in", "Section", "[", "reference", "]", ".", "section", ":", "Related", "Work", "Our", "work", "is", "related", "to", "several", "research", "areas", ",", "including", "deep", "learning", "models", "for", "text", "matching", ",", "factoid", "question", "answering", ",", "answer", "ranking", "in", "CQA", "and", "answer", "passage", "/", "sentence", "retrieval", ".", "Deep", "Learning", "Models", "for", "Text", "Matching", ".", "Recently", "there", "have", "been", "many", "deep", "learning", "models", "proposed", "for", "text", "matching", "and", "ranking", ".", "Such", "deep", "learning", "models", "include", "DSSM", ",", "CDSSM", ",", "ARC", "-", "I", "/", "ARC", "-", "II", ",", "DCNN", ",", "DeepMatch", ",", "MultiGranCNN", "and", "MatchPyramid", ".", "DSSM", "performs", "a", "non", "-", "linear", "projection", "to", "map", "the", "query", "and", "the", "documents", "to", "a", "common", "semantic", "space", ".", "The", "neural", "network", "models", "are", "trained", "using", "clickthrough", "data", "such", "that", "the", "conditional", "likelihood", "of", "the", "clicked", "document", "given", "the", "query", "is", "maximized", ".", "DeepMatch", "uses", "a", "topic", "model", "to", "construct", "the", "interactions", "between", "two", "texts", "and", "then", "makes", "different", "levels", "of", "abstractions", "with", "a", "deep", "architecture", "to", "model", "the", "relationships", "between", "topics", ".", "ARC", "-", "I", "and", "ARC", "-", "II", "are", "two", "different", "architectures", "proposed", "by", "Hu", "et", ".", "al", ".", "for", "matching", "natural", "language", "sentences", ".", "ARC", "-", "I", "firstly", "finds", "the", "representation", "of", "each", "sentence", "and", "then", "compares", "the", "representations", "of", "the", "two", "sentences", "with", "a", "multi", "-", "layer", "perceptron", "(", "MLP", ")", ".", "The", "drawback", "of", "ARC", "-", "I", "is", "that", "it", "defers", "the", "interaction", "between", "two", "sentences", "until", "their", "individual", "representation", "matures", "in", "the", "convolution", "model", ",", "and", "therefore", "has", "the", "risk", "of", "losing", "details", ",", "which", "could", "be", "important", "for", "the", "matching", "task", ".", "On", "the", "other", "hand", ",", "ARC", "-", "II", "is", "built", "directly", "on", "the", "interaction", "space", "between", "two", "sentences", ".", "Thus", "ARC", "-", "II", "makes", "two", "sentences", "meet", "before", "their", "own", "high", "-", "level", "representations", "mature", ",", "while", "still", "retaining", "the", "space", "for", "individual", "development", "of", "abstraction", "of", "each", "sentence", ".", "Our", "aNMM", "architecture", "adopts", "a", "similar", "design", "with", "ARC", "-", "II", "in", "the", "QA", "matching", "matrix", "where", "we", "build", "neural", "networks", "directly", "on", "the", "interaction", "of", "sentence", "term", "pairs", ".", "However", ",", "we", "adopt", "value", "-", "shared", "weights", "instead", "of", "position", "-", "shared", "weights", "as", "in", "the", "CNN", "used", "by", "ARC", "-", "II", ".", "We", "also", "add", "attention", "scheme", "to", "learn", "question", "term", "importance", ".", "Factoid", "Question", "Answering", ".", "There", "have", "been", "many", "previous", "studies", "on", "factoid", "question", "answering", ",", "most", "of", "which", "use", "the", "benchmark", "data", "from", "TREC", "QA", "track", ".", "Yih", "et", ".", "al", ".", "formulated", "answer", "sentence", "selection", "as", "a", "semantic", "matching", "problem", "with", "a", "latent", "word", "-", "alignment", "structure", "and", "conducted", "a", "series", "of", "experimental", "studies", "on", "leveraging", "proposed", "lexical", "semantic", "models", ".", "Iyyer", "et", ".", "al", ".", "introduced", "a", "recursive", "neural", "network", "(", "RNN", ")", "model", "that", "can", "reason", "over", "text", "that", "contains", "very", "few", "individual", "words", "by", "modeling", "textual", "compositionality", ".", "Yu", "et", "al", ".", "proposed", "an", "approach", "for", "answer", "sentence", "selection", "via", "distributed", "representations", ",", "and", "learned", "to", "match", "questions", "with", "answers", "by", "considering", "their", "semantic", "encoding", ".", "They", "combined", "the", "learning", "results", "of", "their", "model", "with", "word", "overlap", "features", "by", "training", "a", "logistic", "regression", "classifier", ".", "Wang", "and", "Nyberg", "proposed", "a", "method", "which", "uses", "a", "stacked", "bidirectional", "Long", "-", "Short", "Term", "Memory", "(", "BLSTM", ")", "network", "to", "sequentially", "read", "words", "from", "question", "and", "answer", "sentences", ",", "and", "then", "output", "their", "relevance", "scores", ".", "Their", "system", "needs", "to", "combine", "the", "stacked", "BLSTM", "relevance", "model", "with", "a", "BM25", "score", "to", "achieve", "good", "performance", ".", "Severyn", "and", "Moschitti", "presented", "a", "convolutional", "neural", "network", "architecture", "for", "re", "-", "ranking", "pairs", "of", "short", "texts", ",", "where", "they", "learned", "the", "optimal", "representation", "of", "text", "pairs", "and", "a", "similarity", "function", "to", "relate", "them", "in", "a", "supervised", "way", "from", "the", "available", "training", "data", ".", "They", "also", "need", "to", "combine", "additional", "features", "into", "their", "model", "to", "outperform", "previous", "methods", ".", "Unlike", "the", "previous", "research", ",", "our", "method", "can", "outperform", "previous", "methods", "using", "feature", "engineering", "without", "combining", "any", "additional", "features", ".", "With", "an", "additional", "simple", "feature", "like", "QL", ",", "our", "model", "is", "significantly", "better", "than", "the", "previous", "state", "-", "of", "-", "the", "-", "art", "methods", "including", "deep", "learning", "methods", ".", "Answer", "Ranking", "in", "CQA", ".", "There", "is", "also", "previous", "research", "on", "ranking", "answers", "from", "community", "question", "answering", "(", "CQA", ")", "sites", ".", "Surdeanu", "et", "al", ".", "investigated", "a", "wide", "range", "of", "feature", "types", "such", "as", "similarity", "features", ",", "translation", "features", ",", "density", "/", "frequency", "features", "for", "ranking", "answers", "to", "non", "-", "factoid", "questions", "in", "Yahoo", "!", "Answers", ".", "Jansen", "et", "al", ".", "presented", "an", "answer", "re", "-", "ranking", "model", "for", "non", "-", "factoid", "questions", "that", "integrate", "lexical", "semantics", "with", "discourse", "information", "driven", "by", "two", "representations", "of", "discourse", ".", "Xue", "et", "al", ".", "proposed", "a", "retrieval", "model", "that", "combines", "a", "translation", "-", "based", "language", "model", "for", "the", "question", "part", "with", "a", "query", "likelihood", "approach", "for", "the", "answer", "part", ".", "Questions", "from", "CQA", "sites", "are", "mostly", "non", "-", "factoid", "questions", ".", "Our", "research", "is", "closer", "to", "factoid", "questions", "such", "as", "questions", "in", "TREC", "QA", "data", ".", "Answer", "Passage", "/", "Sentence", "Retrieval", ".", "Our", "work", "is", "also", "related", "to", "previous", "research", "on", "answer", "passage", "/", "sentence", "retrieval", ".", "Tymoshenko", "and", "Moschitti", "studied", "the", "use", "of", "syntactic", "and", "semantic", "structures", "obtained", "with", "shallow", "and", "deeper", "syntactic", "parsers", "in", "the", "answer", "passage", "re", "-", "ranking", "task", ".", "Keikha", "et", "al", ".", "developed", "an", "annotated", "data", "set", "for", "non", "-", "factoid", "answer", "finding", "using", "TREC", "GOV2", "collections", "and", "topics", ".", "They", "annotated", "passage", "-", "level", "answers", ",", "revisited", "several", "passage", "retrieval", "models", "with", "this", "data", ",", "and", "came", "to", "the", "conclusion", "that", "the", "current", "methods", "are", "not", "effective", "for", "this", "task", ".", "Yang", "et", "al", ".", "developed", "effective", "methods", "for", "answer", "sentence", "retrieval", "using", "this", "annotated", "data", "by", "combining", "semantic", "features", ",", "context", "features", "and", "basic", "text", "matching", "features", "with", "a", "learning", "to", "rank", "approach", ".", "Our", "model", "is", "built", "on", "attention", "-", "based", "neural", "matching", "model", "with", "value", "-", "shared", "weighting", "schema", ".", "Unlike", "learning", "to", "rank", "approaches", "with", "feature", "engineering", ",", "our", "model", "can", "achieve", "good", "performance", "for", "ranking", "answers", "without", "any", "additional", "manual", "feature", "engineering", ",", "preprocessing", "of", "NLP", "parsers", "and", "external", "resources", "like", "knowledge", "bases", ".", "section", ":", "Attention", "-", "based", "Neural", "Matching", "Model", "In", "this", "section", "we", "present", "the", "proposed", "model", "referred", "as", "aNMM", "(", "attention", "-", "based", "Neural", "Matching", "Model", ")", ",", "which", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "Before", "we", "introduce", "our", "model", ",", "we", "firstly", "define", "some", "terminologies", ".", "subsection", ":", "Terminology", "Short", "Answer", "Text", ":", "We", "use", "Short", "Answer", "Text", "to", "refer", "to", "a", "short", "fact", ",", "answer", "sentences", "or", "answer", "passages", "that", "can", "address", "users", "’", "information", "needs", "in", "the", "issued", "questions", ".", "This", "is", "the", "ranking", "object", "in", "this", "paper", "and", "includes", "answers", "in", "various", "lengths", ".", "In", "the", "experiments", "of", "this", "paper", ",", "we", "mainly", "focus", "on", "ranking", "answer", "sentences", "that", "contain", "correct", "answer", "facts", "as", "in", "TREC", "QA", "data", ".", "QA", "Matching", "Matrix", ":", "We", "use", "QA", "Matching", "Matrix", "to", "refer", "to", "a", "matrix", "which", "represents", "the", "semantic", "matching", "information", "of", "term", "pairs", "from", "a", "question", "and", "answer", "pair", ".", "Given", "a", "question", "with", "length", "and", "an", "answer", "with", "length", ",", "a", "QA", "matching", "matrix", "is", "an", "by", "matrix", ",", "where", "denote", "the", "semantic", "similarity", "between", "term", "and", "term", "measured", "by", "the", "cosine", "similarity", "of", "the", "corresponding", "word", "embeddings", "of", "terms", ".", "If", "and", "are", "the", "same", "term", ",", "we", "assign", "as", ".", "QA", "Matching", "Vector", ":", "We", "use", "QA", "Matching", "Vector", "to", "refer", "to", "a", "row", "in", "the", "QA", "matching", "matrix", ".", "As", "presented", "before", ",", "the", "-", "th", "row", "of", "the", "QA", "matching", "matrix", "contains", "the", "semantic", "similarity", "between", "and", "all", "terms", "in", "answer", ".", "subsection", ":", "Model", "Overview", "Our", "method", "contains", "three", "steps", "as", "follows", ":", "We", "construct", "QA", "matching", "matrix", "for", "each", "question", "and", "answer", "pair", "with", "pre", "-", "trained", "word", "embeddings", ".", "We", "then", "employ", "a", "deep", "neural", "network", "with", "value", "-", "shared", "weighting", "scheme", "in", "the", "first", "layer", ",", "and", "fully", "connected", "layers", "in", "the", "rest", "to", "learn", "hierarchical", "abstraction", "of", "the", "semantic", "matching", "between", "question", "and", "answer", "terms", ".", "Finally", ",", "we", "employ", "a", "question", "attention", "network", "to", "learn", "question", "term", "importance", "and", "produce", "the", "final", "ranking", "score", ".", "We", "propose", "two", "neural", "matching", "model", "architectures", "and", "compare", "the", "effectivenesses", "of", "them", ".", "We", "firstly", "describe", "a", "basic", "version", "of", "the", "architecture", ",", "which", "is", "referred", "to", "as", "aNMM", "-", "1", ".", "In", "the", "following", "sections", ",", "we", "will", "explain", "in", "detail", "the", "two", "major", "designs", "of", "aNMM", "-", "1", ",", "i.e.", ",", "value", "-", "shared", "weights", "and", "question", "attention", "network", ".", "subsection", ":", "Value", "-", "shared", "Weighting", "We", "first", "train", "word", "embeddings", "with", "the", "Word2Vec", "tool", "by", "Mikolov", "et", "al", ".", "with", "the", "English", "Wikipedia", "dump", "to", "construct", "QA", "matching", "matrices", ".", "Given", "a", "question", "sentence", "and", "an", "answer", "sentence", ",", "we", "compute", "the", "dot", "product", "of", "the", "normalized", "word", "embeddings", "of", "all", "term", "pairs", "to", "construct", "the", "QA", "matching", "matrix", "as", "defined", "in", "Section", "[", "reference", "]", ".", "A", "major", "problem", "with", "the", "QA", "matching", "matrix", "is", "the", "variable", "size", "due", "to", "the", "different", "lengths", "of", "answers", "for", "a", "given", "question", ".", "To", "solve", "this", "problem", ",", "one", "can", "use", "CNN", "with", "pooling", "strategy", "to", "handle", "the", "variable", "size", ".", "However", ",", "as", "we", "have", "mentioned", "before", ",", "CNNs", "basically", "use", "position", "-", "shared", "weighting", "scheme", "which", "may", "not", "fit", "semantic", "matching", "between", "questions", "and", "answers", ".", "Important", "question", "terms", "and", "semantically", "similar", "answer", "words", "could", "appear", "anywhere", "in", "questions", "/", "answers", "due", "to", "the", "complex", "linguistic", "property", "of", "natural", "languages", ".", "Thus", "we", "adopt", "the", "following", "method", "to", "handle", "the", "various", "length", "problem", ":", "Value", "-", "shared", "Weights", ":", "For", "this", "method", ",", "the", "assumption", "is", "that", "matching", "signals", "in", "different", "ranges", "play", "different", "roles", "in", "deciding", "the", "final", "ranking", "score", ".", "Thus", "we", "introduce", "the", "value", "-", "shared", "weighting", "scheme", "to", "learn", "the", "importance", "of", "different", "levels", "of", "matching", "signals", ".", "The", "comparison", "between", "the", "position", "-", "shared", "weight", "and", "value", "-", "shared", "weight", "is", "shown", "in", "Figure", "[", "reference", "]", ".", "We", "can", "see", "that", "for", "position", "-", "shared", "weights", ",", "the", "weight", "associated", "with", "a", "node", "only", "depends", "on", "its", "position", "or", "relative", "location", "as", "specified", "by", "the", "filters", "in", "CNN", ".", "However", "in", "our", "model", ",", "the", "weight", "associated", "with", "a", "node", "depends", "on", "its", "value", ".", "The", "value", "of", "a", "node", "denotes", "the", "strength", "of", "the", "matching", "signal", "between", "term", "pairs", "of", "questions", "and", "answers", "from", "the", "QA", "matching", "matrix", ",", "as", "explained", "in", "Section", "[", "reference", "]", ".", "Such", "a", "setting", "enables", "us", "to", "use", "the", "learned", "weights", "to", "encode", "how", "to", "combine", "different", "levels", "of", "matching", "signals", ".", "After", "this", "step", ",", "the", "size", "of", "the", "hidden", "representation", "becomes", "fixed", "and", "we", "can", "use", "normal", "fully", "connected", "layers", "to", "learn", "higher", "level", "representations", ".", "We", "use", "the", "term", "bin", "to", "denote", "a", "specific", "range", "of", "matching", "signals", ".", "since", ",", "if", "we", "set", "the", "size", "of", "bins", "as", ",", "then", "we", "have", "bins", "where", "there", "is", "a", "separate", "bin", "for", "to", "denote", "exact", "match", "of", "terms", ".", "Specifically", ",", "value", "-", "shared", "weights", "are", "adopted", "in", "the", "forward", "propagation", "prediction", "process", "from", "the", "input", "layer", "to", "the", "hidden", "layer", "over", "each", "question", "term", "in", "aNMM", "-", "1", "as", "follows", ":", "Input", "Layer", "to", "Hidden", "Layer", ".", "Let", "denote", "a", "dimensional", "model", "parameter", "from", "input", "layer", "to", "hidden", "layer", ".", "denotes", "the", "sum", "of", "all", "matching", "signals", "within", "the", "-", "th", "value", "range", "or", "bin", ".", "For", "each", "QA", "matching", "vector", "of", "a", "given", "query", ",", "the", "combined", "score", "after", "the", "activation", "function", "of", "the", "-", "th", "node", "in", "hidden", "layer", "is", "defined", "as", "where", "is", "the", "index", "of", "question", "words", "in", ".", "We", "use", "the", "sigmoid", "function", "as", "the", "activation", "function", ",", "which", "is", "commonly", "adopted", "in", "many", "neural", "network", "architectures", ".", "subsection", ":", "Question", "Attention", "Network", "In", "addition", "to", "value", "-", "shared", "weighting", ",", "another", "model", "component", "of", "aNMM", "-", "1", "is", "the", "question", "attention", "network", ".", "In", "a", "committee", "of", "neural", "networks", "which", "consists", "of", "multiple", "networks", ",", "we", "need", "to", "combine", "the", "output", "of", "these", "networks", "to", "output", "a", "final", "decision", "vector", ".", "The", "question", "attention", "network", "uses", "the", "gating", "function", "to", "control", "the", "output", "of", "each", "network", "in", "this", "process", ".", "Specifically", ",", "in", "aNMM", "-", "1", "we", "use", "the", "softmax", "gate", "function", "to", "combine", "the", "output", "of", "multiple", "networks", "where", "each", "network", "corresponds", "to", "a", "question", "term", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "We", "feed", "the", "dot", "product", "of", "query", "word", "embedding", "and", "model", "parameter", "to", "the", "softmax", "function", "to", "represent", "the", "query", "term", "importance", ".", "In", "this", "setting", ",", "we", "can", "directly", "compare", "the", "relative", "term", "importance", "of", "query", "words", "within", "the", "same", "query", "with", "softmax", "function", ".", "We", "also", "tried", "sigmoid", "gate", "function", ",", "but", "this", "did", "not", "perform", "as", "well", "as", "softmax", "gate", "function", ".", "Softmax", "gate", "function", "is", "used", "in", "the", "forward", "propagation", "process", "from", "the", "hidden", "layer", "to", "the", "output", "layer", "as", "follows", ":", "Hidden", "Layer", "to", "Output", "Layer", ".", "From", "the", "hidden", "layer", "to", "the", "output", "layer", ",", "we", "add", "a", "softmax", "gate", "function", "to", "learn", "question", "attention", ".", "Let", "denote", "a", "dimensional", "vector", "which", "is", "a", "model", "parameter", ".", "We", "feed", "the", "dot", "product", "of", "query", "word", "embedding", "and", "to", "the", "softmax", "function", "to", "represent", "the", "query", "term", "importance", "as", "shown", "in", "Equation", "[", "reference", "]", ".", "Note", "that", "we", "normalize", "the", "query", "word", "embedding", "before", "computing", "the", "dot", "product", ".", "Unlike", "previous", "models", "like", "CNNs", "and", "BLSTM", ",", "which", "learn", "the", "semantic", "match", "score", "between", "questions", "and", "answers", "through", "representation", "learning", "from", "matching", "matrix", "or", "question", "/", "answer", "pair", "sequences", ",", "aNMM", "achieves", "this", "by", "combining", "semantic", "matching", "signals", "of", "term", "pairs", "in", "questions", "and", "answers", "weighted", "by", "the", "output", "of", "question", "attention", "network", ",", "where", "softmax", "gate", "functions", "help", "discriminate", "the", "term", "importance", "or", "attention", "on", "different", "question", "terms", ".", "subsection", ":", "Model", "Training", "For", "aNMM", "-", "1", ",", "the", "model", "parameters", "contain", "two", "sets", ":", "1", ")", "The", "value", "-", "shared", "weights", "for", "combining", "matching", "signals", "from", "the", "input", "layer", "to", "the", "hidden", "layer", ".", "2", ")", "The", "parameters", "in", "the", "gating", "function", "from", "the", "hidden", "layer", "to", "the", "output", "layer", ".", "To", "learn", "the", "model", "parameters", "from", "the", "training", "data", ",", "we", "adopt", "a", "pair", "-", "wise", "learning", "strategy", "with", "a", "large", "margin", "objective", ".", "Firstly", "we", "construct", "triples", "from", "the", "training", "data", ",", "with", "matched", "with", "better", "than", "with", ".", "We", "have", "the", "ranking", "-", "based", "loss", "as", "the", "objective", "function", "as", "following", ":", "where", "denote", "the", "predicted", "matching", "score", "for", "QA", "pair", ".", "During", "training", "stage", ",", "we", "will", "scan", "all", "the", "triples", "in", "training", "data", ".", "Given", "a", "triple", ",", "we", "will", "compute", ".", "If", ",", "we", "will", "skip", "this", "triple", ".", "Otherwise", ",", "we", "need", "to", "update", "model", "parameters", "with", "back", "propagation", "algorithm", "to", "minimize", "the", "objective", "function", ".", "Under", "softmax", "gate", "function", "setting", ",", "the", "gradients", "of", "w.r.t", ".", "from", "hidden", "layer", "to", "the", "output", "layer", "is", "shown", "in", "Equation", "[", "reference", "]", "where", "can", "be", "derived", "as", "The", "gradient", "of", "w.r.t", ".", "from", "input", "layer", "to", "hidden", "layer", "is", "shown", "in", "Equation", "[", "reference", "]", ".", "With", "the", "formulas", "of", "gradients", ",", "we", "can", "perform", "stochastic", "gradient", "descent", "to", "learn", "model", "parameters", ".", "We", "use", "mini", "-", "batch", "gradient", "descent", "to", "achieve", "more", "robust", "performance", "on", "the", "ranking", "task", ".", "For", "the", "learning", "rate", ",", "we", "adopt", "adaptive", "learning", "rate", ":", ",", "where", "will", "approach", "with", "more", "iterations", ".", "Such", "a", "setting", "has", "better", "guarantee", "for", "convergence", ".", "subsection", ":", "Extension", "to", "Deep", "Neural", "Networks", "with", "Multiple", "Sets", "of", "Value", "-", "shared", "Weights", "In", "aNMM", "-", "1", ",", "we", "can", "only", "use", "one", "set", "of", "value", "-", "shared", "weights", "for", "each", "QA", "matching", "vector", ".", "We", "further", "propose", "a", "more", "flexible", "neural", "network", "architecture", "which", "could", "enable", "us", "to", "use", "multiple", "sets", "of", "value", "-", "shared", "weights", "for", "each", "QA", "matching", "vector", ",", "leading", "to", "multiple", "intermediate", "nodes", "in", "the", "first", "hidden", "layer", ",", "as", "shown", "in", "Figure", "[", "reference", "]", "by", "the", "yellow", "color", ".", "We", "refer", "to", "this", "extended", "model", "as", "aNMM", "-", "2", ".", "The", "model", "architecture", "shown", "in", "Figure", "[", "reference", "]", "is", "corresponding", "to", "aNMM", "-", "2", ".", "subsubsection", ":", "Forward", "Propagation", "Prediction", "For", "aNMM", "-", "2", ",", "we", "add", "a", "hidden", "layer", "in", "the", "neural", "network", "where", "we", "learn", "multiple", "combined", "scores", "from", "the", "input", "layer", ".", "With", "this", "hidden", "layer", ",", "we", "define", "multiple", "weight", "vectors", "as", ".", "Thus", "becomes", "a", "two", "dimensional", "matrix", ".", "The", "formula", "for", "the", "forward", "propagation", "prediction", "is", "as", "follows", ":", "where", "and", "denote", "the", "softmax", "gate", "function", ".", "is", "the", "number", "of", "nodes", "in", "hidden", "layer", "1", ".", "is", "the", "model", "parameter", "from", "hidden", "layer", "1", "to", "hidden", "layer", "2", ",", "where", "we", "feed", "the", "linear", "combination", "of", "outputs", "of", "nodes", "in", "hidden", "layer", "1", "to", "an", "extra", "activation", "function", "comparing", "with", "Equation", "[", "reference", "]", ".", "Then", "from", "hidden", "layer", "2", "to", "output", "layer", ",", "we", "sum", "over", "all", "outputs", "of", "nodes", "in", "hidden", "layer", "2", "weighted", "by", "the", "outputs", "of", "softmax", "gate", "functions", ",", "which", "also", "form", "the", "question", "attention", "network", ".", "subsubsection", ":", "Back", "Propagation", "for", "Model", "Training", "For", "aNMM", "-", "2", ",", "we", "have", "three", "sets", "of", "model", "parameters", ":", "1", ")", "from", "input", "layer", "to", "hidden", "layer", "1", ";", "2", ")", "from", "hidden", "layer", "1", "to", "hidden", "layer", "2", ";", "3", ")", "from", "hidden", "layer", "2", "to", "output", "layer", ".", "All", "three", "sets", "of", "parameters", "are", "updated", "through", "back", "propagation", ".", "The", "definition", "of", "the", "objective", "function", "is", "the", "same", "as", "Equation", "[", "reference", "]", ".", "The", "back", "propagation", "process", "for", "model", "parameter", "learning", "is", "described", "as", "follows", ":", "From", "hidden", "layer", "2", "to", "output", "layer", ".", "The", "gradients", "of", "the", "objective", "function", "w.r.t", ".", "is", "as", "following", ":", "Where", "From", "hidden", "layer", "1", "to", "hidden", "layer", "2", ".", "The", "gradients", "of", "the", "objective", "function", "w.r.t", ".", "is", "as", "following", ":", "Where", ".", "From", "input", "layer", "to", "hidden", "layer", "1", ".", "The", "gradients", "of", "the", "objective", "function", "w.r.t", ".", "is", "as", "following", ":", "Where", "Initially", "we", "will", "randomly", "give", "the", "values", "of", "model", "parameters", ".", "Then", "we", "will", "use", "back", "propagation", "to", "update", "the", "model", "parameters", ".", "When", "the", "learning", "process", "converge", ",", "we", "use", "the", "learned", "model", "parameters", "for", "prediction", "to", "rank", "short", "answer", "texts", ".", "section", ":", "Experiments", "subsection", ":", "Data", "Set", "and", "Experiment", "Settings", "We", "use", "the", "TREC", "QA", "data", "set", "created", "by", "Wang", "et", ".", "al", ".", "from", "TREC", "QA", "track", "8", "-", "13", "data", ",", "with", "candidate", "answers", "automatically", "selected", "from", "each", "question", "’s", "document", "pool", "using", "a", "combination", "of", "overlapping", "non", "-", "stop", "word", "counts", "and", "pattern", "matching", ".", "This", "data", "set", "is", "one", "of", "the", "most", "widely", "used", "benchmarks", "for", "answer", "re", "-", "ranking", ".", "Table", "[", "reference", "]", "shows", "the", "statistics", "of", "this", "data", "set", ".", "The", "dataset", "contains", "a", "set", "of", "factoid", "questions", "with", "candidate", "answers", "which", "are", "limited", "to", "a", "single", "sentence", ".", "There", "are", "two", "training", "data", "sets", ":", "TRAIN", "and", "TRAIN", "-", "ALL", ".", "Answers", "in", "TRAIN", "have", "manual", "judgments", "for", "the", "answer", "correctness", ".", "The", "manual", "judgment", "of", "candidate", "answer", "sentences", "is", "provided", "for", "the", "entire", "TREC", "13", "set", "and", "for", "a", "part", "of", "questions", "from", "TREC", "8", "-", "12", ".", "TRAIN", "-", "ALL", "is", "another", "training", "set", "with", "much", "larger", "number", "of", "questions", ".", "The", "correctness", "of", "candidate", "answer", "sentences", "in", "TRAIN", "-", "ALL", "is", "identified", "by", "matching", "answer", "sentences", "with", "answer", "pattern", "regular", "expressions", "provided", "by", "TREC", ".", "This", "data", "set", "is", "more", "noisy", ",", "however", "it", "provides", "many", "more", "QA", "pairs", "for", "model", "training", ".", "There", "is", "a", "DEV", "set", "for", "hyper", "-", "parameter", "optimization", "and", "TEST", "set", "for", "model", "testing", ".", "We", "use", "the", "same", "train", "/", "dev", "/", "test", "partition", "in", "our", "experiments", "to", "directly", "compare", "our", "results", "with", "previous", "works", ".", "For", "data", "preprocess", ",", "we", "perform", "tokenization", "without", "stemming", ".", "We", "maintain", "stop", "words", "during", "the", "model", "training", "stage", ".", "Word", "Embeddings", ".", "We", "obtain", "pre", "-", "trained", "word", "embeddings", "with", "the", "Word2Vec", "tool", "by", "Mikolov", "et", "al", ".", "with", "the", "English", "Wikipedia", "dump", ".", "We", "use", "the", "skip", "-", "gram", "model", "with", "window", "size", "and", "filter", "words", "with", "frequency", "less", "than", "following", "the", "common", "practice", "in", "many", "neural", "embedding", "models", ".", "For", "the", "word", "vector", "dimension", ",", "we", "tune", "it", "as", "a", "hyper", "-", "parameter", "on", "the", "validation", "data", "starting", "from", "to", ".", "Embeddings", "for", "words", "not", "present", "are", "randomly", "initialized", "with", "sampled", "numbers", "from", "uniform", "distribution", "U", "[-", "0.25", ",", "0.25", "]", ",", "which", "follows", "the", "same", "setting", "as", ".", "Model", "Hyper", "-", "parameters", ".", "For", "the", "setting", "of", "hyper", "-", "parameters", ",", "we", "set", "the", "number", "of", "bins", "as", ",", "word", "embedding", "dimension", "as", "for", "aNNM", "-", "1", ",", "the", "number", "of", "bins", "as", ",", "word", "embedding", "dimension", "as", "for", "aNNM", "-", "2", "after", "we", "tune", "hyper", "-", "parameters", "on", "the", "provided", "DEV", "set", "of", "TREC", "QA", "data", ".", "subsection", ":", "Evaluation", "and", "Metrics", "For", "evaluation", ",", "we", "rank", "answer", "sentences", "with", "the", "predicted", "score", "of", "each", "method", "and", "compare", "the", "rank", "list", "with", "the", "ground", "truth", "to", "compute", "metrics", ".", "We", "choose", "Mean", "Average", "Precision", "(", "MAP", ")", "and", "Mean", "Reciprocal", "Rank", "(", "MRR", ")", ",", "which", "are", "commonly", "used", "in", "information", "retrieval", "and", "question", "answering", ",", "as", "the", "metric", "to", "evaluate", "our", "model", ".", "The", "definition", "of", "MRR", "is", "as", "follows", ":", "where", "is", "the", "position", "of", "the", "first", "correct", "answer", "in", "the", "rank", "list", ".", "Thus", "MRR", "is", "only", "based", "on", "the", "rank", "of", "the", "first", "correct", "answer", ".", "It", "is", "more", "suitable", "for", "the", "cases", "where", "the", "rank", "of", "the", "first", "correct", "answer", "is", "emphasized", "or", "each", "question", "only", "have", "one", "correct", "answer", ".", "On", "the", "other", "hand", ",", "MAP", "looks", "at", "the", "ranks", "of", "all", "correct", "answers", ".", "It", "is", "computed", "as", "following", ":", "where", "is", "the", "average", "precision", "for", "each", "query", ".", "Thus", "MAP", "is", "the", "average", "performance", "on", "all", "correct", "answers", ".", "We", "use", "the", "official", "scripts", "for", "computing", "these", "metrics", ".", "subsection", ":", "Model", "Learning", "Results", "In", "this", "section", ",", "we", "give", "some", "qualitative", "analysis", "and", "visualization", "of", "our", "model", "learning", "results", ".", "Specifically", ",", "we", "analyze", "the", "learned", "value", "-", "shard", "weights", "and", "question", "term", "importance", "by", "aNMM", ".", "subsubsection", ":", "Value", "-", "shared", "Weight", "We", "take", "the", "learned", "value", "-", "shared", "weights", "of", "aNMM", "-", "1", "as", "the", "example", ".", "Figure", "[", "reference", "]", "shows", "the", "learned", "value", "-", "shared", "weights", "by", "aNMM", "-", "1", ".", "In", "aNMM", "-", "1", ",", "for", "each", "QA", "matching", "vector", ",", "there", "is", "only", "one", "bin", "node", ".", "Thus", "the", "learned", "value", "-", "shared", "weights", "for", "aNMM", "-", "1", "is", "a", "one", "dimension", "vector", ".", "For", "aNMM", "-", "1", ",", "we", "set", "the", "number", "of", "bins", "as", "as", "presented", "in", "Section", "[", "reference", "]", ".", "Note", "that", "the", "x", "-", "axis", "is", "the", "index", "of", "bin", "range", "and", "the", "y", "-", "axis", "is", "the", "value", "-", "shared", "weights", "corresponding", "to", "each", "bin", "range", ".", "The", "range", "of", "match", "signals", "is", "[", "-", "1", ",", "1", "]", "from", "the", "left", "to", "the", "right", ".", "We", "make", "the", "following", "observations", ":", "(", "1", ")", "The", "exact", "match", "signal", "which", "is", "corresponding", "to", "in", "the", "last", "bin", "is", "tied", "with", "a", "very", "large", "weight", ",", "which", "shows", "that", "exact", "match", "information", "is", "very", "important", ".", "(", "2", ")", "For", "positive", "matching", "score", "from", ",", "which", "is", "corresponding", "to", "bin", "index", ",", "the", "learned", "value", "-", "shared", "weights", "are", "different", "for", "matching", "score", "range", "(", "bin", "index", ")", "and", "matching", "score", "range", "(", "bin", "index", ")", ".", "We", "can", "observe", "many", "positive", "value", "-", "shared", "weights", "for", "matching", "score", "range", "and", "negative", "value", "-", "shared", "weights", "for", "matching", "score", "range", ".", "This", "makes", "sense", "since", "high", "semantic", "matching", "scores", "are", "positive", "indicators", "on", "answer", "correctness", ",", "whereas", "low", "semantic", "matching", "scores", "indicate", "that", "the", "candidate", "answer", "sentences", "contain", "irrelevant", "terms", ".", "(", "3", ")", "For", "negative", "matching", "scores", "from", ",", "we", "can", "see", "there", "is", "not", "a", "lot", "of", "differences", "between", "value", "-", "shared", "weights", "for", "different", "ranges", ".", "A", "major", "reason", "is", "that", "most", "similarity", "scores", "based", "on", "word", "embeddings", "are", "positive", ".", "Therefore", ",", "we", "can", "remove", "bins", "corresponding", "to", "negative", "matching", "scores", "to", "reduce", "the", "dimension", "of", "value", "-", "shared", "weight", "vectors", ",", "which", "can", "help", "improve", "the", "efficiency", "of", "the", "model", "training", "process", ".", "We", "will", "show", "more", "quantitative", "results", "on", "the", "comparison", "between", "value", "-", "shared", "weights", "and", "position", "-", "shared", "weights", "in", "CNN", "in", "Section", "[", "reference", "]", ".", "subsubsection", ":", "Question", "Term", "Importance", "Next", "we", "analyze", "the", "learned", "question", "term", "importance", "of", "our", "model", ".", "Due", "to", "the", "space", "limit", ",", "we", "also", "use", "the", "learned", "question", "term", "importance", "of", "aNMM", "-", "1", "as", "an", "example", ".", "Table", "[", "reference", "]", "shows", "the", "examples", "of", "learned", "question", "term", "importance", "by", "aNMM", "-", "1", ".", "We", "also", "visualize", "the", "question", "term", "importance", "in", "Figure", "[", "reference", "]", ".", "Based", "on", "the", "results", "in", "the", "table", "and", "the", "figure", ",", "we", "can", "clearly", "see", "that", "aNMM", "-", "1", "learns", "reasonable", "term", "importance", ".", "For", "instance", ",", "with", "the", "question", "attention", "network", ",", "aNMM", "-", "1", "discovers", "important", "terms", "like", "“", "khmer", "”", ",", "“", "rouge", "”", ",", "“", "power", "”", "as", "for", "the", "question", "“", "When", "did", "the", "khmer", "rouge", "come", "into", "power", "?", "”", ".", "Terms", "like", "“", "age", "”", ",", "“", "rossinin", "”", ",", "“", "stop", "”", ",", "“", "writing”", ",", "“opera", "”", "are", "highlighted", "for", "the", "question", "“", "At", "what", "age", "did", "rossini", "stop", "writing", "opera", "?", "”", ".", "For", "the", "question", "“", "Where", "was", "the", "first", "burger", "king", "restaurant", "opened", "?", "”", "mentioned", "in", "Section", "[", "reference", "]", ",", "“", "burger", "”", ",", "“", "king", "”", ",", "“", "opened", "”", "are", "treated", "as", "important", "question", "terms", ".", "An", "interesting", "question", "is", "how", "the", "learned", "term", "importance", "compare", "with", "traditional", "IR", "term", "weighting", "methods", "such", "as", "IDF", ".", "We", "design", "an", "experiment", "to", "compare", "aNMM", "-", "1", "/", "aNMM", "-", "2", "with", "aNMM", "-", "IDF", ",", "which", "is", "a", "degenerate", "version", "of", "our", "model", "where", "we", "use", "IDF", "to", "directly", "replace", "the", "output", "of", "question", "attention", "network", ".", "In", "this", "case", ",", "in", "Equation", "[", "reference", "]", "is", "replaced", "by", "the", "IDF", "of", "the", "j", "-", "th", "question", "term", ".", "Table", "[", "reference", "]", "shows", "the", "results", ".", "We", "find", "that", "if", "we", "replace", "the", "output", "of", "question", "attention", "network", "of", "aNMM", "with", "IDF", ",", "it", "will", "decrease", "the", "answer", "ranking", "performance", ",", "especially", "on", "TRAIN", "data", ".", "Thus", ",", "we", "can", "see", "that", "with", "the", "optimization", "process", "in", "the", "back", "propagation", "training", "process", ",", "aNMM", "can", "learn", "better", "question", "term", "weighting", "score", "than", "heuristic", "term", "weighting", "functions", "like", "IDF", ".", "subsection", ":", "Experimental", "Results", "for", "Ranking", "Answers", "subsubsection", ":", "Learning", "without", "Combining", "Additional", "Features", "Our", "first", "experimental", "setting", "is", "ranking", "answer", "sentences", "directly", "by", "the", "predicted", "score", "from", "aNMM", "without", "combining", "any", "additional", "features", ".", "This", "will", "enable", "us", "to", "answer", "RQ1", "proposed", "in", "Section", "[", "reference", "]", ".", "Table", "[", "reference", "]", "shows", "the", "results", "of", "TREC", "QA", "on", "TRAIN", "and", "TRAIN", "-", "ALL", "without", "combining", "additional", "features", ".", "In", "this", "table", ",", "we", "compare", "the", "results", "of", "aNMM", "with", "other", "previous", "deep", "learning", "methods", "including", "CNN", "and", "LSTM", ".", "We", "summarize", "our", "observations", "as", "follows", ":", "(", "1", ")", "Both", "aNMM", "-", "1", "and", "aNMM", "-", "2", "show", "significant", "improvements", "for", "MAP", "and", "MRR", "on", "TRAIN", "and", "TRAIN", "-", "ALL", "data", "sets", "comparing", "with", "previous", "deep", "learning", "methods", ".", "Specifically", ",", "if", "we", "compare", "the", "results", "of", "aNMM", "-", "1", "with", "the", "strongest", "deep", "learning", "baseline", "method", "by", "Severyn", "et", "al", ".", "based", "on", "CNN", ",", "we", "can", "see", "aNMM", "-", "1", "outperform", "CNN", "for", "%", "in", "MAP", "on", "TRAIN", ",", "%", "in", "MAP", "on", "TRAIN", "-", "ALL", ".", "For", "MRR", ",", "we", "can", "also", "observe", "similar", "significant", "improvements", "of", "aNMM", "-", "1", ".", "These", "results", "show", "that", "with", "the", "value", "-", "shared", "weight", "scheme", "instead", "of", "the", "position", "-", "shared", "weight", "scheme", "in", "CNN", "and", "term", "importance", "learning", "with", "question", "attention", "network", ",", "aNMM", "can", "predict", "ranking", "scores", "with", "much", "higher", "accuracy", "comparing", "with", "previous", "deep", "learning", "models", "for", "ranking", "answers", ".", "(", "2", ")", "If", "we", "compare", "the", "results", "of", "aNMM", "-", "1", "and", "aNMM", "-", "2", ",", "we", "can", "see", "their", "results", "are", "very", "close", ".", "aNMM", "-", "1", "has", "slightly", "better", "performance", "than", "aNMM", "-", "2", ".", "This", "result", "indicates", "that", "adding", "one", "more", "hidden", "layer", "to", "incorporate", "multiple", "bin", "nodes", "does", "not", "necessarily", "increase", "the", "performance", "for", "answer", "ranking", "in", "TREC", "QA", "data", ".", "From", "the", "perspective", "of", "model", "efficiency", ",", "aNMM", "-", "1", "could", "be", "a", "better", "choice", "since", "it", "can", "be", "trained", "much", "faster", "with", "good", "prediction", "accuracy", ".", "However", ",", "for", "larger", "training", "data", "sets", "than", "TREC", "QA", "data", ",", "aNMM", "-", "2", "could", "have", "better", "performance", "since", "it", "has", "more", "model", "parameters", "and", "is", "suitable", "for", "fitting", "larger", "training", "data", "set", ".", "We", "leave", "the", "study", "of", "impact", "of", "the", "number", "of", "hidden", "layers", "in", "aNMM", "to", "future", "work", ".", "Table", "[", "reference", "]", "shows", "the", "comparison", "between", "aNMM", "with", "previous", "methods", "using", "feature", "engineering", "on", "TRAIN", "-", "ALL", "without", "combining", "additional", "features", ".", "We", "find", "that", "both", "aNMM", "-", "1", "and", "aNMM", "-", "2", "achieve", "better", "performance", "comparing", "with", "other", "methods", "using", "feature", "engineering", ".", "Specifically", ",", "comparing", "the", "results", "of", "aNMM", "-", "1", "with", "the", "strongest", "baseline", "by", "Yih", "et", "al", ".", "based", "on", "enhanced", "lexical", "semantic", "models", ",", "aNMM", "-", "1", "achieves", "%", "gain", "for", "MAP", "and", "%", "gain", "for", "MRR", ".", "These", "results", "show", "that", "it", "is", "possible", "to", "build", "a", "uniform", "deep", "learning", "model", "such", "that", "it", "can", "achieve", "better", "performance", "than", "methods", "using", "feature", "engineering", ".", "To", "the", "best", "of", "our", "knowledge", ",", "aNMM", "is", "the", "first", "deep", "learning", "model", "that", "can", "achieve", "good", "performance", "comparing", "with", "previous", "methods", "either", "based", "on", "deep", "learning", "models", "or", "feature", "engineering", "for", "ranking", "answers", "without", "any", "additional", "features", ",", "syntactic", "parsers", "and", "external", "resources", "except", "for", "pre", "-", "trained", "word", "embeddings", ".", "subsubsection", ":", "Learning", "with", "Combining", "Additional", "Features", "Our", "second", "experimental", "setting", "is", "to", "address", "RQ2", "proposed", "in", "Section", "[", "reference", "]", ",", "where", "we", "ask", "whether", "our", "model", "can", "outperform", "the", "state", "-", "of", "-", "the", "-", "art", "performance", "achieved", "by", "CNN", "and", "LSTM", "for", "answer", "ranking", "when", "combining", "additional", "features", ".", "We", "combine", "the", "predicted", "score", "from", "aNMM", "-", "1", "and", "aNMM", "-", "2", "with", "the", "Query", "Likelihood", "(", "QL", ")", "score", "using", "LambdaMART", "following", "a", "similar", "approach", "to", ".", "We", "use", "the", "implementation", "of", "LambdaMART", "in", "jforests", "We", "compare", "the", "results", "with", "previous", "deep", "learning", "models", "with", "additional", "features", ".", "Table", "[", "reference", "]", "shows", "the", "results", "on", "TRAIN", "and", "TRAIN", "-", "ALL", "when", "combining", "additional", "features", ".", "We", "can", "see", "that", "with", "combined", "features", ",", "both", "aNMM", "-", "1", "and", "aNMM", "-", "2", "have", "better", "performance", ".", "aNMM", "-", "1", "also", "outperforms", "CNN", "by", "Severyn", "et", "al", ".", "which", "is", "the", "current", "state", "-", "of", "-", "the", "-", "art", "method", "for", "ranking", "answers", "in", "terms", "of", "both", "MAP", "and", "MRR", "on", "TRAIN", "and", "TRAIN", "-", "ALL", ".", "We", "also", "tried", "to", "combine", "aNMM", "score", "with", "other", "additional", "features", "such", "as", "word", "overlap", "features", ",", "IDF", "weighted", "word", "overlap", "features", "and", "BM25", "as", "in", "previous", "research", ".", "The", "results", "were", "either", "similar", "or", "worse", "than", "combining", "aNMM", "score", "with", "QL", ".", "For", "aNMM", ",", "the", "gains", "after", "combining", "additional", "features", "are", "not", "as", "large", "as", "neural", "network", "models", "like", "CNN", "in", "and", "LSTM", "in", ".", "We", "think", "the", "reasons", "for", "this", "are", "two", "-", "fold", ":", "(", "1", ")", "The", "QA", "matching", "matrix", "in", "aNMM", "model", "can", "capture", "exact", "match", "information", "by", "assigning", "to", "matrix", "elements", "if", "the", "corresponding", "answer", "term", "and", "question", "term", "are", "the", "same", ".", "This", "exact", "match", "information", "include", "match", "between", "numbers", "and", "proper", "nouns", ",", "which", "are", "highlighted", "in", "previous", "research", "work", "as", "especially", "important", "for", "factoid", "questions", "answering", ",", "where", "most", "of", "the", "questions", "are", "of", "type", "what", ",", "when", ",", "who", "that", "are", "looking", "for", "answers", "containing", "numbers", "or", "proper", "nouns", ".", "Within", "aNMM", "architecture", ",", "this", "problem", "has", "already", "been", "handled", "with", "QA", "matching", "matrix", ".", "Thus", "incorporating", "word", "overlap", "features", "will", "not", "help", "much", "for", "improving", "the", "performance", "of", "aNMM", ".", "(", "2", ")", "In", "addition", "to", "exact", "match", "information", ",", "aNMM", "could", "also", "learn", "question", "term", "importance", "like", "IDF", "information", "through", "question", "attention", "network", ".", "Instead", "of", "empirically", "designing", "heuristic", "functions", "like", "IDF", ",", "aNMM", "can", "get", "learning", "based", "question", "term", "importance", "score", ".", "As", "analyzed", "in", "Section", "[", "reference", "]", ",", "with", "the", "optimization", "process", "in", "the", "back", "propagation", "training", "process", ",", "aNMM", "can", "learn", "similar", "or", "even", "better", "term", "weighting", "score", "than", "IDF", ".", "Thus", "combining", "aNMM", "score", "with", "features", "like", "IDF", "weighted", "word", "overlap", "features", "and", "BM25", "may", "not", "increase", "the", "performance", "of", "aNMM", "by", "a", "large", "margin", "as", "the", "case", "in", "related", "research", "works", ".", "subsubsection", ":", "Results", "Summary", "[", "b", "]", "0.48", "[", "b", "]", "0.48", "Finally", "we", "summarize", "the", "results", "of", "previously", "published", "systems", "on", "the", "QA", "answer", "ranking", "task", "in", "Table", "[", "reference", "]", ".", "We", "can", "see", "aNMM", "trained", "with", "TRAIN", "-", "ALL", "set", "beats", "all", "the", "previous", "state", "-", "of", "-", "the", "art", "systems", "including", "both", "methods", "using", "feature", "engineering", "and", "deep", "learning", "models", ".", "These", "results", "are", "very", "promising", "since", "aNMM", "requires", "no", "manual", "feature", "engineering", ",", "no", "expensive", "processing", "by", "various", "NLP", "parsers", "and", "no", "external", "results", "like", "large", "scale", "knowledge", "base", "except", "for", "pre", "-", "trained", "word", "embeddings", ".", "Furthermore", ",", "even", "without", "combining", "additional", "features", ",", "aNMM", "still", "performs", "well", "for", "answer", "ranking", ",", "showing", "significant", "improvements", "over", "previous", "deep", "learning", "model", "with", "no", "additional", "features", "and", "linguistic", "feature", "engineering", "methods", ".", "subsection", ":", "Parameter", "Sensitivity", "Analysis", "We", "perform", "parameter", "sensitivity", "analysis", "of", "our", "proposed", "model", "aNMM", ".", "We", "focus", "on", "aNMM", "-", "1", "as", "the", "example", "due", "to", "the", "space", "limitation", ".", "For", "aNMM", "-", "1", ",", "we", "fix", "the", "number", "of", "bins", "as", "and", "change", "the", "dimension", "of", "word", "vectors", ".", "Similarly", ",", "we", "fix", "the", "dimension", "of", "word", "vectors", "as", "and", "vary", "the", "number", "of", "bins", ".", "Figure", "[", "reference", "]", "shows", "the", "change", "of", "MAP", "and", "MRR", "on", "the", "validation", "data", "as", "we", "vary", "the", "hyper", "-", "parameters", ".", "We", "summarize", "our", "observations", "as", "follows", ":", "(", "1", ")", "For", "word", "vector", "dimension", ",", "the", "range", "is", "a", "good", "choice", "as", "much", "lower", "or", "higher", "word", "vector", "dimensions", "will", "hurt", "the", "performance", ".", "The", "choice", "of", "word", "vector", "dimension", "also", "depends", "on", "the", "size", "of", "training", "corpus", ".", "Larger", "corpus", "requires", "higher", "dimension", "of", "word", "vectors", "to", "embed", "terms", "in", "vocabulary", ".", "(", "2", ")", "For", "the", "number", "of", "bins", ",", "we", "can", "see", "that", "MAP", "and", "MRR", "will", "decrease", "as", "the", "bin", "number", "increase", ".", "Too", "many", "bins", "will", "increase", "the", "model", "complexity", ",", "which", "leads", "aNMM", "to", "be", "more", "likely", "to", "overfit", "the", "training", "data", ".", "Thus", "choosing", "suitable", "number", "of", "bins", "by", "optimizing", "hyper", "-", "parameter", "on", "validation", "data", "can", "help", "improve", "the", "performance", "of", "aNMM", ".", "section", ":", "Conclusions", "and", "Future", "Work", "In", "this", "paper", ",", "we", "propose", "an", "attention", "based", "neural", "matching", "model", "for", "ranking", "short", "answer", "text", ".", "We", "adopt", "value", "-", "shared", "weighting", "scheme", "instead", "of", "position", "-", "shared", "weighting", "scheme", "for", "combing", "different", "matching", "signals", "and", "incorporate", "question", "term", "importance", "learning", "using", "a", "question", "attention", "network", ".", "We", "perform", "a", "thorough", "experimental", "study", "with", "the", "TREC", "QA", "dataset", "from", "TREC", "QA", "tracks", "8", "-", "13", "and", "show", "promising", "results", ".", "Unlike", "previous", "methods", "including", "CNN", "as", "in", "and", "LSTM", "as", "in", ",", "which", "only", "show", "inferior", "results", "without", "combining", "additional", "features", ",", "our", "model", "can", "achieve", "better", "performance", "than", "the", "state", "-", "of", "-", "art", "method", "using", "linguistic", "feature", "engineering", "without", "additional", "features", ".", "With", "a", "simple", "additional", "feature", ",", "our", "method", "can", "achieve", "the", "new", "state", "-", "of", "-", "the", "-", "art", "performance", "among", "current", "existing", "methods", ".", "For", "further", "work", ",", "we", "will", "study", "other", "deep", "learning", "architectures", "for", "answer", "ranking", "and", "extend", "our", "work", "to", "include", "non", "-", "factoid", "question", "answering", "data", "sets", ".", "section", ":", "Acknowledgments", "This", "work", "was", "supported", "in", "part", "by", "the", "Center", "for", "Intelligent", "Information", "Retrieval", ",", "in", "part", "by", "NSF", "IIS", "-", "1160894", ",", "and", "in", "part", "by", "NSF", "grant", "#", "IIS", "-", "1419693", ".", "Any", "opinions", ",", "findings", "and", "conclusions", "or", "recommendations", "expressed", "in", "this", "material", "are", "those", "of", "the", "authors", "and", "do", "not", "necessarily", "reflect", "those", "of", "the", "sponsor", ".", "bibliography", ":", "References"], "entity": [[4, 8, ["Ranking", "Short", "Answer", "Texts"], "Task"], [9, 15, ["Attention", "-", "Based", "Neural", "Matching", "Model"], "Method"], [24, 26, ["feature", "engineering"], "Method"], [27, 30, ["deep", "learning", "approaches"], "Method"], [32, 35, ["convolutional", "neural", "networks"], "Method"], [36, 37, ["CNNs"], "Method"], [39, 45, ["Long", "Short", "-", "Term", "Memory", "Models"], "Method"], [46, 47, ["LSTMs"], "Method"], [53, 59, ["semantic", "matching", "of", "questions", "and", "answers"], "Task"], [80, 82, ["BM25", "scores"], "Metric"], [96, 99, ["linguistic", "feature", "engineering"], "Method"], [107, 112, ["attention", "based", "neural", "matching", "model"], "Method"], [113, 117, ["ranking", "short", "answer", "text"], "Task"], [120, 125, ["value", "-", "shared", "weighting", "scheme"], "Method"], [127, 132, ["position", "-", "shared", "weighting", "scheme"], "Method"], [139, 143, ["question", "term", "importance", "learning"], "Method"], [168, 171, ["neural", "network", "models"], "Method"], [194, 195, ["aNMM"], "Method"], [232, 234, ["concept_desc¿Information", "systems"], "Method"], [234, 236, ["Retrieval", "models"], "Method"], [259, 261, ["concept_desc¿Information", "systems"], "Method"], [261, 263, ["Question", "answering¡"], "Task"], [287, 291, ["ranking", "systems", "Question", "answering"], "Task"], [294, 296, ["Question", "answering"], "Task"], [297, 298, ["QA"], "Task"], [333, 336, ["advanced", "web", "search"], "Task"], [340, 342, ["QA", "systems"], "Method"], [344, 348, ["learning", "to", "rank", "approach"], "Method"], [392, 395, ["web", "correlation", "features"], "Method"], [404, 405, ["accuracy"], "Metric"], [412, 415, ["manual", "feature", "engineering"], "Method"], [436, 438, ["NLP", "parsers"], "Method"], [457, 460, ["deep", "learning", "approaches"], "Method"], [463, 465, ["semantic", "match"], "Task"], [478, 481, ["neural", "network", "models"], "Method"], [483, 486, ["convolutional", "neural", "networks"], "Method"], [487, 488, ["CNNs"], "Method"], [490, 496, ["Long", "Short", "-", "Term", "Memory", "Models"], "Method"], [497, 498, ["LSTMs"], "Method"], [531, 535, ["answer", "sentence", "selection", "task"], "Task"], [554, 556, ["deep", "models"], "Method"], [560, 561, ["CNNs"], "Method"], [562, 563, ["LSTMs"], "Method"], [577, 578, ["BM25"], "Method"], [609, 612, ["linguistic", "feature", "engineering"], "Method"], [632, 635, ["deep", "learning", "models"], "Method"], [646, 648, ["feature", "engineering"], "Method"], [668, 670, ["question", "answering"], "Task"], [682, 685, ["deep", "learning", "architectures"], "Method"], [686, 688, ["answer", "ranking"], "Task"], [701, 705, ["question", "/", "answer", "matching"], "Task"], [709, 710, ["CNNs"], "Method"], [711, 715, ["question", "/", "answer", "matching"], "Task"], [718, 719, ["CNNs"], "Method"], [723, 725, ["computer", "vision"], "Task"], [726, 727, ["CV"], "Task"], [736, 739, ["local", "perceptive", "filters"], "Method"], [758, 760, ["semantic", "matching"], "Task"], [794, 795, ["LSTMs"], "Method"], [797, 802, ["question", "/", "answer", "matching", "problem"], "Task"], [903, 906, ["text", "matching", "models"], "Method"], [919, 920, ["CNNs"], "Method"], [937, 938, ["LSTMs"], "Method"], [958, 961, ["deep", "learning", "architectures"], "Method"], [962, 964, ["ranking", "answers"], "Task"], [968, 973, ["attention", "based", "neural", "matching", "model"], "Method"], [974, 975, ["aNMM"], "Method"], [993, 996, ["Deep", "neural", "network"], "Method"], [997, 1001, ["value", "-", "shared", "weights"], "Method"], [1006, 1011, ["value", "-", "shared", "weighting", "scheme"], "Method"], [1012, 1015, ["deep", "neural", "networks"], "Method"], [1020, 1025, ["position", "-", "shared", "weighting", "scheme"], "Method"], [1026, 1027, ["CNNs"], "Method"], [1033, 1035, ["semantic", "matching"], "Task"], [1056, 1058, ["attention", "scheme"], "Method"], [1065, 1067, ["attention", "scheme"], "Method"], [1073, 1075, ["gating", "function"], "Method"], [1103, 1106, ["TREC", "QA", "dataset"], "Material"], [1126, 1130, ["answer", "re", "-", "ranking"], "Task"], [1135, 1136, ["CNNs"], "Method"], [1137, 1138, ["LSTMs"], "Method"], [1163, 1166, ["linguistic", "feature", "engineering"], "Method"], [1171, 1174, ["deep", "learning", "models"], "Method"], [1190, 1191, ["QL"], "Metric"], [1210, 1212, ["ranking", "answers"], "Task"], [1214, 1215, ["metrics"], "Metric"], [1323, 1326, ["deep", "learning", "models"], "Method"], [1327, 1329, ["text", "matching"], "Task"], [1334, 1336, ["answer", "ranking"], "Task"], [1337, 1338, ["CQA"], "Task"], [1339, 1344, ["answer", "passage", "/", "sentence", "retrieval"], "Task"], [1345, 1348, ["Deep", "Learning", "Models"], "Method"], [1349, 1351, ["Text", "Matching"], "Task"], [1357, 1360, ["deep", "learning", "models"], "Method"], [1362, 1366, ["text", "matching", "and", "ranking"], "Task"], [1368, 1371, ["deep", "learning", "models"], "Method"], [1372, 1373, ["DSSM"], "Method"], [1374, 1375, ["CDSSM"], "Method"], [1376, 1379, ["ARC", "-", "I"], "Method"], [1380, 1383, ["ARC", "-", "II"], "Method"], [1384, 1385, ["DCNN"], "Method"], [1386, 1387, ["DeepMatch"], "Method"], [1388, 1389, ["MultiGranCNN"], "Method"], [1390, 1391, ["MatchPyramid"], "Method"], [1392, 1393, ["DSSM"], "Method"], [1395, 1399, ["non", "-", "linear", "projection"], "Method"], [1413, 1416, ["neural", "network", "models"], "Method"], [1436, 1437, ["DeepMatch"], "Method"], [1439, 1441, ["topic", "model"], "Method"], [1457, 1459, ["deep", "architecture"], "Method"], [1466, 1469, ["ARC", "-", "I"], "Method"], [1470, 1473, ["ARC", "-", "II"], "Method"], [1485, 1489, ["matching", "natural", "language", "sentences"], "Task"], [1490, 1493, ["ARC", "-", "I"], "Method"], [1511, 1515, ["multi", "-", "layer", "perceptron"], "Method"], [1516, 1517, ["MLP"], "Method"], [1522, 1525, ["ARC", "-", "I"], "Method"], [1541, 1543, ["convolution", "model"], "Method"], [1559, 1561, ["matching", "task"], "Task"], [1567, 1570, ["ARC", "-", "II"], "Method"], [1582, 1585, ["ARC", "-", "II"], "Method"], [1620, 1623, ["ARC", "-", "II"], "Method"], [1625, 1628, ["QA", "matching", "matrix"], "Task"], [1631, 1633, ["neural", "networks"], "Method"], [1659, 1660, ["CNN"], "Method"], [1662, 1665, ["ARC", "-", "II"], "Method"], [1669, 1671, ["attention", "scheme"], "Method"], [1677, 1680, ["Factoid", "Question", "Answering"], "Task"], [1710, 1713, ["answer", "sentence", "selection"], "Task"], [1715, 1718, ["semantic", "matching", "problem"], "Task"], [1735, 1738, ["lexical", "semantic", "models"], "Method"], [1746, 1749, ["recursive", "neural", "network"], "Method"], [1750, 1751, ["RNN"], "Method"], [1766, 1768, ["textual", "compositionality"], "Method"], [1777, 1780, ["answer", "sentence", "selection"], "Task"], [1781, 1783, ["distributed", "representations"], "Method"], [1794, 1796, ["semantic", "encoding"], "Method"], [1812, 1815, ["logistic", "regression", "classifier"], "Method"], [1826, 1832, ["bidirectional", "Long", "-", "Short", "Term", "Memory"], "Method"], [1833, 1834, ["BLSTM"], "Method"], [1877, 1881, ["convolutional", "neural", "network", "architecture"], "Method"], [1882, 1886, ["re", "-", "ranking", "pairs"], "Task"], [1894, 1896, ["optimal", "representation"], "Method"], [1943, 1945, ["feature", "engineering"], "Method"], [1957, 1958, ["QL"], "Metric"], [1976, 1979, ["deep", "learning", "methods"], "Method"], [1980, 1982, ["Answer", "Ranking"], "Task"], [1983, 1984, ["CQA"], "Task"], [1991, 1993, ["ranking", "answers"], "Task"], [1994, 1997, ["community", "question", "answering"], "Task"], [1998, 1999, ["CQA"], "Task"], [2021, 2025, ["density", "/", "frequency", "features"], "Method"], [2026, 2028, ["ranking", "answers"], "Task"], [2044, 2049, ["answer", "re", "-", "ranking", "model"], "Method"], [2050, 2054, ["non", "-", "factoid", "questions"], "Task"], [2064, 2067, ["representations", "of", "discourse"], "Method"], [2074, 2076, ["retrieval", "model"], "Method"], [2079, 2084, ["translation", "-", "based", "language", "model"], "Method"], [2090, 2093, ["query", "likelihood", "approach"], "Method"], [2100, 2102, ["CQA", "sites"], "Material"], [2124, 2129, ["Answer", "Passage", "/", "Sentence", "Retrieval"], "Task"], [2139, 2144, ["answer", "passage", "/", "sentence", "retrieval"], "Task"], [2158, 2163, ["shallow", "and", "deeper", "syntactic", "parsers"], "Method"], [2165, 2171, ["answer", "passage", "re", "-", "ranking", "task"], "Task"], [2182, 2187, ["non", "-", "factoid", "answer", "finding"], "Task"], [2203, 2206, ["passage", "retrieval", "models"], "Method"], [2234, 2237, ["answer", "sentence", "retrieval"], "Task"], [2249, 2253, ["basic", "text", "matching", "features"], "Method"], [2255, 2259, ["learning", "to", "rank", "approach"], "Method"], [2265, 2271, ["attention", "-", "based", "neural", "matching", "model"], "Method"], [2272, 2277, ["value", "-", "shared", "weighting", "schema"], "Method"], [2279, 2280, ["learning"], "Method"], [2281, 2283, ["rank", "approaches"], "Method"], [2284, 2286, ["feature", "engineering"], "Method"], [2294, 2296, ["ranking", "answers"], "Task"], [2299, 2302, ["manual", "feature", "engineering"], "Method"], [2305, 2307, ["NLP", "parsers"], "Method"], [2316, 2322, ["Attention", "-", "based", "Neural", "Matching", "Model"], "Method"], [2332, 2333, ["aNMM"], "Method"], [2334, 2340, ["attention", "-", "based", "Neural", "Matching", "Model"], "Method"], [2425, 2428, ["ranking", "answer", "sentences"], "Task"], [2439, 2442, ["QA", "Matching", "Matrix"], "Method"], [2445, 2448, ["QA", "Matching", "Matrix"], "Method"], [2481, 2484, ["QA", "matching", "matrix"], "Method"], [2486, 2488, ["by", "matrix"], "Method"], [2593, 2595, ["word", "embeddings"], "Method"], [2600, 2603, ["deep", "neural", "network"], "Method"], [2604, 2609, ["value", "-", "shared", "weighting", "scheme"], "Method"], [2615, 2618, ["fully", "connected", "layers"], "Method"], [2623, 2629, ["hierarchical", "abstraction", "of", "the", "semantic", "matching"], "Task"], [2652, 2654, ["ranking", "score"], "Metric"], [2658, 2662, ["neural", "matching", "model", "architectures"], "Method"], [2684, 2687, ["aNMM", "-", "1"], "Method"], [2703, 2706, ["aNMM", "-", "1"], "Method"], [2709, 2713, ["value", "-", "shared", "weights"], "Method"], [2720, 2724, ["Value", "-", "shared", "Weighting"], "Method"], [2727, 2729, ["word", "embeddings"], "Method"], [2731, 2733, ["Word2Vec", "tool"], "Method"], [2740, 2743, ["English", "Wikipedia", "dump"], "Material"], [2791, 2794, ["QA", "matching", "matrix"], "Method"], [2818, 2822, ["CNN", "with", "pooling", "strategy"], "Method"], [2836, 2837, ["CNNs"], "Method"], [2839, 2844, ["position", "-", "shared", "weighting", "scheme"], "Method"], [2889, 2892, ["various", "length", "problem"], "Task"], [2918, 2920, ["ranking", "score"], "Metric"], [2925, 2930, ["value", "-", "shared", "weighting", "scheme"], "Method"], [2990, 2991, ["filters"], "Method"], [2992, 2993, ["CNN"], "Method"], [3072, 3074, ["hidden", "representation"], "Method"], [3081, 3084, ["fully", "connected", "layers"], "Method"], [3086, 3089, ["higher", "level", "representations"], "Method"], [3143, 3147, ["forward", "propagation", "prediction", "process"], "Method"], [3160, 3163, ["aNMM", "-", "1"], "Method"], [3169, 3171, ["Hidden", "Layer"], "Method"], [3254, 3257, ["neural", "network", "architectures"], "Method"], [3260, 3263, ["Question", "Attention", "Network"], "Method"], [3266, 3270, ["value", "-", "shared", "weighting"], "Method"], [3272, 3274, ["model", "component"], "Method"], [3275, 3278, ["aNMM", "-", "1"], "Method"], [3286, 3290, ["committee", "of", "neural", "networks"], "Method"], [3334, 3337, ["aNMM", "-", "1"], "Method"], [3438, 3441, ["forward", "propagation", "process"], "Method"], [3475, 3477, ["question", "attention"], "Task"], [3492, 3498, ["dot", "product", "of", "query", "word", "embedding"], "Method"], [3535, 3536, ["CNNs"], "Method"], [3537, 3538, ["BLSTM"], "Method"], [3550, 3552, ["representation", "learning"], "Method"], [3562, 3563, ["aNMM"], "Method"], [3604, 3606, ["Model", "Training"], "Method"], [3607, 3610, ["aNMM", "-", "1"], "Method"], [3668, 3673, ["pair", "-", "wise", "learning", "strategy"], "Method"], [3698, 3702, ["ranking", "-", "based", "loss"], "Metric"], [3704, 3706, ["objective", "function"], "Metric"], [3712, 3715, ["predicted", "matching", "score"], "Metric"], [3755, 3757, ["model", "parameters"], "Method"], [3758, 3761, ["back", "propagation", "algorithm"], "Method"], [3764, 3766, ["objective", "function"], "Metric"], [3825, 3828, ["stochastic", "gradient", "descent"], "Method"], [3835, 3840, ["mini", "-", "batch", "gradient", "descent"], "Method"], [3847, 3849, ["ranking", "task"], "Task"], [3852, 3854, ["learning", "rate"], "Metric"], [3857, 3860, ["adaptive", "learning", "rate"], "Metric"], [3876, 3877, ["convergence"], "Metric"], [3882, 3885, ["Deep", "Neural", "Networks"], "Method"], [3894, 3897, ["aNMM", "-", "1"], "Method"], [3921, 3924, ["neural", "network", "architecture"], "Method"], [3973, 3976, ["aNMM", "-", "2"], "Method"], [3989, 3992, ["aNMM", "-", "2"], "Method"], [3995, 3998, ["Forward", "Propagation", "Prediction"], "Method"], [3999, 4002, ["aNMM", "-", "2"], "Method"], [4010, 4012, ["neural", "network"], "Method"], [4025, 4027, ["hidden", "layer"], "Method"], [4046, 4049, ["forward", "propagation", "prediction"], "Task"], [4149, 4151, ["Back", "Propagation"], "Method"], [4152, 4154, ["Model", "Training"], "Task"], [4155, 4158, ["aNMM", "-", "2"], "Method"], [4206, 4208, ["back", "propagation"], "Method"], [4225, 4228, ["back", "propagation", "process"], "Method"], [4229, 4232, ["model", "parameter", "learning"], "Task"], [4317, 4319, ["back", "propagation"], "Method"], [4338, 4339, ["prediction"], "Task"], [4400, 4402, ["pattern", "matching"], "Method"], [4415, 4419, ["answer", "re", "-", "ranking"], "Task"], [4472, 4474, ["answer", "correctness"], "Metric"], [4487, 4490, ["TREC", "13", "set"], "Material"], [4502, 4505, ["TRAIN", "-", "ALL"], "Material"], [4539, 4540, ["TREC"], "Material"], [4556, 4558, ["model", "training"], "Task"], [4565, 4569, ["hyper", "-", "parameter", "optimization"], "Method"], [4573, 4575, ["model", "testing"], "Task"], [4599, 4601, ["data", "preprocess"], "Task"], [4604, 4607, ["tokenization", "without", "stemming"], "Method"], [4618, 4620, ["Word", "Embeddings"], "Task"], [4630, 4632, ["Word2Vec", "tool"], "Method"], [4639, 4642, ["English", "Wikipedia", "dump"], "Material"], [4646, 4650, ["skip", "-", "gram", "model"], "Method"], [4666, 4669, ["neural", "embedding", "models"], "Method"], [4746, 4749, ["aNNM", "-", "1"], "Method"], [4761, 4764, ["aNNM", "-", "2"], "Method"], [4782, 4783, ["Evaluation"], "Metric"], [4784, 4785, ["Metrics"], "Metric"], [4810, 4811, ["metrics"], "Metric"], [4814, 4817, ["Mean", "Average", "Precision"], "Metric"], [4818, 4819, ["MAP"], "Metric"], [4821, 4824, ["Mean", "Reciprocal", "Rank"], "Metric"], [4825, 4826, ["MRR"], "Metric"], [4833, 4835, ["information", "retrieval"], "Task"], [4836, 4838, ["question", "answering"], "Task"], [4850, 4851, ["MRR"], "Metric"], [4870, 4871, ["MRR"], "Metric"], [4914, 4915, ["MAP"], "Metric"], [4933, 4935, ["average", "precision"], "Metric"], [4940, 4941, ["MAP"], "Metric"], [4962, 4964, ["Model", "Learning"], "Method"], [4972, 4974, ["qualitative", "analysis"], "Task"], [4975, 4976, ["visualization"], "Task"], [4978, 4980, ["model", "learning"], "Task"], [4997, 4998, ["aNMM"], "Method"], [5014, 5017, ["aNMM", "-", "1"], "Method"], [5033, 5036, ["aNMM", "-", "1"], "Method"], [5038, 5041, ["aNMM", "-", "1"], "Method"], [5063, 5066, ["aNMM", "-", "1"], "Method"], [5073, 5076, ["aNMM", "-", "1"], "Method"], [5230, 5233, ["matching", "score", "range"], "Metric"], [5240, 5243, ["matching", "score", "range"], "Metric"], [5249, 5252, ["semantic", "matching", "scores"], "Metric"], [5256, 5258, ["answer", "correctness"], "Metric"], [5260, 5264, ["low", "semantic", "matching", "scores"], "Metric"], [5308, 5310, ["similarity", "scores"], "Metric"], [5312, 5314, ["word", "embeddings"], "Method"], [5347, 5350, ["model", "training", "process"], "Method"], [5371, 5372, ["CNN"], "Method"], [5410, 5413, ["aNMM", "-", "1"], "Method"], [5430, 5433, ["aNMM", "-", "1"], "Method"], [5463, 5466, ["aNMM", "-", "1"], "Method"], [5480, 5483, ["aNMM", "-", "1"], "Method"], [5601, 5605, ["IR", "term", "weighting", "methods"], "Method"], [5607, 5608, ["IDF"], "Method"], [5615, 5618, ["aNMM", "-", "1"], "Method"], [5619, 5622, ["aNMM", "-", "2"], "Method"], [5623, 5626, ["aNMM", "-", "IDF"], "Method"], [5638, 5639, ["IDF"], "Method"], [5662, 5663, ["IDF"], "Method"], [5692, 5693, ["aNMM"], "Method"], [5694, 5695, ["IDF"], "Method"], [5700, 5703, ["answer", "ranking", "performance"], "Metric"], [5717, 5719, ["optimization", "process"], "Method"], [5721, 5725, ["back", "propagation", "training", "process"], "Method"], [5726, 5727, ["aNMM"], "Method"], [5730, 5734, ["question", "term", "weighting", "score"], "Metric"], [5735, 5739, ["heuristic", "term", "weighting", "functions"], "Method"], [5740, 5741, ["IDF"], "Method"], [5747, 5749, ["Ranking", "Answers"], "Task"], [5751, 5752, ["Learning"], "Task"], [5770, 5771, ["aNMM"], "Method"], [5799, 5801, ["TREC", "QA"], "Material"], [5802, 5803, ["TRAIN"], "Material"], [5804, 5807, ["TRAIN", "-", "ALL"], "Material"], [5821, 5822, ["aNMM"], "Method"], [5825, 5828, ["deep", "learning", "methods"], "Method"], [5829, 5830, ["CNN"], "Method"], [5831, 5832, ["LSTM"], "Method"], [5844, 5847, ["aNMM", "-", "1"], "Method"], [5848, 5851, ["aNMM", "-", "2"], "Method"], [5855, 5856, ["MAP"], "Metric"], [5857, 5858, ["MRR"], "Metric"], [5859, 5860, ["TRAIN"], "Material"], [5861, 5866, ["TRAIN", "-", "ALL", "data", "sets"], "Material"], [5869, 5872, ["deep", "learning", "methods"], "Method"], [5881, 5884, ["aNMM", "-", "1"], "Method"], [5887, 5891, ["deep", "learning", "baseline", "method"], "Method"], [5898, 5899, ["CNN"], "Method"], [5903, 5906, ["aNMM", "-", "1"], "Method"], [5907, 5908, ["CNN"], "Method"], [5911, 5912, ["MAP"], "Metric"], [5913, 5914, ["TRAIN"], "Material"], [5917, 5918, ["MAP"], "Metric"], [5919, 5922, ["TRAIN", "-", "ALL"], "Material"], [5924, 5925, ["MRR"], "Metric"], [5934, 5937, ["aNMM", "-", "1"], "Method"], [5944, 5949, ["value", "-", "shared", "weight", "scheme"], "Method"], [5952, 5957, ["position", "-", "shared", "weight", "scheme"], "Method"], [5958, 5959, ["CNN"], "Method"], [5960, 5963, ["term", "importance", "learning"], "Method"], [5968, 5969, ["aNMM"], "Method"], [5971, 5973, ["ranking", "scores"], "Metric"], [5976, 5977, ["accuracy"], "Metric"], [5980, 5983, ["deep", "learning", "models"], "Method"], [5984, 5986, ["ranking", "answers"], "Task"], [5996, 5999, ["aNMM", "-", "1"], "Method"], [6000, 6003, ["aNMM", "-", "2"], "Method"], [6013, 6016, ["aNMM", "-", "1"], "Method"], [6021, 6024, ["aNMM", "-", "2"], "Method"], [6046, 6048, ["answer", "ranking"], "Task"], [6057, 6059, ["model", "efficiency"], "Metric"], [6060, 6063, ["aNMM", "-", "1"], "Method"], [6077, 6079, ["prediction", "accuracy"], "Metric"], [6141, 6142, ["aNMM"], "Method"], [6146, 6148, ["feature", "engineering"], "Method"], [6149, 6152, ["TRAIN", "-", "ALL"], "Material"], [6161, 6164, ["aNMM", "-", "1"], "Method"], [6165, 6168, ["aNMM", "-", "2"], "Method"], [6176, 6178, ["feature", "engineering"], "Method"], [6185, 6188, ["aNMM", "-", "1"], "Method"], [6199, 6203, ["enhanced", "lexical", "semantic", "models"], "Method"], [6204, 6207, ["aNMM", "-", "1"], "Method"], [6211, 6212, ["MAP"], "Metric"], [6216, 6217, ["MRR"], "Metric"], [6228, 6232, ["uniform", "deep", "learning", "model"], "Method"], [6242, 6244, ["feature", "engineering"], "Method"], [6252, 6253, ["aNMM"], "Method"], [6256, 6259, ["deep", "learning", "model"], "Method"], [6271, 6274, ["deep", "learning", "models"], "Method"], [6275, 6277, ["feature", "engineering"], "Method"], [6278, 6280, ["ranking", "answers"], "Task"], [6285, 6287, ["syntactic", "parsers"], "Method"], [6339, 6340, ["CNN"], "Method"], [6341, 6342, ["LSTM"], "Method"], [6343, 6345, ["answer", "ranking"], "Task"], [6353, 6355, ["predicted", "score"], "Metric"], [6356, 6359, ["aNMM", "-", "1"], "Method"], [6360, 6363, ["aNMM", "-", "2"], "Method"], [6365, 6367, ["Query", "Likelihood"], "Metric"], [6368, 6369, ["QL"], "Metric"], [6372, 6373, ["LambdaMART"], "Method"], [6384, 6385, ["LambdaMART"], "Method"], [6386, 6387, ["jforests"], "Method"], [6393, 6396, ["deep", "learning", "models"], "Method"], [6408, 6409, ["TRAIN"], "Material"], [6427, 6430, ["aNMM", "-", "1"], "Method"], [6431, 6434, ["aNMM", "-", "2"], "Method"], [6438, 6441, ["aNMM", "-", "1"], "Method"], [6443, 6444, ["CNN"], "Method"], [6462, 6464, ["ranking", "answers"], "Task"], [6468, 6469, ["MAP"], "Metric"], [6470, 6471, ["MRR"], "Metric"], [6472, 6473, ["TRAIN"], "Material"], [6474, 6477, ["TRAIN", "-", "ALL"], "Material"], [6483, 6484, ["aNMM"], "Method"], [6501, 6502, ["BM25"], "Method"], [6516, 6517, ["aNMM"], "Method"], [6519, 6520, ["QL"], "Metric"], [6522, 6523, ["aNMM"], "Method"], [6535, 6538, ["neural", "network", "models"], "Method"], [6539, 6541, ["CNN", "in"], "Method"], [6542, 6544, ["LSTM", "in"], "Method"], [6560, 6563, ["QA", "matching", "matrix"], "Method"], [6564, 6565, ["aNMM"], "Method"], [6611, 6614, ["factoid", "questions", "answering"], "Task"], [6640, 6641, ["aNMM"], "Method"], [6650, 6653, ["QA", "matching", "matrix"], "Method"], [6668, 6669, ["aNMM"], "Method"], [6680, 6681, ["aNMM"], "Method"], [6699, 6701, ["heuristic", "functions"], "Method"], [6702, 6703, ["IDF"], "Method"], [6704, 6705, ["aNMM"], "Method"], [6709, 6713, ["question", "term", "importance", "score"], "Metric"], [6724, 6726, ["optimization", "process"], "Method"], [6728, 6732, ["back", "propagation", "training", "process"], "Method"], [6733, 6734, ["aNMM"], "Method"], [6740, 6743, ["term", "weighting", "score"], "Metric"], [6744, 6745, ["IDF"], "Method"], [6748, 6749, ["aNMM"], "Method"], [6753, 6758, ["IDF", "weighted", "word", "overlap", "features"], "Method"], [6759, 6760, ["BM25"], "Method"], [6766, 6767, ["aNMM"], "Method"], [6802, 6806, ["QA", "answer", "ranking", "task"], "Task"], [6815, 6816, ["aNMM"], "Method"], [6837, 6839, ["feature", "engineering"], "Method"], [6840, 6843, ["deep", "learning", "models"], "Method"], [6850, 6851, ["aNMM"], "Method"], [6853, 6856, ["manual", "feature", "engineering"], "Method"], [6862, 6864, ["NLP", "parsers"], "Method"], [6889, 6890, ["aNMM"], "Method"], [6894, 6896, ["answer", "ranking"], "Task"], [6902, 6905, ["deep", "learning", "model"], "Method"], [6910, 6914, ["linguistic", "feature", "engineering", "methods"], "Method"], [6917, 6920, ["Parameter", "Sensitivity", "Analysis"], "Method"], [6922, 6925, ["parameter", "sensitivity", "analysis"], "Task"], [6929, 6930, ["aNMM"], "Method"], [6934, 6937, ["aNMM", "-", "1"], "Method"], [6947, 6950, ["aNMM", "-", "1"], "Method"], [6991, 6992, ["MAP"], "Metric"], [6993, 6994, ["MRR"], "Metric"], [7082, 7083, ["MAP"], "Metric"], [7084, 7085, ["MRR"], "Metric"], [7099, 7101, ["model", "complexity"], "Metric"], [7104, 7105, ["aNMM"], "Method"], [7135, 7136, ["aNMM"], "Method"], [7150, 7155, ["attention", "based", "neural", "matching", "model"], "Method"], [7156, 7160, ["ranking", "short", "answer", "text"], "Task"], [7163, 7168, ["value", "-", "shared", "weighting", "scheme"], "Method"], [7170, 7175, ["position", "-", "shared", "weighting", "scheme"], "Method"], [7182, 7186, ["question", "term", "importance", "learning"], "Method"], [7200, 7203, ["TREC", "QA", "dataset"], "Material"], [7219, 7220, ["CNN"], "Method"], [7223, 7224, ["LSTM"], "Method"], [7252, 7255, ["linguistic", "feature", "engineering"], "Method"], [7292, 7295, ["deep", "learning", "architectures"], "Method"], [7296, 7298, ["answer", "ranking"], "Task"], [7304, 7309, ["non", "-", "factoid", "question", "answering"], "Task"], [7325, 7328, ["Intelligent", "Information", "Retrieval"], "Task"], [2, 3, ["aNMM"], "Method"], [19, 21, ["question", "answering"], "Task"], [144, 146, ["question", "attention"], "Task"], [152, 153, ["TREC"], "Material"], [162, 163, ["aNMM"], "Method"], [177, 179, ["question", "answering"], "Task"], [284, 286, ["Retrieval", "models"], "Method"], [538, 539, ["TREC"], "Material"], [746, 747, ["CV"], "Task"], [1107, 1108, ["TREC"], "Material"], [1250, 1251, ["aNMM"], "Method"], [1261, 1263, ["question", "attention"], "Task"], [1288, 1289, ["TREC"], "Material"], [1331, 1333, ["question", "answering"], "Task"], [1613, 1614, ["aNMM"], "Method"], [1689, 1691, ["question", "answering"], "Task"], [1700, 1701, ["TREC"], "Material"], [1860, 1861, ["BLSTM"], "Method"], [2120, 2121, ["TREC"], "Material"], [2188, 2189, ["TREC"], "Material"], [2435, 2436, ["TREC"], "Material"], [2640, 2642, ["question", "attention"], "Task"], [2714, 2716, ["question", "attention"], "Task"], [3280, 3282, ["question", "attention"], "Task"], [3313, 3315, ["question", "attention"], "Task"], [3582, 3584, ["question", "attention"], "Task"], [4143, 4145, ["question", "attention"], "Task"], [4358, 4359, ["TREC"], "Material"], [4370, 4371, ["TREC"], "Material"], [4458, 4459, ["TRAIN"], "Material"], [4460, 4461, ["TRAIN"], "Material"], [4466, 4467, ["TRAIN"], "Material"], [4497, 4498, ["TREC"], "Material"], [4523, 4524, ["TRAIN"], "Material"], [4776, 4777, ["TREC"], "Material"], [4958, 4959, ["metrics"], "Metric"], [5204, 5207, ["matching", "score", "range"], "Metric"], [5212, 5215, ["matching", "score", "range"], "Metric"], [5476, 5478, ["question", "attention"], "Task"], [5645, 5647, ["question", "attention"], "Task"], [5688, 5690, ["question", "attention"], "Task"], [5706, 5707, ["TRAIN"], "Material"], [5964, 5966, ["question", "attention"], "Task"], [6049, 6050, ["TREC"], "Material"], [6088, 6089, ["TREC"], "Material"], [6092, 6093, ["aNMM"], "Method"], [6128, 6129, ["aNMM"], "Method"], [6410, 6411, ["TRAIN"], "Material"], [6691, 6693, ["question", "attention"], "Task"], [6818, 6819, ["TRAIN"], "Material"], [7188, 7190, ["question", "attention"], "Task"], [7204, 7205, ["TREC"], "Material"]]} {"doc_id": "3842ee1e0fdfeff936b5c49973ff21adfaaf3929", "text": ["document", ":", "Adversarial", "Discriminative", "Domain", "Adaptation", "Adversarial", "learning", "methods", "are", "a", "promising", "approach", "to", "training", "robust", "deep", "networks", ",", "and", "can", "generate", "complex", "samples", "across", "diverse", "domains", ".", "They", "also", "can", "improve", "recognition", "despite", "the", "presence", "of", "domain", "shift", "or", "dataset", "bias", ":", "several", "adversarial", "approaches", "to", "unsupervised", "domain", "adaptation", "have", "recently", "been", "introduced", ",", "which", "reduce", "the", "difference", "between", "the", "training", "and", "test", "domain", "distributions", "and", "thus", "improve", "generalization", "performance", ".", "Prior", "generative", "approaches", "show", "compelling", "visualizations", ",", "but", "are", "not", "optimal", "on", "discriminative", "tasks", "and", "can", "be", "limited", "to", "smaller", "shifts", ".", "Prior", "discriminative", "approaches", "could", "handle", "larger", "domain", "shifts", ",", "but", "imposed", "tied", "weights", "on", "the", "model", "and", "did", "not", "exploit", "a", "GAN", "-", "based", "loss", ".", "We", "first", "outline", "a", "novel", "generalized", "framework", "for", "adversarial", "adaptation", ",", "which", "subsumes", "recent", "state", "-", "of", "-", "the", "-", "art", "approaches", "as", "special", "cases", ",", "and", "we", "use", "this", "generalized", "view", "to", "better", "relate", "the", "prior", "approaches", ".", "We", "propose", "a", "previously", "unexplored", "instance", "of", "our", "general", "framework", "which", "combines", "discriminative", "modeling", ",", "untied", "weight", "sharing", ",", "and", "a", "GAN", "loss", ",", "which", "we", "call", "Adversarial", "Discriminative", "Domain", "Adaptation", "(", "ADDA", ")", ".", "We", "show", "that", "ADDA", "is", "more", "effective", "yet", "considerably", "simpler", "than", "competing", "domain", "-", "adversarial", "methods", ",", "and", "demonstrate", "the", "promise", "of", "our", "approach", "by", "exceeding", "state", "-", "of", "-", "the", "-", "art", "unsupervised", "adaptation", "results", "on", "standard", "cross", "-", "domain", "digit", "classification", "tasks", "and", "a", "new", "more", "difficult", "cross", "-", "modality", "object", "classification", "task", ".", "section", ":", "Introduction", "Deep", "convolutional", "networks", ",", "when", "trained", "on", "large", "-", "scale", "datasets", ",", "can", "learn", "representations", "which", "are", "generically", "usefull", "across", "a", "variety", "of", "tasks", "and", "visual", "domains", ".", "However", ",", "due", "to", "a", "phenomenon", "known", "as", "dataset", "bias", "or", "domain", "shift", ",", "recognition", "models", "trained", "along", "with", "these", "representations", "on", "one", "large", "dataset", "do", "not", "generalize", "well", "to", "novel", "datasets", "and", "tasks", ".", "The", "typical", "solution", "is", "to", "further", "fine", "-", "tune", "these", "networks", "on", "task", "-", "specific", "datasets", "—", "however", ",", "it", "is", "often", "prohibitively", "difficult", "and", "expensive", "to", "obtain", "enough", "labeled", "data", "to", "properly", "fine", "-", "tune", "the", "large", "number", "of", "parameters", "employed", "by", "deep", "multilayer", "networks", ".", "Domain", "adaptation", "methods", "attempt", "to", "mitigate", "the", "harmful", "effects", "of", "domain", "shift", ".", "Recent", "domain", "adaptation", "methods", "learn", "deep", "neural", "transformations", "that", "map", "both", "domains", "into", "a", "common", "feature", "space", ".", "This", "is", "generally", "achieved", "by", "optimizing", "the", "representation", "to", "minimize", "some", "measure", "of", "domain", "shift", "such", "as", "maximum", "mean", "discrepancy", "or", "correlation", "distances", ".", "An", "alternative", "is", "to", "reconstruct", "the", "target", "domain", "from", "the", "source", "representation", ".", "Adversarial", "adaptation", "methods", "have", "become", "an", "increasingly", "popular", "incarnation", "of", "this", "type", "of", "approach", "which", "seeks", "to", "minimize", "an", "approximate", "domain", "discrepancy", "distance", "through", "an", "adversarial", "objective", "with", "respect", "to", "a", "domain", "discriminator", ".", "These", "methods", "are", "closely", "related", "to", "generative", "adversarial", "learning", ",", "which", "pits", "two", "networks", "against", "each", "other", "—", "a", "generator", "and", "a", "discriminator", ".", "The", "generator", "is", "trained", "to", "produce", "images", "in", "a", "way", "that", "confuses", "the", "discriminator", ",", "which", "in", "turn", "tries", "to", "distinguish", "them", "from", "real", "image", "examples", ".", "In", "domain", "adaptation", ",", "this", "principle", "has", "been", "employed", "to", "ensure", "that", "the", "network", "can", "not", "distinguish", "between", "the", "distributions", "of", "its", "training", "and", "test", "domain", "examples", ".", "However", ",", "each", "algorithm", "makes", "different", "design", "choices", "such", "as", "whether", "to", "use", "a", "generator", ",", "which", "loss", "function", "to", "employ", ",", "or", "whether", "to", "share", "weights", "across", "domains", ".", "For", "example", ",", "share", "weights", "and", "learn", "a", "symmetric", "mapping", "of", "both", "source", "and", "target", "images", "to", "the", "shared", "feature", "space", ",", "while", "decouple", "some", "layers", "thus", "learning", "a", "partially", "asymmetric", "mapping", ".", "In", "this", "work", ",", "we", "propose", "a", "novel", "unified", "framework", "for", "adversarial", "domain", "adaptation", ",", "allowing", "us", "to", "effectively", "examine", "the", "different", "factors", "of", "variation", "between", "the", "existing", "approaches", "and", "clearly", "view", "the", "similarities", "they", "each", "share", ".", "Our", "framework", "unifies", "design", "choices", "such", "as", "weight", "-", "sharing", ",", "base", "models", ",", "and", "adversarial", "losses", "and", "subsumes", "previous", "work", ",", "while", "also", "facilitating", "the", "design", "of", "novel", "instantiations", "that", "improve", "upon", "existing", "ones", ".", "In", "particular", ",", "we", "observe", "that", "generative", "modeling", "of", "input", "image", "distributions", "is", "not", "necessary", ",", "as", "the", "ultimate", "task", "is", "to", "learn", "a", "discriminative", "representation", ".", "On", "the", "other", "hand", ",", "asymmetric", "mappings", "can", "better", "model", "the", "difference", "in", "low", "level", "features", "than", "symmetric", "ones", ".", "We", "therefore", "propose", "a", "previously", "unexplored", "unsupervised", "adversarial", "adaptation", "method", ",", "Adversarial", "Discriminative", "Domain", "Adaptation", "(", "ADDA", ")", ",", "illustrated", "in", "Figure", "[", "reference", "]", ".", "ADDA", "first", "learns", "a", "discriminative", "representation", "using", "the", "labels", "in", "the", "source", "domain", "and", "then", "a", "separate", "encoding", "that", "maps", "the", "target", "data", "to", "the", "same", "space", "using", "an", "asymmetric", "mapping", "learned", "through", "a", "domain", "-", "adversarial", "loss", ".", "Our", "approach", "is", "simple", "yet", "surprisingly", "powerful", "and", "achieves", "state", "-", "of", "-", "the", "-", "art", "visual", "adaptation", "results", "on", "the", "MNIST", ",", "USPS", ",", "and", "SVHN", "digits", "datasets", ".", "We", "also", "test", "its", "potential", "to", "bridge", "the", "gap", "between", "even", "more", "difficult", "cross", "-", "modality", "shifts", ",", "without", "requiring", "instance", "constraints", ",", "by", "transferring", "object", "classifiers", "from", "RGB", "color", "images", "to", "depth", "observations", ".", "section", ":", "Related", "work", "There", "has", "been", "extensive", "prior", "work", "on", "domain", "transfer", "learning", ",", "see", "e.g.", ",", ".", "Recent", "work", "has", "focused", "on", "transferring", "deep", "neural", "network", "representations", "from", "a", "labeled", "source", "datasets", "to", "a", "target", "domain", "where", "labeled", "data", "is", "sparse", "or", "non", "-", "existent", ".", "In", "the", "case", "of", "unlabeled", "target", "domains", "(", "the", "focus", "of", "this", "paper", ")", "the", "main", "strategy", "has", "been", "to", "guide", "feature", "learning", "by", "minimizing", "the", "difference", "between", "the", "source", "and", "target", "feature", "distributions", ".", "Several", "methods", "have", "used", "the", "Maximum", "Mean", "Discrepancy", "(", "MMD", ")", "loss", "for", "this", "purpose", ".", "MMD", "computes", "the", "norm", "of", "the", "difference", "between", "two", "domain", "means", ".", "The", "DDC", "method", "used", "MMD", "in", "addition", "to", "the", "regular", "classification", "loss", "on", "the", "source", "to", "learn", "a", "representation", "that", "is", "both", "discriminative", "and", "domain", "invariant", ".", "The", "Deep", "Adaptation", "Network", "(", "DAN", ")", "applied", "MMD", "to", "layers", "embedded", "in", "a", "reproducing", "kernel", "Hilbert", "space", ",", "effectively", "matching", "higher", "order", "statistics", "of", "the", "two", "distributions", ".", "In", "contrast", ",", "the", "deep", "Correlation", "Alignment", "(", "CORAL", ")", "method", "proposed", "to", "match", "the", "mean", "and", "covariance", "of", "the", "two", "distributions", ".", "Other", "methods", "have", "chosen", "an", "adversarial", "loss", "to", "minimize", "domain", "shift", ",", "learning", "a", "representation", "that", "is", "simultaneously", "discriminative", "of", "source", "labels", "while", "not", "being", "able", "to", "distinguish", "between", "domains", ".", "proposed", "adding", "a", "domain", "classifier", "(", "a", "single", "fully", "connected", "layer", ")", "that", "predicts", "the", "binary", "domain", "label", "of", "the", "inputs", "and", "designed", "a", "domain", "confusion", "loss", "to", "encourage", "its", "prediction", "to", "be", "as", "close", "as", "possible", "to", "a", "uniform", "distribution", "over", "binary", "labels", ".", "The", "gradient", "reversal", "algorithm", "(", "ReverseGrad", ")", "proposed", "in", "also", "treats", "domain", "invariance", "as", "a", "binary", "classification", "problem", ",", "but", "directly", "maximizes", "the", "loss", "of", "the", "domain", "classifier", "by", "reversing", "its", "gradients", ".", "DRCN", "takes", "a", "similar", "approach", "but", "also", "learns", "to", "reconstruct", "target", "domain", "images", ".", "In", "related", "work", ",", "adversarial", "learning", "has", "been", "explored", "for", "generative", "tasks", ".", "The", "Generative", "Adversarial", "Network", "(", "GAN", ")", "method", "is", "a", "generative", "deep", "model", "that", "pits", "two", "networks", "against", "one", "another", ":", "a", "generative", "model", "G", "that", "captures", "the", "data", "distribution", "and", "a", "discriminative", "model", "D", "that", "distinguishes", "between", "samples", "drawn", "from", "G", "and", "images", "drawn", "from", "the", "training", "data", "by", "predicting", "a", "binary", "label", ".", "The", "networks", "are", "trained", "jointly", "using", "backprop", "on", "the", "label", "prediction", "loss", "in", "a", "mini", "-", "max", "fashion", ":", "simultaneously", "update", "G", "to", "minimize", "the", "loss", "while", "also", "updating", "D", "to", "maximize", "the", "loss", "(", "fooling", "the", "discriminator", ")", ".", "The", "advantage", "of", "GAN", "over", "other", "generative", "methods", "is", "that", "there", "is", "no", "need", "for", "complex", "sampling", "or", "inference", "during", "training", ";", "the", "downside", "is", "that", "it", "may", "be", "difficult", "to", "train", ".", "GANs", "have", "been", "applied", "to", "generate", "natural", "images", "of", "objects", ",", "such", "as", "digits", "and", "faces", ",", "and", "have", "been", "extended", "in", "several", "ways", ".", "The", "BiGAN", "approach", "extends", "GANs", "to", "also", "learn", "the", "inverse", "mapping", "from", "the", "image", "data", "back", "into", "the", "latent", "space", ",", "and", "shows", "that", "this", "can", "learn", "features", "useful", "for", "image", "classification", "tasks", ".", "The", "conditional", "generative", "adversarial", "net", "(", "CGAN", ")", "is", "an", "extension", "of", "the", "GAN", "where", "both", "networks", "G", "and", "D", "receive", "an", "additional", "vector", "of", "information", "as", "input", ".", "This", "might", "contain", ",", "say", ",", "information", "about", "the", "class", "of", "the", "training", "example", ".", "The", "authors", "apply", "CGAN", "to", "generate", "a", "(", "possibly", "multi", "-", "modal", ")", "distribution", "of", "tag", "-", "vectors", "conditional", "on", "image", "features", ".", "Recently", "the", "CoGAN", "approach", "applied", "GANs", "to", "the", "domain", "transfer", "problem", "by", "training", "two", "GANs", "to", "generate", "the", "source", "and", "target", "images", "respectively", ".", "The", "approach", "achieves", "a", "domain", "invariant", "feature", "space", "by", "tying", "the", "high", "-", "level", "layer", "parameters", "of", "the", "two", "GANs", ",", "and", "shows", "that", "the", "same", "noise", "input", "can", "generate", "a", "corresponding", "pair", "of", "images", "from", "the", "two", "distributions", ".", "Domain", "adaptation", "was", "performed", "by", "training", "a", "classifier", "on", "the", "discriminator", "output", "and", "applied", "to", "shifts", "between", "the", "MNIST", "and", "USPS", "digit", "datasets", ".", "However", ",", "this", "approach", "relies", "on", "the", "generators", "finding", "a", "mapping", "from", "the", "shared", "high", "-", "level", "layer", "feature", "space", "to", "full", "images", "in", "both", "domains", ".", "This", "can", "work", "well", "for", "say", "digits", "which", "can", "be", "difficult", "in", "the", "case", "of", "more", "distinct", "domains", ".", "In", "this", "paper", ",", "we", "observe", "that", "modeling", "the", "image", "distributions", "is", "not", "strictly", "necessary", "to", "achieve", "domain", "adaptation", ",", "as", "long", "as", "the", "latent", "feature", "space", "is", "domain", "invariant", ",", "and", "propose", "a", "discriminative", "approach", ".", "section", ":", "Generalized", "adversarial", "adaptation", "We", "present", "a", "general", "framework", "for", "adversarial", "unsupervised", "adaptation", "methods", ".", "In", "unsupervised", "adaptation", ",", "we", "assume", "access", "to", "source", "images", "and", "labels", "drawn", "from", "a", "source", "domain", "distribution", ",", "as", "well", "as", "target", "images", "drawn", "from", "a", "target", "distribution", ",", "where", "there", "are", "no", "label", "observations", ".", "Our", "goal", "is", "to", "learn", "a", "target", "representation", ",", "and", "classifier", "that", "can", "correctly", "classify", "target", "images", "into", "one", "of", "categories", "at", "test", "time", ",", "despite", "the", "lack", "of", "in", "domain", "annotations", ".", "Since", "direct", "supervised", "learning", "on", "the", "target", "is", "not", "possible", ",", "domain", "adaptation", "instead", "learns", "a", "source", "representation", "mapping", ",", ",", "along", "with", "a", "source", "classifier", ",", ",", "and", "then", "learns", "to", "adapt", "that", "model", "for", "use", "in", "the", "target", "domain", ".", "In", "adversarial", "adaptive", "methods", ",", "the", "main", "goal", "is", "to", "regularize", "the", "learning", "of", "the", "source", "and", "target", "mappings", ",", "and", ",", "so", "as", "to", "minimize", "the", "distance", "between", "the", "empirical", "source", "and", "target", "mapping", "distributions", ":", "and", ".", "If", "this", "is", "the", "case", "then", "the", "source", "classification", "model", ",", ",", "can", "be", "directly", "applied", "to", "the", "target", "representations", ",", "elimating", "the", "need", "to", "learn", "a", "separate", "target", "classifier", "and", "instead", "setting", ",", ".", "The", "source", "classification", "model", "is", "then", "trained", "using", "the", "standard", "supervised", "loss", "below", ":", "We", "are", "now", "able", "to", "describe", "our", "full", "general", "framework", "view", "of", "adversarial", "adaptation", "approaches", ".", "We", "note", "that", "all", "approaches", "minimize", "source", "and", "target", "representation", "distances", "through", "alternating", "minimization", "between", "two", "functions", ".", "First", "a", "domain", "discriminator", ",", ",", "which", "classifies", "whether", "a", "data", "point", "is", "drawn", "from", "the", "source", "or", "the", "target", "domain", ".", "Thus", ",", "is", "optimized", "according", "to", "a", "standard", "supervised", "loss", ",", "where", "the", "labels", "indicate", "the", "origin", "domain", ",", "defined", "below", ":", "Second", ",", "the", "source", "and", "target", "mappings", "are", "optimized", "according", "to", "a", "constrained", "adversarial", "objective", ",", "whose", "particular", "instantiation", "may", "vary", "across", "methods", ".", "Thus", ",", "we", "can", "derive", "a", "generic", "formulation", "for", "domain", "adversarial", "techniques", "below", ":", "In", "the", "next", "sections", ",", "we", "demonstrate", "the", "value", "of", "our", "framework", "by", "positioning", "recent", "domain", "adversarial", "approaches", "within", "our", "framework", ".", "We", "describe", "the", "potential", "mapping", "structure", ",", "mapping", "optimization", "constraints", "(", ")", "choices", "and", "finally", "choices", "of", "adversarial", "mapping", "loss", ",", ".", "subsection", ":", "Source", "and", "target", "mappings", "In", "the", "case", "of", "learning", "a", "source", "mapping", "alone", "it", "is", "clear", "that", "supervised", "training", "through", "a", "latent", "space", "discriminative", "loss", "using", "the", "known", "labels", "results", "in", "the", "best", "representation", "for", "final", "source", "recognition", ".", "However", ",", "given", "that", "our", "target", "domain", "is", "unlabeled", ",", "it", "remains", "an", "open", "question", "how", "best", "to", "minimize", "the", "distance", "between", "the", "source", "and", "target", "mappings", ".", "Thus", "the", "first", "choice", "to", "be", "made", "is", "in", "the", "particular", "parameterization", "of", "these", "mappings", ".", "Because", "unsupervised", "domain", "adaptation", "generally", "considers", "target", "discriminative", "tasks", "such", "as", "classification", ",", "previous", "adaptation", "methods", "have", "generally", "relied", "on", "adapting", "discriminative", "models", "between", "domains", ".", "With", "a", "discriminative", "base", "model", ",", "input", "images", "are", "mapped", "into", "a", "feature", "space", "that", "is", "useful", "for", "a", "discriminative", "task", "such", "as", "image", "classification", ".", "For", "example", ",", "in", "the", "case", "of", "digit", "classification", "this", "may", "be", "the", "standard", "LeNet", "model", ".", "However", ",", "Liu", "and", "Tuzel", "achieve", "state", "of", "the", "art", "results", "on", "unsupervised", "MNIST", "-", "USPS", "using", "two", "generative", "adversarial", "networks", ".", "These", "generative", "models", "use", "random", "noise", "as", "input", "to", "generate", "samples", "in", "image", "space", "—", "generally", ",", "an", "intermediate", "feature", "of", "an", "adversarial", "discriminator", "is", "then", "used", "as", "a", "feature", "for", "training", "a", "task", "-", "specific", "classifier", ".", "Once", "the", "mapping", "parameterization", "is", "determined", "for", "the", "source", ",", "we", "must", "decide", "how", "to", "parametrize", "the", "target", "mapping", ".", "In", "general", ",", "the", "target", "mapping", "almost", "always", "matches", "the", "source", "in", "terms", "of", "the", "specific", "functional", "layer", "(", "architecture", ")", ",", "but", "different", "methods", "have", "proposed", "various", "regularization", "techniques", ".", "All", "methods", "initialize", "the", "target", "mapping", "parameters", "with", "the", "source", ",", "but", "different", "methods", "choose", "different", "constraints", "between", "the", "source", "and", "target", "mappings", ",", ".", "The", "goal", "is", "to", "make", "sure", "that", "the", "target", "mapping", "is", "set", "so", "as", "to", "minimize", "the", "distance", "between", "the", "source", "and", "target", "domains", "under", "their", "respective", "mappings", ",", "while", "crucially", "also", "maintaining", "a", "target", "mapping", "that", "is", "category", "discriminative", ".", "Consider", "a", "layered", "representations", "where", "each", "layer", "parameters", "are", "denoted", "as", ",", "or", ",", "for", "a", "given", "set", "of", "equivalent", "layers", ",", ".", "Then", "the", "space", "of", "constraints", "explored", "in", "the", "literature", "can", "be", "described", "through", "layerwise", "equality", "constraints", "as", "follows", ":", "where", "each", "individual", "layer", "can", "be", "constrained", "independently", ".", "A", "very", "common", "form", "of", "constraint", "is", "source", "and", "target", "layerwise", "equality", ":", "It", "is", "also", "common", "to", "leave", "layers", "unconstrained", ".", "These", "equality", "constraints", "can", "easily", "be", "imposed", "within", "a", "convolutional", "network", "framework", "through", "weight", "sharing", ".", "For", "many", "prior", "adversarial", "adaptation", "methods", ",", "all", "layers", "are", "constrained", ",", "thus", "enforcing", "exact", "source", "and", "target", "mapping", "consistency", ".", "Learning", "a", "symmetric", "transformation", "reduces", "the", "number", "of", "parameters", "in", "the", "model", "and", "ensures", "that", "the", "mapping", "used", "for", "the", "target", "is", "discriminative", "at", "least", "when", "applied", "to", "the", "source", "domain", ".", "However", ",", "this", "may", "make", "the", "optimization", "poorly", "conditioned", ",", "since", "the", "same", "network", "must", "handle", "images", "from", "two", "separate", "domains", ".", "An", "alternative", "approach", "is", "instead", "to", "learn", "an", "asymmetric", "transformation", "with", "only", "a", "subset", "of", "the", "layers", "constrained", ",", "thus", "enforcing", "partial", "alignment", ".", "Rozantsev", "et", "al", ".", "showed", "that", "partially", "shared", "weights", "can", "lead", "to", "effective", "adaptation", "in", "both", "supervised", "and", "unsupervised", "settings", ".", "As", "a", "result", ",", "some", "recent", "methods", "have", "favored", "untying", "weights", "(", "fully", "or", "partially", ")", "between", "the", "two", "domains", ",", "allowing", "models", "to", "learn", "parameters", "for", "each", "domain", "individually", ".", "subsection", ":", "Adversarial", "losses", "Once", "we", "have", "decided", "on", "a", "parametrization", "of", ",", "we", "employ", "an", "adversarial", "loss", "to", "learn", "the", "actual", "mapping", ".", "There", "are", "various", "different", "possible", "choices", "of", "adversarial", "loss", "functions", ",", "each", "of", "which", "have", "their", "own", "unique", "use", "cases", ".", "All", "adversarial", "losses", "train", "the", "adversarial", "discriminator", "using", "a", "standard", "classification", "loss", ",", ",", "previously", "stated", "in", "Equation", "[", "reference", "]", ".", "However", ",", "they", "differ", "in", "the", "loss", "used", "to", "train", "the", "mapping", ",", ".", "The", "gradient", "reversal", "layer", "of", "optimizes", "the", "mapping", "to", "maximize", "the", "discriminator", "loss", "directly", ":", "This", "optimization", "corresponds", "to", "the", "true", "minimax", "objective", "for", "generative", "adversarial", "networks", ".", "However", ",", "this", "objective", "can", "be", "problematic", ",", "since", "early", "on", "during", "training", "the", "discriminator", "converges", "quickly", ",", "causing", "the", "gradient", "to", "vanish", ".", "When", "training", "GANs", ",", "rather", "than", "directly", "using", "the", "minimax", "loss", ",", "it", "is", "typical", "to", "train", "the", "generator", "with", "the", "standard", "loss", "function", "with", "inverted", "labels", ".", "This", "splits", "the", "optimization", "into", "two", "independent", "objectives", ",", "one", "for", "the", "generator", "and", "one", "for", "the", "discriminator", ",", "where", "remains", "unchanged", ",", "but", "becomes", ":", "This", "objective", "has", "the", "same", "fixed", "-", "point", "properties", "as", "the", "minimax", "loss", "but", "provides", "stronger", "gradients", "to", "the", "target", "mapping", ".", "We", "refer", "to", "this", "modified", "loss", "function", "as", "the", "“", "GAN", "loss", "function", "”", "for", "the", "remainder", "of", "this", "paper", ".", "Note", "that", ",", "in", "this", "setting", ",", "we", "use", "independent", "mappings", "for", "source", "and", "target", "and", "learn", "only", "adversarially", ".", "This", "mimics", "the", "GAN", "setting", ",", "where", "the", "real", "image", "distribution", "remains", "fixed", ",", "and", "the", "generating", "distribution", "is", "learned", "to", "match", "it", ".", "The", "GAN", "loss", "function", "is", "the", "standard", "choice", "in", "the", "setting", "where", "the", "generator", "is", "attempting", "to", "mimic", "another", "unchanging", "distribution", ".", "However", ",", "in", "the", "setting", "where", "both", "distributions", "are", "changing", ",", "this", "objective", "will", "lead", "to", "oscillation", "—", "when", "the", "mapping", "converges", "to", "its", "optimum", ",", "the", "discriminator", "can", "simply", "flip", "the", "sign", "of", "its", "prediction", "in", "response", ".", "Tzeng", "et", "al", ".", "instead", "proposed", "the", "domain", "confusion", "objective", ",", "under", "which", "the", "mapping", "is", "trained", "using", "a", "cross", "-", "entropy", "loss", "function", "against", "a", "uniform", "distribution", ":", "This", "loss", "ensures", "that", "the", "adversarial", "discriminator", "views", "the", "two", "domains", "identically", ".", "section", ":", "Adversarial", "discriminative", "domain", "adaptation", "The", "benefit", "of", "our", "generalized", "framework", "for", "domain", "adversarial", "methods", "is", "that", "it", "directly", "enables", "the", "development", "of", "novel", "adaptive", "methods", ".", "In", "fact", ",", "designing", "a", "new", "method", "has", "now", "been", "simplified", "to", "the", "space", "of", "making", "three", "design", "choices", ":", "whether", "to", "use", "a", "generative", "or", "discriminative", "base", "model", ",", "whether", "to", "tie", "or", "untie", "the", "weights", ",", "and", "which", "adversarial", "learning", "objective", "to", "use", ".", "In", "light", "of", "this", "view", "we", "can", "summarize", "our", "method", ",", "adversarial", "discriminative", "domain", "adaptation", "(", "ADDA", ")", ",", "as", "well", "as", "its", "connection", "to", "prior", "work", ",", "according", "to", "our", "choices", "(", "see", "Table", "[", "reference", "]", "“", "ADDA", "”", ")", ".", "Specifically", ",", "we", "use", "a", "discriminative", "base", "model", ",", "unshared", "weights", ",", "and", "the", "standard", "GAN", "loss", ".", "We", "illustrate", "our", "overall", "sequential", "training", "procedure", "in", "Figure", "[", "reference", "]", ".", "First", ",", "we", "choose", "a", "discriminative", "base", "model", ",", "as", "we", "hypothesize", "that", "much", "of", "the", "parameters", "required", "to", "generate", "convincing", "in", "-", "domain", "samples", "are", "irrelevant", "for", "discriminative", "adaptation", "tasks", ".", "Most", "prior", "adversarial", "adaptive", "methods", "optimize", "directly", "in", "a", "discriminative", "space", "for", "this", "reason", ".", "One", "counter", "-", "example", "is", "CoGANs", ".", "However", ",", "this", "method", "has", "only", "shown", "dominance", "in", "settings", "where", "the", "source", "and", "target", "domain", "are", "very", "similar", "such", "as", "MNIST", "and", "USPS", ",", "and", "in", "our", "experiments", "we", "have", "had", "difficulty", "getting", "it", "to", "converge", "for", "larger", "distribution", "shifts", ".", "Next", ",", "we", "choose", "to", "allow", "independent", "source", "and", "target", "mappings", "by", "untying", "the", "weights", ".", "This", "is", "a", "more", "flexible", "learing", "paradigm", "as", "it", "allows", "more", "domain", "specific", "feature", "extraction", "to", "be", "learned", ".", "However", ",", "note", "that", "the", "target", "domain", "has", "no", "label", "access", ",", "and", "thus", "without", "weight", "sharing", "a", "target", "model", "may", "quickly", "learn", "a", "degenerate", "solution", "if", "we", "do", "not", "take", "care", "with", "proper", "initialization", "and", "training", "procedures", ".", "Therefore", ",", "we", "use", "the", "pre", "-", "trained", "source", "model", "as", "an", "intitialization", "for", "the", "target", "representation", "space", "and", "fix", "the", "source", "model", "during", "adversarial", "training", ".", "In", "doing", "so", ",", "we", "are", "effectively", "learning", "an", "asymmetric", "mapping", ",", "in", "which", "we", "modify", "the", "target", "model", "so", "as", "to", "match", "the", "source", "distribution", ".", "This", "is", "most", "similar", "to", "the", "original", "generative", "adversarial", "learning", "setting", ",", "where", "a", "generated", "space", "is", "updated", "until", "it", "is", "indistinguishable", "with", "a", "fixed", "real", "space", ".", "Therefore", ",", "we", "choose", "the", "inverted", "label", "GAN", "loss", "described", "in", "the", "previous", "section", ".", "Our", "proposed", "method", ",", "ADDA", ",", "thus", "corresponds", "to", "the", "following", "unconstrained", "optimization", ":", "We", "choose", "to", "optimize", "this", "objective", "in", "stages", ".", "We", "begin", "by", "optimizing", "over", "and", "by", "training", "using", "the", "labeled", "source", "data", ",", "and", ".", "Because", "we", "have", "opted", "to", "leave", "fixed", "while", "learning", ",", "we", "can", "thus", "optimize", "and", "without", "revisiting", "the", "first", "objective", "term", ".", "A", "summary", "of", "this", "entire", "training", "process", "is", "provided", "in", "Figure", "[", "reference", "]", ".", "We", "note", "that", "the", "unified", "framework", "presented", "in", "the", "previous", "section", "has", "enabled", "us", "to", "compare", "prior", "domain", "adversarial", "methods", "and", "make", "informed", "decisions", "about", "the", "different", "factors", "of", "variation", ".", "Through", "this", "framework", "we", "are", "able", "to", "motivate", "a", "novel", "domain", "adaptation", "method", ",", "ADDA", ",", "and", "offer", "insight", "into", "our", "design", "decisions", ".", "In", "the", "next", "section", "we", "demonstrate", "promising", "results", "on", "unsupervised", "adaptation", "benchmark", "tasks", ",", "studying", "adaptation", "across", "digits", "and", "across", "modalities", ".", "section", ":", "Experiments", "We", "now", "evaluate", "ADDA", "for", "unsupervised", "classification", "adaptation", "across", "four", "different", "domain", "shifts", ".", "We", "explore", "three", "digits", "datasets", "of", "varying", "difficulty", ":", "MNIST", ",", "USPS", ",", "and", "SVHN", ".", "We", "additionally", "evaluate", "on", "the", "NYUD", "dataset", "to", "study", "adaptation", "across", "modalities", ".", "Example", "images", "from", "all", "experimental", "datasets", "are", "provided", "in", "Figure", "[", "reference", "]", ".", "For", "the", "case", "of", "digit", "adaptation", ",", "we", "compare", "against", "multiple", "state", "-", "of", "-", "the", "-", "art", "unsupervised", "adaptation", "methods", ",", "all", "based", "upon", "domain", "adversarial", "learning", "objectives", ".", "In", "3", "of", "4", "of", "our", "experimental", "setups", ",", "our", "method", "outperforms", "all", "competing", "approaches", ",", "and", "in", "the", "last", "domain", "shift", "studied", ",", "our", "approach", "outperforms", "all", "but", "one", "competing", "approach", ".", "We", "also", "validate", "our", "model", "on", "a", "real", "-", "world", "modality", "adaptation", "task", "using", "the", "NYU", "depth", "dataset", ".", "Despite", "a", "large", "domain", "shift", "between", "the", "RGB", "and", "depth", "modalities", ",", "ADDA", "learns", "a", "useful", "depth", "representation", "without", "any", "labeled", "depth", "data", "and", "improves", "over", "the", "nonadaptive", "baseline", "by", "over", "50", "%", "(", "relative", ")", ".", "subsection", ":", "MNIST", ",", "USPS", ",", "and", "SVHN", "digits", "datasets", "We", "experimentally", "validate", "our", "proposed", "method", "in", "an", "unsupervised", "adaptation", "task", "between", "the", "MNIST", ",", "USPS", ",", "and", "SVHN", "digits", "datasets", ",", "which", "consist", "10", "classes", "of", "digits", ".", "Example", "images", "from", "each", "dataset", "are", "visualized", "in", "Figure", "[", "reference", "]", "and", "Table", "[", "reference", "]", ".", "For", "adaptation", "between", "MNIST", "and", "USPS", ",", "we", "follow", "the", "training", "protocol", "established", "in", ",", "sampling", "2000", "images", "from", "MNIST", "and", "1800", "from", "USPS", ".", "For", "adaptation", "between", "SVHN", "and", "MNIST", ",", "we", "use", "the", "full", "training", "sets", "for", "comparison", "against", ".", "All", "experiments", "are", "performed", "in", "the", "unsupervised", "settings", ",", "where", "labels", "in", "the", "target", "domain", "are", "withheld", ",", "and", "we", "consider", "adaptation", "in", "three", "directions", ":", "MNIST", "USPS", ",", "USPS", "MNIST", ",", "and", "SVHN", "MNIST", ".", "For", "these", "experiments", ",", "we", "use", "the", "simple", "modified", "LeNet", "architecture", "provided", "in", "the", "Caffe", "source", "code", ".", "When", "training", "with", "ADDA", ",", "our", "adversarial", "discriminator", "consists", "of", "3", "fully", "connected", "layers", ":", "two", "layers", "with", "500", "hidden", "units", "followed", "by", "the", "final", "discriminator", "output", ".", "Each", "of", "the", "500", "-", "unit", "layers", "uses", "a", "ReLU", "activation", "function", ".", "Results", "of", "our", "experiment", "are", "provided", "in", "Table", "[", "reference", "]", ".", "On", "the", "easier", "MNIST", "and", "USPS", "shifts", "ADDA", "achieves", "comparable", "performance", "to", "the", "current", "state", "-", "of", "-", "the", "-", "art", ",", "CoGANs", ",", "despite", "being", "a", "considerably", "simpler", "model", ".", "This", "provides", "compelling", "evidence", "that", "the", "machinery", "required", "to", "generate", "images", "is", "largely", "irrelevant", "to", "enabling", "effective", "adaptation", ".", "Additionally", ",", "we", "show", "convincing", "results", "on", "the", "challenging", "SVHN", "and", "MNIST", "task", "in", "comparison", "to", "other", "methods", ",", "indicating", "that", "our", "method", "has", "the", "potential", "to", "generalize", "to", "a", "variety", "of", "settings", ".", "In", "contrast", ",", "we", "were", "unable", "to", "get", "CoGANs", "to", "converge", "on", "SVHN", "and", "MNIST", "—", "because", "the", "domains", "are", "so", "disparate", ",", "we", "were", "unable", "to", "train", "coupled", "generators", "for", "them", ".", "subsection", ":", "Modality", "adaptation", "bathtub", "bed", "bookshelf", "box", "chair", "counter", "desk", "door", "dresser", "garbage", "bin", "lamp", "monitor", "night", "stand", "pillow", "sink", "sofa", "table", "television", "toilet", "overall", "We", "use", "the", "NYU", "depth", "dataset", ",", "which", "contains", "bounding", "box", "annotations", "for", "19", "object", "classes", "in", "1449", "images", "from", "indoor", "scenes", ".", "The", "dataset", "is", "split", "into", "a", "train", "(", "381", "images", ")", ",", "val", "(", "414", "images", ")", "and", "test", "(", "654", ")", "sets", ".", "To", "perform", "our", "cross", "-", "modality", "adaptation", ",", "we", "first", "crop", "out", "tight", "bounding", "boxes", "around", "instances", "of", "these", "19", "classes", "present", "in", "the", "dataset", "and", "evaluate", "on", "a", "19", "-", "way", "classification", "task", "over", "object", "crops", ".", "In", "order", "to", "ensure", "that", "the", "same", "instance", "is", "not", "seen", "in", "both", "domains", ",", "we", "use", "the", "RGB", "images", "from", "the", "train", "split", "as", "the", "source", "domain", "and", "the", "depth", "images", "from", "the", "val", "split", "as", "the", "target", "domain", ".", "This", "corresponds", "to", "2", ",", "186", "labeled", "source", "images", "and", "2", ",", "401", "unlabeled", "target", "images", ".", "Figure", "[", "reference", "]", "visualizes", "samples", "from", "each", "of", "the", "two", "domains", ".", "We", "consider", "the", "task", "of", "adaptation", "between", "these", "RGB", "and", "HHA", "encoded", "depth", "images", ",", "using", "them", "as", "source", "and", "target", "domains", "respectively", ".", "Because", "the", "bounding", "boxes", "are", "tight", "and", "relatively", "low", "resolution", ",", "accurate", "classification", "is", "quite", "difficult", ",", "even", "when", "evaluating", "in", "-", "domain", ".", "In", "addition", ",", "the", "dataset", "has", "very", "few", "examples", "for", "certain", "classes", ",", "such", "as", "toilet", "and", "bathtub", ",", "which", "directly", "translates", "to", "reduced", "classification", "performance", ".", "For", "this", "experiment", ",", "our", "base", "architecture", "is", "the", "VGG", "-", "16", "architecture", ",", "initializing", "from", "weights", "pretrained", "on", "ImageNet", ".", "This", "network", "is", "then", "fully", "fine", "-", "tuned", "on", "the", "source", "domain", "for", "20", ",", "000", "iterations", "using", "a", "batch", "size", "of", "128", ".", "When", "training", "with", "ADDA", ",", "the", "adversarial", "discriminator", "consists", "of", "three", "additional", "fully", "connected", "layers", ":", "1024", "hidden", "units", ",", "2048", "hidden", "units", ",", "then", "the", "adversarial", "discriminator", "output", ".", "With", "the", "exception", "of", "the", "output", ",", "these", "additionally", "fully", "connected", "layers", "use", "a", "ReLU", "activation", "function", ".", "ADDA", "training", "then", "proceeds", "for", "another", "20", ",", "000", "iterations", ",", "again", "with", "a", "batch", "size", "of", "128", ".", "We", "find", "that", "our", "method", ",", "ADDA", ",", "greatly", "improves", "classification", "accuracy", "for", "this", "task", ".", "For", "certain", "categories", ",", "like", "counter", ",", "classification", "accuracy", "goes", "from", "2.9", "%", "under", "the", "source", "only", "baseline", "up", "to", "44.7", "%", "after", "adaptation", ".", "In", "general", ",", "average", "accuracy", "across", "all", "classes", "improves", "significantly", "from", "13.9", "%", "to", "21.1", "%", ".", "However", ",", "not", "all", "classes", "improve", ".", "Three", "classes", "have", "no", "correctly", "labeled", "target", "images", "before", "adaptation", ",", "and", "adaptation", "is", "unable", "to", "recover", "performance", "on", "these", "classes", ".", "Additionally", ",", "the", "classes", "of", "pillow", "and", "nightstand", "suffer", "performance", "loss", "after", "adaptation", ".", "For", "additional", "insight", "on", "what", "effect", "ADDA", "has", "on", "classification", ",", "Figure", "[", "reference", "]", "plots", "confusion", "matrices", "before", "adaptation", ",", "after", "adaptation", ",", "and", "in", "the", "hypothetical", "best", "-", "case", "scenario", "where", "the", "target", "labels", "are", "present", ".", "Examining", "the", "confusion", "matrix", "for", "the", "source", "only", "baseline", "reveals", "that", "the", "domain", "shift", "is", "quite", "large", "—", "as", "a", "result", ",", "the", "network", "is", "poorly", "conditioned", "and", "incorrectly", "predicts", "pillow", "for", "the", "majority", "of", "the", "dataset", ".", "This", "tendency", "to", "output", "pillow", "also", "explains", "why", "the", "source", "only", "model", "achieves", "such", "abnormally", "high", "accuracy", "on", "the", "pillow", "class", ",", "despite", "poor", "performance", "on", "the", "rest", "of", "the", "classes", ".", "In", "contrast", ",", "the", "classifier", "trained", "using", "ADDA", "predicts", "a", "much", "wider", "variety", "of", "classes", ".", "This", "leads", "to", "decreased", "accuracy", "for", "the", "pillow", "class", ",", "but", "significantly", "higher", "accuracies", "for", "many", "of", "the", "other", "classes", ".", "Additionally", ",", "comparison", "with", "the", "“", "train", "on", "target", "”", "model", "reveals", "that", "many", "of", "the", "mistakes", "the", "ADDA", "model", "makes", "are", "reasonable", ",", "such", "as", "confusion", "between", "the", "chair", "and", "table", "classes", ",", "indicating", "that", "the", "ADDA", "model", "is", "learning", "a", "useful", "representation", "on", "depth", "images", ".", "section", ":", "Conclusion", "We", "have", "proposed", "a", "unified", "framework", "for", "unsupervised", "domain", "adaptation", "techniques", "based", "on", "adversarial", "learning", "objectives", ".", "Our", "framework", "provides", "a", "simplified", "and", "cohesive", "view", "by", "which", "we", "may", "understand", "and", "connect", "the", "similarities", "and", "differences", "between", "recently", "proposed", "adaptation", "methods", ".", "Through", "this", "comparison", ",", "we", "are", "able", "to", "understand", "the", "benefits", "and", "key", "ideas", "from", "each", "approach", "and", "to", "combine", "these", "strategies", "into", "a", "new", "adaptation", "method", ",", "ADDA", ".", "We", "present", "evaluation", "across", "four", "domain", "shifts", "for", "our", "unsupervised", "adaptation", "approach", ".", "Our", "method", "generalizes", "well", "across", "a", "variety", "of", "tasks", ",", "achieving", "strong", "results", "on", "benchmark", "adaptation", "datasets", "as", "well", "as", "a", "challenging", "cross", "-", "modality", "adaptation", "task", ".", "Additional", "analysis", "indicates", "that", "the", "representations", "learned", "via", "ADDA", "resemble", "features", "learned", "with", "supervisory", "data", "in", "the", "target", "domain", "much", "more", "closely", "than", "unadapted", "features", ",", "providing", "further", "evidence", "that", "ADDA", "is", "effective", "at", "partially", "undoing", "the", "effects", "of", "domain", "shift", ".", "bibliography", ":", "References"], "entity": [[2, 6, ["Adversarial", "Discriminative", "Domain", "Adaptation"], "Method"], [6, 9, ["Adversarial", "learning", "methods"], "Method"], [15, 18, ["robust", "deep", "networks"], "Task"], [32, 33, ["recognition"], "Task"], [44, 46, ["adversarial", "approaches"], "Method"], [47, 50, ["unsupervised", "domain", "adaptation"], "Task"], [69, 70, ["generalization"], "Task"], [73, 75, ["generative", "approaches"], "Method"], [84, 86, ["discriminative", "tasks"], "Task"], [95, 97, ["discriminative", "approaches"], "Method"], [125, 127, ["generalized", "framework"], "Method"], [128, 130, ["adversarial", "adaptation"], "Task"], [171, 173, ["discriminative", "modeling"], "Method"], [174, 177, ["untied", "weight", "sharing"], "Method"], [186, 190, ["Adversarial", "Discriminative", "Domain", "Adaptation"], "Method"], [191, 192, ["ADDA"], "Method"], [197, 198, ["ADDA"], "Method"], [206, 210, ["domain", "-", "adversarial", "methods"], "Method"], [227, 229, ["unsupervised", "adaptation"], "Method"], [232, 238, ["cross", "-", "domain", "digit", "classification", "tasks"], "Task"], [243, 249, ["cross", "-", "modality", "object", "classification", "task"], "Task"], [253, 256, ["Deep", "convolutional", "networks"], "Method"], [295, 297, ["recognition", "models"], "Method"], [359, 362, ["deep", "multilayer", "networks"], "Method"], [363, 366, ["Domain", "adaptation", "methods"], "Method"], [377, 380, ["domain", "adaptation", "methods"], "Method"], [381, 384, ["deep", "neural", "transformations"], "Method"], [411, 414, ["maximum", "mean", "discrepancy"], "Metric"], [415, 417, ["correlation", "distances"], "Metric"], [428, 430, ["source", "representation"], "Method"], [431, 434, ["Adversarial", "adaptation", "methods"], "Method"], [451, 454, ["domain", "discrepancy", "distance"], "Metric"], [462, 464, ["domain", "discriminator"], "Method"], [471, 474, ["generative", "adversarial", "learning"], "Method"], [484, 485, ["generator"], "Method"], [487, 488, ["discriminator"], "Method"], [517, 519, ["domain", "adaptation"], "Task"], [558, 559, ["generator"], "Method"], [582, 584, ["symmetric", "mapping"], "Method"], [615, 617, ["unified", "framework"], "Method"], [618, 621, ["adversarial", "domain", "adaptation"], "Task"], [652, 655, ["weight", "-", "sharing"], "Method"], [656, 658, ["base", "models"], "Method"], [660, 662, ["adversarial", "losses"], "Method"], [687, 693, ["generative", "modeling", "of", "input", "image", "distributions"], "Method"], [705, 707, ["discriminative", "representation"], "Method"], [734, 738, ["unsupervised", "adversarial", "adaptation", "method"], "Method"], [739, 743, ["Adversarial", "Discriminative", "Domain", "Adaptation"], "Method"], [744, 745, ["ADDA"], "Method"], [754, 755, ["ADDA"], "Method"], [758, 760, ["discriminative", "representation"], "Method"], [783, 785, ["asymmetric", "mapping"], "Method"], [788, 792, ["domain", "-", "adversarial", "loss"], "Method"], [809, 811, ["visual", "adaptation"], "Task"], [814, 815, ["MNIST"], "Material"], [816, 817, ["USPS"], "Material"], [819, 820, ["SVHN"], "Material"], [848, 850, ["object", "classifiers"], "Method"], [869, 872, ["domain", "transfer", "learning"], "Task"], [883, 887, ["deep", "neural", "network", "representations"], "Method"], [910, 913, ["unlabeled", "target", "domains"], "Task"], [927, 929, ["feature", "learning"], "Task"], [946, 949, ["Maximum", "Mean", "Discrepancy"], "Metric"], [950, 951, ["MMD"], "Metric"], [957, 958, ["MMD"], "Metric"], [970, 972, ["DDC", "method"], "Method"], [973, 974, ["MMD"], "Metric"], [997, 1000, ["Deep", "Adaptation", "Network"], "Method"], [1001, 1002, ["DAN"], "Method"], [1004, 1005, ["MMD"], "Metric"], [1010, 1014, ["reproducing", "kernel", "Hilbert", "space"], "Method"], [1030, 1032, ["Correlation", "Alignment"], "Method"], [1033, 1034, ["CORAL"], "Method"], [1053, 1055, ["adversarial", "loss"], "Method"], [1057, 1059, ["domain", "shift"], "Task"], [1082, 1084, ["domain", "classifier"], "Method"], [1087, 1090, ["fully", "connected", "layer"], "Method"], [1125, 1128, ["gradient", "reversal", "algorithm"], "Method"], [1129, 1130, ["ReverseGrad"], "Method"], [1135, 1137, ["domain", "invariance"], "Task"], [1139, 1142, ["binary", "classification", "problem"], "Task"], [1150, 1152, ["domain", "classifier"], "Method"], [1157, 1158, ["DRCN"], "Method"], [1175, 1177, ["adversarial", "learning"], "Method"], [1181, 1183, ["generative", "tasks"], "Task"], [1185, 1188, ["Generative", "Adversarial", "Network"], "Method"], [1189, 1190, ["GAN"], "Method"], [1194, 1197, ["generative", "deep", "model"], "Method"], [1206, 1208, ["generative", "model"], "Method"], [1216, 1219, ["discriminative", "model", "D"], "Method"], [1245, 1246, ["backprop"], "Method"], [1248, 1251, ["label", "prediction", "loss"], "Task"], [1253, 1257, ["mini", "-", "max", "fashion"], "Method"], [1282, 1283, ["GAN"], "Method"], [1285, 1287, ["generative", "methods"], "Method"], [1297, 1298, ["inference"], "Task"], [1299, 1300, ["training"], "Task"], [1312, 1313, ["GANs"], "Method"], [1318, 1322, ["natural", "images", "of", "objects"], "Task"], [1338, 1340, ["BiGAN", "approach"], "Method"], [1341, 1342, ["GANs"], "Method"], [1367, 1370, ["image", "classification", "tasks"], "Task"], [1372, 1376, ["conditional", "generative", "adversarial", "net"], "Method"], [1377, 1378, ["CGAN"], "Method"], [1384, 1385, ["GAN"], "Method"], [1418, 1419, ["CGAN"], "Method"], [1440, 1442, ["CoGAN", "approach"], "Method"], [1443, 1444, ["GANs"], "Method"], [1446, 1449, ["domain", "transfer", "problem"], "Task"], [1452, 1453, ["GANs"], "Method"], [1481, 1482, ["GANs"], "Method"], [1502, 1504, ["Domain", "adaptation"], "Method"], [1509, 1510, ["classifier"], "Method"], [1520, 1521, ["MNIST"], "Material"], [1522, 1523, ["USPS"], "Material"], [1589, 1591, ["domain", "adaptation"], "Task"], [1606, 1608, ["discriminative", "approach"], "Method"], [1611, 1614, ["Generalized", "adversarial", "adaptation"], "Method"], [1620, 1624, ["adversarial", "unsupervised", "adaptation", "methods"], "Method"], [1626, 1628, ["unsupervised", "adaptation"], "Task"], [1668, 1670, ["target", "representation"], "Method"], [1672, 1673, ["classifier"], "Method"], [1697, 1699, ["supervised", "learning"], "Task"], [1706, 1708, ["domain", "adaptation"], "Task"], [1711, 1714, ["source", "representation", "mapping"], "Method"], [1719, 1721, ["source", "classifier"], "Method"], [1738, 1741, ["adversarial", "adaptive", "methods"], "Method"], [1783, 1786, ["source", "classification", "model"], "Method"], [1804, 1806, ["target", "classifier"], "Method"], [1812, 1815, ["source", "classification", "model"], "Method"], [1821, 1823, ["supervised", "loss"], "Method"], [1837, 1840, ["adversarial", "adaptation", "approaches"], "Method"], [1853, 1855, ["alternating", "minimization"], "Method"], [1861, 1863, ["domain", "discriminator"], "Method"], [1915, 1918, ["constrained", "adversarial", "objective"], "Method"], [1933, 1935, ["generic", "formulation"], "Method"], [1936, 1939, ["domain", "adversarial", "techniques"], "Method"], [1956, 1959, ["domain", "adversarial", "approaches"], "Method"], [1997, 1999, ["source", "mapping"], "Task"], [2004, 2006, ["supervised", "training"], "Method"], [2008, 2012, ["latent", "space", "discriminative", "loss"], "Method"], [2023, 2025, ["source", "recognition"], "Task"], [2071, 2074, ["unsupervised", "domain", "adaptation"], "Task"], [2076, 2079, ["target", "discriminative", "tasks"], "Task"], [2081, 2082, ["classification"], "Task"], [2084, 2086, ["adaptation", "methods"], "Method"], [2091, 2093, ["discriminative", "models"], "Method"], [2098, 2101, ["discriminative", "base", "model"], "Method"], [2115, 2117, ["discriminative", "task"], "Task"], [2119, 2121, ["image", "classification"], "Task"], [2129, 2131, ["digit", "classification"], "Task"], [2136, 2138, ["LeNet", "model"], "Method"], [2157, 2160, ["generative", "adversarial", "networks"], "Method"], [2162, 2164, ["generative", "models"], "Method"], [2183, 2185, ["adversarial", "discriminator"], "Method"], [2194, 2198, ["task", "-", "specific", "classifier"], "Method"], [2247, 2249, ["regularization", "techniques"], "Method"], [2318, 2320, ["layered", "representations"], "Method"], [2398, 2401, ["convolutional", "network", "framework"], "Method"], [2402, 2404, ["weight", "sharing"], "Method"], [2407, 2411, ["prior", "adversarial", "adaptation", "methods"], "Method"], [2428, 2430, ["symmetric", "transformation"], "Method"], [2464, 2465, ["optimization"], "Task"], [2488, 2490, ["asymmetric", "transformation"], "Method"], [2517, 2518, ["adaptation"], "Task"], [2520, 2524, ["supervised", "and", "unsupervised", "settings"], "Task"], [2572, 2574, ["adversarial", "loss"], "Method"], [2602, 2604, ["adversarial", "losses"], "Method"], [2606, 2608, ["adversarial", "discriminator"], "Method"], [2611, 2613, ["classification", "loss"], "Method"], [2634, 2635, ["mapping"], "Task"], [2638, 2641, ["gradient", "reversal", "layer"], "Method"], [2644, 2645, ["mapping"], "Method"], [2658, 2660, ["minimax", "objective"], "Metric"], [2661, 2664, ["generative", "adversarial", "networks"], "Method"], [2679, 2680, ["discriminator"], "Method"], [2691, 2692, ["GANs"], "Method"], [2698, 2700, ["minimax", "loss"], "Method"], [2707, 2708, ["generator"], "Method"], [2720, 2721, ["optimization"], "Task"], [2729, 2730, ["generator"], "Method"], [2734, 2735, ["discriminator"], "Method"], [2879, 2880, ["discriminator"], "Method"], [2898, 2901, ["domain", "confusion", "objective"], "Task"], [2905, 2906, ["mapping"], "Method"], [2910, 2915, ["cross", "-", "entropy", "loss", "function"], "Method"], [2925, 2927, ["adversarial", "discriminator"], "Method"], [2935, 2939, ["Adversarial", "discriminative", "domain", "adaptation"], "Task"], [2946, 2949, ["domain", "adversarial", "methods"], "Method"], [2958, 2960, ["adaptive", "methods"], "Method"], [2985, 2990, ["generative", "or", "discriminative", "base", "model"], "Method"], [3001, 3004, ["adversarial", "learning", "objective"], "Method"], [3018, 3022, ["adversarial", "discriminative", "domain", "adaptation"], "Method"], [3023, 3024, ["ADDA"], "Method"], [3046, 3047, ["ADDA"], "Method"], [3055, 3058, ["discriminative", "base", "model"], "Method"], [3072, 3075, ["sequential", "training", "procedure"], "Method"], [3086, 3089, ["discriminative", "base", "model"], "Method"], [3109, 3112, ["discriminative", "adaptation", "tasks"], "Task"], [3115, 3118, ["adversarial", "adaptive", "methods"], "Method"], [3156, 3157, ["MNIST"], "Material"], [3158, 3159, ["USPS"], "Material"], [3198, 3200, ["learing", "paradigm"], "Method"], [3204, 3208, ["domain", "specific", "feature", "extraction"], "Task"], [3227, 3229, ["weight", "sharing"], "Method"], [3230, 3232, ["target", "model"], "Method"], [3246, 3250, ["initialization", "and", "training", "procedures"], "Method"], [3259, 3261, ["source", "model"], "Method"], [3263, 3264, ["intitialization"], "Method"], [3272, 3274, ["source", "model"], "Method"], [3275, 3277, ["adversarial", "training"], "Task"], [3287, 3289, ["asymmetric", "mapping"], "Method"], [3312, 3316, ["generative", "adversarial", "learning", "setting"], "Method"], [3352, 3353, ["ADDA"], "Method"], [3359, 3361, ["unconstrained", "optimization"], "Task"], [3395, 3396, ["learning"], "Task"], [3440, 3444, ["prior", "domain", "adversarial", "methods"], "Method"], [3465, 3468, ["domain", "adaptation", "method"], "Method"], [3469, 3470, ["ADDA"], "Method"], [3488, 3492, ["unsupervised", "adaptation", "benchmark", "tasks"], "Task"], [3494, 3495, ["adaptation"], "Task"], [3507, 3508, ["ADDA"], "Method"], [3509, 3512, ["unsupervised", "classification", "adaptation"], "Task"], [3527, 3528, ["MNIST"], "Material"], [3529, 3530, ["USPS"], "Material"], [3532, 3533, ["SVHN"], "Material"], [3539, 3541, ["NYUD", "dataset"], "Material"], [3565, 3567, ["digit", "adaptation"], "Task"], [3579, 3582, ["unsupervised", "adaptation", "methods"], "Method"], [3586, 3590, ["domain", "adversarial", "learning", "objectives"], "Method"], [3631, 3637, ["real", "-", "world", "modality", "adaptation", "task"], "Task"], [3639, 3642, ["NYU", "depth", "dataset"], "Material"], [3655, 3656, ["ADDA"], "Method"], [3659, 3661, ["depth", "representation"], "Method"], [3682, 3683, ["MNIST"], "Material"], [3684, 3685, ["USPS"], "Material"], [3698, 3701, ["unsupervised", "adaptation", "task"], "Task"], [3703, 3704, ["MNIST"], "Material"], [3705, 3706, ["USPS"], "Material"], [3708, 3709, ["SVHN"], "Material"], [3738, 3739, ["adaptation"], "Task"], [3740, 3741, ["MNIST"], "Material"], [3742, 3743, ["USPS"], "Material"], [3756, 3757, ["MNIST"], "Material"], [3760, 3761, ["USPS"], "Material"], [3763, 3764, ["adaptation"], "Task"], [3765, 3768, ["SVHN", "and", "MNIST"], "Material"], [3785, 3787, ["unsupervised", "settings"], "Task"], [3800, 3801, ["adaptation"], "Task"], [3805, 3807, ["MNIST", "USPS"], "Material"], [3808, 3810, ["USPS", "MNIST"], "Material"], [3812, 3814, ["SVHN", "MNIST"], "Material"], [3823, 3826, ["modified", "LeNet", "architecture"], "Method"], [3829, 3832, ["Caffe", "source", "code"], "Method"], [3836, 3837, ["ADDA"], "Method"], [3839, 3841, ["adversarial", "discriminator"], "Method"], [3844, 3847, ["fully", "connected", "layers"], "Method"], [3864, 3868, ["500", "-", "unit", "layers"], "Method"], [3870, 3873, ["ReLU", "activation", "function"], "Method"], [3889, 3890, ["MNIST"], "Material"], [3891, 3892, ["USPS"], "Material"], [3893, 3894, ["ADDA"], "Method"], [3908, 3909, ["CoGANs"], "Method"], [3934, 3935, ["adaptation"], "Task"], [3978, 3979, ["CoGANs"], "Method"], [3982, 3983, ["SVHN"], "Material"], [3984, 3985, ["MNIST"], "Material"], [3998, 4000, ["coupled", "generators"], "Method"], [4005, 4007, ["Modality", "adaptation"], "Task"], [4032, 4035, ["NYU", "depth", "dataset"], "Material"], [4079, 4083, ["cross", "-", "modality", "adaptation"], "Method"], [4105, 4110, ["19", "-", "way", "classification", "task"], "Task"], [4111, 4113, ["object", "crops"], "Task"], [4190, 4191, ["adaptation"], "Task"], [4193, 4194, ["RGB"], "Material"], [4195, 4199, ["HHA", "encoded", "depth", "images"], "Material"], [4221, 4222, ["classification"], "Task"], [4229, 4232, ["in", "-", "domain"], "Task"], [4257, 4259, ["classification", "performance"], "Metric"], [4269, 4273, ["VGG", "-", "16", "architecture"], "Method"], [4279, 4280, ["ImageNet"], "Material"], [4308, 4309, ["ADDA"], "Method"], [4311, 4313, ["adversarial", "discriminator"], "Method"], [4317, 4320, ["fully", "connected", "layers"], "Method"], [4344, 4347, ["fully", "connected", "layers"], "Method"], [4349, 4352, ["ReLU", "activation", "function"], "Method"], [4378, 4379, ["ADDA"], "Method"], [4382, 4384, ["classification", "accuracy"], "Metric"], [4395, 4397, ["classification", "accuracy"], "Metric"], [4411, 4412, ["adaptation"], "Method"], [4417, 4418, ["accuracy"], "Metric"], [4449, 4450, ["adaptation"], "Method"], [4471, 4472, ["adaptation"], "Task"], [4479, 4480, ["ADDA"], "Method"], [4482, 4483, ["classification"], "Task"], [4501, 4505, ["best", "-", "case", "scenario"], "Task"], [4514, 4516, ["confusion", "matrix"], "Metric"], [4559, 4562, ["source", "only", "model"], "Method"], [4566, 4567, ["accuracy"], "Metric"], [4586, 4587, ["classifier"], "Method"], [4589, 4590, ["ADDA"], "Method"], [4602, 4603, ["accuracy"], "Metric"], [4611, 4612, ["accuracies"], "Metric"], [4625, 4630, ["train", "on", "target", "”", "model"], "Method"], [4683, 4686, ["adversarial", "learning", "objectives"], "Method"], [4709, 4711, ["adaptation", "methods"], "Method"], [4737, 4739, ["adaptation", "method"], "Method"], [4740, 4741, ["ADDA"], "Method"], [4751, 4754, ["unsupervised", "adaptation", "approach"], "Method"], [4777, 4782, ["cross", "-", "modality", "adaptation", "task"], "Task"], [4791, 4792, ["ADDA"], "Method"], [4813, 4814, ["ADDA"], "Method"], [115, 116, ["GAN"], "Method"], [180, 181, ["GAN"], "Method"], [851, 852, ["RGB"], "Material"], [2152, 2153, ["MNIST"], "Material"], [2154, 2155, ["USPS"], "Material"], [2775, 2776, ["GAN"], "Method"], [2809, 2810, ["GAN"], "Method"], [2831, 2832, ["GAN"], "Method"], [3065, 3066, ["GAN"], "Method"], [3340, 3341, ["GAN"], "Method"], [3650, 3651, ["RGB"], "Material"], [3687, 3688, ["SVHN"], "Material"], [3945, 3946, ["SVHN"], "Material"], [3947, 3948, ["MNIST"], "Material"], [4132, 4133, ["RGB"], "Material"], [4331, 4333, ["adversarial", "discriminator"], "Method"], [4353, 4354, ["ADDA"], "Method"], [4637, 4638, ["ADDA"], "Method"], [4656, 4657, ["ADDA"], "Method"], [4677, 4680, ["unsupervised", "domain", "adaptation"], "Task"]]} {"doc_id": "38cc89399dd6f5aaab1654f27ab3c9eeade12a36", "text": ["document", ":", "Exploiting", "temporal", "information", "for", "3D", "human", "pose", "estimation", "In", "this", "work", ",", "we", "address", "the", "problem", "of", "3D", "human", "pose", "estimation", "from", "a", "sequence", "of", "2D", "human", "poses", ".", "Although", "the", "recent", "success", "of", "deep", "networks", "has", "led", "many", "state", "-", "of", "-", "the", "-", "art", "methods", "for", "3D", "pose", "estimation", "to", "train", "deep", "networks", "end", "-", "to", "-", "end", "to", "predict", "from", "images", "directly", ",", "the", "top", "-", "performing", "approaches", "have", "shown", "the", "effectiveness", "of", "dividing", "the", "task", "of", "3D", "pose", "estimation", "into", "two", "steps", ":", "using", "a", "state", "-", "of", "-", "the", "-", "art", "2D", "pose", "estimator", "to", "estimate", "the", "2D", "pose", "from", "images", "and", "then", "mapping", "them", "into", "3D", "space", ".", "They", "also", "showed", "that", "a", "low", "-", "dimensional", "representation", "like", "2D", "locations", "of", "a", "set", "of", "joints", "can", "be", "discriminative", "enough", "to", "estimate", "3D", "pose", "with", "high", "accuracy", ".", "However", ",", "estimation", "of", "3D", "pose", "for", "individual", "frames", "leads", "to", "temporally", "incoherent", "estimates", "due", "to", "independent", "error", "in", "each", "frame", "causing", "jitter", ".", "Therefore", ",", "in", "this", "work", "we", "utilize", "the", "temporal", "information", "across", "a", "sequence", "of", "2D", "joint", "locations", "to", "estimate", "a", "sequence", "of", "3D", "poses", ".", "We", "designed", "a", "sequence", "-", "to", "-", "sequence", "network", "composed", "of", "layer", "-", "normalized", "LSTM", "units", "with", "shortcut", "connections", "connecting", "the", "input", "to", "the", "output", "on", "the", "decoder", "side", "and", "imposed", "temporal", "smoothness", "constraint", "during", "training", ".", "We", "found", "that", "the", "knowledge", "of", "temporal", "consistency", "improves", "the", "best", "reported", "result", "on", "Human3.6", "M", "dataset", "by", "approximately", "and", "helps", "our", "network", "to", "recover", "temporally", "consistent", "3D", "poses", "over", "a", "sequence", "of", "images", "even", "when", "the", "2D", "pose", "detector", "fails", ".", "section", ":", "Introduction", "The", "task", "of", "estimating", "3D", "human", "pose", "from", "2D", "representations", "like", "monocular", "images", "or", "videos", "is", "an", "open", "research", "problem", "among", "the", "computer", "vision", "and", "graphics", "community", "for", "a", "long", "time", ".", "An", "understanding", "of", "human", "posture", "and", "limb", "articulation", "is", "important", "for", "high", "level", "computer", "vision", "tasks", "such", "as", "human", "action", "or", "activity", "recognition", ",", "sports", "analysis", ",", "augmented", "and", "virtual", "reality", ".", "A", "2D", "representation", "of", "human", "pose", ",", "which", "is", "considered", "to", "be", "much", "easier", "to", "estimate", ",", "can", "be", "used", "for", "these", "tasks", ".", "However", ",", "2D", "poses", "can", "be", "ambiguous", "because", "of", "occlusion", "and", "foreshortening", ".", "Additionally", "poses", "that", "are", "totally", "different", "can", "appear", "to", "be", "similar", "in", "2D", "because", "of", "the", "way", "they", "are", "projected", "as", "shown", "in", "Figure", "[", "reference", "]", ".", "The", "depth", "information", "in", "3D", "representation", "of", "human", "pose", "makes", "it", "free", "from", "such", "ambiguities", "and", "hence", "can", "improve", "performance", "for", "higher", "level", "tasks", ".", "Moreover", ",", "3D", "pose", "can", "be", "very", "useful", "in", "computer", "animation", ",", "where", "the", "articulated", "pose", "of", "a", "person", "in", "3D", "can", "be", "used", "to", "accurately", "model", "human", "posture", "and", "movement", ".", "However", ",", "3D", "pose", "estimation", "is", "an", "ill", "-", "posed", "problem", "because", "of", "the", "inherent", "ambiguity", "in", "back", "-", "projecting", "a", "2D", "view", "of", "an", "object", "to", "the", "3D", "space", "maintaining", "its", "structure", ".", "Since", "the", "3D", "pose", "of", "a", "person", "can", "be", "projected", "in", "an", "infinite", "number", "of", "ways", "on", "a", "2D", "plane", ",", "the", "mapping", "from", "a", "2D", "pose", "to", "3D", "is", "not", "unique", ".", "Moreover", ",", "obtaining", "a", "dataset", "for", "3D", "pose", "is", "difficult", "and", "expensive", ".", "Unlike", "the", "2D", "pose", "datasets", "where", "the", "users", "can", "manually", "label", "the", "keypoints", "by", "mouse", "clicks", ",", "3D", "pose", "datasets", "require", "a", "complicated", "laboratory", "setup", "with", "motion", "capture", "sensors", "and", "cameras", ".", "Hence", ",", "there", "is", "a", "lack", "of", "motion", "capture", "datasets", "for", "images", "in", "-", "the", "-", "wild", ".", "Over", "the", "years", ",", "different", "techniques", "have", "been", "used", "to", "address", "the", "problem", "of", "3D", "pose", "estimation", ".", "Earlier", "methods", "used", "to", "focus", "on", "extracting", "features", ",", "invariant", "to", "factors", "such", "as", "background", "scenes", ",", "lighting", ",", "and", "skin", "color", "from", "images", "and", "mapping", "them", "into", "3D", "human", "pose", ".", "With", "the", "success", "of", "deep", "networks", ",", "recent", "methods", "tend", "to", "focus", "on", "training", "a", "deep", "convolutional", "neural", "network", "(", "CNN", ")", "end", "-", "to", "-", "end", "to", "estimate", "3D", "poses", "from", "images", "directly", ".", "Some", "approaches", "divided", "the", "3D", "pose", "estimation", "task", "into", "first", "predicting", "the", "joint", "locations", "in", "2D", "using", "2D", "pose", "estimators", "and", "then", "back", "-", "projecting", "them", "to", "estimate", "the", "3D", "joint", "locations", ".", "These", "results", "suggest", "the", "effectiveness", "of", "decoupling", "the", "task", "of", "3D", "pose", "estimation", "where", "2D", "pose", "estimator", "abstracts", "the", "complexities", "in", "the", "image", ".", "In", "this", "paper", ",", "we", "also", "adopt", "the", "decoupled", "approach", "to", "3D", "pose", "estimation", ".", "However", ",", "predicting", "3D", "pose", "for", "each", "frame", "individually", "can", "lead", "to", "jitter", "in", "videos", "because", "the", "errors", "in", "each", "frame", "are", "independent", "of", "each", "other", ".", "Therefore", ",", "we", "designed", "a", "sequence", "-", "to", "-", "sequence", "network", "with", "shortcut", "connections", "on", "the", "decoder", "side", "that", "predicts", "a", "sequence", "of", "temporally", "consistent", "3D", "poses", "given", "a", "sequence", "of", "2D", "poses", ".", "Each", "unit", "of", "our", "network", "is", "a", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "unit", "with", "layer", "normalization", "and", "recurrent", "dropout", ".", "We", "also", "imposed", "a", "temporal", "smoothness", "constraint", "on", "the", "predicted", "3D", "poses", "during", "training", "to", "ensure", "that", "our", "predictions", "are", "smooth", "over", "a", "sequence", ".", "Our", "network", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "result", "on", "the", "Human3.6", "M", "dataset", "improving", "the", "previous", "best", "result", "by", "approximately", ".", "We", "also", "obtained", "the", "lowest", "error", "for", "every", "action", "class", "in", "Human3.6", "M", "dataset", ".", "Moreover", ",", "we", "observed", "that", "our", "network", "predicted", "meaningful", "3D", "poses", "on", "Youtube", "videos", ",", "even", "when", "the", "detections", "from", "the", "2D", "pose", "detector", "were", "extremely", "noisy", "or", "meaningless", ".", "This", "shows", "the", "effectiveness", "of", "using", "temporal", "information", ".", "In", "short", "our", "contributions", "in", "this", "work", "are", ":", "Designing", "an", "efficient", "sequence", "-", "to", "-", "sequence", "network", "that", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "results", "for", "every", "action", "class", "of", "Human3.6", "M", "dataset", "and", "can", "be", "trained", "very", "fast", ".", "Exploiting", "the", "ability", "of", "sequence", "-", "to", "-", "sequence", "networks", "to", "take", "into", "account", "the", "events", "in", "the", "past", ",", "to", "predict", "temporally", "consistent", "3D", "poses", ".", "Effectively", "imposing", "temporal", "consistency", "constraint", "on", "the", "predicted", "3D", "poses", "during", "training", "so", "that", "the", "errors", "in", "the", "predictions", "are", "distributed", "smoothly", "over", "the", "sequence", ".", "Using", "only", "the", "previous", "frames", "to", "understand", "temporal", "context", "so", "that", "it", "can", "be", "deployed", "online", "and", "real", "-", "time", ".", "section", ":", "Related", "Work", "paragraph", ":", "Representation", "of", "3D", "pose", "Both", "model", "-", "based", "and", "model", "-", "free", "representations", "of", "3D", "human", "pose", "have", "been", "used", "in", "the", "past", ".", "The", "most", "common", "model", "-", "based", "representation", "is", "a", "skeleton", "defined", "by", "a", "kinematic", "tree", "of", "a", "set", "of", "joints", ",", "parameterized", "by", "the", "offset", "and", "rotational", "parameters", "of", "each", "joint", "relative", "to", "its", "parent", ".", "Several", "3D", "pose", "methods", "have", "used", "this", "representation", ".", "Others", "model", "3D", "pose", "as", "a", "sparse", "linear", "combination", "of", "an", "over", "-", "complete", "dictionary", "of", "basis", "poses", ".", "However", ",", "we", "have", "chosen", "a", "model", "-", "free", "representation", "of", "3D", "pose", ",", "where", "a", "3D", "pose", "is", "simply", "a", "set", "of", "3D", "joint", "locations", "relative", "to", "the", "root", "node", "like", "several", "recent", "approaches", ".", "This", "representation", "is", "much", "simpler", "and", "low", "-", "dimensional", ".", "paragraph", ":", "Estimating", "3D", "pose", "from", "2D", "joints", "Lee", "and", "Chen", "were", "the", "first", "to", "infer", "3D", "joint", "locations", "from", "their", "2D", "projections", "given", "the", "bone", "lengths", "using", "a", "binary", "decision", "tree", "where", "each", "branch", "corresponds", "to", "two", "possible", "states", "of", "a", "joint", "relative", "to", "its", "parent", ".", "Jiang", "used", "the", "2D", "joint", "locations", "to", "estimate", "a", "set", "of", "hypothesis", "3D", "poses", "using", "Taylor", "’s", "algorithm", "and", "used", "them", "to", "query", "a", "large", "database", "of", "motion", "capture", "data", "to", "find", "the", "nearest", "neighbor", ".", "Gupta", "et", "al", ".", "and", "Chen", "and", "Ramanan", "also", "used", "this", "idea", "of", "using", "the", "detected", "2D", "pose", "to", "query", "a", "large", "database", "of", "exemplar", "poses", "to", "find", "the", "nearest", "nearest", "neighbor", "3D", "pose", ".", "Another", "common", "approach", "to", "estimating", "3D", "joint", "locations", "given", "the", "2D", "pose", "is", "to", "separate", "the", "camera", "pose", "variability", "from", "the", "intrinsic", "deformation", "of", "the", "human", "body", ",", "the", "latter", "of", "which", "is", "modeled", "by", "learning", "an", "over", "-", "complete", "dictionary", "of", "basis", "3D", "poses", "from", "a", "large", "database", "of", "motion", "capture", "data", ".", "A", "valid", "3D", "pose", "is", "defined", "by", "a", "sparse", "linear", "combination", "of", "the", "bases", "and", "by", "transforming", "the", "points", "using", "transformation", "matrix", "representing", "camera", "extrinsic", "parameters", ".", "Moreno", "-", "Nouguer", "used", "the", "pair", "-", "wise", "distance", "matrix", "of", "2D", "joints", "to", "learn", "a", "distance", "matrix", "for", "3D", "joints", ",", "which", "they", "found", "invariant", "up", "to", "a", "rigid", "similarity", "transform", "with", "the", "ground", "truth", "3D", "and", "used", "multi", "-", "dimensional", "scaling", "(", "MDS", ")", "with", "pose", "-", "priors", "to", "rule", "out", "the", "ambiguities", ".", "Martinez", "et", "al", ".", "designed", "a", "fully", "connected", "network", "with", "shortcut", "connections", "every", "two", "linear", "layers", "to", "estimate", "3D", "joint", "locations", "relative", "to", "the", "root", "node", "in", "the", "camera", "coordinate", "space", ".", "paragraph", ":", "Deep", "network", "based", "methods", "With", "the", "success", "of", "deep", "networks", ",", "many", "have", "designed", "networks", "that", "can", "be", "trained", "end", "-", "to", "-", "end", "to", "predict", "3D", "poses", "from", "images", "directly", ".", "Li", "et", "al", ".", "and", "Park", "et", "al", ".", "designed", "CNNs", "to", "jointly", "predict", "2D", "and", "3D", "poses", ".", "Mehta", "et", "al", ".", "and", "Sun", "et", "al", ".", "used", "transfer", "learning", "to", "transfer", "the", "knowledge", "learned", "for", "2D", "human", "pose", "estimation", "to", "the", "task", "of", "3D", "pose", "estimation", ".", "Pavlakos", "et", "al", ".", "extended", "the", "stacked", "-", "hourglass", "network", "originally", "designed", "to", "predict", "2D", "heatmaps", "of", "each", "joint", "to", "make", "it", "predict", "3D", "volumetric", "heatmaps", ".", "Tome", "et", "al", ".", "also", "extended", "a", "2D", "pose", "estimator", "called", "Convolutional", "Pose", "Machine", "(", "CPM", ")", "to", "make", "it", "predict", "3D", "pose", ".", "Rogesz", "and", "Schmid", "and", "Varol", "et", "al", ".", "augmented", "the", "training", "data", "with", "synthetic", "images", "and", "trained", "CNNs", "to", "predict", "3D", "poses", "from", "real", "images", ".", "Sun", "et", "al", ".", "designed", "a", "unified", "network", "that", "can", "regress", "both", "2D", "and", "3D", "poses", "at", "the", "same", "time", "given", "an", "image", ".", "Hence", "during", "training", "time", ",", "in", "-", "the", "-", "wild", "images", "which", "do", "not", "have", "any", "ground", "truth", "3D", "poses", "can", "be", "combined", "with", "the", "data", "with", "ground", "truth", "3D", "poses", ".", "A", "similar", "idea", "of", "exploiting", "in", "-", "the", "-", "wild", "images", "to", "learn", "pose", "structure", "was", "used", "by", "Fang", "et", "al", ".", ".", "They", "learned", "a", "pose", "grammar", "that", "encodes", "the", "possible", "human", "pose", "configurations", ".", "paragraph", ":", "Using", "temporal", "information", "Since", "estimating", "poses", "for", "each", "frame", "individually", "leads", "to", "incoherent", "and", "jittery", "predictions", "over", "a", "sequence", ",", "many", "approaches", "tried", "to", "exploit", "temporal", "information", ".", "Andriluka", "et", "al", ".", "used", "tracking", "-", "by", "-", "detection", "to", "associate", "2D", "poses", "detected", "in", "each", "frame", "individually", "and", "used", "them", "to", "retrieve", "3D", "pose", ".", "Tekin", "et", "al", ".", "used", "a", "CNN", "to", "first", "align", "bounding", "boxes", "of", "successive", "frames", "so", "that", "the", "person", "in", "the", "image", "is", "always", "at", "the", "center", "of", "the", "box", "and", "then", "extracted", "3D", "HOG", "features", "densely", "over", "the", "spatio", "-", "temporal", "volume", "from", "which", "they", "regress", "the", "3D", "pose", "of", "the", "central", "frame", ".", "Mehta", "et", "al", ".", "implemented", "a", "real", "-", "time", "system", "for", "3D", "pose", "estimation", "that", "applies", "temporal", "filtering", "across", "2D", "and", "3D", "poses", "from", "previous", "frames", "to", "predict", "a", "temporally", "consistent", "3D", "pose", ".", "Lin", "et", "al", ".", "performed", "a", "multi", "-", "stage", "sequential", "refinement", "using", "LSTMs", "to", "predict", "3D", "pose", "sequences", "using", "previously", "predicted", "2D", "pose", "representations", "and", "3D", "pose", ".", "We", "focus", "on", "predicting", "temporally", "consistent", "3D", "poses", "by", "learning", "the", "temporal", "context", "of", "a", "sequence", "using", "a", "form", "of", "sequence", "-", "to", "-", "sequence", "network", ".", "Unlike", "Lin", "et", "al", ".", "our", "method", "does", "not", "need", "multiple", "stages", "of", "refinement", ".", "It", "is", "simpler", "and", "requires", "fewer", "parameters", "to", "train", ",", "leading", "to", "much", "improved", "performance", ".", "section", ":", "Our", "Approach", "paragraph", ":", "Network", "Design", "We", "designed", "a", "sequence", "-", "to", "-", "sequence", "network", "with", "LSTM", "units", "and", "residual", "connections", "on", "the", "decoder", "side", "to", "predict", "a", "temporally", "coherent", "sequence", "of", "3D", "poses", "given", "a", "sequence", "of", "2D", "joint", "locations", ".", "Figure", "[", "reference", "]", "shows", "the", "architecture", "of", "our", "network", ".", "The", "motivation", "behind", "using", "a", "sequence", "-", "to", "-", "sequence", "network", "comes", "from", "its", "application", "on", "the", "task", "of", "Neural", "Machine", "Translation", "(", "NMT", ")", "by", "Sutskever", "et", "al", ".", ",", "where", "their", "model", "translates", "a", "sentence", "in", "one", "language", "to", "a", "sentence", "in", "another", "language", "e.g.", "English", "to", "French", ".", "In", "a", "language", "translation", "model", ",", "the", "input", "and", "output", "sentences", "can", "have", "different", "lengths", ".", "Although", "our", "case", "is", "analogous", "to", "the", "NMT", ",", "the", "input", "and", "output", "sequences", "always", "have", "the", "same", "length", "while", "the", "input", "vectors", "to", "the", "encoder", "and", "decoder", "have", "different", "dimensions", ".", "The", "encoder", "side", "of", "our", "network", "takes", "a", "sequence", "of", "2D", "poses", "and", "encodes", "them", "in", "a", "fixed", "size", "high", "dimensional", "vector", "in", "the", "hidden", "state", "of", "its", "final", "LSTM", "unit", ".", "Since", "the", "LSTMs", "are", "excellent", "in", "memorizing", "events", "and", "information", "from", "the", "past", ",", "the", "encoded", "vector", "stores", "the", "2D", "pose", "information", "of", "all", "the", "frames", ".", "The", "initial", "state", "of", "the", "decoder", "is", "initialized", "by", "the", "final", "state", "of", "the", "encoder", ".", "A", "token", "is", "passed", "as", "initial", "input", "to", "the", "decoder", ",", "which", "in", "our", "case", "is", "a", "vector", "of", "ones", ",", "telling", "it", "to", "start", "decoding", ".", "Given", "a", "3D", "pose", "estimate", "at", "a", "time", "step", "each", "decoder", "unit", "predicts", "the", "3D", "pose", "for", "next", "time", "step", ".", "Note", "that", "the", "order", "of", "the", "input", "sequence", "is", "reversed", "as", "recommended", "by", "Sutskever", "et", "al", ".", ".", "The", "shortcut", "connections", "on", "the", "decoder", "side", "cause", "each", "decoder", "unit", "to", "estimate", "the", "amount", "of", "perturbation", "in", "the", "3D", "pose", "from", "the", "previous", "frame", "instead", "of", "having", "to", "estimate", "the", "actual", "3D", "pose", "for", "each", "frame", ".", "As", "suggested", "by", "He", "et", "al", ".", ",", "such", "a", "mapping", "is", "easier", "to", "learn", "for", "the", "network", ".", "We", "use", "layer", "normalization", "and", "recurrent", "dropout", "to", "regularize", "our", "network", ".", "Ba", "et", "al", ".", "came", "up", "with", "the", "idea", "of", "layer", "normalization", "which", "estimates", "the", "normalization", "statistics", "(", "mean", "and", "standard", "deviation", ")", "from", "the", "summed", "inputs", "to", "the", "recurrent", "neurons", "of", "hidden", "layer", "on", "a", "single", "training", "example", "to", "regularize", "the", "RNN", "units", ".", "Similarly", ",", "Zaremba", "et", "al", ".", "proposed", "the", "idea", "of", "applying", "dropout", "only", "on", "the", "non", "-", "recurrent", "connections", "of", "the", "network", "with", "a", "certain", "probability", "while", "always", "keeping", "the", "recurrent", "connections", "intact", "because", "they", "are", "necessary", "for", "the", "recurrent", "units", "to", "remember", "the", "information", "from", "the", "past", ".", "paragraph", ":", "Loss", "function", "Given", "a", "sequence", "of", "2D", "joint", "locations", "as", "input", ",", "our", "network", "predicts", "a", "sequence", "of", "3D", "joint", "locations", "relative", "to", "the", "root", "node", "(", "central", "hip", ")", ".", "We", "predict", "each", "3D", "pose", "in", "the", "camera", "coordinate", "space", "instead", "of", "predicting", "them", "in", "an", "arbitrary", "global", "frame", "as", "suggested", "by", "Martinez", "et", "al", ".", ".", "We", "impose", "a", "temporal", "smoothness", "constraint", "on", "the", "predicted", "3D", "joint", "locations", "to", "ensure", "that", "the", "prediction", "of", "each", "joint", "in", "one", "frame", "does", "not", "differ", "too", "much", "from", "its", "previous", "frame", ".", "Because", "the", "2D", "pose", "detectors", "work", "on", "individual", "frames", ",", "even", "with", "the", "minimal", "movement", "of", "the", "subject", "in", "the", "image", ",", "the", "detections", "from", "successive", "frames", "may", "vary", ",", "particularly", "for", "the", "joints", "which", "move", "fast", "or", "are", "prone", "to", "occlusion", ".", "Hence", ",", "we", "made", "an", "assumption", "that", "the", "subject", "does", "not", "move", "too", "much", "in", "successive", "frames", "given", "the", "frame", "rate", "is", "high", "enough", ".", "Therefore", ",", "we", "added", "the", "L2", "norm", "of", "the", "first", "order", "derivative", "on", "the", "3D", "joint", "locations", "with", "respect", "to", "time", "to", "our", "loss", "function", "during", "training", ".", "This", "constraint", "helps", "us", "to", "estimate", "3D", "poses", "reliably", "even", "when", "the", "2D", "pose", "detector", "fails", "for", "a", "few", "frames", "within", "the", "temporal", "window", "without", "any", "post", "-", "processing", ".", "Empirically", "we", "found", "that", "certain", "joints", "are", "more", "difficult", "to", "estimate", "accurately", "e.g.", "wrist", ",", "ankle", ",", "elbow", "compared", "to", "others", ".", "To", "address", "this", "issue", ",", "we", "partitioned", "the", "joints", "into", "three", "disjoint", "sets", ",", "and", "based", "on", "their", "contribution", "to", "overall", "error", ".", "We", "observed", "that", "the", "joints", "connected", "to", "the", "torso", "and", "the", "head", "e.g.", "hips", ",", "shoulders", ",", "neck", "are", "always", "predicted", "with", "high", "accuracy", "compared", "to", "those", "joints", "belonging", "to", "the", "limbs", "and", "therefore", "put", "them", "in", "the", "set", ".", "The", "joints", "of", "the", "limbs", ",", "especially", "the", "joints", "on", "the", "arms", ",", "are", "always", "more", "difficult", "to", "predict", "due", "to", "their", "high", "range", "of", "motion", "and", "occlusion", ".", "We", "put", "the", "knees", "and", "the", "ankles", "in", "the", "set", "and", "the", "elbow", "and", "wrist", "in", ".", "We", "multiply", "the", "derivatives", "of", "each", "set", "of", "joints", "with", "different", "scalar", "values", "based", "on", "their", "contribution", "to", "the", "overall", "error", ".", "Therefore", "our", "loss", "function", "consists", "of", "the", "sum", "of", "two", "separate", "terms", ":", "Mean", "Squared", "Error", "(", "MSE", ")", "of", "different", "sequences", "of", "3D", "joint", "locations", ";", "and", "the", "mean", "of", "the", "L2", "norm", "of", "the", "first", "order", "derivative", "of", "sequences", "of", "3D", "joint", "locations", "with", "respect", "to", "time", ",", "where", "the", "joints", "are", "divided", "into", "three", "disjoint", "sets", ".", "The", "MSE", "over", "sequences", ",", "each", "of", "time", "-", "steps", ",", "of", "3D", "joint", "locations", "is", "given", "by", "Here", ",", "denotes", "the", "estimated", "3D", "joint", "locations", "while", "denotes", "3D", "ground", "truth", ".", "The", "mean", "of", "L2", "norm", "of", "the", "first", "order", "derivative", "of", "sequences", "of", "3D", "joint", "locations", ",", "each", "of", "length", ",", "with", "respect", "to", "time", "is", "given", "by", "In", "the", "above", "equation", ",", ",", "and", "denotes", "the", "predicted", "3D", "locations", "of", "joints", "belonging", "to", "the", "sets", ",", "and", "respectively", ".", "The", "and", "are", "scalar", "hyper", "-", "parameters", "to", "control", "the", "significance", "of", "the", "derivatives", "of", "3D", "locations", "of", "each", "of", "the", "three", "set", "of", "joints", ".", "A", "higher", "weight", "is", "assigned", "to", "the", "set", "of", "joints", "which", "are", "generally", "predicted", "with", "higher", "error", ".", "The", "overall", "loss", "function", "for", "our", "network", "is", "given", "as", "Here", "and", "are", "scalar", "hyper", "-", "parameters", "regulating", "the", "importance", "of", "each", "of", "the", "two", "terms", "in", "the", "loss", "function", ".", "section", ":", "Experimental", "Evaluation", "paragraph", ":", "Datasets", "and", "protocols", "We", "perform", "quantitative", "evaluation", "on", "the", "Human", "3.6", "M", "dataset", "and", "on", "the", "HumanEva", "dataset", ".", "Human", "3.6", "M", ",", "to", "the", "best", "of", "our", "knowledge", ",", "is", "the", "largest", "publicly", "available", "dataset", "for", "human", "3D", "pose", "estimation", ".", "The", "dataset", "contains", "3.6", "million", "images", "of", "7", "different", "professional", "actors", "performing", "15", "everyday", "activities", "like", "walking", ",", "eating", ",", "sitting", ",", "making", "a", "phone", "call", ".", "The", "dataset", "consists", "of", "2D", "and", "3D", "joint", "locations", "for", "each", "corresponding", "image", ".", "Each", "video", "is", "captured", "using", "4", "different", "calibrated", "high", "resolution", "cameras", ".", "In", "addition", "to", "2D", "and", "3D", "pose", "ground", "truth", ",", "the", "dataset", "also", "provides", "ground", "truth", "for", "bounding", "boxes", ",", "the", "camera", "parameters", ",", "the", "body", "proportion", "of", "all", "the", "actors", "and", "high", "resolution", "body", "scans", "or", "meshes", "of", "each", "actor", ".", "HumanEva", ",", "on", "the", "other", "hand", ",", "is", "a", "much", "smaller", "dataset", ".", "It", "has", "been", "largely", "used", "to", "benchmark", "previous", "work", "over", "the", "last", "decade", ".", "Most", "of", "the", "methods", "report", "results", "on", "two", "different", "actions", "and", "on", "three", "actors", ".", "For", "qualitative", "evaluation", ",", "we", "used", "the", "some", "videos", "from", "Youtube", "and", "the", "Human3.6", "M", "dataset", ".", "We", "follow", "the", "standard", "protocols", "of", "the", "Human3.6", "M", "dataset", "used", "in", "the", "literature", ".", "We", "used", "subjects", "1", ",", "5", ",", "6", ",", "7", ",", "and", "8", "for", "training", ",", "and", "subjects", "9", "and", "11", "for", "testing", "and", "the", "error", "is", "evaluated", "on", "the", "predicted", "3D", "pose", "without", "any", "transformation", ".", "We", "refer", "this", "as", "protocol", "#", "1", ".", "Another", "common", "approach", "used", "by", "many", "to", "evaluate", "their", "methods", "is", "to", "align", "the", "predicted", "3D", "pose", "with", "the", "ground", "truth", "using", "a", "similarity", "transformation", "(", "Procrustes", "analysis", ")", ".", "We", "refer", "this", "as", "protocol", "#", "2", ".", "We", "use", "the", "average", "error", "per", "joint", "in", "millimeters", "between", "the", "estimated", "and", "the", "ground", "truth", "3D", "pose", "relative", "to", "the", "root", "node", "as", "the", "error", "metric", ".", "For", "the", "HumanEva", "dataset", ",", "we", "report", "results", "on", "each", "subject", "and", "action", "separately", "after", "performing", "rigid", "alignment", "with", "the", "ground", "truth", "data", ",", "following", "the", "protocol", "used", "by", "the", "previous", "methods", ".", "paragraph", ":", "2D", "detections", "We", "fine", "-", "tuned", "a", "model", "of", "stacked", "-", "hourglass", "network", ",", "initially", "trained", "on", "the", "MPII", "dataset", "(", "a", "benchmark", "dataset", "for", "2D", "pose", "estimation", ")", ",", "on", "the", "images", "of", "the", "Human3.6", "M", "dataset", "to", "obtain", "2D", "pose", "estimations", "for", "each", "image", ".", "We", "used", "the", "bounding", "box", "information", "provided", "with", "the", "dataset", "to", "first", "compute", "the", "center", "of", "the", "person", "in", "the", "image", "and", "then", "cropped", "a", "region", "across", "the", "person", "and", "resized", "it", "to", ".", "We", "fine", "-", "tuned", "the", "network", "for", "250", "iterations", "and", "used", "a", "batch", "size", "of", "3", "and", "a", "learning", "rate", "of", ".", "paragraph", ":", "Baselines", "Since", "many", "of", "the", "previous", "methods", "are", "based", "on", "single", "frame", "predictions", ",", "we", "used", "two", "baselines", "for", "comparison", ".", "To", "show", "that", "our", "method", "is", "much", "better", "than", "naive", "post", "processing", ",", "we", "applied", "a", "mean", "filter", "and", "a", "median", "filter", "on", "the", "3D", "pose", "predictions", "of", "Martinez", "et", "al", ".", ".", "We", "used", "a", "window", "size", "of", "5", "frames", "and", "a", "stride", "length", "of", "1", "to", "apply", "the", "filters", ".", "Although", "non", "-", "rigid", "structure", "from", "motion", "(", "NRSFM", ")", "is", "one", "of", "the", "most", "general", "approaches", "for", "any", "3D", "reconstruction", "problem", "from", "a", "sequence", "of", "2D", "correspondences", ",", "we", "did", "not", "use", "it", "as", "a", "baseline", "because", "Zhou", "et", "al", ".", "did", "not", "find", "NRSFM", "techniques", "to", "be", "effective", "for", "3D", "human", "pose", "estimation", ".", "They", "found", "that", "the", "NRSFM", "techniques", "do", "not", "work", "well", "with", "slow", "camera", "motion", ".", "Since", "the", "videos", "in", "the", "Human3.6", "M", "dataset", "are", "captured", "by", "stationary", "cameras", ",", "the", "subjects", "in", "the", "dataset", "do", "not", "rotate", "that", "much", "to", "provide", "alternative", "views", "for", "NRSFM", "algorithm", "to", "perform", "well", ".", "Another", "reason", "is", "that", "human", "pose", "reconstruction", "is", "a", "specialized", "problem", "in", "which", "constraints", "from", "human", "body", "structure", "apply", ".", "paragraph", ":", "Data", "pre", "-", "processing", "We", "normalized", "the", "3D", "ground", "truth", "poses", ",", "the", "noisy", "2D", "pose", "estimates", "from", "stacked", "-", "hourglass", "network", "and", "the", "2D", "ground", "truth", "by", "subtracting", "the", "mean", "and", "dividing", "by", "standard", "deviation", ".", "We", "do", "not", "predict", "the", "3D", "location", "of", "the", "root", "joint", "i.e.", "central", "hip", "joint", "and", "hence", "zero", "center", "the", "3D", "joint", "locations", "relative", "to", "the", "global", "position", "of", "the", "root", "node", ".", "To", "obtain", "the", "ground", "truth", "3D", "poses", "in", "camera", "coordinate", "space", ",", "an", "inverse", "rigid", "body", "transformation", "is", "applied", "on", "the", "the", "ground", "truth", "3D", "poses", "in", "global", "coordinate", "space", "using", "the", "given", "camera", "parameters", ".", "To", "generate", "both", "training", "and", "test", "sequences", ",", "we", "translated", "a", "sliding", "window", "of", "length", "by", "one", "frame", ".", "Hence", "there", "is", "an", "overlap", "between", "the", "sequences", ".", "This", "gives", "us", "more", "data", "to", "train", "on", ",", "which", "is", "always", "an", "advantage", "for", "deep", "learning", "systems", ".", "During", "test", "time", ",", "we", "initially", "predict", "the", "first", "frames", "of", "the", "sequence", "and", "slide", "the", "window", "by", "a", "stride", "length", "of", "1", "to", "predict", "the", "next", "frame", "using", "the", "previous", "frames", ".", "paragraph", ":", "Training", "details", "We", "trained", "our", "network", "for", "100", "epochs", ",", "where", "each", "epoch", "makes", "a", "complete", "pass", "over", "the", "entire", "Human", "3.6", "M", "dataset", ".", "We", "used", "the", "Adam", "optimizer", "for", "training", "the", "network", "with", "a", "learning", "rate", "of", "which", "is", "decayed", "exponentially", "per", "iteration", ".", "The", "weights", "of", "the", "LSTM", "units", "are", "initialized", "by", "Xavier", "uniform", "initializer", ".", "We", "used", "a", "mini", "-", "batch", "batch", "size", "of", "32", "i.e.", "32", "sequences", ".", "For", "most", "of", "our", "experiments", "we", "used", "a", "sequence", "length", "of", "5", ",", "because", "it", "allows", "faster", "training", "with", "high", "accuracy", ".", "We", "experimented", "with", "different", "sequence", "lengths", "and", "found", "sequence", "length", "4", ",", "5", "and", "6", "to", "generally", "give", "better", "results", ",", "which", "we", "will", "discuss", "in", "detail", "in", "the", "results", "section", ".", "We", "trained", "a", "single", "model", "for", "all", "the", "action", "classes", ".", "Our", "code", "is", "implemented", "in", "Tensorflow", ".", "We", "perform", "cross", "-", "validation", "on", "the", "training", "set", "to", "select", "the", "hyper", "-", "parameter", "values", "and", "of", "our", "loss", "function", "to", "and", "respectively", ".", "Similarly", ",", "using", "cross", "-", "validation", ",", "the", "three", "hyper", "-", "parameters", "of", "the", "temporal", "consistency", "constraint", "and", ",", "are", "set", "to", "and", "respectively", ".", "A", "single", "training", "step", "for", "sequences", "of", "length", "5", "takes", "only", "34", "ms", "approximately", ",", "while", "a", "forward", "pass", "takes", "only", "about", "16ms", "on", "NVIDIA", "Titan", "X", "GPU", ".", "Therefore", "given", "the", "2D", "joint", "locations", "from", "a", "pose", "detector", ",", "our", "network", "takes", "about", "3.2ms", "to", "predict", "3D", "pose", "per", "frame", ".", "subsection", ":", "Quantitative", "results", "paragraph", ":", "Evaluation", "on", "estimated", "2D", "pose", "As", "mentioned", "before", ",", "we", "used", "a", "sequence", "length", "of", "5", "to", "perform", "both", "qualitative", "and", "quantitative", "evaluation", "of", "our", "network", ".", "The", "results", "on", "Human3.6", "M", "dataset", "under", "protocol", "#", "1", "are", "shown", "in", "Table", "[", "reference", "]", ".", "From", "the", "table", "we", "observe", "that", "our", "model", "achieves", "the", "lowest", "error", "for", "every", "action", "class", "under", "protocol", "#", "1", ",", "unlike", "many", "of", "the", "previous", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "Note", "that", "we", "train", "a", "single", "model", "for", "all", "the", "action", "classes", "unlike", "many", "other", "methods", "which", "trained", "a", "model", "for", "each", "action", "class", ".", "Our", "network", "significantly", "improves", "the", "state", "-", "of", "-", "the", "-", "art", "result", "of", "Sun", "et", "al", ".", "by", "approximately", "(", "by", "mm", ")", ".", "The", "results", "under", "protocol", "#", "2", ",", "which", "aligns", "the", "predictions", "to", "the", "ground", "truth", "using", "a", "rigid", "body", "similarity", "transform", "before", "computing", "the", "error", ",", "is", "reported", "in", "Table", "[", "reference", "]", ".", "Our", "network", "improves", "the", "reported", "state", "-", "of", "-", "the", "-", "art", "results", "by", "(", "by", "mm", ")", "and", "achieves", "the", "lowest", "error", "for", "each", "action", "in", "protocol", "#", "2", "as", "well", ".", "From", "the", "results", ",", "we", "observe", "the", "effectiveness", "of", "exploiting", "temporal", "information", "across", "multiple", "sequences", ".", "By", "using", "the", "information", "of", "temporal", "context", ",", "our", "network", "reduced", "the", "overall", "error", "in", "estimating", "3D", "joint", "locations", ",", "especially", "on", "actions", "like", "phone", ",", "photo", ",", "sit", "and", "sitting", "down", "on", "which", "most", "previous", "methods", "did", "not", "perform", "well", "due", "to", "heavy", "occlusion", ".", "We", "also", "observe", "that", "our", "method", "outperforms", "both", "the", "baselines", "by", "a", "large", "margin", "on", "both", "the", "protocols", ".", "This", "shows", "that", "our", "method", "learned", "the", "temporal", "context", "of", "the", "sequences", "and", "predicted", "temporally", "consistent", "3D", "poses", ",", "which", "naive", "post", "-", "processing", "techniques", "like", "temporal", "mean", "and", "median", "filters", "over", "frame", "-", "wise", "prediction", "failed", "to", "do", ".", "Like", "most", "previous", "methods", ",", "we", "report", "the", "results", "on", "action", "classes", "Walking", "and", "Jogging", "of", "the", "HumanEva", "dataset", "in", "Table", "[", "reference", "]", ".", "We", "obtained", "the", "lowest", "error", "in", "four", "of", "the", "six", "cases", "and", "the", "lowest", "average", "error", "for", "the", "two", "actions", ".", "We", "also", "obtained", "the", "second", "best", "result", "on", "subject", "2", "of", "action", "Walking", ".", "However", ",", "HumanEva", "is", "a", "smaller", "dataset", "than", "Human3.6", "M", "and", "the", "same", "subjects", "appear", "in", "both", "training", "and", "testing", ".", "paragraph", ":", "Evaluation", "on", "2D", "ground", "truth", "As", "suggested", "by", "Martinez", "et", "al", ".", ",", "we", "also", "found", "that", "the", "more", "accurate", "the", "2D", "joint", "locations", "are", ",", "the", "better", "are", "the", "estimates", "for", "3D", "pose", ".", "We", "trained", "our", "model", "on", "ground", "truth", "2D", "poses", "for", "a", "sequence", "length", "of", "5", ".", "The", "results", "under", "protocol", "#", "1", "are", "reported", "in", "Table", "[", "reference", "]", ".", "As", "seen", "from", "the", "table", ",", "our", "model", "improves", "the", "lower", "bound", "error", "of", "Martinez", "et", "al", ".", "by", "almost", ".", "The", "results", "on", "ground", "truth", "2D", "joint", "input", "for", "protocol", "#", "2", "are", "reported", "in", "Table", "[", "reference", "]", ".", "When", "there", "is", "no", "noise", "in", "2D", "joint", "locations", ",", "our", "network", "performs", "better", "than", "the", "models", "by", "Martinez", "et", "al", ".", "and", "Moreno", "-", "Nouguer", ".", "These", "results", "suggest", "that", "the", "information", "of", "temporal", "consistency", "from", "previous", "frames", "is", "a", "valuable", "cue", "for", "the", "task", "of", "estimating", "3D", "pose", "even", "when", "the", "detections", "are", "noise", "free", ".", "paragraph", ":", "Robustness", "to", "noise", "We", "carried", "out", "some", "experiments", "to", "test", "the", "tolerance", "of", "our", "model", "to", "different", "levels", "of", "noise", "in", "the", "input", "data", "by", "training", "our", "network", "on", "2D", "ground", "truth", "poses", "and", "testing", "on", "inputs", "corrupted", "by", "different", "levels", "of", "Gaussian", "noise", ".", "Table", "[", "reference", "]", "shows", "how", "our", "final", "model", "compares", "against", "the", "models", "by", "Moreno", "-", "Nouguer", "and", "Martinez", "et", "al", ".", ".", "Our", "network", "is", "significantly", "more", "robust", "than", "Moreno", "-", "Nouguer", "’s", "model", ".", "When", "compared", "against", "Martinez", "et", "al", ".", "our", "network", "performs", "better", "when", "the", "level", "of", "input", "noise", "is", "low", "i.e.", "standard", "deviation", "less", "than", "or", "equal", "to", "10", ".", "However", ",", "for", "higher", "levels", "of", "noise", "our", "network", "performs", "slightly", "worse", "than", "Martinez", "et", "al", ".", ".", "We", "would", "like", "to", "attribute", "the", "cause", "of", "this", "to", "the", "temporal", "smoothness", "constraint", "imposed", "during", "training", "which", "distributes", "the", "error", "of", "individual", "frames", "over", "the", "entire", "sequence", ".", "However", ",", "its", "usefulness", "can", "be", "observed", "in", "the", "qualitative", "results", "(", "See", "Figure", "[", "reference", "]", "and", "Figure", "[", "reference", "]", ")", ".", "paragraph", ":", "Ablative", "analysis", "To", "show", "the", "usefulness", "of", "each", "component", "and", "design", "decision", "of", "our", "network", ",", "we", "perform", "an", "ablative", "analysis", ".", "We", "follow", "protocol", "#", "1", "for", "performing", "ablative", "analysis", "and", "trained", "a", "single", "model", "for", "all", "the", "actions", ".", "The", "results", "are", "reported", "in", "Table", "[", "reference", "]", ".", "We", "observe", "that", "the", "biggest", "improvement", "in", "result", "is", "due", "the", "the", "residual", "connections", "on", "the", "decoder", "side", ",", "which", "agrees", "with", "the", "hypothesis", "of", "He", "et", "al", ".", ".", "Removing", "the", "residual", "connections", "massively", "increases", "the", "error", "by", "mm", ".", "When", "we", "do", "not", "apply", "layer", "normalization", "on", "LSTM", "units", ",", "the", "error", "increases", "by", "mm", ".", "On", "the", "other", "hand", "when", "dropout", "is", "not", "performed", ",", "the", "error", "raises", "by", "mm", ".", "When", "both", "layer", "normalization", "and", "recurrent", "dropout", "are", "not", "used", "the", "results", "get", "worse", "by", "mm", ".", "Although", "the", "temporal", "consistency", "constraint", "may", "seem", "to", "have", "less", "impact", "(", "only", "mm", ")", "quantitatively", "on", "the", "performance", "of", "our", "network", ",", "it", "ensures", "that", "the", "predictions", "over", "a", "sequence", "are", "smooth", "and", "temporally", "consistent", "which", "is", "apparent", "from", "our", "qualitative", "results", "as", "seen", "in", "Figure", "[", "reference", "]", "and", "Figure", "[", "reference", "]", ".", "To", "show", "the", "effectiveness", "of", "our", "model", "on", "detections", "from", "different", "2D", "pose", "detectors", ",", "we", "also", "experimented", "with", "the", "detections", "from", "CPM", "and", "from", "stacked", "-", "hourglass", "(", "SH", ")", "module", "which", "is", "not", "fine", "-", "tuned", "on", "Human3.6", "M", "dataset", ".", "We", "observe", "that", "even", "for", "the", "non", "-", "fine", "tuned", "stacked", "hourglass", "detections", ",", "our", "model", "achieves", "the", "state", "-", "of", "-", "the", "-", "art", "results", ".", "For", "detections", "from", "CPM", ",", "our", "model", "achieves", "competitive", "accuracy", "for", "the", "predictions", ".", "paragraph", ":", "Performance", "on", "different", "sequence", "lengths", "The", "results", "reported", "so", "far", "have", "been", "for", "input", "and", "output", "sequences", "of", "length", "5", ".", "We", "carried", "out", "experiments", "to", "see", "how", "our", "network", "performs", "for", "different", "sequence", "lengths", "ranging", "from", "2", "to", "10", ".", "The", "results", "are", "shown", "in", "Figure", "[", "reference", "]", ".", "As", "can", "be", "seen", ",", "the", "performance", "of", "our", "network", "remains", "stable", "for", "sequences", "of", "varying", "lengths", ".", "Even", "for", "a", "sequence", "length", "of", "2", ",", "which", "only", "considers", "the", "previous", "and", "the", "current", "frame", ",", "our", "model", "generates", "very", "good", "results", ".", "Particularly", "the", "best", "results", "were", "obtained", "for", "length", "4", ",", "5", "and", "6", ".", "However", ",", "we", "chose", "sequence", "length", "5", "for", "carrying", "out", "our", "experiments", "as", "a", "compromise", "between", "training", "time", "and", "accuracy", ".", "subsection", ":", "Qualitative", "Analysis", "We", "provide", "qualitative", "results", "on", "some", "videos", "of", "Human3.6", "M", "and", "Youtube", ".", "We", "apply", "the", "model", "trained", "on", "the", "Human3.6", "M", "dataset", "on", "some", "videos", "gathered", "from", "Youtube", ",", "The", "bounding", "box", "for", "each", "person", "in", "the", "Youtube", "video", "is", "labeled", "manually", "and", "for", "Human3.6", "M", "the", "ground", "truth", "bounding", "box", "is", "used", ".", "The", "2D", "poses", "are", "detected", "using", "the", "stacked", "-", "hourglass", "model", "fine", "-", "tuned", "on", "Human3.6", "M", "data", ".", "The", "qualitative", "result", "for", "Youtube", "videos", "is", "shown", "in", "Figure", "[", "reference", "]", "and", "for", "Human3.6", "M", "in", "Figure", "[", "reference", "]", ".", "The", "real", "advantage", "of", "using", "the", "temporal", "smoothness", "constraint", "during", "training", "is", "apparent", "in", "these", "figures", ".", "For", "Figure", "[", "reference", "]", ",", "we", "can", "see", "that", "even", "when", "the", "2D", "pose", "estimator", "breaks", "or", "generates", "extremely", "noisy", "detections", ",", "our", "system", "can", "recover", "temporally", "coherent", "3D", "poses", "by", "exploiting", "the", "temporal", "consistency", "information", ".", "A", "similar", "trend", "can", "also", "be", "found", "for", "Human3.6", "M", "videos", "in", "Figure", "[", "reference", "]", ",", "particularly", "for", "the", "action", "sitting", "down", "of", "subject", "11", ".", "We", "have", "provided", "more", "qualitative", "results", "in", "the", "supplementary", "material", ".", "section", ":", "Conclusion", "Both", "the", "quantitative", "and", "qualitative", "results", "for", "our", "network", "show", "the", "effectiveness", "of", "exploiting", "temporal", "information", "over", "multiple", "sequences", "to", "estimate", "3D", "poses", "which", "are", "temporally", "smooth", ".", "Our", "network", "achieved", "the", "best", "accuracy", "till", "date", "on", "all", "of", "the", "15", "action", "classes", "in", "the", "Human3.6", "M", "dataset", ".", "Particularly", ",", "most", "of", "the", "previous", "methods", "struggled", "with", "actions", "which", "have", "a", "high", "degree", "of", "occlusion", "like", "taking", "photo", ",", "talking", "on", "the", "phone", ",", "sitting", "and", "sitting", "down", ".", "Our", "network", "has", "significantly", "better", "results", "on", "these", "actions", ".", "Additionally", "we", "found", "that", "our", "network", "is", "reasonably", "robust", "to", "noisy", "2D", "poses", ".", "Although", "the", "contribution", "of", "temporal", "smoothness", "constraint", "is", "not", "apparent", "in", "the", "ablative", "analysis", "in", "Table", "[", "reference", "]", ",", "its", "effectiveness", "is", "clearly", "visible", "in", "the", "qualitative", "results", ",", "particularly", "on", "challenging", "Youtube", "videos", "(", "see", "Figure", "[", "reference", "]", ")", ".", "Our", "network", "effectively", "demonstrates", "the", "power", "of", "using", "temporal", "context", "information", "which", "we", "achieved", "using", "a", "sequence", "-", "to", "-", "sequence", "network", "that", "can", "be", "trained", "efficiently", "in", "a", "reasonably", "quick", "time", ".", "Also", "our", "network", "makes", "predictions", "from", "2D", "poses", "at", "3ms", "per", "frame", "on", "average", "which", "suggests", "that", ",", "given", "the", "2D", "pose", "detector", "is", "real", "time", ",", "our", "network", "can", "be", "applied", "in", "real", "-", "time", "scenarios", ".", "bibliography", ":", "References"], "entity": [[6, 10, ["3D", "human", "pose", "estimation"], "Task"], [19, 23, ["3D", "human", "pose", "estimation"], "Task"], [36, 38, ["deep", "networks"], "Method"], [55, 57, ["deep", "networks"], "Method"], [98, 101, ["2D", "pose", "estimator"], "Method"], [121, 125, ["low", "-", "dimensional", "representation"], "Method"], [143, 144, ["accuracy"], "Metric"], [162, 163, ["error"], "Metric"], [197, 203, ["sequence", "-", "to", "-", "sequence", "network"], "Method"], [245, 248, ["Human3.6", "M", "dataset"], "Material"], [268, 271, ["2D", "pose", "detector"], "Method"], [279, 283, ["estimating", "3D", "human", "pose"], "Task"], [284, 286, ["2D", "representations"], "Method"], [298, 303, ["computer", "vision", "and", "graphics", "community"], "Task"], [309, 316, ["understanding", "of", "human", "posture", "and", "limb", "articulation"], "Task"], [319, 324, ["high", "level", "computer", "vision", "tasks"], "Task"], [326, 331, ["human", "action", "or", "activity", "recognition"], "Task"], [332, 334, ["sports", "analysis"], "Task"], [335, 339, ["augmented", "and", "virtual", "reality"], "Task"], [341, 346, ["2D", "representation", "of", "human", "pose"], "Method"], [409, 414, ["3D", "representation", "of", "human", "pose"], "Task"], [427, 429, ["level", "tasks"], "Task"], [439, 441, ["computer", "animation"], "Task"], [471, 473, ["posed", "problem"], "Task"], [535, 537, ["3D", "pose"], "Task"], [646, 648, ["deep", "networks"], "Method"], [657, 661, ["deep", "convolutional", "neural", "network"], "Method"], [662, 663, ["CNN"], "Method"], [671, 673, ["3D", "poses"], "Task"], [694, 697, ["2D", "pose", "estimators"], "Method"], [724, 727, ["2D", "pose", "estimator"], "Method"], [742, 744, ["decoupled", "approach"], "Method"], [781, 787, ["sequence", "-", "to", "-", "sequence", "network"], "Method"], [788, 790, ["shortcut", "connections"], "Method"], [817, 822, ["Long", "Short", "-", "Term", "Memory"], "Method"], [823, 824, ["LSTM"], "Method"], [827, 829, ["layer", "normalization"], "Method"], [830, 832, ["recurrent", "dropout"], "Method"], [872, 875, ["Human3.6", "M", "dataset"], "Material"], [888, 889, ["error"], "Metric"], [894, 897, ["Human3.6", "M", "dataset"], "Material"], [910, 912, ["Youtube", "videos"], "Material"], [919, 922, ["2D", "pose", "detector"], "Method"], [949, 955, ["sequence", "-", "to", "-", "sequence", "network"], "Method"], [971, 974, ["Human3.6", "M", "dataset"], "Material"], [985, 991, ["sequence", "-", "to", "-", "sequence", "networks"], "Method"], [1066, 1078, ["model", "-", "based", "and", "model", "-", "free", "representations", "of", "3D", "human", "pose"], "Method"], [1088, 1092, ["model", "-", "based", "representation"], "Method"], [1136, 1139, ["sparse", "linear", "combination"], "Method"], [1258, 1261, ["Taylor", "’s", "algorithm"], "Method"], [1318, 1322, ["estimating", "3D", "joint", "locations"], "Task"], [1324, 1326, ["2D", "pose"], "Task"], [1376, 1379, ["sparse", "linear", "combination"], "Method"], [1424, 1427, ["rigid", "similarity", "transform"], "Method"], [1434, 1438, ["multi", "-", "dimensional", "scaling"], "Method"], [1439, 1440, ["MDS"], "Method"], [1457, 1460, ["fully", "connected", "network"], "Method"], [1485, 1489, ["Deep", "network", "based", "methods"], "Method"], [1493, 1495, ["deep", "networks"], "Method"], [1510, 1513, ["predict", "3D", "poses"], "Task"], [1527, 1528, ["CNNs"], "Method"], [1546, 1548, ["transfer", "learning"], "Method"], [1554, 1558, ["2D", "human", "pose", "estimation"], "Task"], [1589, 1592, ["3D", "volumetric", "heatmaps"], "Task"], [1600, 1603, ["2D", "pose", "estimator"], "Method"], [1604, 1607, ["Convolutional", "Pose", "Machine"], "Method"], [1608, 1609, ["CPM"], "Method"], [1634, 1635, ["CNNs"], "Method"], [1637, 1639, ["3D", "poses"], "Task"], [1649, 1651, ["unified", "network"], "Method"], [1704, 1710, ["in", "-", "the", "-", "wild", "images"], "Task"], [1725, 1727, ["pose", "grammar"], "Method"], [1770, 1775, ["tracking", "-", "by", "-", "detection"], "Method"], [1798, 1799, ["CNN"], "Method"], [1853, 1857, ["real", "-", "time", "system"], "Method"], [1863, 1865, ["temporal", "filtering"], "Method"], [1887, 1892, ["multi", "-", "stage", "sequential", "refinement"], "Method"], [1893, 1894, ["LSTMs"], "Method"], [1912, 1917, ["predicting", "temporally", "consistent", "3D", "poses"], "Task"], [1929, 1935, ["sequence", "-", "to", "-", "sequence", "network"], "Method"], [1973, 1975, ["Network", "Design"], "Method"], [1978, 1984, ["sequence", "-", "to", "-", "sequence", "network"], "Method"], [1988, 1990, ["residual", "connections"], "Method"], [2027, 2033, ["sequence", "-", "to", "-", "sequence", "network"], "Method"], [2041, 2044, ["Neural", "Machine", "Translation"], "Task"], [2045, 2046, ["NMT"], "Task"], [2075, 2078, ["language", "translation", "model"], "Method"], [2096, 2097, ["NMT"], "Task"], [2155, 2156, ["LSTMs"], "Method"], [2205, 2206, ["decoder"], "Method"], [2221, 2222, ["decoding"], "Task"], [2233, 2235, ["decoder", "unit"], "Method"], [2271, 2273, ["decoder", "unit"], "Method"], [2321, 2323, ["layer", "normalization"], "Method"], [2324, 2326, ["recurrent", "dropout"], "Method"], [2341, 2343, ["layer", "normalization"], "Method"], [2349, 2353, ["mean", "and", "standard", "deviation"], "Metric"], [2363, 2365, ["hidden", "layer"], "Method"], [2373, 2375, ["RNN", "units"], "Method"], [2387, 2388, ["dropout"], "Method"], [2427, 2429, ["Loss", "function"], "Method"], [2520, 2523, ["2D", "pose", "detectors"], "Method"], [2580, 2582, ["frame", "rate"], "Metric"], [2609, 2611, ["loss", "function"], "Method"], [2626, 2629, ["2D", "pose", "detector"], "Method"], [2687, 2688, ["error"], "Metric"], [2712, 2713, ["accuracy"], "Metric"], [2795, 2796, ["error"], "Metric"], [2799, 2801, ["loss", "function"], "Method"], [2810, 2813, ["Mean", "Squared", "Error"], "Metric"], [2814, 2815, ["MSE"], "Metric"], [2829, 2831, ["L2", "norm"], "Metric"], [2858, 2859, ["MSE"], "Metric"], [2890, 2894, ["mean", "of", "L2", "norm"], "Metric"], [2981, 2982, ["error"], "Metric"], [2985, 2987, ["loss", "function"], "Metric"], [3029, 3033, ["Human", "3.6", "M", "dataset"], "Material"], [3036, 3038, ["HumanEva", "dataset"], "Material"], [3039, 3042, ["Human", "3.6", "M"], "Material"], [3157, 3158, ["HumanEva"], "Material"], [3209, 3210, ["Youtube"], "Material"], [3212, 3215, ["Human3.6", "M", "dataset"], "Material"], [3223, 3226, ["Human3.6", "M", "dataset"], "Material"], [3256, 3257, ["error"], "Metric"], [3299, 3301, ["similarity", "transformation"], "Method"], [3302, 3304, ["Procrustes", "analysis"], "Method"], [3317, 3319, ["average", "error"], "Metric"], [3339, 3341, ["error", "metric"], "Metric"], [3344, 3346, ["HumanEva", "dataset"], "Material"], [3358, 3360, ["rigid", "alignment"], "Task"], [3377, 3379, ["2D", "detections"], "Task"], [3395, 3397, ["MPII", "dataset"], "Material"], [3402, 3405, ["2D", "pose", "estimation"], "Task"], [3412, 3415, ["Human3.6", "M", "dataset"], "Material"], [3417, 3420, ["2D", "pose", "estimations"], "Task"], [3476, 3478, ["learning", "rate"], "Metric"], [3492, 3495, ["single", "frame", "predictions"], "Method"], [3512, 3515, ["naive", "post", "processing"], "Method"], [3519, 3521, ["mean", "filter"], "Method"], [3523, 3525, ["median", "filter"], "Method"], [3556, 3562, ["non", "-", "rigid", "structure", "from", "motion"], "Method"], [3563, 3565, ["NRSFM", ")"], "Method"], [3574, 3577, ["3D", "reconstruction", "problem"], "Task"], [3600, 3602, ["NRSFM", "techniques"], "Method"], [3606, 3610, ["3D", "human", "pose", "estimation"], "Task"], [3615, 3617, ["NRSFM", "techniques"], "Method"], [3631, 3634, ["Human3.6", "M", "dataset"], "Material"], [3655, 3657, ["NRSFM", "algorithm"], "Method"], [3665, 3668, ["human", "pose", "reconstruction"], "Task"], [3670, 3672, ["specialized", "problem"], "Task"], [3683, 3687, ["Data", "pre", "-", "processing"], "Task"], [3696, 3700, ["noisy", "2D", "pose", "estimates"], "Method"], [3766, 3770, ["inverse", "rigid", "body", "transformation"], "Method"], [3832, 3835, ["deep", "learning", "systems"], "Method"], [3891, 3895, ["Human", "3.6", "M", "dataset"], "Material"], [3899, 3901, ["Adam", "optimizer"], "Method"], [3907, 3909, ["learning", "rate"], "Metric"], [3926, 3929, ["Xavier", "uniform", "initializer"], "Method"], [3961, 3962, ["training"], "Task"], [3964, 3965, ["accuracy"], "Metric"], [4014, 4015, ["Tensorflow"], "Method"], [4018, 4021, ["cross", "-", "validation"], "Metric"], [4044, 4047, ["cross", "-", "validation"], "Method"], [4068, 4070, ["training", "step"], "Method"], [4083, 4085, ["forward", "pass"], "Method"], [4103, 4105, ["pose", "detector"], "Method"], [4113, 4115, ["3D", "pose"], "Task"], [4126, 4129, ["estimated", "2D", "pose"], "Task"], [4154, 4157, ["Human3.6", "M", "dataset"], "Material"], [4180, 4181, ["error"], "Metric"], [4271, 4275, ["rigid", "body", "similarity", "transform"], "Method"], [4278, 4279, ["error"], "Metric"], [4304, 4305, ["mm"], "Method"], [4310, 4311, ["error"], "Metric"], [4350, 4351, ["error"], "Metric"], [4352, 4356, ["estimating", "3D", "joint", "locations"], "Task"], [4423, 4427, ["post", "-", "processing", "techniques"], "Method"], [4428, 4433, ["temporal", "mean", "and", "median", "filters"], "Method"], [4434, 4438, ["frame", "-", "wise", "prediction"], "Method"], [4452, 4455, ["action", "classes", "Walking"], "Task"], [4456, 4457, ["Jogging"], "Material"], [4459, 4461, ["HumanEva", "dataset"], "Material"], [4471, 4472, ["error"], "Metric"], [4481, 4483, ["average", "error"], "Metric"], [4499, 4501, ["action", "Walking"], "Task"], [4504, 4505, ["HumanEva"], "Material"], [4510, 4512, ["Human3.6", "M"], "Material"], [4527, 4530, ["2D", "ground", "truth"], "Task"], [4557, 4559, ["3D", "pose"], "Task"], [4600, 4603, ["lower", "bound", "error"], "Metric"], [4691, 4692, ["Robustness"], "Metric"], [4768, 4771, ["Nouguer", "’s", "model"], "Method"], [4839, 4840, ["error"], "Metric"], [4874, 4876, ["Ablative", "analysis"], "Method"], [4893, 4895, ["ablative", "analysis"], "Method"], [4903, 4905, ["ablative", "analysis"], "Task"], [4962, 4963, ["error"], "Metric"], [4971, 4973, ["layer", "normalization"], "Method"], [4978, 4979, ["error"], "Metric"], [4981, 4982, ["mm"], "Method"], [4994, 4995, ["error"], "Metric"], [5001, 5003, ["layer", "normalization"], "Method"], [5004, 5006, ["recurrent", "dropout"], "Method"], [5014, 5015, ["mm"], "Method"], [5080, 5081, ["detections"], "Task"], [5083, 5086, ["2D", "pose", "detectors"], "Method"], [5094, 5095, ["CPM"], "Method"], [5097, 5100, ["stacked", "-", "hourglass"], "Method"], [5101, 5102, ["SH"], "Method"], [5111, 5114, ["Human3.6", "M", "dataset"], "Material"], [5123, 5128, ["fine", "tuned", "stacked", "hourglass", "detections"], "Method"], [5145, 5146, ["CPM"], "Method"], [5150, 5152, ["competitive", "accuracy"], "Metric"], [5154, 5155, ["predictions"], "Task"], [5282, 5284, ["training", "time"], "Metric"], [5285, 5286, ["accuracy"], "Metric"], [5289, 5291, ["Qualitative", "Analysis"], "Method"], [5299, 5301, ["Human3.6", "M"], "Material"], [5302, 5303, ["Youtube"], "Material"], [5311, 5314, ["Human3.6", "M", "dataset"], "Material"], [5329, 5331, ["Youtube", "video"], "Material"], [5336, 5338, ["Human3.6", "M"], "Material"], [5369, 5371, ["Youtube", "videos"], "Material"], [5380, 5382, ["Human3.6", "M"], "Material"], [5418, 5421, ["2D", "pose", "estimator"], "Method"], [5517, 5518, ["accuracy"], "Metric"], [5529, 5532, ["Human3.6", "M", "dataset"], "Material"], [5600, 5602, ["ablative", "analysis"], "Method"], [5621, 5623, ["Youtube", "videos"], "Material"], [5647, 5653, ["sequence", "-", "to", "-", "sequence", "network"], "Method"], [5684, 5687, ["2D", "pose", "detector"], "Method"], [5697, 5701, ["real", "-", "time", "scenarios"], "Task"], [50, 52, ["3D", "pose"], "Task"], [82, 84, ["3D", "pose"], "Task"], [139, 141, ["3D", "pose"], "Task"], [149, 151, ["3D", "pose"], "Task"], [208, 209, ["LSTM"], "Method"], [432, 434, ["3D", "pose"], "Task"], [464, 466, ["3D", "pose"], "Task"], [498, 500, ["3D", "pose"], "Task"], [559, 561, ["3D", "pose"], "Task"], [606, 608, ["3D", "pose"], "Task"], [681, 683, ["3D", "pose"], "Task"], [720, 722, ["3D", "pose"], "Task"], [745, 747, ["3D", "pose"], "Task"], [752, 754, ["3D", "pose"], "Task"], [1063, 1065, ["3D", "pose"], "Task"], [1122, 1124, ["3D", "pose"], "Task"], [1132, 1134, ["3D", "pose"], "Task"], [1160, 1162, ["3D", "pose"], "Task"], [1165, 1167, ["3D", "pose"], "Task"], [1198, 1200, ["3D", "pose"], "Task"], [1311, 1313, ["3D", "pose"], "Task"], [1370, 1372, ["3D", "pose"], "Task"], [1562, 1564, ["3D", "pose"], "Task"], [1572, 1575, ["stacked", "-", "hourglass"], "Method"], [1614, 1616, ["3D", "pose"], "Task"], [1789, 1791, ["3D", "pose"], "Task"], [1840, 1842, ["3D", "pose"], "Task"], [1858, 1860, ["3D", "pose"], "Task"], [1878, 1880, ["3D", "pose"], "Task"], [1896, 1898, ["3D", "pose"], "Task"], [1906, 1908, ["3D", "pose"], "Task"], [1985, 1986, ["LSTM"], "Method"], [2150, 2151, ["LSTM"], "Method"], [2225, 2227, ["3D", "pose"], "Task"], [2237, 2239, ["3D", "pose"], "Task"], [2281, 2283, ["3D", "pose"], "Task"], [2294, 2296, ["3D", "pose"], "Task"], [2461, 2463, ["3D", "pose"], "Task"], [3058, 3060, ["3D", "pose"], "Task"], [3120, 3122, ["3D", "pose"], "Task"], [3262, 3264, ["3D", "pose"], "Task"], [3291, 3293, ["3D", "pose"], "Task"], [3330, 3332, ["3D", "pose"], "Task"], [3386, 3389, ["stacked", "-", "hourglass"], "Method"], [3527, 3529, ["3D", "pose"], "Task"], [3701, 3704, ["stacked", "-", "hourglass"], "Method"], [3921, 3922, ["LSTM"], "Method"], [4679, 4681, ["3D", "pose"], "Task"], [4974, 4975, ["LSTM"], "Method"], [5353, 5356, ["stacked", "-", "hourglass"], "Method"], [5361, 5363, ["Human3.6", "M"], "Material"], [5451, 5453, ["Human3.6", "M"], "Material"]]} {"doc_id": "38e2f851b705faa0d0a698ed9885bd6834440073", "text": ["Probabilistic", "Model", "-", "Agnostic", "Meta", "-", "Learning", "section", ":", "Abstract", "Meta", "-", "learning", "for", "few", "-", "shot", "learning", "entails", "acquiring", "a", "prior", "over", "previous", "tasks", "and", "experiences", ",", "such", "that", "new", "tasks", "be", "learned", "from", "small", "amounts", "of", "data", ".", "However", ",", "a", "critical", "challenge", "in", "few", "-", "shot", "learning", "is", "task", "ambiguity", ":", "even", "when", "a", "powerful", "prior", "can", "be", "meta", "-", "learned", "from", "a", "large", "number", "of", "prior", "tasks", ",", "a", "small", "dataset", "for", "a", "new", "task", "can", "simply", "be", "too", "ambiguous", "to", "acquire", "a", "single", "model", "(", "e.g.", ",", "a", "classifier", ")", "for", "that", "task", "that", "is", "accurate", ".", "In", "this", "paper", ",", "we", "propose", "a", "probabilistic", "meta", "-", "learning", "algorithm", "that", "can", "sample", "models", "for", "a", "new", "task", "from", "a", "model", "distribution", ".", "Our", "approach", "extends", "model", "-", "agnostic", "meta", "-", "learning", ",", "which", "adapts", "to", "new", "tasks", "via", "gradient", "descent", ",", "to", "incorporate", "a", "parameter", "distribution", "that", "is", "trained", "via", "a", "variational", "lower", "bound", ".", "At", "meta", "-", "test", "time", ",", "our", "algorithm", "adapts", "via", "a", "simple", "procedure", "that", "injects", "noise", "into", "gradient", "descent", ",", "and", "at", "meta", "-", "training", "time", ",", "the", "model", "is", "trained", "such", "that", "this", "stochastic", "adaptation", "procedure", "produces", "samples", "from", "the", "approximate", "model", "posterior", ".", "Our", "experimental", "results", "show", "that", "our", "method", "can", "sample", "plausible", "classifiers", "and", "regressors", "in", "ambiguous", "few", "-", "shot", "learning", "problems", ".", "section", ":", "Introduction", "Learning", "from", "a", "few", "examples", "is", "a", "key", "aspect", "of", "human", "intelligence", ".", "One", "way", "to", "make", "it", "possible", "to", "acquire", "solutions", "to", "complex", "tasks", "from", "only", "a", "few", "examples", "is", "to", "leverage", "past", "experience", "to", "learn", "a", "prior", "over", "tasks", ".", "The", "process", "of", "learning", "this", "prior", "entails", "discovering", "the", "shared", "structure", "across", "different", "tasks", "from", "the", "same", "family", ",", "such", "as", "commonly", "occurring", "visual", "features", "or", "semantic", "cues", ".", "Structure", "is", "useful", "insofar", "as", "it", "yields", "efficient", "learning", "of", "new", "tasks", "-", "a", "mechanism", "known", "as", "learning", "-", "to", "-", "learn", ",", "or", "meta", "-", "learning", "[", "reference", "]", ".", "However", ",", "when", "the", "end", "goal", "of", "few", "-", "shot", "meta", "-", "learning", "is", "to", "learn", "solutions", "to", "new", "tasks", "from", "small", "amounts", "of", "data", ",", "a", "critical", "issue", "that", "must", "be", "dealt", "with", "is", "task", "ambiguity", ":", "even", "with", "the", "best", "possible", "prior", ",", "there", "might", "simply", "not", "be", "enough", "information", "in", "the", "examples", "for", "a", "new", "task", "to", "resolve", "that", "task", "with", "high", "certainty", ".", "It", "is", "therefore", "quite", "desireable", "to", "develop", "few", "-", "shot", "meta", "-", "learning", "methods", "that", "can", "propose", "multiple", "potential", "solutions", "to", "an", "ambiguous", "few", "-", "shot", "learning", "problem", ".", "Such", "a", "method", "could", "be", "used", "to", "evaluate", "uncertainty", "(", "by", "measuring", "agreement", "between", "the", "samples", ")", ",", "perform", "active", "learning", ",", "or", "elicit", "direct", "human", "supervision", "about", "which", "sample", "is", "preferable", ".", "For", "example", ",", "in", "safety", "-", "critical", "applications", ",", "such", "as", "few", "-", "shot", "medical", "image", "classification", ",", "uncertainty", "is", "crucial", "for", "determining", "if", "the", "learned", "classifier", "should", "be", "trusted", ".", "When", "learning", "from", "such", "small", "amounts", "of", "data", ",", "uncertainty", "estimation", "can", "also", "help", "predict", "if", "additional", "data", "would", "be", "beneficial", "for", "learning", "and", "improving", "the", "estimate", "of", "the", "rewards", ".", "Finally", ",", "while", "we", "do", "not", "experiment", "with", "this", "in", "this", "paper", ",", "we", "expect", "that", "modeling", "this", "ambiguity", "will", "be", "helpful", "for", "reinforcement", "learning", "problems", ",", "where", "it", "can", "be", "used", "to", "aid", "in", "exploration", ".", "While", "recognizing", "and", "accounting", "for", "ambiguity", "is", "an", "important", "aspect", "of", "the", "few", "-", "shot", "learning", "problem", ",", "it", "is", "particularly", "challenging", "to", "model", "when", "scaling", "to", "high", "-", "dimensional", "data", ",", "large", "function", "approximators", ",", "and", "multimodal", "task", "structure", ".", "Representing", "distributions", "over", "functions", "is", "relatively", "straightforward", "when", "using", "simple", "function", "approximators", ",", "such", "as", "linear", "functions", ",", "and", "has", "been", "done", "extensively", "in", "early", "few", "-", "shot", "learning", "approaches", "using", "Bayesian", "models", "[", "reference", "][", "reference", "]", ".", "But", "this", "problem", "becomes", "substantially", "more", "challenging", "when", "reasoning", "over", "high", "-", "dimensional", "function", "approximators", "such", "as", "deep", "neural", "networks", ",", "since", "explicitly", "representing", "expressive", "distributions", "over", "thousands", "or", "millions", "of", "parameters", "if", "often", "intractable", ".", "As", "a", "result", ",", "recent", "more", "scalable", "approaches", "to", "few", "-", "shot", "learning", "have", "focused", "on", "acquiring", "deterministic", "learning", "algorithms", "that", "disregard", "ambiguity", "over", "the", "underlying", "function", ".", "Can", "we", "develop", "an", "approach", "that", "has", "the", "benefits", "of", "both", "classes", "of", "few", "-", "shot", "learning", "methods", "-", "scalability", "and", "uncertainty", "awareness", "?", "To", "do", "so", ",", "we", "build", "upon", "tools", "in", "amortized", "variational", "inference", "for", "developing", "a", "probabilistic", "meta", "-", "learning", "approach", ".", "In", "particular", ",", "our", "method", "builds", "on", "model", "-", "agnostic", "meta", "-", "learning", "(", "MAML", ")", "[", "reference", "]", ",", "a", "few", "shot", "metalearning", "algorithm", "that", "uses", "standard", "gradient", "descent", "to", "adapt", "the", "model", "at", "meta", "-", "test", "time", "to", "a", "new", "few", "-", "shot", "task", ",", "and", "trains", "the", "model", "parameters", "at", "meta", "-", "training", "time", "to", "enable", "rapid", "adaptation", ",", "essentially", "optimizing", "for", "a", "neural", "network", "initialization", "that", "is", "well", "-", "suited", "for", "few", "shot", "learning", ".", "MAML", "can", "be", "shown", "to", "retain", "the", "generality", "of", "black", "-", "box", "meta", "-", "learners", "such", "as", "RNNs", "[", "reference", "]", ",", "while", "being", "applicable", "to", "standard", "neural", "network", "architectures", ".", "Our", "approach", "extends", "MAML", "to", "model", "a", "distribution", "over", "prior", "model", "parameters", ",", "which", "leads", "to", "an", "appealing", "simple", "stochastic", "adaptation", "procedure", "that", "simply", "injects", "noise", "into", "gradient", "descent", "at", "meta", "-", "test", "time", ".", "The", "meta", "-", "training", "procedure", "then", "optimizes", "for", "this", "simple", "inference", "process", "to", "produce", "samples", "from", "an", "approximate", "model", "posterior", ".", "The", "primary", "contribution", "of", "this", "paper", "is", "a", "reframing", "of", "MAML", "as", "a", "graphical", "model", "inference", "problem", ",", "where", "variational", "inference", "can", "provide", "us", "with", "a", "principled", "and", "natural", "mechanism", "for", "modeling", "uncertainty", "and", "ambiguity", ".", "Our", "approach", "enables", "sampling", "multiple", "potential", "solutions", "to", "a", "few", "-", "shot", "learning", "problem", "at", "meta", "-", "test", "time", ",", "and", "our", "experiments", "show", "that", "this", "ability", "can", "be", "utilized", "to", "sample", "multiple", "possible", "regressors", "for", "an", "ambiguous", "regression", "problem", ",", "as", "well", "as", "multiple", "possible", "classifiers", "for", "ambiguous", "few", "-", "shot", "attribute", "classification", "tasks", ".", "section", ":", "Related", "Work", "Hierarchical", "Bayesian", "models", "are", "a", "long", "-", "standing", "approach", "for", "few", "-", "shot", "learning", "that", "naturally", "allow", "for", "the", "ability", "to", "reason", "about", "uncertainty", "over", "functions", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "While", "these", "approaches", "have", "been", "demonstrated", "on", "simple", "few", "-", "shot", "image", "classification", "datasets", "[", "reference", "]", ",", "they", "have", "yet", "to", "scale", "to", "the", "more", "complex", "problems", ",", "such", "as", "the", "experiments", "in", "this", "paper", ".", "A", "number", "of", "works", "have", "approached", "the", "problem", "of", "few", "-", "shot", "learning", "from", "a", "meta", "-", "learning", "perspective", "[", "reference", "][", "reference", "]", ",", "including", "black", "-", "box", "[", "reference", "][", "reference", "][", "reference", "]", "and", "optimization", "-", "based", "approaches", "[", "reference", "][", "reference", "]", ".", "While", "these", "approaches", "scale", "to", "large", "-", "scale", "image", "datasets", "[", "reference", "]", "and", "visual", "reinforcement", "learning", "problems", "[", "reference", "]", ",", "they", "typically", "lack", "the", "ability", "to", "reason", "about", "uncertainty", ".", "Our", "work", "is", "most", "related", "to", "methods", "that", "combine", "deep", "networks", "and", "probabilistic", "methods", "for", "few", "-", "shot", "learning", "[", "reference", "][", "reference", "][", "reference", "]", ".", "One", "approach", "that", "considers", "hierarchical", "Bayesian", "models", "for", "few", "-", "shot", "learning", "is", "the", "neural", "statistician", "[", "reference", "]", ",", "which", "uses", "an", "explicit", "task", "variable", "to", "model", "task", "distributions", ".", "Our", "method", "is", "fully", "model", "agnostic", ",", "and", "directly", "samples", "model", "weights", "for", "each", "task", "for", "any", "network", "architecture", ".", "Our", "experiments", "show", "that", "our", "approach", "improves", "on", "MAML", "[", "reference", "]", ",", "which", "outperforms", "the", "model", "by", "Edwards", "and", "Storkey", "[", "reference", "]", ".", "Other", "work", "that", "considers", "model", "uncertainty", "in", "the", "few", "-", "shot", "learning", "setting", "is", "the", "LLAMA", "method", "[", "reference", "]", ",", "which", "also", "builds", "on", "the", "MAML", "algorithm", ".", "LLAMA", "makes", "use", "of", "a", "local", "Laplace", "approximation", "for", "modeling", "the", "task", "parameters", "(", "post", "-", "update", "parameters", ")", ",", "which", "introduces", "the", "need", "to", "approximate", "a", "high", "dimensional", "covariance", "matrix", ".", "We", "instead", "propose", "a", "method", "that", "approximately", "infers", "the", "pre", "-", "update", "parameters", ",", "which", "we", "make", "tractable", "through", "a", "choice", "of", "approximate", "posterior", "parameterized", "by", "gradient", "operations", ".", "Bayesian", "neural", "networks", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "have", "been", "studied", "extensively", "as", "a", "way", "to", "incorporate", "uncertainty", "into", "deep", "networks", ".", "Although", "exact", "inference", "in", "Bayesian", "neural", "networks", "is", "impractical", ",", "approximations", "based", "on", "backpropagation", "and", "sampling", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "have", "been", "effective", "in", "incorporating", "uncertainty", "into", "the", "weights", "of", "generic", "networks", ".", "Our", "approach", "differs", "from", "these", "methods", "in", "that", "we", "explicitly", "train", "a", "hierarchical", "Bayesian", "model", "over", "weights", ",", "where", "a", "posterior", "task", "-", "specific", "parameter", "distribution", "is", "inferred", "at", "meta", "-", "test", "time", "conditioned", "on", "a", "learned", "weight", "prior", "and", "a", "(", "few", "-", "shot", ")", "training", "set", ",", "while", "conventional", "Bayesian", "neural", "networks", "directly", "learn", "only", "the", "posterior", "weight", "distribution", "for", "a", "single", "task", ".", "Our", "method", "draws", "on", "amortized", "variational", "inference", "methods", "[", "reference", "][", "reference", "][", "reference", "]", "to", "make", "this", "possible", ",", "but", "the", "key", "modification", "is", "that", "the", "model", "and", "inference", "networks", "share", "the", "same", "parameters", ".", "The", "resulting", "method", "corresponds", "structurally", "to", "a", "Bayesian", "version", "of", "model", "-", "agnostic", "meta", "-", "learning", "[", "reference", "]", ".", "into", "the", "center", "model", "after", "performing", "inference", "over", "φi", ".", "We", "find", "it", "beneficial", "to", "introduce", "additional", "dependencies", "of", "the", "prior", "on", "the", "training", "data", "to", "compensate", "for", "using", "the", "MAP", "estimate", "to", "approximate", "p", "(", "φi", ")", ",", "as", "shown", "on", "the", "right", ".", "section", ":", "Preliminaries", "In", "the", "meta", "-", "learning", "problem", "setting", "that", "we", "consider", ",", "the", "goal", "is", "to", "learn", "models", "that", "can", "learn", "new", "tasks", "from", "small", "amounts", "of", "data", ".", "To", "do", "so", ",", "meta", "-", "learning", "algorithms", "require", "a", "set", "of", "meta", "-", "training", "and", "meta", "-", "testing", "tasks", "drawn", "from", "some", "distribution", "p", "(", "T", ")", ".", "The", "key", "assumption", "of", "learning", "-", "to", "-", "learn", "is", "that", "the", "tasks", "in", "this", "distribution", "share", "common", "structure", "that", "can", "be", "exploited", "for", "faster", "learning", "of", "new", "tasks", ".", "Thus", ",", "the", "goal", "of", "the", "meta", "-", "learning", "process", "is", "to", "discover", "that", "structure", ".", "In", "this", "section", ",", "we", "will", "introduce", "notation", "and", "overview", "the", "model", "-", "agnostic", "meta", "-", "learning", "(", "MAML", ")", "algorithm", "[", "reference", "]", ".", "Meta", "-", "learning", "algorithms", "proceed", "by", "sampling", "data", "from", "a", "given", "task", ",", "and", "splitting", "the", "sampled", "data", "into", "a", "set", "of", "a", "few", "datapoints", ",", "D", "where", "φ", "i", "is", "used", "to", "denote", "the", "parameters", "updated", "by", "gradient", "descent", "and", "where", "the", "loss", "corresponds", "to", "negative", "log", "likelihood", "of", "the", "data", ".", "In", "particular", ",", "in", "the", "case", "of", "supervised", "classification", "with", "inputs", "{", "x", "j", "}", ",", "their", "corresponding", "labels", "{", "y", "j", "}", ",", "and", "a", "classifier", "f", "θ", ",", "we", "will", "denote", "the", "negative", "log", "likelihood", "of", "the", "data", "under", "the", "classifier", "as", "L", "(", "θ", ",", "D", ")", "=", "−", "(", "xj", ",", "yj", ")", "∈D", "log", "p", "(", "y", "j", "|x", "j", ",", "θ", ")", ".", "This", "corresponds", "to", "the", "cross", "entropy", "loss", "function", ".", "section", ":", "Method", "Our", "goal", "is", "to", "build", "a", "meta", "-", "learning", "method", "that", "can", "handle", "the", "uncertainty", "and", "ambiguity", "that", "occurs", "when", "learning", "from", "small", "amounts", "of", "data", ",", "while", "scaling", "to", "highly", "-", "expressive", "function", "approximators", "such", "as", "neural", "networks", ".", "To", "do", "so", ",", "we", "set", "up", "a", "graphical", "model", "for", "the", "few", "-", "shot", "learning", "problem", ".", "In", "particular", ",", "we", "want", "a", "hierarchical", "Bayesian", "model", "that", "includes", "random", "variables", "for", "the", "prior", "distribution", "over", "function", "parameters", ",", "θ", ",", "the", "distribution", "over", "parameters", "for", "a", "particular", "task", ",", "φ", "i", ",", "and", "the", "task", "training", "and", "test", "datapoints", ".", "This", "graphical", "model", "is", "illustrated", "in", "Figure", "1", "(", "left", ")", ",", "where", "tasks", "are", "indexed", "over", "i", "and", "datapoints", "are", "indexed", "over", "j.", "We", "will", "use", "the", "shorthand", "x", ".", "Therefore", ",", "posterior", "inference", "over", "φ", "i", "must", "take", "into", "account", "both", "the", "evidence", "(", "training", "set", ")", "and", "the", "prior", "imposed", "by", "p", "(", "θ", ")", "and", "p", "(", "φ", "i", "|θ", ")", ".", "Conventional", "MAML", "can", "be", "interpreted", "as", "approximating", "maximum", "a", "posteriori", "inference", "under", "a", "simplified", "model", "where", "p", "(", "θ", ")", "is", "a", "delta", "function", ",", "and", "inference", "is", "performed", "by", "running", "gradient", "descent", "on", "log", "p", "(", "y", "tr", "|x", "tr", ",", "φ", "i", ")", "for", "a", "fixed", "number", "of", "iterations", "starting", "from", "φ", "[", "reference", "]", ".", "The", "corresponding", "distribution", "p", "(", "φ", "i", "|θ", ")", "is", "approximately", "Gaussian", ",", "with", "a", "mean", "that", "depends", "on", "the", "step", "size", "and", "number", "of", "gradient", "steps", ".", "When", "p", "(", "θ", ")", "is", "not", "deterministic", ",", "we", "must", "make", "a", "further", "approximation", "to", "account", "for", "the", "random", "variable", "θ", ".", "One", "way", "we", "can", "do", "this", "is", "by", "using", "structured", "variational", "inference", ".", "In", "structured", "variational", "inference", ",", "we", "approximate", "the", "distribution", "over", "the", "hidden", "variables", "θ", "and", "φ", "i", "for", "each", "task", "with", "some", "approximate", "distribution", "q", "i", "(", "θ", ",", "φ", "i", ")", ".", "There", "are", "two", "reasonable", "choices", "we", "can", "make", "for", "q", "i", "(", "θ", ",", "φ", "i", ")", ".", "First", ",", "we", "can", "approximate", "it", "as", "a", "product", "of", "independent", "marginals", ",", "according", "to", "However", ",", "this", "approximation", "does", "not", "permit", "uncertainty", "to", "propagate", "effectively", "from", "θ", "to", "φ", "i", ".", "A", "more", "expressive", "approximation", "is", "the", "structured", "variational", "approximation", "We", "can", "further", "avoid", "storing", "a", "separate", "variational", "distribution", "q", "i", "(", "φ", "i", "|θ", ")", "and", "q", "i", "(", "θ", ")", "for", "each", "task", "T", "i", "by", "employing", "an", "amortized", "variational", "inference", "technique", "[", "reference", "][", "reference", "][", "reference", "]", ",", "where", "we", "instead", "set", ",", "where", "q", "ψ", "is", "defined", "by", "some", "function", "approximator", "with", "parameters", "ψ", "that", "takes", "x", "tr", "i", ",", "y", "tr", "i", "as", "input", ",", "and", "the", "same", "q", "ψ", "is", "used", "for", "all", "tasks", ".", "Similarly", ",", "we", "can", "define", "We", "can", "now", "write", "down", "the", "variational", "lower", "bound", "on", "the", "log", "-", "likelihood", "as", "The", "likelihood", "terms", "on", "the", "first", "line", "can", "be", "evaluated", "efficiently", ":", "given", "a", "sample", "θ", ",", "φ", "i", "∼", "q", "(", "θ", ",", "φ", "i", "|x", ",", "the", "training", "and", "test", "likelihoods", "simply", "correspond", "to", "the", "loss", "of", "the", "network", "with", "parameters", "φ", "i", ".", "The", "prior", "p", "(", "θ", ")", "can", "be", "chosen", "to", "be", "Gaussian", ",", "with", "a", "learned", "mean", "and", "(", "diagonal", ")", "covariance", "to", "provide", "for", "flexibility", "to", "choose", "the", "prior", "parameters", ".", "This", "corresponds", "to", "a", "Bayesian", "version", "of", "the", "MAML", "algorithm", ".", "We", "will", "define", "these", "parameters", "as", "µ", "θ", "and", "σ", "2", "θ", ".", "Lastly", ",", "p", "(", "φ", "i", "|θ", ")", "must", "be", "chosen", ".", "This", "choice", "is", "more", "delicate", ".", "One", "way", "to", "ensure", "a", "tractable", "likelihood", "is", "to", "use", "a", "Gaussian", "with", "mean", "θ", ".", "This", "choice", "is", "reasonable", ",", "because", "it", "encourages", "φ", "i", "to", "stay", "close", "to", "the", "prior", "parameters", "φ", "i", ",", "but", "we", "will", "see", "in", "the", "next", "section", "how", "a", "more", "expressive", "implicit", "conditional", "can", "be", "obtained", "using", "gradient", "descent", ",", "resulting", "in", "a", "procedure", "that", "more", "closely", "resembles", "the", "original", "MAML", "algorithm", "while", "still", "modeling", "the", "uncertainty", ".", "Lastly", ",", "we", "must", "choose", "a", "form", "for", "the", "inference", "networks", "q", "ψ", "(", "φ", "i", "|θ", ",", "They", "must", "be", "chosen", "so", "that", "their", "entropies", "on", "the", "second", "line", "of", "the", "above", "equation", "are", "tractable", ".", "Furthermore", ",", "note", "that", "both", "of", "these", "distributions", "model", "very", "high", "-", "dimensional", "random", "variables", ":", "a", "deep", "neural", "network", "can", "have", "hundreds", "of", "thousands", "or", "millions", "of", "parameters", ".", "So", "while", "we", "can", "use", "an", "arbitrary", "function", "approximator", ",", "we", "would", "like", "to", "find", "a", "scalable", "solution", ".", "One", "convenient", "solution", "is", "to", "allow", "q", "ψ", "to", "reuse", "the", "learned", "mean", "of", "the", "prior", "µ", "θ", ".", "We", "observe", "that", "adapting", "the", "parameters", "with", "gradient", "descent", "is", "a", "good", "way", "to", "update", "them", "to", "a", "given", "training", "set", "x", "where", "v", "q", "is", "a", "learned", "(", "diagonal", ")", "covariance", ",", "and", "the", "mean", "has", "an", "additional", "parameter", "beyond", "µ", "θ", ",", "which", "is", "a", "\"", "learning", "rate", "\"", "vector", "γ", "q", "that", "is", "pointwise", "multiplied", "with", "the", "gradient", ".", "While", "this", "choice", "may", "at", "first", "seem", "arbitrary", ",", "there", "is", "a", "simple", "intuition", ":", "the", "inference", "network", "should", "produce", "a", "sample", "of", "θ", "that", "is", "close", "to", "the", "posterior", "p", "(", "θ|x", "A", "reasonable", "way", "to", "arrive", "at", "a", "value", "of", "θ", "close", "to", "this", "posterior", "is", "to", "adapt", "it", "to", "both", "the", "training", "set", "and", "test", "set", ".", "[", "reference", "]", "Note", "that", "this", "is", "only", "done", "during", "meta", "-", "training", ".", "It", "remains", "to", "choose", "q", "ψ", "(", "φ", "i", "|θ", ",", ",", "which", "can", "also", "be", "formulated", "as", "a", "conditional", "Gaussian", "with", "mean", "given", "by", "applying", "gradient", "descent", ".", "Although", "this", "variational", "distribution", "is", "substantially", "more", "compact", "in", "terms", "of", "parameters", "than", "a", "separate", "neural", "network", ",", "it", "only", "provides", "estimates", "of", "the", "posterior", "during", "meta", "-", "training", ".", "At", "meta", "-", "test", "time", ",", "we", "must", "obtain", "the", "posterior", "p", "(", "φ", "i", "|x", ".", "We", "can", "train", "a", "separate", "set", "of", "inference", "networks", "to", "perform", "this", "operation", ",", "potentially", "also", "using", "gradient", "descent", "within", "the", "inference", "network", ".", "However", ",", "these", "networks", "do", "not", "receive", "any", "gradient", "information", "during", "meta", "-", "training", ",", "and", "may", "not", "work", "well", "in", "practice", ".", "In", "the", "next", "section", "we", "propose", "an", "even", "simpler", "and", "more", "practical", "approach", "that", "uses", "only", "a", "single", "inference", "network", "during", "meta", "-", "training", ",", "and", "none", "during", "meta", "-", "testing", ".", "section", ":", "Algorithm", "1", "Meta", "-", "training", ",", "differences", "from", "MAML", "in", "red", "Require", ":", "p", "(", "T", ")", ":", "distribution", "over", "tasks", "1", ":", "initialize", "Θ", ":", "=", "{", "µ", "θ", ",", "σ", "2", "θ", ",", "vq", ",", "γp", ",", "γq", "}", "2", ":", "while", "not", "done", "do", "3", ":", "Sample", "batch", "of", "tasks", "Ti", "∼", "p", "(", "T", ")", "4", ":", "for", "all", "Ti", "do", "5", ":", "Evaluate", "Compute", "adapted", "parameters", "with", "gradient", "descent", ":", "Compute", "∇Θ", "Update", "Θ", "using", "Adam", "section", ":", "Algorithm", "2", "Meta", "-", "testing", "section", ":", "Probabilistic", "Model", "-", "Agnostic", "Meta", "-", "Learning", "Approach", "with", "Hybrid", "Inference", "To", "formulate", "a", "simpler", "variational", "meta", "-", "learning", "procedure", ",", "we", "recall", "the", "probabilistic", "interpretation", "of", "MAML", ":", "as", "discussed", "by", "Grant", "et", "al", ".", "[", "reference", "]", ",", "MAML", "can", "be", "interpreted", "as", "approximate", "inference", "for", "the", "posterior", "p", "(", "y", "where", "we", "use", "the", "maximum", "a", "posteriori", "(", "MAP", ")", "value", "φ", "i", ".", "It", "can", "be", "shown", "that", ",", "for", "likelihoods", "that", "are", "Gaussian", "in", "φ", "i", ",", "gradient", "descent", "for", "a", "fixed", "number", "of", "iterations", "using", "x", "tr", "i", ",", "y", "tr", "i", "corresponds", "exactly", "to", "maximum", "a", "posteriori", "inference", "under", "a", "Gaussian", "prior", "p", "(", "φ", "i", "|θ", ")", "[", "reference", "]", ".", "In", "the", "case", "of", "non", "-", "Gaussian", "likelihoods", ",", "the", "equivalence", "is", "only", "locally", "approximate", ",", "and", "the", "exact", "form", "of", "the", "prior", "p", "(", "φ", "i", "|θ", ")", "is", "intractable", ".", "However", ",", "in", "practice", "this", "implicit", "prior", "can", "actually", "be", "preferable", "to", "an", "explicit", "(", "and", "simple", ")", "Gaussian", "prior", ",", "since", "it", "incorporates", "the", "rich", "nonlinear", "structure", "of", "the", "neural", "network", "parameter", "manifold", ",", "and", "produces", "good", "performance", "in", "practice", "[", "reference", "][", "reference", "]", ".", "We", "can", "interpret", "this", "MAP", "approximation", "as", "inferring", "an", "approximate", "posterior", "on", "φ", "i", "of", "the", "form", "p", "(", "φ", "i", "|x", "where", "φ", "i", "is", "obtained", "via", "gradient", "descent", "on", "the", "training", "set", "x", "tr", "i", ",", "y", "tr", "i", "starting", "from", "θ", ".", "Incorporating", "this", "approximate", "inference", "procedure", "transforms", "the", "graphical", "model", "in", "Figure", "1", "(", "a", ")", "into", "the", "one", "in", "Figure", "1", "(", "b", ")", ",", "where", "there", "is", "now", "a", "factor", "over", "p", "(", "φ", "i", "|x", "tr", "i", ",", "y", "tr", "i", ",", "θ", ")", ".", "While", "this", "is", "a", "crude", "approximation", "to", "the", "likelihood", ",", "it", "provides", "us", "with", "an", "empirically", "effective", "and", "simple", "tool", "that", "greatly", "simplifies", "the", "variational", "inference", "procedure", "described", "in", "the", "previous", "section", ",", "in", "the", "case", "where", "we", "aim", "to", "model", "a", "distribution", "over", "the", "global", "parameters", "p", "(", "θ", ")", ".", "After", "using", "gradient", "descent", "to", "estimate", "p", "(", "φ", "i", "|", "x", "is", "not", "observed", ".", "Thus", ",", "we", "can", "now", "write", "down", "a", "variational", "lower", "bound", "for", "the", "logarithm", "of", "the", "approximate", "likelihood", "on", "the", "second", "line", ",", "which", "is", "given", "by", "In", "this", "bound", ",", "we", "essentially", "perform", "approximate", "inference", "via", "MAP", "on", "φ", "i", "to", "obtain", "p", "(", "φ", "i", "|x", "To", "evaluate", "the", "variational", "lower", "bound", "during", "training", ",", "we", "can", "use", "the", "following", "procedure", ":", "first", ",", "we", "evaluate", "the", "mean", "by", "starting", "from", "µ", "θ", "and", "taking", "one", "(", "or", "more", ")", "gradient", "steps", "on", "log", "p", "(", "y", "test", "i", "|x", "test", "i", ",", "θ", "current", ")", ",", "where", "θ", "current", "starts", "at", "µ", "θ", ".", "We", "then", "add", "noise", "with", "variance", "v", "q", ",", "which", "is", "made", "differentiable", "via", "the", "reparameterization", "trick", "[", "reference", "]", ".", "We", "then", "take", "additional", "gradient", "steps", "on", "the", "training", "likelihood", "log", "p", "(", "y", "tr", "i", "|x", "tr", "i", ",", "θ", "current", ")", ".", "This", "accounts", "for", "the", "MAP", "inference", "procedure", "on", "φ", "i", ".", "Training", "of", "µ", "θ", ",", "σ", "2", "θ", ",", "and", "v", "q", "is", "performed", "by", "backpropagating", "gradients", "through", "this", "entire", "procedure", "with", "respect", "to", "the", "variational", "lower", "bound", ",", "which", "includes", "a", "term", "for", "the", "likelihood", "tr", ",", "y", "tr", ",", "φ", "i", ")", "and", "the", "KL", "-", "divergence", "between", "the", "sample", "θ", "∼", "q", "ψ", "and", "the", "prior", "p", "(", "θ", ")", ".", "This", "meta", "-", "training", "procedure", "is", "detailed", "in", "Algorithm", "1", ".", "At", "meta", "-", "test", "time", ",", "the", "inference", "procedure", "is", "much", "simpler", ".", "The", "test", "labels", "are", "not", "available", ",", "so", "we", "simply", "sample", "θ", "∼", "p", "(", "θ", ")", "and", "perform", "MAP", "inference", "on", "φ", "i", "using", "the", "training", "set", ",", "which", "corresponds", "to", "gradient", "steps", "on", "log", "p", "(", "y", "tr", "i", "|x", "tr", "i", ",", "θ", "current", ")", ",", "where", "θ", "current", "starts", "at", "the", "sampled", "θ", ".", "This", "meta", "-", "testing", "procedure", "is", "detailed", "in", "Algorithm", "2", ".", "section", ":", "Adding", "Additional", "Dependencies", "In", "the", "transformed", "graphical", "model", ",", "the", "training", "data", "x", "tr", "i", ",", "y", "tr", "i", "and", "the", "prior", "θ", "are", "conditionally", "independent", ".", "However", ",", "since", "we", "have", "only", "a", "crude", "approximation", "to", "p", "(", "φ", "i", "|", "x", "tr", "i", ",", "y", "tr", "i", ",", "θ", ")", ",", "this", "independence", "often", "does", "n't", "actually", "hold", ".", "We", "can", "allow", "the", "model", "to", "compensate", "for", "this", "approximation", "by", "additionally", "conditioning", "the", "learned", "prior", "p", "(", "θ", ")", "on", "the", "training", "data", ".", "In", "this", "case", ",", "the", "learned", "\"", "prior", "\"", "has", "the", "form", "p", "(", "θ", "i", "|x", "tr", "i", ",", "y", "tr", "i", ")", ",", "where", "θ", "i", "is", "now", "task", "-", "specific", ",", "but", "with", "global", "parameters", "µ", "θ", "and", "σ", "2", "θ", ".", "We", "thus", "obtain", "the", "modified", "graphical", "model", "in", "Figure", "1", "(", "c", ")", ".", "Similarly", "to", "the", "inference", "network", "q", "ψ", ",", "we", "parameterize", "the", "learned", "prior", "as", "follows", ":", "With", "this", "new", "form", "for", "distribution", "over", "θ", ",", "the", "variational", "training", "objective", "uses", "the", "likelihood", "term", "log", "p", "(", "θ", "i", "|x", "In", "our", "experiments", ",", "we", "find", "that", "this", "more", "expressive", "distribution", "often", "leads", "to", "better", "performance", ".", "section", ":", "Experiments", "The", "goal", "of", "our", "experimental", "evaluation", "is", "to", "answer", "the", "following", "questions", ":", "(", "1", ")", "can", "our", "approach", "enable", "sampling", "from", "the", "distribution", "over", "potential", "functions", "underlying", "the", "training", "data", "?", ",", "(", "2", ")", "does", "our", "approach", "improve", "upon", "the", "MAML", "algorithm", "when", "there", "is", "ambiguity", "over", "the", "class", "of", "functions", "?", ",", "and", "(", "3", ")", "can", "our", "approach", "scale", "to", "deep", "convolutional", "networks", "?", "We", "study", "two", "illustrative", "toy", "examples", "and", "a", "realistic", "ambiguous", "few", "-", "shot", "image", "classification", "problem", ".", "For", "the", "both", "experimental", "domains", ",", "we", "compare", "MAML", "to", "our", "probabilistic", "approach", ".", "We", "will", "refer", "to", "our", "version", "of", "MAML", "as", "a", "PLATIPUS", "(", "Probabilistic", "LATent", "model", "for", "Incorporating", "Priors", "and", "Uncertainty", "in", "few", "-", "Shot", "learning", ")", ",", "due", "to", "its", "unusual", "combination", "of", "two", "approximate", "inference", "methods", ":", "amortized", "inference", "and", "MAP", ".", "Both", "PLATIPUS", "and", "MAML", "use", "the", "same", "neural", "network", "architecture", "and", "the", "same", "number", "of", "inner", "gradient", "steps", ".", "We", "additionally", "provide", "a", "comparison", "on", "the", "MiniImagenet", "benchmark", "and", "specify", "the", "hyperparameters", "in", "the", "supplementary", "appendix", ".", ",", "and", "Gaussian", "noise", "with", "a", "standard", "deviation", "of", "0.3", "is", "added", "to", "the", "labels", ".", "We", "trained", "both", "MAML", "and", "PLATIPUS", "for", "5", "-", "shot", "regression", ".", "In", "Figure", "2", ",", "we", "show", "the", "qualitative", "performance", "of", "both", "methods", ",", "where", "the", "ground", "truth", "underlying", "function", "is", "shown", "in", "gray", "and", "the", "datapoints", "in", "D", "tr", "are", "shown", "as", "purple", "triangles", ".", "We", "show", "the", "function", "f", "φi", "learned", "by", "MAML", "in", "black", ".", "For", "PLATIPUS", ",", "we", "sample", "10", "sets", "of", "parameters", "from", "p", "(", "φ", "i", "|θ", ")", "and", "plot", "the", "resulting", "functions", "in", "different", "colors", ".", "In", "the", "top", "row", ",", "we", "can", "see", "that", "PLATIPUS", "allows", "the", "model", "to", "effectively", "reason", "over", "the", "set", "of", "functions", "underlying", "the", "provided", "datapoints", ",", "with", "increased", "variance", "in", "parts", "of", "the", "function", "where", "there", "is", "more", "uncertainty", ".", "Further", ",", "we", "see", "that", "PLATIPUS", "is", "able", "to", "capture", "the", "multimodal", "structure", ",", "as", "the", "curves", "are", "all", "linear", "or", "sinusoidal", ".", "A", "particularly", "useful", "application", "of", "uncertainty", "estimates", "in", "few", "-", "shot", "learning", "is", "estimating", "when", "more", "data", "would", "be", "helpful", ".", "In", "particular", ",", "seeing", "a", "large", "variance", "in", "a", "particular", "part", "of", "the", "input", "space", "suggests", "that", "more", "data", "would", "be", "helpful", "for", "learning", "the", "function", "in", "that", "part", "of", "the", "input", "space", ".", "On", "the", "bottom", "of", "Figure", "2", ",", "we", "show", "the", "results", "for", "a", "single", "task", "at", "meta", "-", "test", "time", "with", "increasing", "numbers", "of", "training", "datapoints", ".", "Even", "though", "the", "model", "was", "only", "trained", "on", "training", "set", "sizes", "of", "5", "datapoints", ",", "we", "observe", "that", "PLATIPUS", "is", "able", "to", "effectively", "reduce", "its", "uncertainty", "as", "more", "and", "more", "datapoints", "are", "available", ".", "This", "suggests", "that", "the", "uncertainty", "provided", "by", "PLATIPUS", "can", "be", "used", "for", "approximately", "gauging", "when", "more", "data", "would", "be", "helpful", "for", "learning", "a", "new", "task", ".", "consisting", "of", "both", "positive", "and", "negative", "examples", ".", "We", "plot", "the", "results", "using", "the", "same", "scheme", "as", "before", ",", "except", "that", "we", "plot", "the", "decision", "boundary", "(", "rather", "than", "the", "regression", "function", ")", "and", "visualize", "the", "single", "positive", "datapoint", "with", "a", "green", "plus", ".", "As", "seen", "in", "Figure", "3", ",", "we", "see", "that", "PLATIPUS", "captures", "a", "broad", "distribution", "over", "possible", "decision", "boundaries", ",", "all", "of", "which", "are", "roughly", "circular", ".", "MAML", "provides", "a", "single", "decision", "boundary", "of", "average", "size", ".", "Ambiguous", "image", "classification", ".", "The", "ambiguity", "illustrated", "in", "the", "previous", "settings", "is", "common", "in", "real", "world", "tasks", "where", "images", "can", "share", "multiple", "attributes", ".", "We", "study", "an", "ambiguous", "extension", "to", "the", "celebA", "attribute", "classification", "task", ".", "Our", "meta", "-", "training", "dataset", "is", "formed", "by", "sampling", "two", "attributes", "at", "random", "to", "form", "a", "positive", "class", "and", "taking", "the", "same", "number", "of", "random", "examples", "without", "either", "attribute", "to", "from", "the", "negative", "classes", ".", "To", "evaluate", "the", "ability", "to", "capture", "multiple", "decision", "boundaries", "while", "simultaneously", "obtaining", "good", "performance", ",", "we", "evaluate", "our", "method", "as", "follows", ":", "We", "sample", "from", "a", "test", "set", "of", "three", "attributes", "and", "a", "corresponding", "set", "of", "images", "with", "those", "attributes", ".", "Since", "the", "tasks", "involve", "classifying", "images", "that", "have", "two", "attributes", ",", "this", "task", "is", "ambiguous", ",", "and", "there", "are", "three", "possible", "combinations", "of", "two", "attributes", "that", "explain", "the", "training", "set", ".", "We", "sample", "models", "from", "our", "prior", "as", "described", "in", "Section", "4", "and", "assign", "each", "of", "the", "sampled", "models", "to", "one", "of", "the", "three", "possible", "tasks", "based", "on", "its", "log", "-", "likelihood", ".", "If", "each", "of", "the", "three", "possible", "tasks", "is", "assigned", "a", "nonzero", "number", "of", "samples", ",", "this", "means", "that", "the", "model", "effectively", "covers", "all", "three", "possible", "modes", "that", "explain", "the", "ambiguous", "training", "set", ".", "We", "can", "measure", "coverage", "and", "accuracy", "from", "this", "protocol", ".", "The", "coverage", "score", "indicates", "the", "average", "number", "of", "tasks", "(", "between", "1", "and", "3", ")", "that", "receive", "at", "least", "one", "sample", "for", "each", "ambiguous", "training", "set", ",", "and", "the", "accuracy", "score", "is", "the", "average", "number", "of", "correct", "classifications", "on", "these", "tasks", "(", "according", "to", "the", "sampled", "models", "assigned", "to", "them", ")", ".", "A", "highly", "random", "method", "will", "achieve", "good", "coverage", "but", "poor", "accuracy", ",", "while", "a", "deterministic", "method", "will", "have", "a", "coverage", "of", "1", ".", "Our", "results", "are", "summarized", "in", "Table", "5", "and", "Fig", ".", "4", ".", "The", "accuracy", "of", "our", "method", "is", "comparable", "to", "standard", ",", "deterministic", "MAML", ".", "However", ",", "the", "deterministic", "algorithm", "only", "ever", "captures", "one", "mode", "observes", "five", "positives", "that", "share", "three", "attributes", ",", "and", "five", "negatives", ".", "A", "classifier", "that", "uses", "any", "two", "attributes", "can", "correctly", "classify", "the", "training", "set", ".", "On", "the", "right", ",", "we", "show", "each", "of", "the", "possible", "two", "-", "attribute", "tasks", "that", "this", "training", "set", "can", "correspond", "to", ",", "and", "illustrate", "the", "labels", "(", "positive", "indicated", "by", "red", "border", ")", "assigned", "by", "the", "best", "sample", "for", "that", "task", ".", "We", "see", "that", "the", "different", "samples", "are", "able", "to", "make", "reasonable", "predictions", "with", "no", "hats", "(", "2nd", "column", ")", "or", "pay", "attention", "to", "them", "(", "1st", "and", "3rd", "column", ")", ",", "and", "can", "effectively", "capture", "the", "three", "possible", "explanations", ".", "for", "each", "ambiguous", "task", ",", "where", "the", "maximum", "is", "three", ".", "Our", "method", "on", "average", "captures", "between", "two", "and", "three", "modes", ".", "The", "qualitative", "analysis", "in", "Figure", "4", "illustrates", "3", "an", "example", "ambiguous", "training", "set", ",", "example", "images", "for", "the", "three", "possible", "two", "-", "attribute", "pairs", "that", "can", "correspond", "to", "this", "training", "set", ",", "and", "the", "classifications", "made", "by", "different", "sampled", "classifiers", "trained", "on", "the", "ambiguous", "training", "set", ".", "Note", "that", "the", "different", "samples", "each", "pay", "attention", "to", "different", "attributes", ",", "indicating", "that", "PLATIPUS", "is", "effective", "at", "capturing", "the", "different", "modes", "of", "the", "task", ".", "section", ":", "Discussion", "and", "Future", "Work", "We", "introduced", "an", "algorithm", "for", "few", "-", "shot", "meta", "-", "learning", "that", "enables", "simple", "and", "effective", "sampling", "of", "models", "for", "new", "tasks", "at", "meta", "-", "test", "time", ".", "Our", "algorithm", ",", "PLATIPUS", ",", "adapts", "to", "new", "tasks", "by", "running", "gradient", "descent", "with", "injected", "noise", ".", "During", "meta", "-", "training", ",", "the", "model", "parameters", "are", "optimized", "with", "respect", "to", "a", "variational", "lower", "bound", "on", "the", "likelihood", "for", "the", "meta", "-", "training", "tasks", ",", "so", "as", "to", "enable", "this", "simple", "adaptation", "procedure", "to", "produce", "approximate", "samples", "from", "the", "model", "posterior", "when", "conditioned", "on", "a", "few", "-", "shot", "training", "set", ".", "This", "approach", "has", "a", "number", "of", "benefits", ".", "The", "adaptation", "procedure", "is", "exceedingly", "simple", ",", "and", "the", "method", "can", "be", "applied", "to", "any", "standard", "model", "architecture", ".", "The", "algorithm", "introduces", "a", "modest", "number", "of", "additional", "parameters", ":", "besides", "the", "initial", "model", "weights", ",", "we", "must", "learn", "a", "variance", "on", "each", "parameter", "for", "the", "inference", "network", "and", "prior", ",", "and", "the", "number", "of", "parameters", "scales", "only", "linearly", "with", "the", "number", "of", "model", "weights", ".", "Our", "experimental", "results", "show", "that", "our", "method", "can", "be", "used", "to", "effectively", "sample", "diverse", "solutions", "to", "both", "regression", "and", "classification", "tasks", "at", "meta", "-", "test", "time", ",", "including", "for", "task", "families", "that", "have", "multi", "-", "modal", "task", "distributions", ".", "Although", "our", "approach", "is", "simple", "and", "broadly", "applicable", ",", "it", "has", "a", "number", "of", "potential", "limitations", "that", "could", "be", "addressed", "in", "future", "work", ".", "First", ",", "the", "current", "form", "of", "the", "method", "provides", "a", "relatively", "impoverished", "estimator", "of", "posterior", "variance", ",", "which", "might", "be", "less", "effective", "at", "gauging", "uncertainty", "in", "settings", "where", "different", "tasks", "have", "very", "different", "degrees", "of", "ambiguity", ".", "In", "these", "cases", ",", "finding", "a", "way", "to", "make", "the", "variance", "dependent", "on", "the", "few", "-", "shot", "training", "set", "might", "produce", "better", "results", ",", "and", "investigating", "how", "to", "do", "this", "without", "adding", "a", "large", "number", "of", "additional", "parameters", "would", "be", "an", "interesting", "direction", "for", "future", "work", ".", "Another", "exciting", "direction", "for", "future", "research", "would", "be", "to", "study", "how", "our", "approach", "could", "be", "applied", "in", "settings", "where", "ambiguity", "and", "uncertainty", "can", "directly", "guide", "data", "acquisition", ",", "so", "as", "to", "devise", "better", "few", "-", "shot", "active", "learning", "and", "reinforcement", "learning", "algorithms", ".", "section", ":", "Appendix", "A", "Ambiguous", "CelebA", "Details", "To", "construct", "our", "ambiguous", "few", "-", "shot", "variant", "of", "CelebA", ",", "we", "take", "the", "entire", "base", "set", "of", "attributes", "holding", "out", "10", "attributes", "for", "testing", ".", "We", "consider", "every", "combination", "of", "2", "attributes", ",", "discarding", "those", "with", "insufficient", "numbers", "of", "examples", ".", "This", "leave", "us", "with", "a", "total", "of", "387", "training", "tasks", "and", "43", "testing", "attributes", ".", "We", "partition", "our", "meta", "-", "training", "set", "and", "meta", "-", "validation", "set", "to", "337", "/", "50", "respectively", ".", "During", "meta", "-", "training", ",", "we", "sample", "2", "random", "attributes", "to", "construct", "a", "positive", "class", "and", "randomly", "sample", "examples", "with", "neither", "attribute", "as", "negative", "examples", ".", "During", "testing", "of", "our", "approach", ",", "we", "sample", "3", "attributes", "from", "the", "test", "set", ",", "and", "sample", "the", "3", "corresponding", "2", "-", "uples", "to", "form", "the", "test", "task", ".", "The", "training", "attributes", "are", ":", "section", ":", "B", "Experimental", "Details", "In", "the", "illustrative", "experiments", ",", "we", "use", "a", "fully", "connected", "network", "with", "3", "ReLU", "layers", "of", "size", "100", ".", "For", "CelebA", ",", "we", "adapt", "the", "base", "convolutional", "architecture", "described", "in", "Finn", "et", "al", ".", "[", "reference", "]", "which", "we", "refer", "the", "readers", "to", "for", "more", "detail", ".", "Our", "approximate", "posterior", "and", "prior", "have", "dimensionality", "matching", "the", "underlying", "model", ".", "We", "tune", "our", "approach", "over", "the", "inner", "learning", "rate", "α", ",", "a", "weight", "on", "the", "D", "KL", ",", "the", "scale", "of", "the", "initialization", "of", "µ", "θ", ",", "σ", "2", "θ", ",", "v", "q", ",", "γ", "p", ",", "γ", "q", ",", "with", "early", "stopping", "on", "the", "validation", "set", ".", "At", "meta", "-", "test", "time", ",", "we", "evaluate", "our", "approach", "by", "taking", "10", "samples", "from", "the", "prior", "before", "determining", "the", "assignments", ".", "The", "assignments", "are", "made", "based", "on", "the", "complete", "likelihood", "of", "the", "testing", "examples", "(", "including", "the", "negatives", ")", ".", "section", ":", "C", "MiniImagenet", "Comparison", "We", "provide", "an", "additional", "comparison", "on", "the", "MiniImagenet", "dataset", ".", "Since", "this", "benchmark", "does", "not", "contain", "a", "large", "amount", "of", "ambiguity", ",", "we", "do", "not", "aim", "to", "show", "state", "-", "of", "-", "the", "-", "art", "performance", ".", "Instead", ",", "our", "goal", "with", "this", "experiment", "is", "to", "compare", "our", "approach", "on", "to", "MAML", "and", "prior", "methods", "that", "build", "upon", "MAML", "on", "this", "standard", "benchmark", ".", "Since", "our", "goal", "is", "to", "compare", "algorithms", ",", "rather", "than", "achieving", "maximal", "performance", ",", "we", "decouple", "the", "effect", "of", "the", "meta", "-", "learning", "algorithm", "and", "the", "architecture", "used", "by", "using", "the", "standard", "4", "-", "block", "convolutional", "architecture", "used", "by", "Vinyals", "et", "al", ".", "[", "reference", "]", ",", "Ravi", "and", "Larochelle", "[", "reference", "]", ",", "Finn", "et", "al", ".", "[", "reference", "]", "and", "others", ".", "We", "note", "that", "better", "performance", "can", "likely", "be", "achieved", "by", "tuning", "the", "architecture", ".", "The", "results", ",", "in", "Table", "2", "indicate", "that", "our", "method", "slightly", "outperforms", "MAML", "and", "achieves", "comparable", "performance", "to", "a", "number", "of", "other", "prior", "methods", ".", "MiniImagenet", "5", "-", "way", ",", "1", "-", "shot", "Accuracy", "MAML", "[", "reference", "]", "48.70", "±", "1.84", "%", "LLAMA", "[", "reference", "]", "49.40", "±", "1.83", "%", "Reptile", "[", "reference", "]", "49.97", "±", "0.32", "%", "PLATIPUS", "(", "ours", ")", "50.13", "±", "1.86", "%", "Meta", "-", "SGD", "[", "reference", "]", "50.71", "±", "1.87", "%", "matching", "nets", "[", "reference", "]", "43.56", "±", "0.84", "%", "meta", "-", "learner", "LSTM", "[", "reference", "]", "43.44", "±", "0.77", "%", "SNAIL", "[", "reference", "]", "*", "45.10", "±", "0.00", "%", "prototypical", "networks", "[", "reference", "]", "46.61", "±", "0.78", "%", "mAP", "-", "DLM", "[", "reference", "]", "49.82", "±", "0.78", "%", "GNN", "[", "reference", "]", "50.33", "±", "0.36", "%", "Relation", "Net", "[", "reference", "]", "50.44", "±", "0.82", "%", "Table", "2", ":", "Comparison", "between", "our", "approach", "and", "prior", "MAML", "-", "based", "methods", "(", "top", ")", ",", "and", "other", "prior", "few", "-", "shot", "learning", "techniques", "on", "the", "5", "-", "way", ",", "1", "-", "shot", "MiniImagenet", "benchmark", ".", "Our", "approach", "gives", "a", "small", "boost", "over", "MAML", ",", "and", "is", "comparable", "to", "other", "approaches", ".", "We", "bold", "the", "approaches", "that", "are", "above", "the", "highest", "confidence", "interval", "lower", "-", "bound", ".", "*", "Accuracy", "using", "comparable", "network", "architecture", ".", "section", ":", "section", ":", "Acknowledgments", "We", "thank", "Marvin", "Zhang", "and", "Dibya", "Ghosh", "for", "comments", "on", "an", "earlier", "draft", "of", "this", "paper", ".", "This", "research", "was", "supported", "by", "an", "NSF", "Graduate", "Research", "Fellowship", ",", "NSF", "IIS", "-", "1651843", ",", "the", "Office", "of", "Naval", "Research", ",", "and", "NVIDIA", ".", "section", ":"], "entity": [[0, 7, ["Probabilistic", "Model", "-", "Agnostic", "Meta", "-", "Learning"], "Method"], [10, 13, ["Meta", "-", "learning"], "Method"], [14, 18, ["few", "-", "shot", "learning"], "Task"], [46, 50, ["few", "-", "shot", "learning"], "Task"], [51, 53, ["task", "ambiguity"], "Task"], [93, 94, ["classifier"], "Method"], [109, 114, ["probabilistic", "meta", "-", "learning", "algorithm"], "Method"], [130, 136, ["model", "-", "agnostic", "meta", "-", "learning"], "Method"], [143, 145, ["gradient", "descent"], "Method"], [177, 179, ["gradient", "descent"], "Method"], [182, 186, ["meta", "-", "training", "time"], "Task"], [194, 197, ["stochastic", "adaptation", "procedure"], "Method"], [201, 204, ["approximate", "model", "posterior"], "Method"], [214, 216, ["plausible", "classifiers"], "Method"], [217, 218, ["regressors"], "Method"], [219, 225, ["ambiguous", "few", "-", "shot", "learning", "problems"], "Task"], [239, 241, ["human", "intelligence"], "Task"], [300, 301, ["Structure"], "Method"], [308, 312, ["learning", "of", "new", "tasks"], "Task"], [317, 322, ["learning", "-", "to", "-", "learn"], "Task"], [324, 327, ["meta", "-", "learning"], "Method"], [338, 344, ["few", "-", "shot", "meta", "-", "learning"], "Method"], [366, 368, ["task", "ambiguity"], "Task"], [405, 412, ["few", "-", "shot", "meta", "-", "learning", "methods"], "Method"], [420, 426, ["ambiguous", "few", "-", "shot", "learning", "problem"], "Task"], [446, 448, ["active", "learning"], "Method"], [464, 468, ["safety", "-", "critical", "applications"], "Task"], [471, 477, ["few", "-", "shot", "medical", "image", "classification"], "Task"], [478, 479, ["uncertainty"], "Task"], [486, 487, ["classifier"], "Method"], [492, 493, ["learning"], "Task"], [500, 502, ["uncertainty", "estimation"], "Task"], [513, 514, ["learning"], "Task"], [545, 548, ["reinforcement", "learning", "problems"], "Task"], [557, 558, ["exploration"], "Task"], [564, 565, ["ambiguity"], "Task"], [571, 576, ["few", "-", "shot", "learning", "problem"], "Task"], [592, 594, ["function", "approximators"], "Method"], [610, 612, ["function", "approximators"], "Method"], [615, 617, ["linear", "functions"], "Method"], [625, 630, ["few", "-", "shot", "learning", "approaches"], "Method"], [631, 633, ["Bayesian", "models"], "Method"], [649, 654, ["high", "-", "dimensional", "function", "approximators"], "Method"], [656, 659, ["deep", "neural", "networks"], "Method"], [684, 688, ["few", "-", "shot", "learning"], "Task"], [692, 695, ["deterministic", "learning", "algorithms"], "Method"], [716, 721, ["few", "-", "shot", "learning", "methods"], "Method"], [724, 726, ["uncertainty", "awareness"], "Task"], [736, 739, ["amortized", "variational", "inference"], "Method"], [742, 747, ["probabilistic", "meta", "-", "learning", "approach"], "Method"], [755, 761, ["model", "-", "agnostic", "meta", "-", "learning"], "Method"], [762, 763, ["MAML"], "Method"], [769, 773, ["few", "shot", "metalearning", "algorithm"], "Method"], [776, 778, ["gradient", "descent"], "Method"], [790, 794, ["few", "-", "shot", "task"], "Task"], [798, 800, ["model", "parameters"], "Method"], [807, 809, ["rapid", "adaptation"], "Task"], [814, 817, ["neural", "network", "initialization"], "Method"], [823, 826, ["few", "shot", "learning"], "Task"], [827, 828, ["MAML"], "Method"], [836, 842, ["black", "-", "box", "meta", "-", "learners"], "Method"], [844, 845, ["RNNs"], "Method"], [854, 857, ["neural", "network", "architectures"], "Method"], [861, 862, ["MAML"], "Method"], [877, 880, ["stochastic", "adaptation", "procedure"], "Method"], [885, 887, ["gradient", "descent"], "Method"], [894, 898, ["meta", "-", "training", "procedure"], "Method"], [903, 905, ["inference", "process"], "Method"], [924, 925, ["MAML"], "Method"], [927, 931, ["graphical", "model", "inference", "problem"], "Task"], [933, 935, ["variational", "inference"], "Method"], [945, 949, ["modeling", "uncertainty", "and", "ambiguity"], "Task"], [959, 964, ["few", "-", "shot", "learning", "problem"], "Task"], [987, 990, ["ambiguous", "regression", "problem"], "Task"], [996, 997, ["classifiers"], "Method"], [998, 1005, ["ambiguous", "few", "-", "shot", "attribute", "classification", "tasks"], "Task"], [1010, 1013, ["Hierarchical", "Bayesian", "models"], "Method"], [1020, 1024, ["few", "-", "shot", "learning"], "Task"], [1060, 1066, ["few", "-", "shot", "image", "classification", "datasets"], "Task"], [1098, 1102, ["few", "-", "shot", "learning"], "Task"], [1104, 1108, ["meta", "-", "learning", "perspective"], "Method"], [1115, 1118, ["black", "-", "box"], "Method"], [1126, 1130, ["optimization", "-", "based", "approaches"], "Method"], [1150, 1154, ["visual", "reinforcement", "learning", "problems"], "Task"], [1177, 1179, ["deep", "networks"], "Method"], [1180, 1182, ["probabilistic", "methods"], "Method"], [1183, 1187, ["few", "-", "shot", "learning"], "Task"], [1199, 1202, ["hierarchical", "Bayesian", "models"], "Method"], [1203, 1207, ["few", "-", "shot", "learning"], "Task"], [1209, 1211, ["neural", "statistician"], "Method"], [1230, 1232, ["model", "agnostic"], "Method"], [1243, 1245, ["network", "architecture"], "Method"], [1254, 1255, ["MAML"], "Method"], [1275, 1277, ["model", "uncertainty"], "Task"], [1279, 1284, ["few", "-", "shot", "learning", "setting"], "Task"], [1286, 1288, ["LLAMA", "method"], "Method"], [1300, 1301, ["LLAMA"], "Method"], [1305, 1308, ["local", "Laplace", "approximation"], "Method"], [1358, 1360, ["gradient", "operations"], "Method"], [1361, 1364, ["Bayesian", "neural", "networks"], "Method"], [1384, 1386, ["deep", "networks"], "Method"], [1388, 1390, ["exact", "inference"], "Task"], [1391, 1394, ["Bayesian", "neural", "networks"], "Method"], [1400, 1403, ["backpropagation", "and", "sampling"], "Method"], [1422, 1424, ["generic", "networks"], "Method"], [1437, 1440, ["hierarchical", "Bayesian", "model"], "Method"], [1476, 1479, ["Bayesian", "neural", "networks"], "Method"], [1495, 1499, ["amortized", "variational", "inference", "methods"], "Method"], [1520, 1522, ["inference", "networks"], "Method"], [1549, 1551, ["center", "model"], "Method"], [1597, 1602, ["meta", "-", "learning", "problem", "setting"], "Task"], [1627, 1631, ["meta", "-", "learning", "algorithms"], "Method"], [1639, 1643, ["meta", "-", "testing", "tasks"], "Task"], [1646, 1650, ["distribution", "p", "(", "T"], "Method"], [1656, 1661, ["learning", "-", "to", "-", "learn"], "Task"], [1676, 1681, ["faster", "learning", "of", "new", "tasks"], "Task"], [1688, 1692, ["meta", "-", "learning", "process"], "Method"], [1709, 1715, ["model", "-", "agnostic", "meta", "-", "learning"], "Method"], [1716, 1717, ["MAML"], "Method"], [1723, 1727, ["Meta", "-", "learning", "algorithms"], "Method"], [1761, 1763, ["gradient", "descent"], "Method"], [1783, 1785, ["supervised", "classification"], "Task"], [1802, 1805, ["classifier", "f", "θ"], "Method"], [1818, 1819, ["classifier"], "Method"], [1863, 1867, ["meta", "-", "learning", "method"], "Method"], [1890, 1892, ["function", "approximators"], "Method"], [1894, 1896, ["neural", "networks"], "Method"], [1905, 1907, ["graphical", "model"], "Method"], [1909, 1914, ["few", "-", "shot", "learning", "problem"], "Task"], [1921, 1924, ["hierarchical", "Bayesian", "model"], "Method"], [1959, 1961, ["graphical", "model"], "Method"], [1991, 1993, ["posterior", "inference"], "Task"], [2025, 2026, ["MAML"], "Method"], [2031, 2035, ["maximum", "a", "posteriori", "inference"], "Method"], [2050, 2051, ["inference"], "Task"], [2055, 2057, ["gradient", "descent"], "Method"], [2142, 2145, ["structured", "variational", "inference"], "Method"], [2147, 2150, ["structured", "variational", "inference"], "Task"], [2235, 2238, ["structured", "variational", "approximation"], "Method"], [2245, 2247, ["variational", "distribution"], "Method"], [2268, 2272, ["amortized", "variational", "inference", "technique"], "Method"], [2292, 2294, ["function", "approximator"], "Method"], [2331, 2334, ["variational", "lower", "bound"], "Metric"], [2422, 2424, ["Bayesian", "version"], "Method"], [2471, 2475, ["Gaussian", "with", "mean", "θ"], "Method"], [2514, 2516, ["gradient", "descent"], "Method"], [2544, 2548, ["inference", "networks", "q", "ψ"], "Method"], [2589, 2592, ["deep", "neural", "network"], "Method"], [2609, 2611, ["function", "approximator"], "Method"], [2647, 2649, ["gradient", "descent"], "Method"], [2718, 2720, ["inference", "network"], "Method"], [2772, 2775, ["meta", "-", "training"], "Task"], [2802, 2804, ["gradient", "descent"], "Method"], [2807, 2809, ["variational", "distribution"], "Method"], [2820, 2822, ["neural", "network"], "Method"], [2831, 2834, ["meta", "-", "training"], "Task"], [2859, 2861, ["inference", "networks"], "Method"], [2869, 2871, ["gradient", "descent"], "Method"], [2873, 2875, ["inference", "network"], "Method"], [2887, 2890, ["meta", "-", "training"], "Method"], [2917, 2919, ["inference", "network"], "Method"], [2920, 2923, ["meta", "-", "training"], "Task"], [2927, 2930, ["meta", "-", "testing"], "Task"], [2935, 2938, ["Meta", "-", "training"], "Method"], [2941, 2942, ["MAML"], "Method"], [3005, 3007, ["gradient", "descent"], "Method"], [3009, 3010, ["∇Θ"], "Method"], [3011, 3012, ["Θ"], "Method"], [3013, 3014, ["Adam"], "Method"], [3018, 3021, ["Meta", "-", "testing"], "Task"], [3023, 3031, ["Probabilistic", "Model", "-", "Agnostic", "Meta", "-", "Learning", "Approach"], "Method"], [3032, 3034, ["Hybrid", "Inference"], "Method"], [3038, 3043, ["variational", "meta", "-", "learning", "procedure"], "Method"], [3047, 3049, ["probabilistic", "interpretation"], "Method"], [3050, 3051, ["MAML"], "Method"], [3063, 3064, ["MAML"], "Method"], [3068, 3070, ["approximate", "inference"], "Method"], [3072, 3076, ["posterior", "p", "(", "y"], "Task"], [3105, 3107, ["gradient", "descent"], "Method"], [3124, 3128, ["maximum", "a", "posteriori", "inference"], "Task"], [3192, 3194, ["Gaussian", "prior"], "Method"], [3204, 3208, ["neural", "network", "parameter", "manifold"], "Method"], [3249, 3251, ["gradient", "descent"], "Method"], [3268, 3271, ["approximate", "inference", "procedure"], "Method"], [3273, 3275, ["graphical", "model"], "Method"], [3337, 3340, ["variational", "inference", "procedure"], "Method"], [3367, 3369, ["gradient", "descent"], "Method"], [3371, 3375, ["p", "(", "φ", "i"], "Task"], [3389, 3392, ["variational", "lower", "bound"], "Metric"], [3415, 3417, ["approximate", "inference"], "Method"], [3418, 3419, ["MAP"], "Metric"], [3436, 3437, ["training"], "Task"], [3503, 3505, ["reparameterization", "trick"], "Method"], [3513, 3515, ["gradient", "steps"], "Method"], [3544, 3545, ["Training"], "Task"], [3554, 3556, ["v", "q"], "Method"], [3602, 3606, ["prior", "p", "(", "θ"], "Method"], [3609, 3613, ["meta", "-", "training", "procedure"], "Method"], [3620, 3624, ["meta", "-", "test", "time"], "Task"], [3626, 3628, ["inference", "procedure"], "Method"], [3691, 3695, ["meta", "-", "testing", "procedure"], "Method"], [3708, 3711, ["transformed", "graphical", "model"], "Method"], [3839, 3841, ["graphical", "model"], "Method"], [3851, 3855, ["inference", "network", "q", "ψ"], "Method"], [3869, 3872, ["distribution", "over", "θ"], "Task"], [3874, 3877, ["variational", "training", "objective"], "Task"], [3879, 3885, ["likelihood", "term", "log", "p", "(", "θ"], "Method"], [3971, 3974, ["deep", "convolutional", "networks"], "Method"], [3984, 3991, ["ambiguous", "few", "-", "shot", "image", "classification", "problem"], "Task"], [4000, 4001, ["MAML"], "Method"], [4003, 4005, ["probabilistic", "approach"], "Method"], [4013, 4014, ["MAML"], "Method"], [4016, 4017, ["PLATIPUS"], "Method"], [4018, 4031, ["Probabilistic", "LATent", "model", "for", "Incorporating", "Priors", "and", "Uncertainty", "in", "few", "-", "Shot", "learning"], "Method"], [4040, 4043, ["approximate", "inference", "methods"], "Method"], [4044, 4046, ["amortized", "inference"], "Method"], [4047, 4048, ["MAP"], "Metric"], [4050, 4051, ["PLATIPUS"], "Method"], [4052, 4053, ["MAML"], "Method"], [4056, 4059, ["neural", "network", "architecture"], "Method"], [4075, 4077, ["MiniImagenet", "benchmark"], "Material"], [4105, 4106, ["MAML"], "Method"], [4107, 4108, ["PLATIPUS"], "Method"], [4109, 4113, ["5", "-", "shot", "regression"], "Task"], [4157, 4158, ["MAML"], "Method"], [4162, 4163, ["PLATIPUS"], "Method"], [4195, 4196, ["PLATIPUS"], "Method"], [4231, 4232, ["PLATIPUS"], "Method"], [4254, 4256, ["uncertainty", "estimates"], "Task"], [4257, 4261, ["few", "-", "shot", "learning"], "Task"], [4349, 4350, ["PLATIPUS"], "Method"], [4372, 4373, ["PLATIPUS"], "Method"], [4377, 4379, ["approximately", "gauging"], "Task"], [4444, 4445, ["PLATIPUS"], "Method"], [4461, 4462, ["MAML"], "Method"], [4471, 4474, ["Ambiguous", "image", "classification"], "Task"], [4485, 4488, ["real", "world", "tasks"], "Task"], [4498, 4500, ["ambiguous", "extension"], "Task"], [4502, 4506, ["celebA", "attribute", "classification", "task"], "Task"], [4682, 4683, ["coverage"], "Metric"], [4684, 4685, ["accuracy"], "Metric"], [4690, 4692, ["coverage", "score"], "Metric"], [4743, 4745, ["random", "method"], "Method"], [4748, 4749, ["coverage"], "Metric"], [4751, 4752, ["accuracy"], "Metric"], [4755, 4757, ["deterministic", "method"], "Method"], [4777, 4778, ["accuracy"], "Metric"], [4792, 4794, ["deterministic", "algorithm"], "Method"], [4812, 4813, ["classifier"], "Method"], [4909, 4911, ["ambiguous", "task"], "Task"], [4967, 4969, ["sampled", "classifiers"], "Method"], [4990, 4991, ["PLATIPUS"], "Method"], [5013, 5019, ["few", "-", "shot", "meta", "-", "learning"], "Method"], [5024, 5027, ["sampling", "of", "models"], "Method"], [5031, 5035, ["meta", "-", "test", "time"], "Task"], [5039, 5040, ["PLATIPUS"], "Method"], [5047, 5049, ["gradient", "descent"], "Method"], [5054, 5057, ["meta", "-", "training"], "Task"], [5075, 5079, ["meta", "-", "training", "tasks"], "Task"], [5086, 5088, ["adaptation", "procedure"], "Method"], [5115, 5117, ["adaptation", "procedure"], "Method"], [5130, 5132, ["model", "architecture"], "Method"], [5159, 5161, ["inference", "network"], "Method"], [5196, 5200, ["regression", "and", "classification", "tasks"], "Task"], [5201, 5205, ["meta", "-", "test", "time"], "Task"], [5253, 5258, ["impoverished", "estimator", "of", "posterior", "variance"], "Method"], [5351, 5353, ["data", "acquisition"], "Task"], [5359, 5368, ["few", "-", "shot", "active", "learning", "and", "reinforcement", "learning", "algorithms"], "Method"], [5379, 5384, ["ambiguous", "few", "-", "shot", "variant"], "Method"], [5385, 5386, ["CelebA"], "Material"], [5452, 5455, ["meta", "-", "training"], "Task"], [5524, 5527, ["fully", "connected", "network"], "Method"], [5536, 5537, ["CelebA"], "Material"], [5542, 5544, ["convolutional", "architecture"], "Method"], [5581, 5585, ["inner", "learning", "rate", "α"], "Metric"], [5616, 5618, ["early", "stopping"], "Method"], [5667, 5669, ["MiniImagenet", "Comparison"], "Metric"], [5676, 5678, ["MiniImagenet", "dataset"], "Material"], [5720, 5721, ["MAML"], "Method"], [5727, 5728, ["MAML"], "Method"], [5753, 5757, ["meta", "-", "learning", "algorithm"], "Method"], [5765, 5770, ["4", "-", "block", "convolutional", "architecture"], "Method"], [5823, 5824, ["MAML"], "Method"], [5833, 5835, ["prior", "methods"], "Method"], [5836, 5840, ["MiniImagenet", "5", "-", "way"], "Material"], [5869, 5870, ["PLATIPUS"], "Method"], [5877, 5880, ["Meta", "-", "SGD"], "Method"], [5887, 5889, ["matching", "nets"], "Method"], [5896, 5900, ["meta", "-", "learner", "LSTM"], "Method"], [5907, 5908, ["SNAIL"], "Method"], [5925, 5928, ["mAP", "-", "DLM"], "Method"], [5935, 5936, ["GNN"], "Method"], [5971, 5977, ["prior", "few", "-", "shot", "learning", "techniques"], "Method"], [5979, 5988, ["5", "-", "way", ",", "1", "-", "shot", "MiniImagenet", "benchmark"], "Material"], [5996, 5997, ["MAML"], "Method"], [6021, 6022, ["Accuracy"], "Metric"], [6023, 6026, ["comparable", "network", "architecture"], "Method"], [1297, 1298, ["MAML"], "Method"], [1537, 1543, ["model", "-", "agnostic", "meta", "-", "learning"], "Method"], [1577, 1578, ["MAP"], "Metric"], [2426, 2427, ["MAML"], "Method"], [2527, 2528, ["MAML"], "Method"], [3084, 3085, ["MAP"], "Metric"], [3225, 3226, ["MAP"], "Metric"], [3537, 3538, ["MAP"], "Metric"], [3651, 3652, ["MAP"], "Metric"], [3949, 3950, ["MAML"], "Method"], [4718, 4719, ["accuracy"], "Metric"], [4787, 4788, ["MAML"], "Method"], [5374, 5375, ["CelebA"], "Material"], [5845, 5846, ["MAML"], "Method"], [5961, 5962, ["MAML"], "Method"]]} {"doc_id": "3aa21de1a7c97e0458e10ed5730ce160bb436caa", "text": ["document", ":", "Pixel2Mesh", ":", "Generating", "3D", "Mesh", "Models", "from", "Single", "RGB", "Images", "We", "propose", "an", "end", "-", "to", "-", "end", "deep", "learning", "architecture", "that", "produces", "a", "3D", "shape", "in", "triangular", "mesh", "from", "a", "single", "color", "image", ".", "Limited", "by", "the", "nature", "of", "deep", "neural", "network", ",", "previous", "methods", "usually", "represent", "a", "3D", "shape", "in", "volume", "or", "point", "cloud", ",", "and", "it", "is", "non", "-", "trivial", "to", "convert", "them", "to", "the", "more", "ready", "-", "to", "-", "use", "mesh", "model", ".", "Unlike", "the", "existing", "methods", ",", "our", "network", "represents", "3D", "mesh", "in", "a", "graph", "-", "based", "convolutional", "neural", "network", "and", "produces", "correct", "geometry", "by", "progressively", "deforming", "an", "ellipsoid", ",", "leveraging", "perceptual", "features", "extracted", "from", "the", "input", "image", ".", "We", "adopt", "a", "coarse", "-", "to", "-", "fine", "strategy", "to", "make", "the", "whole", "deformation", "procedure", "stable", ",", "and", "define", "various", "of", "mesh", "related", "losses", "to", "capture", "properties", "of", "different", "levels", "to", "guarantee", "visually", "appealing", "and", "physically", "accurate", "3D", "geometry", ".", "Extensive", "experiments", "show", "that", "our", "method", "not", "only", "qualitatively", "produces", "mesh", "model", "with", "better", "details", ",", "but", "also", "achieves", "higher", "3D", "shape", "estimation", "accuracy", "compared", "to", "the", "state", "-", "of", "-", "the", "-", "art", ".", "indicates", "equal", "contributions", ".", "indicates", "corresponding", "author", ".", "section", ":", "Introduction", "Inferring", "3D", "shape", "from", "a", "single", "perspective", "is", "a", "fundamental", "human", "vision", "functionality", "but", "is", "extremely", "challenging", "for", "computer", "vision", ".", "Recently", ",", "great", "success", "has", "been", "achieved", "for", "3d", "shape", "generation", "from", "a", "single", "color", "image", "using", "deep", "learning", "techniques", ".", "Taking", "advantage", "of", "convolutional", "layers", "on", "regular", "grids", "or", "multi", "-", "layer", "perception", ",", "the", "estimated", "3D", "shape", ",", "as", "the", "output", "of", "the", "neural", "network", ",", "is", "represented", "as", "either", "a", "volume", "or", "point", "cloud", ".", "However", ",", "both", "representations", "lose", "important", "surface", "details", ",", "and", "is", "non", "-", "trivial", "to", "reconstruct", "a", "surface", "model", "(", "Fig", ".", "[", "reference", "]", ")", ",", "i.e.", "a", "mesh", ",", "which", "is", "more", "desirable", "for", "many", "real", "applications", "since", "it", "is", "lightweight", ",", "capable", "of", "modelling", "shape", "details", ",", "easy", "to", "deform", "for", "animation", ",", "to", "name", "a", "few", ".", "In", "this", "paper", ",", "we", "push", "along", "the", "direction", "of", "single", "image", "reconstruction", ",", "and", "propose", "an", "algorithm", "to", "extract", "a", "3D", "triangular", "mesh", "from", "a", "single", "color", "image", ".", "Rather", "than", "directly", "synthesizing", ",", "our", "model", "learns", "to", "deform", "a", "mesh", "from", "a", "mean", "shape", "to", "the", "target", "geometry", ".", "This", "benefits", "us", "from", "several", "aspects", ".", "First", ",", "deep", "network", "is", "better", "at", "predicting", "residual", ",", "e.g.", "a", "spatial", "deformation", ",", "rather", "than", "structured", "output", ",", "e.g.", "a", "graph", ".", "Second", ",", "a", "series", "of", "deformations", "can", "be", "added", "up", "together", ",", "which", "allows", "shape", "to", "be", "gradually", "refined", "in", "detail", ".", "It", "also", "enables", "the", "control", "of", "the", "trade", "-", "off", "between", "the", "complexity", "of", "the", "deep", "learning", "model", "and", "the", "quality", "of", "the", "result", ".", "Lastly", ",", "it", "provides", "the", "chance", "to", "encode", "any", "prior", "knowledge", "to", "the", "initial", "mesh", ",", "e.g.", "topology", ".", "As", "a", "pioneer", "study", ",", "in", "this", "work", ",", "we", "specifically", "work", "on", "objects", "that", "can", "be", "approximated", "using", "3D", "mesh", "with", "genus", "0", "by", "deforming", "an", "ellipsoid", "with", "a", "fixed", "size", ".", "In", "practice", ",", "we", "found", "most", "of", "the", "commonly", "seen", "categories", "can", "be", "handled", "well", "under", "this", "setting", ",", "e.g.", "car", ",", "plane", ",", "table", ",", "etc", ".", "To", "achieve", "this", "goal", ",", "there", "are", "several", "inherent", "challenges", ".", "The", "first", "challenge", "is", "how", "to", "represent", "a", "mesh", "model", ",", "which", "is", "essentially", "an", "irregular", "graph", ",", "in", "a", "neural", "network", "and", "still", "be", "capable", "of", "extracting", "shape", "details", "effectively", "from", "a", "given", "color", "image", "represented", "in", "a", "2D", "regular", "grid", ".", "It", "requires", "the", "integration", "of", "the", "knowledge", "learned", "from", "two", "data", "modalities", ".", "On", "the", "3D", "geometry", "side", ",", "we", "directly", "build", "a", "graph", "based", "fully", "convolutional", "network", "(", "GCN", ")", "on", "the", "mesh", "model", ",", "where", "the", "vertices", "and", "edges", "in", "the", "mesh", "are", "directly", "represented", "as", "nodes", "and", "connections", "in", "a", "graph", ".", "Network", "feature", "encoding", "information", "for", "3D", "shape", "is", "saved", "on", "each", "vertex", ".", "Through", "forward", "propagation", ",", "the", "convolutional", "layers", "enable", "feature", "exchanging", "across", "neighboring", "nodes", ",", "and", "eventually", "regress", "the", "3D", "location", "for", "each", "vertex", ".", "On", "the", "2D", "image", "side", ",", "we", "use", "a", "VGG", "-", "16", "like", "architecture", "to", "extract", "features", "as", "it", "has", "been", "demonstrated", "to", "be", "successful", "for", "many", "tasks", ".", "To", "bridge", "these", "two", ",", "we", "design", "a", "perceptual", "feature", "pooling", "layer", "which", "allows", "each", "node", "in", "the", "GCN", "to", "pool", "image", "features", "from", "its", "2D", "projection", "on", "the", "image", ",", "which", "can", "be", "readily", "obtained", "by", "assuming", "known", "camera", "intrinsic", "matrix", ".", "The", "perceptual", "feature", "pooling", "is", "enabled", "once", "after", "several", "convolutions", "(", "i.e.", "a", "deformation", "block", "described", "in", "Sec", ".", "[", "reference", "]", ")", "using", "updated", "3D", "locations", ",", "and", "hence", "the", "image", "features", "from", "correct", "locations", "can", "be", "effectively", "integrated", "with", "3D", "shapes", ".", "Given", "the", "graph", "representation", ",", "the", "next", "challenge", "is", "how", "to", "update", "the", "vertex", "location", "effectively", "towards", "ground", "truth", ".", "In", "practice", ",", "we", "observe", "that", "network", "trained", "to", "directly", "predict", "mesh", "with", "a", "large", "number", "of", "vertices", "is", "likely", "to", "make", "mistake", "in", "the", "beginning", "and", "hard", "to", "fix", "later", ".", "One", "reason", "is", "that", "a", "vertex", "can", "not", "effectively", "retrieve", "features", "from", "other", "vertices", "with", "a", "number", "of", "edges", "away", ",", "i.e.", "the", "limited", "receptive", "field", ".", "To", "solve", "this", "problem", ",", "we", "design", "a", "graph", "unpooling", "layer", ",", "which", "allows", "the", "network", "to", "initiate", "with", "a", "smaller", "number", "of", "vertices", "and", "increase", "during", "the", "forward", "propagation", ".", "With", "fewer", "vertices", "at", "the", "beginning", "stages", ",", "the", "network", "learns", "to", "distribute", "the", "vertices", "around", "to", "the", "most", "representative", "location", ",", "and", "then", "add", "local", "details", "as", "the", "number", "of", "vertices", "increases", "later", ".", "Besides", "the", "graph", "unpooling", "layer", ",", "we", "use", "a", "deep", "GCN", "enhanced", "by", "shortcut", "connections", "as", "the", "backbone", "of", "our", "architecture", ",", "which", "enables", "large", "receptive", "fields", "for", "global", "context", "and", "more", "steps", "of", "movements", ".", "Representing", "the", "shape", "in", "graph", "also", "benefits", "the", "learning", "procedure", ".", "The", "known", "connectivity", "allows", "us", "to", "define", "higher", "order", "loss", "functions", "across", "neighboring", "nodes", ",", "which", "are", "important", "to", "regularize", "3D", "shapes", ".", "Specifically", ",", "we", "define", "a", "surface", "normal", "loss", "to", "favor", "smooth", "surface", ";", "an", "edge", "loss", "to", "encourage", "uniform", "distribution", "of", "mesh", "vertices", "for", "high", "recall", ";", "and", "a", "laplacian", "loss", "to", "prevent", "mesh", "faces", "from", "intersecting", "each", "other", ".", "All", "of", "these", "losses", "are", "essential", "to", "generate", "quality", "appealing", "mesh", "model", ",", "and", "none", "of", "them", "can", "be", "trivially", "defined", "without", "the", "graph", "representation", ".", "The", "contributions", "of", "this", "paper", "are", "mainly", "in", "three", "aspects", ".", "First", ",", "we", "propose", "a", "novel", "end", "-", "to", "-", "end", "neural", "network", "architecture", "that", "generates", "a", "3D", "mesh", "model", "from", "a", "single", "RGB", "image", ".", "Second", ",", "we", "design", "a", "projection", "layer", "which", "incorporates", "perceptual", "image", "features", "into", "the", "3D", "geometry", "represented", "by", "GCN", ".", "Third", ",", "our", "network", "predict", "3D", "geometry", "in", "a", "coarse", "to", "fine", "fashion", ",", "which", "is", "more", "reliable", "and", "easy", "to", "learn", ".", "section", ":", "Related", "Work", "3D", "reconstruction", "has", "been", "well", "studied", "based", "on", "the", "multi", "-", "view", "geometry", "(", "MVG", ")", "in", "the", "literature", ".", "The", "major", "research", "directions", "include", "structure", "from", "motion", "(", "SfM", ")", "for", "large", "-", "scale", "high", "-", "quality", "reconstruction", "and", "simultaneous", "localization", "and", "mapping", "(", "SLAM", ")", "for", "navigation", ".", "Though", "they", "are", "very", "successful", "in", "these", "scenarios", ",", "they", "are", "restricted", "by", "1", ")", "the", "coverage", "that", "the", "multiple", "views", "can", "give", "and", "2", ")", "the", "appearance", "of", "the", "object", "that", "wants", "to", "reconstruct", ".", "The", "former", "restriction", "means", "MVG", "can", "not", "reconstruct", "unseen", "parts", "of", "the", "object", ",", "and", "thus", "it", "usually", "takes", "a", "long", "time", "to", "get", "enough", "views", "for", "a", "good", "reconstruction", ";", "the", "latter", "restriction", "means", "MVG", "can", "not", "reconstruct", "non", "-", "lambertian", "(", "e.g.", "reflective", "or", "transparent", ")", "or", "textureless", "objects", ".", "These", "restrictions", "lead", "to", "the", "trend", "of", "resorting", "to", "learning", "based", "approaches", ".", "Learning", "based", "approaches", "usually", "consider", "single", "or", "few", "images", ",", "as", "it", "largely", "relies", "on", "the", "shape", "priors", "that", "it", "can", "learn", "from", "data", ".", "Early", "works", "can", "be", "traced", "back", "to", "Hoiem", "et", "al", ".", "and", "Saxena", "et", "al", ".", ".", "Most", "recently", ",", "with", "the", "success", "of", "deep", "learning", "architectures", "and", "the", "release", "of", "large", "-", "scale", "3D", "shape", "datasets", "such", "as", "ShapeNet", ",", "learning", "based", "approaches", "have", "achieved", "great", "progress", ".", "Huang", "et", "al", ".", "and", "Su", "et", "al", ".", "retrieve", "shape", "components", "from", "a", "large", "dataset", ",", "assemble", "them", "and", "deform", "the", "assembled", "shape", "to", "fit", "the", "observed", "image", ".", "However", ",", "shape", "retrieval", "from", "images", "itself", "is", "an", "ill", "-", "posed", "problem", ".", "To", "avoid", "this", "problem", ",", "Kar", "et", "al", ".", "learns", "a", "3D", "deformable", "model", "for", "each", "object", "category", "and", "capture", "the", "shape", "variations", "in", "different", "images", ".", "However", ",", "the", "reconstruction", "is", "limited", "to", "the", "popular", "categories", "and", "its", "reconstruction", "result", "is", "usually", "lack", "of", "details", ".", "Another", "line", "of", "research", "is", "to", "directly", "learn", "3D", "shapes", "from", "single", "images", ".", "Restricted", "by", "the", "prevalent", "grid", "-", "based", "deep", "learning", "architectures", ",", "most", "works", "outputs", "3D", "voxels", ",", "which", "are", "usually", "with", "low", "resolutions", "due", "to", "the", "memory", "constraint", "on", "a", "modern", "GPU", ".", "Most", "recently", ",", "Tatarchenko", "et", "al", ".", "have", "proposed", "an", "octree", "representation", ",", "which", "allows", "to", "reconstructing", "higher", "resolution", "outputs", "with", "a", "limited", "memory", "budget", ".", "However", ",", "a", "3D", "voxel", "is", "still", "not", "a", "popular", "shape", "representation", "in", "game", "and", "movie", "industries", ".", "To", "avoid", "drawbacks", "of", "the", "voxel", "representation", ",", "Fan", "et", "al", ".", "propose", "to", "generate", "point", "clouds", "from", "single", "images", ".", "The", "point", "cloud", "representation", "has", "no", "local", "connections", "between", "points", ",", "and", "thus", "the", "point", "positions", "have", "a", "very", "large", "degree", "of", "freedom", ".", "Consequently", ",", "the", "generated", "point", "cloud", "is", "usually", "not", "close", "to", "a", "surface", "and", "can", "not", "be", "used", "to", "recover", "a", "3D", "mesh", "directly", ".", "Besides", "these", "typical", "3D", "representations", ",", "there", "is", "an", "interesting", "work", "which", "uses", "a", "so", "-", "called", "“", "geometry", "image", "”", "to", "represent", "a", "3D", "shape", ".", "Thus", ",", "their", "network", "is", "a", "2D", "convolutional", "neural", "network", "which", "conducts", "an", "image", "to", "image", "mapping", ".", "Our", "works", "are", "mostly", "related", "to", "the", "two", "recent", "works", "and", ".", "However", ",", "the", "former", "adopts", "simple", "silhouette", "supervision", ",", "and", "hence", "does", "not", "perform", "well", "for", "complicated", "objects", "such", "as", "car", ",", "lamp", ",", "etc", ";", "the", "latter", "needs", "a", "large", "model", "repository", "to", "generate", "a", "combined", "model", ".", "Our", "base", "network", "is", "a", "graph", "neural", "network", ";", "this", "architecture", "has", "been", "adopted", "for", "shape", "analysis", ".", "In", "the", "meanwhile", ",", "there", "are", "charting", "-", "based", "methods", "which", "directly", "apply", "convolutions", "on", "surface", "manifolds", "for", "shape", "analysis", ".", "As", "far", "as", "we", "know", ",", "these", "architectures", "have", "never", "been", "adopted", "for", "3D", "reconstruction", "from", "single", "images", ",", "though", "graph", "and", "surface", "manifold", "are", "natural", "representations", "for", "meshed", "objects", ".", "For", "a", "comprehensive", "understanding", "of", "the", "graph", "neural", "network", ",", "the", "charting", "-", "based", "methods", "and", "their", "applications", ",", "please", "refer", "to", "this", "survey", ".", "section", ":", "Method", "subsection", ":", "Preliminary", ":", "Graph", "-", "based", "Convolution", "We", "first", "provide", "some", "background", "about", "graph", "based", "convolution", ";", "more", "detailed", "introduction", "can", "be", "found", "in", ".", "A", "3D", "mesh", "is", "a", "collection", "of", "vertices", ",", "edges", "and", "faces", "that", "defines", "the", "shape", "of", "a", "3D", "object", ";", "it", "can", "be", "represented", "by", "a", "graph", ",", "where", "is", "the", "set", "of", "vertices", "in", "the", "mesh", ",", "is", "the", "set", "of", "edges", "with", "each", "connecting", "two", "vertices", ",", "and", "are", "the", "feature", "vectors", "attached", "on", "vertices", ".", "A", "graph", "based", "convolutional", "layer", "is", "defined", "on", "irregular", "graph", "as", ":", "where", "are", "the", "feature", "vectors", "on", "vertex", "before", "and", "after", "the", "convolution", ",", "and", "is", "the", "neighboring", "vertices", "of", ";", "and", "are", "the", "learnable", "parameter", "matrices", "of", "that", "are", "applied", "to", "all", "vertices", ".", "Note", "that", "is", "shared", "for", "all", "edges", ",", "and", "thus", "(", "[", "reference", "]", ")", "works", "on", "nodes", "with", "different", "vertex", "degrees", ".", "In", "our", "case", ",", "the", "attached", "feature", "vector", "is", "the", "concatenation", "of", "the", "3D", "vertex", "coordinate", ",", "feature", "encoding", "3D", "shape", ",", "and", "feature", "learned", "from", "the", "input", "color", "image", "(", "if", "they", "exist", ")", ".", "Running", "convolutions", "updates", "the", "features", ",", "which", "is", "equivalent", "as", "applying", "a", "deformation", ".", "subsection", ":", "System", "Overview", "Our", "model", "is", "an", "end", "-", "to", "-", "end", "deep", "learning", "framework", "that", "takes", "a", "single", "color", "image", "as", "input", "and", "produces", "a", "3D", "mesh", "model", "in", "camera", "coordinate", ".", "The", "overview", "of", "our", "framework", "is", "illustrated", "in", "Fig", ".", "[", "reference", "]", ".", "The", "whole", "network", "consists", "an", "image", "feature", "network", "and", "a", "cascaded", "mesh", "deformation", "network", ".", "The", "image", "feature", "network", "is", "a", "2D", "CNN", "that", "extract", "perceptual", "feature", "from", "the", "input", "image", ",", "which", "is", "leveraged", "by", "the", "mesh", "deformation", "network", "to", "progressively", "deform", "an", "ellipsoid", "mesh", "into", "the", "desired", "3D", "model", ".", "The", "cascaded", "mesh", "deformation", "network", "is", "a", "graph", "-", "based", "convolution", "network", "(", "GCN", ")", ",", "which", "contains", "three", "deformation", "blocks", "intersected", "by", "two", "graph", "unpooling", "layers", ".", "Each", "deformation", "block", "takes", "an", "input", "graph", "representing", "the", "current", "mesh", "model", "with", "the", "3D", "shape", "feature", "attached", "on", "vertices", ",", "and", "produces", "new", "vertices", "locations", "and", "features", ".", "Whereas", "the", "graph", "unpooling", "layers", "increase", "the", "number", "of", "vertices", "to", "increase", "the", "capacity", "of", "handling", "details", ",", "while", "still", "maintain", "the", "triangular", "mesh", "topology", ".", "Starting", "from", "a", "smaller", "number", "of", "vertices", ",", "our", "model", "learns", "to", "gradually", "deform", "and", "add", "details", "to", "the", "mesh", "model", "in", "a", "coarse", "-", "to", "-", "fine", "fashion", ".", "In", "order", "to", "train", "the", "network", "to", "produce", "stable", "deformation", "and", "generate", "an", "accurate", "mesh", ",", "we", "extend", "the", "Chamfer", "Distance", "loss", "used", "by", "Fan", "et", "al", ".", "with", "three", "other", "mesh", "specific", "loss", "–", "Surface", "normal", "loss", ",", "Laplacian", "regularization", "loss", ",", "and", "Edge", "length", "loss", ".", "The", "remaining", "part", "of", "this", "section", "describes", "details", "of", "these", "components", ".", "subsection", ":", "Initial", "ellipsoid", "Our", "model", "does", "not", "require", "any", "prior", "knowledge", "of", "the", "3D", "shape", ",", "and", "always", "deform", "from", "an", "initial", "ellipsoid", "with", "average", "size", "placed", "at", "the", "common", "location", "in", "the", "camera", "coordinate", ".", "The", "ellipsoid", "is", "centered", "at", "0.8", "m", "in", "front", "of", "the", "camera", "with", "0.2", "m", ",", "0.2", "m", ",", "0.4", "m", "as", "the", "radius", "of", "three", "axis", ".", "The", "mesh", "model", "is", "generated", "by", "implicit", "surface", "algorithm", "in", "Meshlab", "and", "contains", "156", "vertices", ".", "We", "use", "this", "ellipsoid", "to", "initialize", "our", "input", "graph", ",", "where", "the", "initial", "feature", "contains", "only", "the", "3D", "coordinate", "of", "each", "vertex", ".", "subsection", ":", "Mesh", "deformation", "block", "The", "architecture", "of", "mesh", "deformation", "block", "is", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "a", ")", ".", "In", "order", "to", "generate", "3D", "mesh", "model", "that", "is", "consistent", "with", "the", "object", "shown", "in", "the", "input", "image", ",", "the", "deformation", "block", "need", "to", "pool", "feature", "(", ")", "from", "the", "input", "image", ".", "This", "is", "done", "in", "conjunction", "with", "the", "image", "feature", "network", "and", "a", "perceptual", "feature", "pooling", "layer", "given", "the", "location", "of", "vertex", "(", ")", "in", "the", "current", "mesh", "model", ".", "The", "pooled", "perceptual", "feature", "is", "then", "concatenated", "with", "the", "3D", "shape", "feature", "attached", "on", "the", "vertex", "from", "the", "input", "graph", "(", ")", "and", "fed", "into", "a", "series", "of", "graph", "based", "ResNet", "(", "G", "-", "ResNet", ")", ".", "The", "G", "-", "ResNet", "produces", ",", "also", "as", "the", "output", "of", "the", "mesh", "deformation", "block", ",", "the", "new", "coordinates", "(", ")", "and", "3d", "shape", "feature", "(", ")", "for", "each", "vertex", ".", "subsubsection", ":", "Perceptual", "feature", "pooling", "layer", "We", "use", "a", "VGG", "-", "16", "architecture", "up", "to", "layer", "conv5_3", "as", "the", "image", "feature", "network", "as", "it", "has", "been", "widely", "used", ".", "Given", "the", "3D", "coordinate", "of", "a", "vertex", ",", "we", "calculate", "its", "2D", "projection", "on", "input", "image", "plane", "using", "camera", "intrinsics", ",", "and", "then", "pool", "the", "feature", "from", "four", "nearby", "pixels", "using", "bilinear", "interpolation", ".", "In", "particular", ",", "we", "concatenate", "feature", "extracted", "from", "layer", "‘", "conv3_3", "’", ",", "‘", "conv4_3", "’", ",", "and", "‘", "conv5_3", "’", ",", "which", "results", "in", "a", "total", "dimension", "of", "1280", ".", "This", "perceptual", "feature", "is", "then", "concatenated", "with", "the", "128", "-", "dim", "3D", "feature", "from", "the", "input", "mesh", ",", "which", "results", "in", "a", "total", "dimension", "of", "1408", ".", "This", "is", "illustrated", "in", "Fig", ".", "[", "reference", "]", "(", "b", ")", ".", "Note", "that", "in", "the", "first", "block", ",", "the", "perceptual", "feature", "is", "concatenated", "with", "the", "3", "-", "dim", "feature", "(", "coordinate", ")", "since", "there", "is", "no", "learnt", "shape", "feature", "at", "the", "beginning", ".", "subsubsection", ":", "G", "-", "ResNet", "After", "obtaining", "1408", "-", "dim", "feature", "for", "each", "vertex", "representing", "both", "3D", "shape", "and", "2D", "image", "information", ",", "we", "design", "a", "graph", "based", "convolutional", "neural", "network", "to", "predict", "new", "location", "and", "3D", "shape", "feature", "for", "each", "vertex", ".", "This", "requires", "efficient", "exchange", "of", "the", "information", "between", "vertices", ".", "However", ",", "as", "defined", "in", "(", "[", "reference", "]", ")", ",", "each", "convolution", "only", "enables", "the", "feature", "exchanging", "between", "neighboring", "pixels", ",", "which", "severely", "impairs", "the", "efficiency", "of", "information", "exchanging", ".", "This", "is", "equivalent", "as", "the", "small", "receptive", "field", "issue", "on", "2D", "CNN", ".", "To", "solve", "this", "issue", ",", "we", "make", "a", "very", "deep", "network", "with", "shortcut", "connections", "and", "denote", "it", "as", "G", "-", "ResNet", "(", "Fig", ".", "[", "reference", "]", "(", "a", ")", ")", ".", "In", "this", "work", ",", "the", "G", "-", "ResNet", "in", "all", "blocks", "has", "the", "same", "structure", ",", "which", "consists", "of", "14", "graph", "residual", "convolutional", "layers", "with", "128", "channels", ".", "The", "serial", "of", "G", "-", "ResNet", "block", "produces", "a", "new", "128", "-", "dim", "3D", "feature", ".", "In", "addition", "to", "the", "feature", "output", ",", "there", "is", "a", "branch", "which", "applies", "an", "extra", "graph", "convolutional", "layer", "to", "the", "last", "layer", "features", "and", "outputs", "the", "3D", "coordinates", "of", "the", "vertex", ".", "subsection", ":", "Graph", "unpooling", "layer", "The", "goal", "of", "unpooling", "layer", "is", "to", "increase", "the", "number", "of", "vertex", "in", "the", "GCNN", ".", "It", "allows", "us", "to", "start", "from", "a", "mesh", "with", "fewer", "vertices", "and", "add", "more", "only", "when", "necessary", ",", "which", "reduces", "memory", "costs", "and", "produces", "better", "results", ".", "A", "straightforward", "approach", "is", "to", "add", "one", "vertex", "in", "the", "center", "of", "each", "triangle", "and", "connect", "it", "with", "the", "three", "vertices", "of", "the", "triangle", "(", "Fig", ".", "[", "reference", "]", "(", "b", ")", "Face", "-", "based", ")", ".", "However", ",", "this", "causes", "imbalanced", "vertex", "degrees", ",", "i.e.", "number", "of", "edges", "on", "vertex", ".", "Inspired", "by", "the", "vertex", "adding", "strategy", "of", "the", "mesh", "subdivision", "algorithm", "prevalent", "in", "computer", "graphics", ",", "we", "add", "a", "vertex", "at", "the", "center", "of", "each", "edge", "and", "connect", "it", "with", "the", "two", "end", "-", "point", "of", "this", "edge", "(", "Fig", ".", "[", "reference", "]", "(", "a", ")", ")", ".", "The", "3D", "feature", "for", "newly", "added", "vertex", "is", "set", "as", "the", "average", "of", "its", "two", "neighbors", ".", "We", "also", "connect", "three", "vertices", "if", "they", "are", "added", "on", "the", "same", "triangle", "(", "dashed", "line", ".", ")", "Consequently", ",", "we", "create", "4", "new", "triangles", "for", "each", "triangle", "in", "the", "original", "mesh", ",", "and", "the", "number", "of", "vertex", "is", "increased", "by", "the", "number", "of", "edges", "in", "the", "original", "mesh", ".", "This", "edge", "-", "based", "unpooling", "uniformly", "upsamples", "the", "vertices", "as", "shown", "in", "Fig", ".", "[", "reference", "]", "(", "b", ")", "Edge", "-", "based", ".", "subsection", ":", "Losses", "We", "define", "four", "kinds", "of", "losses", "to", "constrain", "the", "property", "of", "the", "output", "shape", "and", "the", "deformation", "procedure", "to", "guarantee", "appealing", "results", ".", "We", "adopt", "the", "Chamfer", "loss", "to", "constrain", "the", "location", "of", "mesh", "vertices", ",", "a", "normal", "loss", "to", "enforce", "the", "consistency", "of", "surface", "normal", ",", "a", "laplacian", "regularization", "to", "maintain", "relative", "location", "between", "neighboring", "vertices", "during", "deformation", ",", "and", "an", "edge", "length", "regularization", "to", "prevent", "outliers", ".", "These", "losses", "are", "applied", "with", "equal", "weight", "on", "both", "the", "intermediate", "and", "final", "mesh", ".", "Unless", "otherwise", "stated", ",", "we", "use", "for", "a", "vertex", "in", "the", "predicted", "mesh", ",", "for", "a", "vertex", "in", "the", "ground", "truth", "mesh", ",", "for", "the", "neighboring", "pixel", "of", ",", "till", "the", "end", "of", "this", "section", ".", "subsubsection", ":", "Chamfer", "loss", "The", "Chamfer", "distance", "measures", "the", "distance", "of", "each", "point", "to", "the", "other", "set", ":", "It", "is", "reasonably", "good", "to", "regress", "the", "vertices", "close", "to", "its", "correct", "position", ",", "however", "is", "not", "sufficient", "to", "produce", "nice", "3D", "mesh", "(", "see", "the", "result", "of", "Fan", "et", "al", ".", "in", "Fig", ".", "[", "reference", "]", ")", ".", "subsubsection", ":", "Normal", "loss", "We", "further", "define", "loss", "on", "surface", "normal", "to", "characterize", "high", "order", "properties", ":", "where", "is", "the", "closest", "vertex", "for", "that", "is", "found", "when", "calculating", "the", "chamfer", "loss", ",", "is", "the", "neighboring", "pixel", "of", ",", "is", "the", "inner", "product", "of", "two", "vectors", ",", "and", "is", "the", "observed", "surface", "normal", "from", "ground", "truth", ".", "Essentially", ",", "this", "loss", "requires", "the", "edge", "between", "a", "vertex", "with", "its", "neighbors", "to", "perpendicular", "to", "the", "observation", "from", "the", "ground", "truth", ".", "One", "may", "find", "that", "this", "loss", "does", "not", "equal", "to", "zero", "unless", "on", "a", "planar", "surface", ".", "However", ",", "optimizing", "this", "loss", "is", "equivalent", "as", "forcing", "the", "normal", "of", "a", "locally", "fitted", "tangent", "plane", "to", "be", "consistent", "with", "the", "observation", ",", "which", "works", "practically", "well", "in", "our", "experiment", ".", "Moreover", ",", "this", "normal", "loss", "is", "fully", "differentiable", "and", "easy", "to", "optimize", ".", "subsubsection", ":", "Regularization", "Even", "with", "the", "Chamfer", "loss", "and", "Normal", "loss", ",", "the", "optimization", "is", "easily", "stucked", "in", "some", "local", "minimum", ".", "More", "specifically", ",", "the", "network", "may", "generate", "some", "super", "large", "deformation", "to", "favor", "some", "local", "consistency", ",", "which", "is", "especially", "harmful", "at", "the", "beginning", "when", "the", "estimation", "is", "far", "from", "ground", "truth", ",", "and", "causes", "flying", "vertices", "(", "Fig", ".", "[", "reference", "]", ")", ".", "paragraph", ":", "Laplacian", "regularization", "To", "handle", "these", "problem", ",", "we", "first", "propose", "a", "Laplacian", "term", "to", "prevent", "the", "vertices", "from", "moving", "too", "freely", ",", "which", "potentially", "avoids", "mesh", "self", "-", "intersection", ".", "The", "laplaician", "term", "serves", "as", "a", "local", "detail", "preserving", "operator", ",", "that", "encourages", "neighboring", "vertices", "to", "have", "the", "same", "movement", ".", "In", "the", "first", "deformation", "block", ",", "it", "acts", "like", "a", "surface", "smoothness", "term", "since", "the", "input", "to", "this", "block", "is", "a", "smooth", "-", "everywhere", "ellipsoid", ";", "starting", "from", "the", "second", "block", ",", "it", "prevents", "the", "3D", "mesh", "model", "from", "deforming", "too", "much", ",", "so", "that", "only", "fine", "-", "grained", "details", "are", "added", "to", "the", "mesh", "model", ".", "To", "calculate", "this", "loss", ",", "we", "first", "define", "a", "laplacian", "coordinate", "for", "each", "vertex", "as", "and", "the", "laplacian", "regularization", "is", "defined", "as", ":", "where", "and", "are", "the", "laplacian", "coordinate", "of", "a", "vertex", "after", "and", "before", "a", "deformation", "block", ".", "paragraph", ":", "Edge", "length", "regularization", ".", "To", "penalize", "flying", "vertices", ",", "which", "ususally", "cause", "long", "edge", ",", "we", "add", "an", "edge", "length", "regularization", "loss", ":", "The", "overall", "loss", "is", "a", "weighted", "sum", "of", "all", "four", "losses", ",", ",", "where", ",", "and", "are", "the", "hyperparameters", "which", "balance", "the", "losses", "and", "fixed", "for", "all", "the", "experiments", ".", "section", ":", "Experiment", "In", "this", "section", ",", "we", "perform", "an", "extensive", "evaluation", "on", "our", "model", ".", "In", "addition", "to", "comparing", "with", "previous", "3D", "shape", "generation", "works", "for", "evaluating", "the", "reconstruction", "accuracy", ",", "we", "also", "analyse", "the", "importance", "of", "each", "component", "in", "our", "model", ".", "Qualitative", "results", "on", "both", "synthetic", "and", "real", "-", "world", "images", "further", "show", "that", "our", "model", "produces", "triangular", "meshes", "with", "smooth", "surfaces", "and", "still", "maintains", "details", "depicted", "in", "the", "input", "images", ".", "subsection", ":", "Experimental", "setup", "subsubsection", ":", "Data", ".", "We", "use", "the", "dataset", "provided", "by", "Choy", "et", "al", ".", ".", "The", "dataset", "contains", "rendering", "images", "of", "50k", "models", "belonging", "to", "13", "object", "categories", "from", "ShapeNet", ",", "which", "is", "a", "collection", "of", "3D", "CAD", "models", "that", "are", "organized", "according", "to", "the", "WordNet", "hierarchy", ".", "A", "model", "is", "rendered", "from", "various", "camera", "viewpoints", ",", "and", "camera", "intrinsic", "and", "extrinsic", "matrices", "are", "recorded", ".", "For", "fair", "comparison", ",", "we", "use", "the", "same", "training", "/", "testing", "split", "as", "in", "Choy", "et", ".", "al", ".", ".", "subsubsection", ":", "Evaluation", "Metric", ".", "We", "adopt", "the", "standard", "3D", "reconstruction", "metric", ".", "We", "first", "uniformly", "sample", "points", "from", "our", "result", "and", "ground", "truth", ".", "We", "calculate", "precision", "and", "recall", "by", "checking", "the", "percentage", "of", "points", "in", "prediction", "or", "ground", "truth", "that", "can", "find", "a", "nearest", "neighbor", "from", "the", "other", "within", "certain", "threshold", ".", "A", "F", "-", "score", "as", "the", "harmonic", "mean", "of", "precision", "and", "recall", "is", "then", "calculated", ".", "Following", "Fan", "et", ".", "al", ".", ",", "we", "also", "report", "the", "Chamfer", "Distance", "(", "CD", ")", "and", "Earth", "Mover", "’s", "Distance", "(", "EMD", ")", ".", "For", "F", "-", "Score", ",", "larger", "is", "better", ".", "For", "CD", "and", "EMD", ",", "smaller", "is", "better", ".", "On", "the", "other", "hand", ",", "we", "realize", "that", "the", "commonly", "used", "evaluation", "metrics", "for", "shape", "generation", "may", "not", "thoroughly", "reflect", "the", "shape", "quality", ".", "They", "often", "capture", "occupancy", "or", "point", "-", "wise", "distance", "rather", "than", "surface", "properties", ",", "such", "as", "continuity", ",", "smoothness", ",", "high", "-", "order", "details", ",", "for", "which", "a", "standard", "evaluation", "metric", "is", "barely", "missing", "in", "literature", ".", "Thus", ",", "we", "recommend", "to", "pay", "attention", "on", "qualitative", "results", "for", "better", "understanding", "of", "these", "aspects", ".", "subsubsection", ":", "Baselines", ".", "We", "compare", "the", "presented", "approach", "to", "the", "most", "recent", "single", "image", "reconstruction", "approaches", ".", "Specifically", ",", "we", "compare", "with", "two", "state", "-", "of", "-", "the", "-", "art", "methods", "-", "Choy", "et", ".", "al", ".", "(", "3D", "-", "R2N2", ")", "producing", "3D", "volume", ",", "and", "Fan", "et", ".", "al", ".", "(", "PSG", ")", "producing", "point", "cloud", ".", "Since", "the", "metrics", "are", "defined", "on", "point", "cloud", ",", "we", "can", "evaluate", "PSG", "directly", "on", "its", "output", ",", "our", "method", "by", "uniformly", "sampling", "point", "on", "surface", ",", "and", "3D", "-", "R2N2", "by", "uniformly", "sampling", "point", "from", "mesh", "created", "using", "the", "Marching", "Cube", "method", ".", "We", "also", "compare", "to", "Neural", "3D", "Mesh", "Renderer", "(", "N3MR", ")", "which", "is", "so", "far", "the", "only", "deep", "learning", "based", "mesh", "generation", "model", "with", "code", "public", "available", ".", "For", "fair", "comparison", ",", "the", "models", "are", "trained", "with", "the", "same", "data", "using", "the", "same", "amount", "of", "time", ".", "subsubsection", ":", "Training", "and", "Runtime", ".", "Our", "network", "receives", "input", "images", "of", "size", ",", "and", "initial", "ellipsoid", "with", "156", "vertices", "and", "462", "edges", ".", "The", "network", "is", "implemented", "in", "Tensorflow", "and", "optimized", "using", "Adam", "with", "weight", "decay", "1e", "-", "5", ".", "The", "batch", "size", "is", "1", ";", "the", "total", "number", "of", "training", "epoch", "is", "50", ";", "the", "learning", "rate", "is", "initialized", "as", "3e", "-", "5", "and", "drops", "to", "1e", "-", "5", "after", "40", "epochs", ".", "The", "total", "training", "time", "is", "72", "hours", "on", "a", "Nvidia", "Titan", "X.", "During", "testing", ",", "our", "model", "takes", "15.58ms", "to", "generate", "a", "mesh", "with", "2466", "vertices", ".", "subsection", ":", "Comparison", "to", "state", "of", "the", "art", "Tab", ".", "[", "reference", "]", "shows", "the", "F", "-", "score", "with", "different", "thresholds", "of", "different", "methods", ".", "Our", "approach", "outperforms", "the", "other", "methods", "in", "all", "categories", "except", "watercraft", ".", "Notably", ",", "our", "results", "are", "significantly", "better", "than", "the", "others", "in", "all", "categories", "under", "a", "smaller", "threshold", ",", "showing", "at", "least", "10", "%", "F", "-", "score", "improvement", ".", "N3MR", "does", "not", "perform", "well", ",", "and", "its", "result", "is", "about", "50", "%", "worse", "than", "ours", ",", "probably", "because", "their", "model", "only", "learns", "from", "limited", "silhouette", "signal", "in", "images", "and", "lacks", "of", "explicit", "handling", "of", "the", "3D", "mesh", ".", "We", "also", "show", "the", "CD", "and", "EMD", "for", "all", "categories", "in", "Tab", ".", "[", "reference", "]", ".", "Our", "approach", "outperforms", "the", "other", "methods", "in", "most", "categories", "and", "achieves", "the", "best", "mean", "score", ".", "The", "major", "competitor", "is", "PSG", ",", "which", "produces", "a", "point", "cloud", "and", "has", "the", "most", "freedom", ";", "this", "freedom", "leads", "to", "smaller", "CD", "and", "EMD", ",", "however", "does", "not", "necessarily", "leads", "to", "a", "better", "mesh", "model", "without", "proper", "regularization", ".", "To", "demonstrate", "this", ",", "we", "show", "the", "qualitative", "results", "to", "analyze", "why", "our", "approach", "outperforms", "the", "others", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "visual", "results", ".", "To", "compare", "the", "quality", "of", "mesh", "model", ",", "we", "convert", "volumetric", "and", "point", "cloud", "to", "mesh", "using", "standard", "approaches", ".", "As", "we", "can", "see", ",", "the", "3D", "volume", "results", "produced", "by", "3D", "-", "R2N2", "lack", "of", "details", "due", "to", "the", "low", "resolution", ",", "e.g.", ",", "the", "legs", "are", "missing", "in", "the", "chair", "example", "as", "shown", "in", "the", "4", "-", "th", "row", "of", "Fig", ".", "[", "reference", "]", ".", "We", "tried", "octree", "based", "solution", "to", "increase", "the", "volume", "resolution", ",", "but", "found", "it", "still", "hard", "to", "recover", "surface", "level", "details", "as", "much", "as", "our", "model", ".", "PSG", "produces", "sparse", "3D", "point", "clouds", ",", "and", "it", "is", "non", "-", "trivial", "to", "recover", "meshes", "from", "them", ".", "This", "is", "due", "to", "the", "applied", "Chamfer", "loss", "acting", "like", "a", "regression", "loss", "which", "gives", "too", "much", "degree", "of", "freedom", "to", "the", "point", "cloud", ".", "N3MR", "produces", "very", "rough", "shape", ",", "which", "might", "be", "sufficient", "for", "some", "rendering", "tasks", ",", "however", "can", "not", "recover", "complicated", "objects", "such", "as", "chairs", "and", "tables", ".", "In", "contrast", ",", "our", "model", "does", "not", "suffer", "from", "these", "issues", "by", "leveraging", "a", "mesh", "representation", ",", "integration", "of", "perceptual", "feature", ",", "and", "carefully", "defined", "losses", "during", "the", "training", ".", "Our", "result", "is", "not", "restricted", "by", "the", "resolution", "due", "to", "the", "limited", "memory", "budget", "and", "contains", "both", "smooth", "continuous", "surface", "and", "local", "details", ".", "subsection", ":", "Ablation", "Study", "Now", "we", "conduct", "controlled", "experiments", "to", "analyse", "the", "importance", "of", "each", "component", "in", "our", "model", ".", "Tab", ".", "[", "reference", "]", "reports", "the", "performance", "of", "each", "model", "by", "removing", "one", "component", "from", "the", "full", "model", ".", "Again", ",", "we", "argue", "that", "these", "commonly", "used", "evaluation", "metrics", "does", "not", "necessarily", "reflect", "the", "quality", "of", "the", "recovered", "3D", "geometry", ".", "For", "example", ",", "the", "model", "with", "no", "edge", "length", "regularization", "achieves", "the", "best", "performance", "across", "all", ",", "however", ",", "in", "fact", "produces", "the", "worst", "mesh", "(", "Fig", ".", "[", "reference", "]", ",", "the", "last", "2nd", "column", ")", ".", "As", "such", ",", "we", "use", "qualitative", "result", "Fig", ".", "[", "reference", "]", "to", "show", "the", "contribution", "of", "each", "component", "in", "our", "system", ".", "subsubsection", ":", "Graph", "Unpooling", "We", "first", "remove", "the", "graph", "unpooling", "layers", ",", "and", "thus", "each", "block", "has", "the", "same", "number", "of", "vertices", "as", "in", "the", "last", "block", "of", "our", "full", "model", ".", "It", "is", "observed", "that", "the", "deformation", "makes", "mistake", "easier", "at", "beginning", ",", "which", "can", "not", "be", "fixed", "later", "on", ".", "Consequently", ",", "there", "are", "some", "obvious", "artifacts", "in", "some", "parts", "of", "the", "objects", ".", "subsubsection", ":", "G", "-", "ResNet", "We", "then", "remove", "the", "shortcut", "connections", "in", "G", "-", "ResNet", ",", "and", "make", "it", "regular", "GCN", ".", "As", "can", "be", "seen", "from", "Tab", ".", "[", "reference", "]", ",", "there", "is", "a", "huge", "performance", "gap", "in", "all", "four", "measurement", "metrics", ",", "which", "means", "the", "failure", "of", "optimizing", "Chamfer", "distance", ".", "The", "main", "reason", "is", "the", "degradation", "problem", "observed", "in", "the", "very", "deep", "2D", "convolutional", "neural", "network", ".", "Such", "problem", "leads", "to", "a", "higher", "training", "error", "(", "and", "thus", "higher", "testing", "error", ")", "when", "adding", "more", "layers", "to", "a", "suitably", "deep", "model", ".", "Essetially", ",", "our", "network", "has", "42", "graph", "convolutional", "layers", ".", "Thus", ",", "this", "phenomenon", "has", "also", "been", "observed", "in", "our", "very", "deep", "graph", "neural", "network", "experiment", ".", "subsubsection", ":", "Loss", "terms", "We", "evaluate", "the", "function", "of", "each", "additional", "terms", "besides", "the", "Chamfer", "loss", ".", "As", "can", "be", "seen", "in", "Fig", ".", "[", "reference", "]", ",", "removing", "normal", "loss", "severely", "impairs", "the", "surface", "smoothness", "and", "local", "details", ",", "e.g.", "seat", "back", ";", "removing", "Laplacian", "term", "causes", "intersecting", "geometry", "because", "the", "local", "topology", "changes", ",", "e.g.", "the", "hand", "held", "of", "the", "chair", ";", "removing", "edge", "length", "term", "causes", "flying", "vertices", "and", "surfaces", ",", "which", "completely", "ruins", "the", "surface", "characteristics", ".", "These", "results", "demonstrate", "that", "all", "the", "components", "presented", "in", "this", "work", "contribute", "to", "the", "final", "performance", ".", "subsubsection", ":", "Number", "of", "Deformation", "Blocks", "We", "now", "analyze", "the", "effects", "of", "the", "number", "of", "blocks", ".", "Figure", "Fig", ".", "[", "reference", "]", "(", "left", ")", "shows", "the", "mean", "F", "-", "score", "(", ")", "and", "CD", "with", "regard", "to", "the", "number", "of", "blocks", ".", "The", "results", "indicate", "that", "increasing", "the", "number", "of", "blocks", "helps", ",", "but", "the", "benefit", "is", "getting", "saturated", "with", "more", "blocks", ",", "e.g.", "from", "3", "to", "4", ".", "In", "our", "experiment", ",", "we", "found", "that", "4", "blocks", "results", "in", "too", "many", "vertices", "and", "edges", ",", "which", "slow", "down", "our", "approach", "dramatically", "even", "though", "it", "provides", "better", "accuracy", "on", "evaluation", "metrics", ".", "Therefore", ",", "we", "use", "3", "blocks", "in", "all", "our", "experiment", "for", "the", "best", "balance", "of", "performance", "and", "efficiency", ".", "Fig", ".", "[", "reference", "]", "(", "right", ")", "shows", "the", "output", "of", "our", "model", "after", "each", "deformation", "block", ".", "Notice", "how", "mesh", "is", "densified", "with", "more", "vertices", "and", "new", "details", "are", "added", ".", "subsection", ":", "Reconstructing", "Real", "-", "World", "images", "Following", "Choy", "et", ".", "al", ".", ",", "we", "test", "our", "network", "on", "the", "Online", "Products", "dataset", "and", "Internet", "images", "for", "qualitative", "evaluation", "on", "real", "images", ".", "We", "use", "the", "model", "trained", "from", "ShapeNet", "dataset", "and", "directly", "run", "on", "real", "images", "without", "finetuning", ",", "and", "show", "results", "in", "Fig", ".", "[", "reference", "]", ".", "As", "can", "be", "seen", ",", "our", "model", "trained", "on", "synthetic", "data", "generalizes", "well", "to", "the", "real", "-", "world", "images", "across", "various", "categories", ".", "section", ":", "Conclusion", "We", "have", "presented", "an", "approach", "to", "extract", "3D", "triangular", "meshes", "from", "singe", "images", ".", "We", "exploit", "the", "key", "advantages", "the", "mesh", "presentation", "can", "bring", "to", "us", ",", "and", "the", "key", "issues", "required", "to", "solve", "for", "success", ".", "The", "former", "includes", "surface", "normal", "constraints", "and", "information", "propagation", "along", "edges", ";", "the", "latter", "includes", "perceptual", "features", "extracted", "from", "images", "as", "a", "guidance", ".", "We", "carefully", "design", "our", "network", "structure", "and", "propose", "a", "very", "deep", "cascaded", "graph", "convolutional", "neural", "network", "with", "“", "shortcut", "”", "connections", ".", "Meshes", "are", "progressively", "refined", "by", "our", "network", "trained", "end", "-", "to", "-", "end", "with", "the", "chamfer", "loss", "and", "normal", "loss", ".", "Our", "results", "are", "significantly", "better", "than", "the", "previous", "state", "-", "of", "-", "the", "-", "art", "using", "other", "shape", "representations", "such", "as", "3D", "volume", "or", "3D", "point", "cloud", ".", "Thus", ",", "we", "believe", "mesh", "representation", "is", "the", "next", "big", "thing", "in", "this", "direction", ",", "and", "we", "hope", "that", "the", "key", "components", "discovered", "in", "our", "work", "can", "support", "follow", "-", "up", "works", "that", "will", "further", "advance", "direct", "3D", "mesh", "reconstruction", "from", "single", "images", ".", "subsubsection", ":", "Future", "work", "Our", "method", "only", "produces", "meshes", "with", "the", "same", "topology", "as", "the", "initial", "mesh", ".", "In", "the", "future", ",", "we", "will", "extend", "our", "approach", "to", "more", "general", "cases", ",", "such", "as", "scene", "level", "reconstruction", ",", "and", "learn", "from", "multiple", "images", "for", "multi", "-", "view", "reconstruction", ".", "subsubsection", ":", "Acknowledgements", "This", "work", "was", "supported", "by", "two", "projects", "from", "NSFC", "(", "#", "61622204", "and", "#", "61572134", ")", ",", "two", "projects", "from", "STCSM", "(", "#", "16JC1420401", "and", "#", "16QA1400500", ")", ",", "Eastern", "Scholar", "(", "TP2017006", ")", ",", "and", "The", "Thousand", "Talents", "Plan", "of", "China", "(", "for", "young", "professionals", ",", "D1410009", ")", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["Pixel2Mesh"], "Method"], [5, 8, ["3D", "Mesh", "Models"], "Method"], [10, 12, ["RGB", "Images"], "Material"], [19, 23, ["end", "deep", "learning", "architecture"], "Method"], [42, 45, ["deep", "neural", "network"], "Method"], [76, 78, ["mesh", "model"], "Method"], [87, 89, ["3D", "mesh"], "Method"], [91, 97, ["graph", "-", "based", "convolutional", "neural", "network"], "Method"], [119, 125, ["coarse", "-", "to", "-", "fine", "strategy"], "Method"], [129, 131, ["deformation", "procedure"], "Method"], [166, 168, ["mesh", "model"], "Method"], [176, 180, ["3D", "shape", "estimation", "accuracy"], "Metric"], [202, 205, ["Inferring", "3D", "shape"], "Task"], [212, 215, ["human", "vision", "functionality"], "Task"], [220, 222, ["computer", "vision"], "Task"], [231, 234, ["3d", "shape", "generation"], "Task"], [240, 243, ["deep", "learning", "techniques"], "Method"], [247, 249, ["convolutional", "layers"], "Method"], [253, 257, ["multi", "-", "layer", "perception"], "Method"], [268, 270, ["neural", "network"], "Method"], [298, 300, ["surface", "model"], "Method"], [318, 320, ["real", "applications"], "Task"], [335, 336, ["animation"], "Task"], [352, 355, ["single", "image", "reconstruction"], "Task"], [363, 366, ["3D", "triangular", "mesh"], "Task"], [402, 404, ["deep", "network"], "Method"], [458, 459, ["complexity"], "Metric"], [461, 464, ["deep", "learning", "model"], "Method"], [570, 572, ["mesh", "model"], "Method"], [577, 579, ["irregular", "graph"], "Method"], [582, 584, ["neural", "network"], "Method"], [628, 633, ["graph", "based", "fully", "convolutional", "network"], "Method"], [634, 635, ["GCN"], "Method"], [638, 640, ["mesh", "model"], "Method"], [674, 676, ["forward", "propagation"], "Method"], [678, 680, ["convolutional", "layers"], "Method"], [706, 711, ["VGG", "-", "16", "like", "architecture"], "Method"], [734, 738, ["perceptual", "feature", "pooling", "layer"], "Method"], [744, 745, ["GCN"], "Method"], [770, 773, ["perceptual", "feature", "pooling"], "Method"], [778, 779, ["convolutions"], "Method"], [815, 817, ["graph", "representation"], "Method"], [839, 840, ["network"], "Method"], [900, 903, ["graph", "unpooling", "layer"], "Method"], [920, 922, ["forward", "propagation"], "Method"], [960, 963, ["graph", "unpooling", "layer"], "Method"], [971, 973, ["shortcut", "connections"], "Method"], [1002, 1004, ["learning", "procedure"], "Method"], [1053, 1054, ["recall"], "Task"], [1077, 1080, ["appealing", "mesh", "model"], "Method"], [1091, 1093, ["graph", "representation"], "Method"], [1111, 1119, ["end", "-", "to", "-", "end", "neural", "network", "architecture"], "Method"], [1122, 1125, ["3D", "mesh", "model"], "Method"], [1128, 1130, ["RGB", "image"], "Material"], [1136, 1138, ["projection", "layer"], "Method"], [1149, 1150, ["GCN"], "Method"], [1178, 1180, ["3D", "reconstruction"], "Task"], [1192, 1194, ["MVG", ")"], "Method"], [1203, 1206, ["structure", "from", "motion"], "Method"], [1207, 1208, ["SfM"], "Method"], [1210, 1217, ["large", "-", "scale", "high", "-", "quality", "reconstruction"], "Method"], [1218, 1222, ["simultaneous", "localization", "and", "mapping"], "Method"], [1223, 1224, ["SLAM"], "Method"], [1226, 1227, ["navigation"], "Task"], [1268, 1269, ["MVG"], "Method"], [1293, 1294, ["reconstruction"], "Task"], [1299, 1300, ["MVG"], "Method"], [1325, 1328, ["learning", "based", "approaches"], "Method"], [1329, 1332, ["Learning", "based", "approaches"], "Method"], [1378, 1381, ["deep", "learning", "architectures"], "Method"], [1393, 1394, ["ShapeNet"], "Method"], [1395, 1398, ["learning", "based", "approaches"], "Method"], [1413, 1415, ["shape", "components"], "Method"], [1435, 1437, ["shape", "retrieval"], "Task"], [1442, 1446, ["ill", "-", "posed", "problem"], "Task"], [1458, 1461, ["3D", "deformable", "model"], "Method"], [1477, 1478, ["reconstruction"], "Task"], [1512, 1518, ["grid", "-", "based", "deep", "learning", "architectures"], "Method"], [1551, 1553, ["octree", "representation"], "Method"], [1570, 1572, ["3D", "voxel"], "Method"], [1577, 1579, ["shape", "representation"], "Method"], [1580, 1584, ["game", "and", "movie", "industries"], "Task"], [1590, 1592, ["voxel", "representation"], "Method"], [1600, 1602, ["point", "clouds"], "Task"], [1607, 1610, ["point", "cloud", "representation"], "Method"], [1658, 1660, ["3D", "representations"], "Method"], [1688, 1692, ["2D", "convolutional", "neural", "network"], "Method"], [1695, 1699, ["image", "to", "image", "mapping"], "Task"], [1743, 1745, ["model", "repository"], "Method"], [1748, 1750, ["combined", "model"], "Method"], [1756, 1759, ["graph", "neural", "network"], "Method"], [1766, 1768, ["shape", "analysis"], "Task"], [1775, 1779, ["charting", "-", "based", "methods"], "Method"], [1782, 1786, ["convolutions", "on", "surface", "manifolds"], "Method"], [1787, 1789, ["shape", "analysis"], "Task"], [1803, 1805, ["3D", "reconstruction"], "Task"], [1810, 1814, ["graph", "and", "surface", "manifold"], "Method"], [1827, 1830, ["graph", "neural", "network"], "Method"], [1832, 1836, ["charting", "-", "based", "methods"], "Method"], [1853, 1857, ["Graph", "-", "based", "Convolution"], "Method"], [1863, 1866, ["graph", "based", "convolution"], "Method"], [1902, 1903, ["graph"], "Method"], [1935, 1939, ["graph", "based", "convolutional", "layer"], "Method"], [2040, 2041, ["convolutions"], "Method"], [2061, 2069, ["end", "-", "to", "-", "end", "deep", "learning", "framework"], "Method"], [2080, 2083, ["3D", "mesh", "model"], "Method"], [2106, 2109, ["image", "feature", "network"], "Method"], [2111, 2115, ["cascaded", "mesh", "deformation", "network"], "Method"], [2117, 2120, ["image", "feature", "network"], "Method"], [2122, 2124, ["2D", "CNN"], "Method"], [2138, 2141, ["mesh", "deformation", "network"], "Method"], [2150, 2152, ["3D", "model"], "Method"], [2154, 2158, ["cascaded", "mesh", "deformation", "network"], "Method"], [2160, 2165, ["graph", "-", "based", "convolution", "network"], "Method"], [2166, 2167, ["GCN"], "Method"], [2177, 2180, ["graph", "unpooling", "layers"], "Method"], [2182, 2184, ["deformation", "block"], "Method"], [2191, 2193, ["mesh", "model"], "Method"], [2212, 2215, ["graph", "unpooling", "layers"], "Method"], [2255, 2257, ["mesh", "model"], "Method"], [2392, 2394, ["mesh", "model"], "Method"], [2397, 2400, ["implicit", "surface", "algorithm"], "Method"], [2401, 2402, ["Meshlab"], "Method"], [2432, 2435, ["Mesh", "deformation", "block"], "Method"], [2438, 2441, ["mesh", "deformation", "block"], "Task"], [2457, 2460, ["3D", "mesh", "model"], "Method"], [2493, 2496, ["image", "feature", "network"], "Method"], [2498, 2502, ["perceptual", "feature", "pooling", "layer"], "Method"], [2512, 2514, ["mesh", "model"], "Method"], [2543, 2546, ["graph", "based", "ResNet"], "Method"], [2547, 2550, ["G", "-", "ResNet"], "Method"], [2553, 2556, ["G", "-", "ResNet"], "Method"], [2564, 2567, ["mesh", "deformation", "block"], "Method"], [2585, 2589, ["Perceptual", "feature", "pooling", "layer"], "Method"], [2592, 2596, ["VGG", "-", "16", "architecture"], "Method"], [2598, 2600, ["layer", "conv5_3"], "Method"], [2602, 2605, ["image", "feature", "network"], "Method"], [2643, 2645, ["bilinear", "interpolation"], "Method"], [2660, 2661, ["conv4_3"], "Method"], [2751, 2754, ["G", "-", "ResNet"], "Method"], [2775, 2780, ["graph", "based", "convolutional", "neural", "network"], "Method"], [2814, 2815, ["convolution"], "Method"], [2830, 2832, ["information", "exchanging"], "Task"], [2843, 2845, ["2D", "CNN"], "Method"], [2855, 2857, ["deep", "network"], "Method"], [2898, 2902, ["graph", "residual", "convolutional", "layers"], "Method"], [2907, 2913, ["serial", "of", "G", "-", "ResNet", "block"], "Method"], [2932, 2933, ["branch"], "Method"], [2937, 2940, ["graph", "convolutional", "layer"], "Method"], [2956, 2959, ["Graph", "unpooling", "layer"], "Method"], [2962, 2964, ["unpooling", "layer"], "Method"], [2973, 2974, ["GCNN"], "Method"], [2995, 2997, ["memory", "costs"], "Metric"], [3035, 3038, ["Face", "-", "based"], "Method"], [3058, 3061, ["vertex", "adding", "strategy"], "Method"], [3063, 3066, ["mesh", "subdivision", "algorithm"], "Method"], [3068, 3070, ["computer", "graphics"], "Task"], [3191, 3194, ["Edge", "-", "based"], "Method"], [3214, 3216, ["deformation", "procedure"], "Method"], [3246, 3248, ["laplacian", "regularization"], "Method"], [3260, 3263, ["edge", "length", "regularization"], "Method"], [3320, 3322, ["Chamfer", "loss"], "Method"], [3378, 3380, ["Normal", "loss"], "Method"], [3507, 3509, ["normal", "loss"], "Method"], [3519, 3520, ["Regularization"], "Method"], [3530, 3531, ["optimization"], "Task"], [3586, 3588, ["Laplacian", "regularization"], "Method"], [3597, 3599, ["Laplacian", "term"], "Method"], [3611, 3615, ["mesh", "self", "-", "intersection"], "Task"], [3617, 3619, ["laplaician", "term"], "Method"], [3622, 3626, ["local", "detail", "preserving", "operator"], "Method"], [3672, 3675, ["3D", "mesh", "model"], "Method"], [3691, 3693, ["mesh", "model"], "Method"], [3735, 3738, ["Edge", "length", "regularization"], "Method"], [3759, 3761, ["overall", "loss"], "Metric"], [3810, 3813, ["3D", "shape", "generation"], "Method"], [3896, 3897, ["ShapeNet"], "Material"], [3903, 3906, ["3D", "CAD", "models"], "Method"], [3955, 3957, ["Evaluation", "Metric"], "Metric"], [3980, 3981, ["precision"], "Metric"], [3982, 3983, ["recall"], "Metric"], [4008, 4011, ["F", "-", "score"], "Metric"], [4013, 4019, ["harmonic", "mean", "of", "precision", "and", "recall"], "Metric"], [4037, 4038, ["CD"], "Method"], [4045, 4046, ["EMD"], "Method"], [4049, 4052, ["F", "-", "Score"], "Metric"], [4058, 4059, ["CD"], "Method"], [4060, 4061, ["EMD"], "Method"], [4077, 4079, ["evaluation", "metrics"], "Metric"], [4080, 4082, ["shape", "generation"], "Task"], [4087, 4089, ["shape", "quality"], "Metric"], [4119, 4121, ["evaluation", "metric"], "Metric"], [4198, 4199, ["PSG"], "Method"], [4216, 4217, ["PSG"], "Method"], [4244, 4247, ["Marching", "Cube", "method"], "Method"], [4252, 4256, ["Neural", "3D", "Mesh", "Renderer"], "Method"], [4257, 4258, ["N3MR"], "Method"], [4265, 4271, ["deep", "learning", "based", "mesh", "generation", "model"], "Method"], [4324, 4325, ["Tensorflow"], "Method"], [4328, 4329, ["Adam"], "Method"], [4330, 4332, ["weight", "decay"], "Method"], [4352, 4354, ["learning", "rate"], "Metric"], [4372, 4374, ["training", "time"], "Metric"], [4379, 4382, ["Nvidia", "Titan", "X."], "Material"], [4412, 4415, ["F", "-", "score"], "Metric"], [4457, 4461, ["F", "-", "score", "improvement"], "Metric"], [4462, 4463, ["N3MR"], "Method"], [4507, 4508, ["EMD"], "Method"], [4531, 4533, ["mean", "score"], "Metric"], [4538, 4539, ["PSG"], "Method"], [4558, 4559, ["EMD"], "Method"], [4568, 4570, ["mesh", "model"], "Method"], [4572, 4573, ["regularization"], "Method"], [4607, 4609, ["mesh", "model"], "Method"], [4672, 4675, ["octree", "based", "solution"], "Method"], [4697, 4698, ["PSG"], "Method"], [4699, 4703, ["sparse", "3D", "point", "clouds"], "Task"], [4722, 4724, ["Chamfer", "loss"], "Method"], [4727, 4729, ["regression", "loss"], "Method"], [4741, 4742, ["N3MR"], "Method"], [4753, 4755, ["rendering", "tasks"], "Task"], [4782, 4784, ["mesh", "representation"], "Method"], [4824, 4826, ["Ablation", "Study"], "Task"], [4859, 4861, ["full", "model"], "Method"], [4870, 4872, ["evaluation", "metrics"], "Metric"], [4891, 4894, ["edge", "length", "regularization"], "Method"], [4947, 4949, ["Graph", "Unpooling"], "Method"], [4974, 4976, ["full", "model"], "Method"], [5013, 5016, ["G", "-", "ResNet"], "Method"], [5053, 5055, ["measurement", "metrics"], "Metric"], [5061, 5064, ["optimizing", "Chamfer", "distance"], "Metric"], [5070, 5072, ["degradation", "problem"], "Task"], [5076, 5081, ["deep", "2D", "convolutional", "neural", "network"], "Method"], [5088, 5090, ["training", "error"], "Metric"], [5094, 5096, ["testing", "error"], "Metric"], [5104, 5106, ["deep", "model"], "Method"], [5113, 5116, ["graph", "convolutional", "layers"], "Method"], [5128, 5132, ["deep", "graph", "neural", "network"], "Method"], [5260, 5264, ["mean", "F", "-", "score"], "Metric"], [5267, 5268, ["CD"], "Metric"], [5331, 5332, ["accuracy"], "Metric"], [5333, 5335, ["evaluation", "metrics"], "Metric"], [5390, 5395, ["Reconstructing", "Real", "-", "World", "images"], "Task"], [5408, 5411, ["Online", "Products", "dataset"], "Material"], [5412, 5414, ["Internet", "images"], "Material"], [5427, 5429, ["ShapeNet", "dataset"], "Material"], [5481, 5484, ["3D", "triangular", "meshes"], "Task"], [5494, 5496, ["mesh", "presentation"], "Task"], [5539, 5541, ["network", "structure"], "Method"], [5545, 5551, ["deep", "cascaded", "graph", "convolutional", "neural", "network"], "Method"], [5553, 5556, ["shortcut", "”", "connections"], "Method"], [5572, 5574, ["chamfer", "loss"], "Method"], [5575, 5577, ["normal", "loss"], "Method"], [5595, 5597, ["shape", "representations"], "Method"], [5610, 5612, ["mesh", "representation"], "Method"], [5684, 5687, ["scene", "level", "reconstruction"], "Task"], [5694, 5698, ["multi", "-", "view", "reconstruction"], "Task"], [968, 969, ["GCN"], "Method"], [1486, 1487, ["reconstruction"], "Task"], [3817, 3818, ["reconstruction"], "Task"], [3962, 3964, ["3D", "reconstruction"], "Task"], [4157, 4160, ["single", "image", "reconstruction"], "Task"], [5031, 5032, ["GCN"], "Method"], [5645, 5646, ["reconstruction"], "Task"]]} {"doc_id": "3acc07f7f8951617276cf99483ed02aeb0a6eeac", "text": ["document", ":", "Curriculum", "Domain", "Adaptation", "for", "Semantic", "Segmentation", "of", "Urban", "Scenes", "During", "the", "last", "half", "decade", ",", "convolutional", "neural", "networks", "(", "CNNs", ")", "have", "triumphed", "over", "semantic", "segmentation", ",", "which", "is", "a", "core", "task", "of", "various", "emerging", "industrial", "applications", "such", "as", "autonomous", "driving", "and", "medical", "imaging", ".", "However", ",", "to", "train", "CNNs", "requires", "a", "huge", "amount", "of", "data", ",", "which", "is", "difficult", "to", "collect", "and", "laborious", "to", "annotate", ".", "Recent", "advances", "in", "computer", "graphics", "make", "it", "possible", "to", "train", "CNN", "models", "on", "photo", "-", "realistic", "synthetic", "data", "with", "computer", "-", "generated", "annotations", ".", "Despite", "this", ",", "the", "domain", "mismatch", "between", "the", "real", "images", "and", "the", "synthetic", "data", "significantly", "decreases", "the", "models", "’", "performance", ".", "Hence", "we", "propose", "a", "curriculum", "-", "style", "learning", "approach", "to", "minimize", "the", "domain", "gap", "in", "semantic", "segmentation", ".", "The", "curriculum", "domain", "adaptation", "solves", "easy", "tasks", "first", "in", "order", "to", "infer", "some", "necessary", "properties", "about", "the", "target", "domain", ";", "in", "particular", ",", "the", "first", "task", "is", "to", "learn", "global", "label", "distributions", "over", "images", "and", "local", "distributions", "over", "landmark", "superpixels", ".", "These", "are", "easy", "to", "estimate", "because", "images", "of", "urban", "traffic", "scenes", "have", "strong", "idiosyncrasies", "(", "e.g.", ",", "the", "size", "and", "spatial", "relations", "of", "buildings", ",", "streets", ",", "cars", ",", "etc", ".", ")", ".", "We", "then", "train", "the", "segmentation", "network", "in", "such", "a", "way", "that", "the", "network", "predictions", "in", "the", "target", "domain", "follow", "those", "inferred", "properties", ".", "In", "experiments", ",", "our", "method", "significantly", "outperforms", "the", "baselines", "as", "well", "as", "the", "only", "known", "existing", "approach", "to", "the", "same", "problem", ".", "1", "]", "Yang", "Zhang", "2", "]", "Philip", "David", "1", "]", "Boqing", "Gong", "[", "1", "]", "Center", "for", "Research", "in", "Computer", "Vision", ",", "University", "of", "Central", "Florida", "[", "2", "]", "Computational", "and", "Information", "Sciences", "Directorate", ",", "U.S.", "Army", "Research", "Laboratory", "yangzhang@knights.ucf.edu", ",", "philip.j.david4.civ@mail.mil", ",", "bgong@crcv.ucf.edu", "section", ":", "Introduction", "This", "paper", "is", "concerned", "with", "domain", "adaptation", "for", "semantic", "image", "segmentation", "of", "urban", "scenes", ",", "i.e.", ",", "assigning", "a", "category", "label", "to", "every", "pixel", "of", "an", "image", "or", "video", "frame", ".", "Our", "interest", "in", "this", "problem", "is", "partially", "due", "to", "the", "exciting", "vision", "of", "autonomous", "driving", ",", "where", "understanding", "complex", "inner", "-", "city", "traffic", "scenes", "is", "an", "essential", "module", "and", "semantic", "segmentation", "is", "one", "of", "its", "key", "constituents", ".", "Machine", "learning", "methods", "for", "automatic", "semantic", "segmentation", "require", "massive", "amounts", "of", "high", "-", "quality", "annotated", "imagery", "in", "order", "to", "produce", "effective", "classifiers", "that", "generalize", "well", "to", "novel", "scenes", ".", "However", ",", "annotating", "training", "imagery", "for", "semantic", "segmentation", "is", "a", "very", "cumbersome", "task", "for", "humans", ".", "Cordts", "et", "al", ".", "report", "that", "the", "annotation", "and", "quality", "control", "take", "more", "than", "1.5", "hours", "on", "a", "single", "image", "of", "the", "Cityscapes", "dataset", ".", "Besides", ",", "it", "is", "very", "difficult", "and", "time", "-", "consuming", "to", "collect", "imagery", "that", "depicts", "the", "large", "number", "of", "variabilities", "possible", "of", "urban", "scenes", "in", "different", "countries", ",", "seasons", ",", "and", "lighting", "conditions", ",", "etc", ".", "To", "overcome", "both", "shortcomings", ",", "simulated", "urban", "environments", "may", "be", "used", "to", "automatically", "generate", "large", "amounts", "of", "annotated", "training", "imagery", ".", "This", ",", "however", ",", "introduces", "a", "new", "problem", ",", "that", "of", "domain", "mismatch", "between", "the", "source", "(", "simulated", ")", "domain", "and", "the", "target", "(", "real", ")", "domain", ".", "Figure", "[", "reference", "]", "illustrates", "some", "examples", "drawn", "from", "the", "synthetic", "SYNTHIA", "dataset", "and", "the", "real", "Cityscapes", "dataset", ".", "It", "is", "readily", "apparent", "that", "there", "are", "significant", "visual", "differences", "between", "the", "two", "datasets", ".", "Domain", "adaptation", "techniques", "may", "be", "used", "by", "machine", "learning", "methods", "to", "bridge", "this", "gap", "between", "the", "two", "domains", ".", "In", "computer", "vision", ",", "learning", "domain", "-", "invariant", "features", "has", "been", "a", "prevalent", "and", "successful", "strategy", "to", "tackle", "the", "discrepancy", "between", "two", "domains", ",", "mainly", "for", "classification", "and", "regression", "problems", ".", "The", "core", "idea", "is", "to", "infer", "a", "new", "feature", "space", "such", "that", "the", "marginal", "distributions", "of", "the", "source", "domain", "(", "S", ")", "and", "the", "target", "domain", "(", "T", ")", "are", "about", "the", "same", ",", "i.e.", ",", ".", "Furthermore", ",", "the", "prediction", "function", "from", "that", "space", "is", "assumed", "to", "be", "the", "same", "across", "the", "domains", "so", "that", "one", "can", "leverage", "the", "rich", "labeled", "data", "in", "the", "source", "domain", "to", "train", "classifiers", "that", "generalize", "well", "to", "the", "target", ".", "It", "is", "hard", "to", "verify", "the", "assumption", ",", "but", "the", "work", "along", "this", "line", "is", "rich", "and", "has", "led", "to", "impressive", "practical", "results", "regardless", ",", "such", "as", "the", "algorithms", "using", "linear", "transformation", ",", "kernel", "methods", ",", "and", "the", "recent", "deep", "learning", "methods", "that", "directly", "extract", "domain", "-", "invariant", "features", "from", "raw", "input", "images", ".", "In", "contrast", "to", "prior", "arts", ",", "the", "semantic", "segmentation", "we", "study", "in", "this", "paper", "is", "a", "highly", "structured", "prediction", "problem", ",", "for", "which", "domain", "adaptation", "is", "only", "sparsely", "explored", "in", "the", "literature", ".", "Under", "structured", "prediction", ",", "can", "we", "still", "achieve", "good", "domain", "adaptation", "results", "by", "following", "the", "above", "principles", "?", "Our", "intuition", "and", "experimental", "studies", "(", "cf", ".", "Section", "[", "reference", "]", ")", "tell", "us", "no", ".", "Learning", "a", "decision", "function", "for", "structured", "prediction", "is", "more", "involved", "than", "classification", "because", "it", "has", "to", "resolve", "the", "predictions", "in", "an", "exponentially", "large", "label", "space", ".", "As", "a", "result", ",", "the", "assumption", "that", "the", "source", "and", "target", "domains", "share", "the", "same", "prediction", "function", "becomes", "less", "likely", "to", "hold", ".", "Besides", ",", "some", "discriminative", "cues", "in", "the", "data", "would", "be", "suppressed", "if", "one", "matches", "the", "feature", "representations", "of", "the", "two", "domains", "without", "taking", "careful", "account", "of", "the", "structured", "labels", ".", "Finally", ",", "data", "instances", "are", "the", "proxy", "to", "measure", "the", "domain", "difference", ".", "However", ",", "it", "is", "not", "immediately", "clear", "what", "comprises", "the", "instances", "in", "semantic", "segmentation", ",", "especially", "given", "that", "the", "top", "-", "performing", "segmentation", "methods", "are", "built", "upon", "deep", "neural", "networks", ".", "Hoffman", "et", "al", ".", "take", "each", "spatial", "unit", "in", "the", "fully", "convolutional", "network", "(", "FCN", ")", "as", "an", "instance", ".", "We", "contend", "that", "such", "instances", "are", "actually", "non", "-", "i.i.d", ".", "in", "either", "individual", "domain", ",", "as", "their", "receptive", "fields", "overlap", "with", "each", "other", ".", "How", "can", "we", "avoid", "the", "assumption", "that", "the", "source", "and", "target", "domains", "share", "the", "same", "prediction", "function", "in", "a", "transformed", "domain", "-", "invariant", "feature", "space", "?", "Our", "proposed", "solution", "draws", "on", "two", "key", "observations", ".", "One", "is", "that", "the", "urban", "traffic", "scene", "images", "have", "strong", "idiosyncrasies", "(", "e.g.", ",", "the", "size", "and", "spatial", "relations", "of", "buildings", ",", "streets", ",", "cars", ",", "etc", ".", ")", ".", "Therefore", ",", "some", "tasks", "are", "“", "easy", "”", "and", ",", "more", "importantly", ",", "suffer", "less", "because", "of", "the", "domain", "discrepancy", ".", "Second", ",", "the", "structured", "output", "in", "semantic", "segmentation", "enables", "convenient", "posterior", "regularization", ",", "as", "opposed", "to", "the", "popular", "(", "e.g.", ",", ")", "regularization", "over", "model", "parameters", ".", "Accordingly", ",", "we", "propose", "a", "curriculum", "-", "style", "domain", "adaptation", "approach", ".", "Recall", "that", ",", "in", "domain", "adaptation", ",", "only", "the", "source", "domain", "supplies", "many", "labeled", "data", "while", "there", "are", "no", "or", "only", "scarce", "labels", "from", "the", "target", ".", "The", "curriculum", "domain", "adaptation", "begins", "with", "the", "easy", "tasks", ",", "in", "order", "to", "gain", "some", "high", "-", "level", "properties", "about", "the", "unknown", "pixel", "-", "level", "labels", "for", "each", "target", "image", ".", "It", "then", "learns", "a", "semantic", "segmentation", "network", "—", "the", "hard", "task", ",", "whose", "predictions", "over", "the", "target", "images", "are", "forced", "to", "follow", "those", "necessary", "properties", "as", "much", "as", "possible", ".", "To", "develop", "the", "easy", "tasks", "in", "the", "curriculum", ",", "we", "consider", "label", "distributions", "over", "both", "holistic", "images", "and", "some", "landmark", "superpixels", "of", "the", "target", "domain", ".", "Take", "the", "former", "for", "instance", ".", "The", "label", "distribution", "of", "an", "image", "indicates", "the", "percentage", "of", "pixels", "that", "belong", "to", "each", "category", ",", "respectively", ".", "We", "argue", "that", "such", "tasks", "are", "easier", ",", "despite", "the", "domain", "mismatch", ",", "than", "assigning", "pixel", "-", "wise", "labels", ".", "Indeed", ",", "we", "may", "directly", "estimate", "the", "label", "distributions", "without", "inferring", "the", "pixel", "-", "wise", "labels", ".", "Moreover", ",", "the", "relative", "sizes", "of", "road", ",", "vehicle", ",", "pedestrian", ",", "etc", ".", "constrain", "the", "shape", "of", "the", "distributions", ",", "effectively", "reducing", "the", "search", "space", ".", "Finally", ",", "models", "to", "estimate", "the", "label", "distributions", "over", "superpixels", "may", "benefit", "from", "the", "urban", "scenes", "’", "canonical", "layout", "that", "transcends", "domains", ",", "e.g.", ",", "buildings", "stand", "beside", "streets", ".", "Why", "and", "when", "are", "the", "seemingly", "simple", "label", "distributions", "useful", "for", "the", "domain", "adaptation", "of", "semantic", "segmentation", "?", "In", "our", "experiments", ",", "we", "find", "that", "the", "segmentation", "networks", "trained", "on", "the", "source", "domain", "perform", "poorly", "on", "many", "target", "images", ",", "giving", "rise", "to", "disproportionate", "label", "assignments", "(", "e.g.", ",", "many", "more", "pixels", "are", "classified", "to", "sidewalks", "than", "to", "streets", ")", ".", "To", "rectify", "this", ",", "the", "image", "-", "level", "label", "distribution", "informs", "the", "segmentation", "network", "how", "to", "update", "the", "predictions", "while", "the", "label", "distributions", "of", "the", "landmark", "superpixels", "tell", "the", "network", "where", "to", "update", ".", "Jointly", ",", "they", "guide", "the", "adaptation", "of", "the", "networks", "to", "the", "target", "domain", "to", ",", "at", "least", ",", "generate", "proportional", "label", "predictions", ".", "Note", "that", "additional", "“", "easy", "tasks", "”", "can", "be", "conveniently", "incorporated", "into", "our", "framework", "in", "the", "future", ".", "Our", "main", "contribution", "is", "on", "the", "proposed", "curriculum", "-", "style", "domain", "adaptation", "for", "the", "semantic", "segmentation", "of", "urban", "scenes", ".", "We", "select", "into", "the", "curriculum", "the", "easy", "and", "useful", "tasks", "of", "inferring", "label", "distributions", "for", "the", "target", "images", "and", "landmark", "superpixels", ",", "in", "order", "to", "gain", "some", "necessary", "properties", "about", "the", "target", "domain", ".", "Built", "upon", "these", ",", "we", "learn", "a", "pixel", "-", "wise", "discriminative", "segmentation", "network", "from", "the", "labeled", "source", "data", "and", ",", "meanwhile", ",", "conduct", "a", "“", "sanity", "check", "”", "to", "ensure", "the", "network", "behavior", "is", "consistent", "with", "the", "previously", "learned", "knowledge", "about", "the", "target", "domain", ".", "Our", "approach", "effectively", "eludes", "the", "assumption", "about", "the", "existence", "of", "a", "common", "prediction", "function", "for", "both", "domains", "in", "a", "transformed", "feature", "space", ".", "It", "readily", "applies", "to", "different", "segmentation", "networks", ",", "as", "it", "does", "not", "change", "the", "network", "architecture", "or", "tax", "any", "intermediate", "layers", ".", "section", ":", "Related", "work", "We", "discuss", "some", "related", "work", "on", "domain", "adaptation", "and", "semantic", "segmentation", ",", "with", "special", "focus", "on", "that", "transferring", "knowledge", "from", "virtual", "images", "to", "real", "photos", ".", "paragraph", ":", "Domain", "adaptation", ".", "Conventional", "machine", "learning", "algorithms", "rely", "on", "the", "assumption", "that", "the", "training", "and", "test", "data", "are", "drawn", "i.i.d", ".", "from", "the", "same", "underlying", "distribution", ".", "However", ",", "it", "is", "often", "the", "case", "that", "there", "exists", "some", "discrepancy", "from", "the", "training", "to", "the", "test", "stage", ".", "Domain", "adaptation", "aims", "to", "rectify", "this", "mismatch", "and", "tune", "the", "models", "toward", "better", "generalization", "at", "testing", ".", "The", "existing", "work", "on", "domain", "adaptation", "mostly", "focuses", "on", "classification", "and", "regression", "problems", ",", "e.g.", ",", "learning", "from", "online", "commercial", "images", "to", "classify", "real", "-", "world", "objects", ",", "and", ",", "more", "recently", ",", "aims", "to", "improve", "the", "adaptability", "of", "deep", "neural", "networks", ".", "Among", "them", ",", "the", "most", "relevant", "work", "to", "ours", "is", "that", "exploring", "simulated", "data", ".", "Sun", "and", "Saenko", "train", "generic", "object", "detectors", "from", "the", "synthetic", "images", ",", "while", "Vazquez", "et", "al", ".", "use", "the", "virtual", "images", "to", "improve", "pedestrian", "detections", "in", "real", "environment", ".", "The", "other", "way", "around", ",", "i.e.", ",", "how", "to", "improve", "the", "quality", "of", "the", "simulated", "images", "using", "the", "real", "ones", ",", "is", "studied", "in", ".", "paragraph", ":", "Semantic", "segmentation", ".", "Semantic", "segmentation", "is", "the", "task", "of", "assigning", "an", "object", "label", "to", "each", "pixel", "of", "an", "image", ".", "Traditional", "methods", "rely", "on", "local", "image", "features", "manually", "designed", "by", "domain", "experts", ".", "After", "the", "pioneering", "work", "that", "introduced", "the", "convolutional", "neural", "network", "(", "CNN", ")", "to", "semantic", "segmentation", ",", "most", "recent", "top", "-", "performing", "methods", "are", "built", "on", "CNNs", ".", "An", "enormous", "amount", "of", "labor", "-", "intensive", "work", "is", "required", "to", "annotate", "the", "many", "images", "that", "are", "needed", "to", "obtain", "accurate", "segmentation", "models", ".", "The", "PASCAL", "VOC2012", "Challenge", "contains", "nearly", "10", ",", "000", "annotated", "images", "for", "the", "segmentation", "competition", ",", "and", "the", "MS", "COCO", "Challenge", "includes", "over", "200", ",", "000", "annotated", "images", ".", "According", "to", ",", "it", "took", "about", "60", "minutes", "to", "manually", "segment", "each", "image", "in", "and", "about", "90", "minutes", "for", "each", "in", ".", "A", "plausible", "approach", "to", "reducing", "the", "human", "workload", "is", "to", "utilize", "weakly", "supervised", "information", "such", "as", "image", "labels", "and", "bounding", "boxes", ".", "We", "instead", "explore", "the", "use", "of", "almost", "effortlessly", "labeled", "virtual", "images", "for", "training", "high", "-", "quality", "segmentation", "networks", ".", "In", ",", "annotating", "a", "synthetic", "image", "took", "only", "7", "seconds", "on", "average", "through", "a", "computer", "game", ".", "For", "the", "urban", "scenes", ",", "we", "use", "the", "SYNTHIA", "dataset", "which", "contains", "images", "of", "a", "virtual", "city", ".", "paragraph", ":", "Domain", "adaptation", "for", "semantic", "segmentation", ".", "Upon", "observing", "the", "obvious", "mismatch", "between", "virtual", "and", "real", "data", ",", "we", "expect", "domain", "adaptation", "to", "enhance", "the", "segmentation", "performance", "on", "real", "images", "by", "networks", "trained", "on", "virtual", "ones", ".", "To", "the", "best", "of", "our", "knowledge", ",", "the", "only", "attempt", "to", "algorithmically", "address", "this", "problem", "is", ".", "While", "it", "regularizes", "the", "intermediate", "layers", "and", "constrains", "the", "output", "of", "the", "network", ",", "we", "propose", "a", "different", "curriculum", "domain", "adaptation", "strategy", ".", "We", "solve", "the", "easy", "task", "first", "and", "then", "use", "the", "learned", "knowledge", "about", "the", "target", "domain", "to", "regularize", "the", "network", "predictions", ".", "section", ":", "Approach", "In", "this", "section", ",", "we", "present", "the", "details", "of", "the", "proposed", "curriculum", "domain", "adaptation", "for", "semantic", "segmentation", "of", "urban", "scene", "images", ".", "Unlike", "previous", "work", "that", "aligns", "the", "domains", "via", "an", "intermediate", "feature", "space", "and", "thereby", "implicitly", "assumes", "the", "existence", "of", "the", "same", "decision", "function", "for", "the", "two", "domains", ",", "it", "is", "our", "intuition", "that", ",", "for", "structured", "prediction", "(", "i.e.", ",", "semantic", "segmentation", "here", ")", ",", "the", "cross", "-", "domain", "generalization", "of", "machine", "learning", "models", "can", "be", "more", "efficiently", "improved", "if", "we", "avoid", "this", "assumption", "and", "instead", "train", "them", "subject", "to", "necessary", "properties", "they", "should", "retain", "in", "the", "target", "domain", ".", "paragraph", ":", "Preliminaries", ".", "In", "particular", ",", "the", "properties", "are", "about", "the", "pixel", "-", "wise", "category", "labels", "of", "an", "arbitrary", "image", "from", "the", "target", "domain", ",", "where", "and", "are", "the", "width", "and", "height", "of", "the", "image", ",", "respectively", ",", "and", "is", "the", "number", "of", "categories", ".", "We", "use", "one", "-", "hot", "vector", "encoding", "for", "the", "groundtruth", "labels", ",", "i.e.", ",", "takes", "the", "value", "of", "0", "or", "1", "and", "the", "latter", "means", "that", "the", "-", "th", "label", "is", "assigned", "by", "a", "human", "annotator", "to", "the", "pixel", "at", ".", "Correspondingly", ",", "the", "prediction", "by", "a", "segmentation", "network", "is", "realized", "by", "a", "softmax", "function", "per", "pixel", ".", "We", "express", "each", "target", "property", "in", "the", "form", "of", "a", "distribution", "over", "the", "categories", ",", "where", "represents", "the", "occupancy", "proportion", "of", "the", "category", "over", "the", "-", "th", "target", "image", "or", "a", "superpixel", "of", "the", "image", ".", "Therefore", ",", "one", "can", "immediately", "calculate", "the", "distribution", "given", "the", "human", "annotations", "to", "the", "image", ".", "For", "instance", ",", "the", "image", "level", "label", "distribution", "is", "expressed", "by", "Similarly", ",", "we", "can", "compute", "the", "target", "property", "/", "distribution", "from", "the", "network", "predictions", "and", "denote", "it", "by", ".", "subsection", ":", "Domain", "adaptation", "using", "the", "target", "properties", "Ideally", ",", "we", "would", "like", "to", "have", "a", "segmentation", "network", "to", "imitate", "human", "annotators", "on", "the", "target", "domain", ".", "Therefore", ",", "necessarily", ",", "the", "properties", "of", "their", "annotation", "results", "should", "be", "the", "same", "too", ".", "We", "capture", "this", "notion", "by", "minimizing", "the", "cross", "entropy", "at", "training", ",", "where", "the", "first", "term", "of", "the", "right", "-", "hand", "side", "is", "the", "entropy", "and", "the", "second", "is", "the", "KL", "-", "divergence", ".", "Given", "a", "mini", "-", "batch", "consisting", "of", "both", "source", "images", "(", ")", "and", "target", "images", "(", ")", ",", "the", "overall", "objective", "function", "for", "training", "the", "cross", "-", "domain", "generalizing", "segmentation", "network", "is", ",", "where", "is", "the", "pixel", "-", "wise", "cross", "-", "entropy", "loss", "defined", "over", "the", "sufficiently", "labeled", "source", "domain", "images", ",", "enforcing", "the", "network", "to", "have", "the", "pixel", "level", "discriminative", "capabilities", ",", "and", "the", "second", "term", "is", "over", "the", "unlabeled", "target", "domain", "images", ",", "hinting", "the", "network", "what", "necessary", "properties", "its", "predictions", "should", "have", "in", "the", "target", "domain", ".", "We", "use", "to", "balance", "the", "two", "strengths", "in", "training", "and", "superscript", "to", "index", "different", "types", "of", "label", "distributions", ".", "Note", "that", "in", "the", "domain", "adaptation", "context", ",", "we", "actually", "can", "not", "directly", "compute", "the", "label", "distribution", "from", "the", "groundtruth", "annotations", "of", "the", "target", "domain", ".", "Nonetheless", ",", "estimating", "them", "using", "the", "labeled", "source", "data", "is", "easier", "than", "assigning", "labels", "to", "every", "single", "pixel", "of", "the", "target", "images", ".", "We", "present", "the", "details", "in", "the", "next", "section", ".", "paragraph", ":", "Remarks", ".", "Mathematically", ",", "the", "objective", "function", "has", "a", "similar", "form", "as", "in", "model", "compression", ".", "We", "thus", "borrow", "some", "concepts", "to", "gain", "more", "intuitive", "understanding", "about", "our", "domain", "adaptation", "procedure", ".", "The", "“", "student", "”", "network", "follows", "a", "curriculum", "to", "learn", "simple", "knowledge", "about", "the", "target", "domain", "before", "it", "addresses", "the", "hard", "one", "of", "semantically", "segmenting", "images", ".", "The", "models", "inferring", "the", "target", "properties", "act", "like", "“", "teachers", "”", ",", "as", "they", "hint", "what", "label", "distributions", "the", "final", "solution", "(", "image", "annotation", ")", "may", "have", "in", "the", "target", "domain", "at", "the", "image", "and", "superpixel", "levels", ".", "Another", "perspective", "is", "to", "understand", "the", "target", "properties", "as", "a", "posterior", "regularization", "for", "the", "network", ".", "The", "posterior", "regularization", "can", "conveniently", "encode", "a", "priori", "knowledge", "into", "the", "objective", "function", ".", "Some", "applications", "include", "weakly", "supervised", "segmentation", "and", "detection", ",", "and", "rule", "-", "regularized", "training", "of", "neural", "networks", ".", "In", "addition", "to", "the", "domain", "adaptation", "setting", "and", "novel", "target", "properties", ",", "another", "key", "distinction", "of", "our", "work", "is", "that", "we", "decouple", "the", "label", "distributions", "from", "the", "network", "predictions", "and", "thus", "avoid", "the", "EM", "type", "of", "optimizations", ".", "Our", "approach", "learns", "the", "segmentation", "network", "with", "almost", "effortless", "changes", "to", "the", "popular", "deep", "learning", "tools", ".", "subsection", ":", "Inferring", "the", "target", "properties", "Thus", "far", "we", "have", "presented", "the", "“", "hard", "”", "task", "in", "the", "curriculum", "domain", "adaptation", ".", "In", "this", "section", ",", "we", "describe", "the", "“", "easy", "”", "ones", ",", "i.e.", ",", "how", "to", "infer", "the", "target", "properties", "without", "accessing", "the", "image", "annotations", "of", "the", "target", "domain", ".", "Our", "contributions", "also", "include", "selecting", "the", "particular", "property", "of", "label", "distributions", "to", "constitute", "the", "simple", "tasks", ".", "subsubsection", ":", "Global", "label", "distributions", "of", "images", "Due", "to", "the", "domain", "disparity", ",", "a", "baseline", "segmentation", "network", "trained", "on", "the", "source", "domain", "(", "i.e.", ",", "using", "the", "first", "term", "of", "eq", ".", "(", "[", "reference", "]", ")", ")", "could", "be", "easily", "crippled", "given", "the", "target", "images", ".", "In", "our", "experiments", ",", "we", "find", "that", "our", "baseline", "network", "constantly", "mistakes", "streets", "for", "sidewalks", "and", "/", "or", "cars", "(", "cf", ".", "Figure", "[", "reference", "]", ")", ".", "Consequently", ",", "the", "predicted", "labels", "for", "the", "pixels", "are", "highly", "disproportionate", ".", "To", "rectify", "this", ",", "we", "employ", "the", "label", "distribution", "over", "the", "global", "image", "as", "our", "first", "property", "(", "cf", ".", "eq", ".", "(", "[", "reference", "]", ")", ")", ".", "Without", "access", "to", "the", "target", "labels", ",", "we", "have", "to", "train", "machine", "learning", "models", "from", "the", "labeled", "source", "images", "to", "estimate", "the", "label", "distribution", "for", "the", "target", "image", ".", "Nonetheless", ",", "we", "argue", "that", "this", "is", "less", "challenging", "than", "generating", "the", "per", "-", "pixel", "predictions", "despite", "that", "both", "tasks", "are", "influenced", "by", "the", "domain", "mismatch", ".", "In", "our", "experiments", ",", "we", "examine", "different", "approaches", "to", "this", "task", ".", "We", "extract", "image", "features", "using", "the", "Inception", "-", "Resnet", "-", "v2", "as", "the", "input", "to", "the", "following", "models", ".", "Although", "multinomial", "logistic", "regression", "(", "LR", ")", "is", "mainly", "used", "for", "classification", ",", "its", "output", "is", "actually", "a", "valid", "distribution", "over", "the", "categories", ".", "For", "our", "purpose", ",", "we", "thus", "train", "it", "by", "replacing", "the", "one", "-", "hot", "vectors", "in", "the", "cross", "-", "entropy", "loss", "with", "the", "groundtruth", "label", "distribution", ",", "which", "is", "calculated", "using", "eq", ".", "(", "[", "reference", "]", ")", "and", "the", "available", "human", "labels", "of", "the", "source", "domain", ".", "Given", "a", "target", "image", ",", "we", "directly", "take", "the", "LR", "’s", "output", "as", "the", "predicted", "label", "distribution", ".", "We", "also", "test", "a", "nonparametric", "method", "by", "simply", "retrieving", "the", "nearest", "neighbors", "(", "NNs", ")", "for", "a", "target", "image", "and", "then", "transferring", "the", "mean", "of", "the", "NNs", "’", "label", "distributions", "to", "the", "target", "image", ".", "We", "use", "the", "distance", "for", "the", "NN", "retrieval", ".", "Finally", ",", "we", "include", "two", "dumb", "predictions", "as", "the", "control", "experiment", ".", "One", "is", ",", "for", "any", "target", "image", ",", "to", "output", "the", "mean", "of", "all", "the", "label", "distributions", "in", "the", "source", "domain", "(", "source", "mean", ")", ",", "and", "the", "other", "is", "to", "output", "a", "uniform", "distribution", ".", "subsubsection", ":", "Local", "label", "distributions", "of", "landmark", "superpixels", "The", "image", "level", "label", "distribution", "globally", "penalizes", "potentially", "disproportional", "segmentation", "output", "on", "the", "target", "domain", ",", "and", "yet", "is", "inadequate", "in", "providing", "spatial", "constraints", ".", "In", "this", "section", ",", "we", "consider", "the", "use", "of", "label", "distributions", "over", "some", "superpixels", "as", "the", "anchors", "to", "drive", "the", "network", "towards", "spatially", "desired", "target", "properties", ".", "Note", "that", "it", "is", "not", "necessary", ",", "and", "is", "even", "harmful", ",", "to", "use", "all", "of", "the", "superpixels", "in", "a", "target", "image", "to", "regularize", "the", "segmentation", "network", ",", "because", "that", "would", "be", "too", "strong", "a", "force", "and", "may", "overrule", "the", "pixel", "-", "wise", "discriminativeness", "revealed", "by", "the", "labeled", "source", "images", ",", "especially", "when", "the", "label", "distributions", "are", "not", "inferred", "accurately", "enough", ".", "In", "order", "to", "have", "the", "dual", "effect", "of", "both", "estimating", "the", "label", "distributions", "of", "superpixels", "and", "filtering", "the", "superpixels", ",", "we", "simplify", "the", "problem", "and", "employ", "a", "linear", "SVM", "in", "this", "work", ".", "In", "particular", ",", "we", "segment", "each", "image", "into", "100", "superpixels", "using", "linear", "spectral", "clustering", ".", "For", "the", "superpixels", "of", "the", "source", "domain", ",", "we", "are", "able", "to", "assign", "a", "single", "dominant", "label", "to", "each", "of", "them", ",", "and", "then", "use", "the", "labels", "and", "the", "corresponding", "features", "extracted", "from", "the", "superpixels", "to", "train", "a", "multi", "-", "class", "SVM", ".", "Given", "a", "test", "superpixel", "of", "a", "target", "image", ",", "the", "multi", "-", "class", "SVM", "returns", "a", "class", "label", "as", "well", "as", "a", "decision", "value", ",", "which", "is", "interpreted", "as", "the", "confidence", "score", "about", "classifying", "this", "superpixel", ".", "We", "keep", "the", "top", "60", "%", "superpixels", ",", "called", "landmark", "superpixels", ",", "in", "the", "target", "domain", "and", "calculate", "their", "label", "distributions", "as", "the", "second", "type", "of", "“", "easy", "”", "tasks", ".", "In", "particular", ",", "the", "class", "label", "of", "a", "landmark", "superpixel", "is", "encoded", "into", "a", "one", "-", "hot", "vector", ",", "which", "serves", "as", "a", "valid", "distribution", "about", "the", "categories", "in", "the", "landmark", "superpixel", ".", "Albeit", "simple", ",", "we", "find", "this", "method", "works", "very", "well", "in", "our", "experiments", ".", "We", "encode", "both", "visual", "and", "contextual", "information", "to", "represent", "a", "superpixel", ".", "First", ",", "we", "use", "the", "FCN", "-", "8s", "pre", "-", "trained", "on", "the", "PASCAL", "CONTEXT", "dataset", ",", "which", "has", "59", "distinct", "classes", ",", "to", "obtain", "59", "detection", "scores", "for", "each", "pixel", ".", "We", "then", "average", "them", "within", "each", "superpixel", ".", "Finally", ",", "we", "represent", "a", "superpixel", "by", "the", "concatenation", "of", "the", "59D", "vectors", "of", "itself", ",", "its", "left", "and", "right", "superpixels", ",", "as", "well", "as", "the", "two", "respectively", "above", "and", "below", "it", ".", "subsection", ":", "Curriculum", "domain", "adaptation", ":", "recapitulation", "We", "recap", "the", "proposed", "curriculum", "domain", "adaptation", "using", "Figure", "[", "reference", "]", "before", "presenting", "the", "experiments", "in", "the", "next", "section", ".", "Our", "main", "idea", "is", "to", "execute", "the", "domain", "adaptation", "step", "by", "step", ",", "starting", "from", "the", "easy", "tasks", "that", "are", "less", "sensitive", "to", "the", "domain", "discrepancy", "than", "the", "semantic", "segmentation", ".", "We", "choose", "the", "labels", "distributions", "over", "global", "images", "and", "local", "landmark", "superpixels", "in", "this", "work", ";", "more", "tasks", "will", "be", "explored", "in", "the", "future", ".", "The", "solutions", "to", "them", "provide", "useful", "gradients", "originating", "from", "the", "target", "domain", "(", "cf", ".", "the", "arrows", "with", "brown", "color", "in", "Figure", "[", "reference", "]", ")", ",", "while", "the", "source", "domain", "feeds", "the", "network", "with", "well", "-", "labeled", "images", "and", "segmentation", "masks", "(", "cf", ".", "the", "dark", "blue", "arrows", "in", "Figure", "[", "reference", "]", ")", ".", "section", ":", "Experiments", "In", "this", "section", ",", "we", "describe", "the", "experimental", "setup", "and", "compare", "the", "results", "of", "our", "approach", ",", "its", "variations", ",", "and", "some", "existing", "baseline", "methods", ".", "subsection", ":", "Segmentation", "network", "and", "optimization", "In", "our", "experiments", ",", "we", "use", "FCN", "-", "8s", "as", "our", "semantic", "segmentation", "network", ".", "We", "initialize", "its", "convolutional", "layers", "with", "VGG", "-", "19", ",", "and", "then", "train", "it", "using", "the", "AdaDelta", "optimizer", "with", "default", "parameters", ".", "Each", "mini", "-", "batch", "is", "comprised", "of", "five", "source", "images", "and", "five", "randomly", "chosen", "target", "images", ".", "When", "we", "train", "the", "baseline", "network", "with", "no", "adaptation", ",", "however", ",", "we", "try", "to", "use", "the", "largest", "possible", "mini", "-", "batch", "that", "includes", "15", "source", "images", ".", "The", "network", "is", "implemented", "in", "Keras", "and", "Theano", ".", "We", "train", "different", "versions", "of", "the", "network", "on", "a", "single", "Tesla", "K40", "GPU", ".", "Unlike", "the", "existing", "deep", "domain", "adaptation", "methods", "which", "introduce", "regularization", "to", "the", "intermediate", "layers", ",", "we", "only", "revise", "the", "loss", "function", "over", "the", "output", ".", "Hence", ",", "our", "curriculum", "domain", "adaptation", "can", "be", "readily", "applied", "to", "other", "segmentation", "networks", "(", "e.g.", ",", ")", ".", "subsection", ":", "Datasets", "and", "evaluation", "We", "use", "the", "publicly", "available", "Cityscpaes", "and", "SYNTHIA", "datasets", "in", "our", "experiments", ".", "Cityscapes", "is", "a", "real", "-", "world", ",", "vehicle", "-", "egocentric", "image", "dataset", "collected", "in", "50", "cities", "in", "Germany", "and", "nearby", "countries", ".", "It", "provides", "four", "disjoint", "subsets", ":", "2", ",", "993", "training", "images", ",", "503", "validation", "image", ",", "1", ",", "531", "test", "images", ",", "and", "20", ",", "021", "auxiliary", "images", ".", "All", "the", "training", ",", "validation", ",", "and", "test", "images", "are", "accurately", "annotated", "with", "per", "pixel", "category", "labels", ",", "while", "the", "auxiliary", "set", "is", "coarsely", "labeled", ".", "There", "are", "34", "distinct", "categories", "in", "the", "dataset", ".", "SYNTHIA", "is", "a", "large", "dataset", "of", "synthetic", "images", "and", "provides", "a", "particular", "subset", ",", "called", "SYNTHIA", "-", "RAND", "-", "CITYSCAPES", ",", "to", "pair", "with", "Cityscapes", ".", "This", "subset", "contains", "9", ",", "400", "images", "that", "are", "automatically", "annotated", "with", "12", "object", "categories", ",", "one", "void", "class", ",", "and", "some", "unnamed", "classes", ".", "Note", "that", "the", "virtual", "city", "used", "to", "generate", "the", "synthetic", "images", "does", "not", "correspond", "to", "any", "of", "the", "real", "cities", "covered", "by", "Cityscapes", ".", "We", "abbreviate", "SYNTHIA", "-", "RAND", "-", "CITYSCAPES", "to", "SYNTHIA", "hereon", ".", "paragraph", ":", "Domain", "idiosyncrasies", ".", "Although", "both", "datasets", "depict", "urban", "scenes", ",", "and", "SYNTHIA", "is", "created", "to", "be", "as", "photo", "-", "realistic", "as", "possible", ",", "they", "are", "mismatched", "domains", "in", "several", "ways", ".", "The", "most", "noticeable", "difference", "is", "probably", "the", "coarse", "-", "grained", "textures", "in", "SYNTHIA", ";", "very", "similar", "texture", "patterns", "repeat", "in", "a", "regular", "manner", "across", "different", "images", ".", "In", "contrast", ",", "the", "Cityscapes", "images", "are", "captured", "by", "high", "-", "quality", "dash", "-", "cameras", ".", "Another", "major", "distinction", "is", "the", "variability", "in", "view", "angles", ".", "Since", "Cityscapes", "images", "are", "recorded", "by", "the", "dash", "cameras", "mounted", "on", "a", "moving", "car", ",", "they", "are", "viewed", "from", "almost", "a", "constant", "angle", "that", "is", "about", "parallel", "to", "the", "ground", ".", "More", "diverse", "view", "angles", "are", "employed", "by", "SYNTHIA", "—", "it", "seems", "like", "some", "cameras", "are", "placed", "on", "the", "buildings", "that", "are", "significantly", "higher", "than", "a", "bus", ".", "Finally", ",", "some", "of", "the", "SYNTHIA", "images", "are", "severely", "shadowed", "by", "extreme", "lighting", "conditions", ",", "while", "we", "find", "no", "such", "conditions", "in", "the", "Cityscapes", "images", ".", "These", "combined", "factors", ",", "among", "others", ",", "make", "domain", "adaptation", "from", "SYNTHIA", "to", "Cityscapes", "a", "very", "challenging", "problem", ".", "Figure", "[", "reference", "]", "shows", "some", "example", "images", "from", "both", "datasets", ".", "We", "pair", "each", "Cityscpaes", "image", "with", "its", "nearest", "neighbor", "in", "SYNTHIA", ",", "retrieved", "by", "the", "Inception", "-", "Resnet", "-", "v2", "features", ".", "However", ",", "the", "cross", "-", "dataset", "nearest", "neighbors", "are", "visually", "very", "different", "from", "the", "query", "images", ",", "verifying", "the", "dramatic", "disparity", "between", "the", "two", "domains", ".", "paragraph", ":", "Experiment", "setup", ".", "Since", "our", "ultimate", "goal", "is", "to", "solve", "the", "semantic", "segmentation", "problem", "for", "real", "images", "of", "urban", "scenes", ",", "we", "take", "Cityscapes", "as", "the", "target", "domain", "and", "SYNTHIA", "as", "the", "source", "domain", ".", "The", "Cityscapes", "validation", "set", "is", "used", "as", "our", "test", "set", ".", "We", "split", "500", "images", "out", "of", "the", "Cityscpaes", "training", "set", "for", "the", "validation", "purpose", "(", "e.g.", ",", "to", "monitor", "the", "convergence", "of", "the", "networks", ")", ".", "In", "training", ",", "we", "randomly", "sample", "mini", "-", "batches", "from", "both", "the", "images", "(", "and", "their", "labels", ")", "of", "SYNTHIA", "and", "the", "remaining", "images", "of", "Cityscapes", "yet", "with", "no", "labels", ".", "As", "in", ",", "we", "manually", "find", "16", "common", "classes", "between", "the", "two", "datasets", ":", "sky", ",", "building", ",", "road", ",", "sidewalk", ",", "fence", ",", "vegetation", ",", "pole", ",", "car", ",", "traffic", "sign", ",", "person", ",", "bicycle", ",", "motorcycle", ",", "traffic", "light", ",", "bus", ",", "wall", ",", "and", "rider", ".", "The", "last", "four", "are", "unnamed", "and", "yet", "labeled", "in", "SYNTHIA", ".", "paragraph", ":", "Evaluation", ".", "We", "use", "the", "evaluation", "code", "released", "along", "with", "the", "Cityscapes", "dataset", "to", "evaluate", "our", "results", ".", "It", "calculates", "the", "PASCAL", "VOC", "intersection", "-", "over", "-", "union", ",", "i.e.", ",", ",", "where", "TP", ",", "FP", ",", "and", "FN", "are", "the", "numbers", "of", "true", "positive", ",", "false", "positive", ",", "and", "false", "negative", "pixels", ",", "respectively", ",", "determined", "over", "the", "whole", "test", "set", ".", "Since", "we", "have", "to", "resize", "the", "images", "before", "feeding", "them", "to", "the", "segmentation", "network", ",", "we", "resize", "the", "output", "segmentation", "mask", "back", "to", "the", "original", "image", "size", "before", "running", "the", "evaluation", "against", "the", "groundtruth", "annotations", ".", "subsection", ":", "Results", "of", "inferring", "global", "label", "distributions", "Before", "presenting", "the", "final", "semantic", "segmentation", "results", ",", "we", "first", "compare", "the", "different", "approaches", "to", "inferring", "the", "global", "label", "distributions", "of", "the", "target", "images", "(", "cf", ".", "Section", "[", "reference", "]", ")", ".", "We", "report", "the", "results", "on", "the", "held", "-", "out", "validation", "images", "of", "Cityscapes", "in", "this", "experiment", ",", "and", "then", "select", "the", "best", "method", "for", "the", "remaining", "experiments", ".", "In", "Table", "[", "reference", "]", ",", "we", "compare", "the", "estimated", "label", "distributions", "with", "the", "groundtruth", "ones", "using", "the", "distance", ",", "the", "smaller", "the", "better", ".", "We", "see", "that", "the", "baseline", "network", "(", "NoAdapt", ")", ",", "which", "is", "directly", "learned", "from", "the", "source", "domain", "without", "any", "adaptation", "methods", ",", "outperforms", "the", "dumb", "uniform", "distribution", "(", "Uniform", ")", "and", "yet", "no", "other", "methods", ".", "This", "confirms", "that", "the", "baseline", "network", "gives", "rise", "to", "severely", "disproportionate", "predictions", "over", "the", "target", "domain", ".", "Another", "dumb", "prediction", "(", "Src", "mean", ")", ",", "i.e.", ",", "using", "the", "mean", "of", "all", "label", "distributions", "over", "the", "source", "domain", "as", "the", "prediction", "for", "the", "target", "images", ",", "however", ",", "performs", "reasonably", "well", ".", "To", "some", "extent", ",", "this", "indicates", "the", "value", "of", "the", "simulated", "source", "domain", "for", "the", "semantic", "segmentation", "task", "of", "urban", "scenes", ".", "Finally", ",", "the", "nearest", "neighbors", "(", "NN", ")", "based", "method", "and", "the", "multinomial", "logistic", "regression", "(", "LR", ")", "(", "cf", ".", "Section", "[", "reference", "]", ")", "perform", "the", "best", ".", "We", "use", "the", "output", "of", "LR", "on", "the", "target", "domain", "in", "our", "remaining", "experiments", ".", "subsection", ":", "Comparison", "results", "bike", "fence", "wall", "t", "-", "sign", "pole", "mbike", "t", "-", "light", "sky", "bus", "rider", "veg", "bldg", "car", "person", "sidewalk", "road", "We", "report", "the", "final", "semantic", "segmentation", "results", "on", "the", "test", "data", "of", "the", "target", "domain", "in", "this", "section", ".", "We", "compare", "our", "approach", "to", "the", "following", "competing", "methods", ".", "We", "directly", "train", "the", "FCN", "-", "8s", "model", "on", "SYNTHIA", "without", "applying", "any", "domain", "adaptation", "methods", ".", "This", "is", "the", "most", "basic", "baseline", "for", "our", "experiments", ".", "Recall", "that", "we", "have", "trained", "a", "multi", "-", "class", "SVM", "using", "the", "dominant", "labels", "of", "the", "superpixels", "in", "the", "source", "domain", ".", "We", "then", "use", "them", "to", "classify", "the", "target", "superpixels", ".", "Since", "we", "keep", "the", "top", "60", "%", "most", "confidently", "classified", "superpixels", "as", "the", "landmarks", "to", "regularize", "our", "segmentation", "network", "during", "training", "(", "cf", ".", "Section", "[", "reference", "]", ")", ",", "it", "is", "also", "interesting", "to", "examine", "the", "classification", "results", "of", "these", "superpixels", ".", "We", "run", "the", "evaluation", "after", "assigning", "the", "void", "class", "label", "to", "the", "other", "pixels", "of", "the", "images", ".", "In", "addition", "to", "the", "IoU", ",", "we", "have", "also", "evaluated", "the", "classification", "results", "of", "the", "superpixels", "by", "accuracy", ".", "We", "find", "that", "the", "classification", "accuracy", "is", "71", "%", "for", "all", "the", "superpixels", "of", "the", "target", "domain", ",", "while", "for", "the", "selected", "60", "%", "landmark", "superpixels", ",", "the", "classification", "accuracy", "is", "more", "than", "88", "%", ".", "Hoffman", "et", "al", ".", "’s", "work", "is", "the", "only", "existing", "one", "addressing", "the", "same", "problem", "as", "ours", ",", "to", "the", "best", "of", "our", "knowledge", ".", "They", "introduce", "a", "pixel", "-", "level", "adversarial", "loss", "to", "the", "intermediate", "layers", "of", "the", "network", "and", "impose", "constraints", "to", "the", "network", "output", ".", "Their", "experimental", "setup", "is", "about", "identical", "to", "ours", "except", "that", "they", "do", "not", "specify", "which", "part", "of", "Cityscapes", "is", "considered", "as", "the", "test", "set", ".", "Nonetheless", ",", "we", "include", "their", "results", "for", "comparison", "to", "put", "our", "work", "in", "a", "better", "perspective", ".", "The", "comparison", "results", "are", "shown", "in", "Table", "[", "reference", "]", ".", "Immediately", ",", "we", "note", "that", "all", "our", "domain", "adaptation", "results", "are", "significantly", "better", "than", "those", "without", "adaptation", "(", "NoAdapt", ")", ".", "We", "denote", "by", "(", "Ours", "(", "I", ")", ")", "the", "network", "trained", "using", "the", "global", "label", "distributions", "over", "the", "target", "images", "(", "and", "the", "labeled", "source", "images", ")", ".", "Although", "one", "may", "wonder", "that", "the", "image", "-", "wise", "label", "distributions", "are", "too", "high", "-", "level", "to", "supervise", "the", "pixel", "-", "wise", "discriminative", "network", ",", "the", "gain", "is", "actually", "significant", ".", "They", "are", "able", "to", "correct", "some", "obvious", "errors", "of", "the", "baseline", "network", ",", "such", "as", "the", "disproportional", "predictions", "about", "road", "and", "sidewalk", "(", "cf", ".", "the", "results", "of", "Ours", "(", "I", ")", "vs.", "NoAdapt", "in", "the", "last", "two", "columns", ")", ".", "It", "is", "interesting", "to", "see", "that", "both", "superpixel", "classification", "-", "based", "segmentation", "results", "(", "SP", "and", "SP", "Lndmk", ")", "are", "also", "better", "than", "the", "baseline", "network", "(", "NoAdapt", ")", ".", "The", "label", "distributions", "obtained", "over", "the", "landmark", "superpixels", "boost", "the", "segmentation", "network", "(", "Ours", "(", "SP", ")", ")", "to", "the", "mean", "IoU", "of", "28.1", "%", ",", "which", "is", "better", "than", "those", "by", "either", "superpixel", "classification", "or", "the", "baseline", "network", "individually", ".", "We", "have", "also", "tried", "to", "use", "the", "label", "distributions", "over", "all", "the", "superpixels", "to", "train", "the", "network", ",", "and", "observe", "little", "improvement", "over", "NoAdapt", ".", "This", "is", "probably", "because", "it", "is", "too", "forceful", "to", "regularize", "the", "network", "output", "at", "every", "single", "superpixel", "especially", "when", "the", "estimated", "label", "distributions", "are", "not", "accurate", "enough", ".", "The", "superpixel", "-", "based", "methods", ",", "including", "Ours", "(", "SP", ")", ",", "miss", "small", "objects", "such", "as", "fences", ",", "traffic", "lights", "(", "t", "-", "light", ")", ",", "and", "traffic", "signs", "(", "t", "-", "sign", ")", ",", "and", "instead", "are", "very", "accurate", "for", "categories", "like", "the", "sky", ",", "road", ",", "and", "building", ",", "that", "typically", "occupy", "larger", "image", "regions", ".", "On", "the", "contrary", ",", "the", "label", "distributions", "on", "the", "images", "give", "rise", "to", "a", "network", "(", "Ours", "(", "I", ")", ")", "that", "performs", "better", "on", "the", "small", "objects", "than", "Ours", "(", "SP", ")", ".", "In", "other", "words", ",", "they", "mutually", "complement", "to", "some", "extent", ".", "Re", "-", "training", "the", "network", "by", "using", "the", "label", "distributions", "over", "both", "global", "images", "and", "local", "landmark", "superpixels", "(", "Ours", "(", "I", "+", "SP", ")", ")", ",", "we", "achieve", "the", "best", "semantic", "segmentation", "results", "on", "the", "target", "domain", ".", "In", "the", "future", "work", ",", "it", "is", "worth", "exploring", "other", "target", "properties", ",", "perhaps", "still", "in", "the", "form", "of", "label", "distributions", ",", "that", "handle", "the", "small", "objects", "well", ",", "in", "order", "to", "further", "complement", "the", "superpixel", "-", "level", "label", "distributions", ".", "paragraph", ":", "Comparison", "with", "FCNs", "in", "the", "wild", ".", "Although", "we", "use", "the", "same", "segmentation", "network", "(", "FCN", "-", "8s", ")", "as", ",", "our", "baseline", "results", "(", "NoAdapt", ")", "are", "better", "than", "those", "reported", "in", ".", "This", "may", "be", "due", "to", "subtle", "differences", "in", "terms", "of", "implementation", "or", "experimental", "setup", ".", "Although", "our", "own", "baseline", "results", "are", "superior", ",", "we", "gain", "larger", "improvements", "(", "7", "%", ")", "over", "them", "than", "the", "performance", "gain", "of", "(", "3", "%", ")", "over", "the", "seemingly", "underperforming", "baseline", "network", "there", ".", "paragraph", ":", "Comparison", "with", "learning", "domain", "-", "invariant", "features", ".", "At", "our", "first", "attempt", "to", "solve", "the", "domain", "adaptation", "problem", "for", "the", "semantic", "segmentation", "of", "urban", "scenes", ",", "we", "tried", "to", "learn", "domain", "invariant", "features", "following", "the", "deep", "domain", "adaptation", "methods", "for", "classification", ".", "In", "particular", ",", "we", "impose", "the", "maximum", "mean", "discrepancy", "over", "the", "layer", "before", "the", "output", ".", "We", "name", "such", "network", "layer", "the", "feature", "layer", ".", "Since", "there", "are", "virtually", "three", "output", "layers", "in", "FCN", "-", "8s", ",", "we", "experiment", "with", "all", "the", "three", "feature", "layers", "correspondingly", ".", "We", "have", "also", "tested", "the", "domain", "adaptation", "by", "reversing", "the", "gradients", "of", "a", "domain", "classifier", ".", "However", ",", "none", "of", "these", "efforts", "lead", "to", "any", "noticeable", "gain", "over", "the", "baseline", "network", "so", "the", "results", "are", "omitted", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "address", "domain", "adaptation", "for", "the", "semantic", "segmentation", "of", "urban", "scenes", ".", "We", "propose", "a", "curriculum", "style", "approach", "to", "this", "problem", ".", "We", "learn", "to", "estimate", "the", "global", "label", "distributions", "of", "the", "images", "and", "local", "label", "distributions", "of", "the", "landmark", "superpixels", "of", "the", "target", "domain", ".", "Such", "tasks", "are", "easier", "to", "solve", "than", "the", "pixel", "-", "wise", "label", "assignment", ".", "Therefore", ",", "we", "use", "their", "results", "to", "effectively", "regularize", "our", "training", "of", "the", "semantic", "segmentation", "network", "such", "that", "its", "predictions", "meet", "the", "inferred", "label", "distributions", "over", "the", "target", "domain", ".", "Our", "method", "outperforms", "several", "competing", "methods", "that", "do", "domain", "adaptation", "from", "simulated", "images", "to", "real", "photos", "of", "urban", "traffic", "scenes", ".", "In", "future", "work", ",", "we", "will", "explore", "more", "target", "properties", "that", "can", "be", "conveniently", "inferred", "to", "enrich", "our", "curriculum", "domain", "adaptation", "framework", ".", "paragraph", ":", "Acknowledgements", ".", "This", "work", "is", "supported", "by", "the", "NSF", "award", "IIS", "#", "1566511", ",", "a", "gift", "from", "Adobe", "Systems", "Inc.", ",", "and", "a", "GPU", "from", "NVIDIA", ".", "We", "thank", "the", "anonymous", "reviewers", "and", "area", "chairs", "for", "their", "insightful", "comments", ".", "bibliography", ":", "References", "bike", "fence", "wall", "t", "-", "sign", "pole", "mbike", "t", "-", "light", "sky", "bus", "rider", "veg", "terrain", "train", "bldg", "car", "person", "truck", "sidewalk", "road", "appendix", ":", "GTA", "Cityscapes", "The", "main", "text", "above", "has", "been", "accepted", "to", "IEEE", "International", "Conference", "on", "Computer", "Vision", "(", "ICCV", ")", "2017", ".", "After", "the", "paper", "submission", ",", "we", "have", "been", "continuously", "working", "on", "the", "project", "and", "have", "got", "more", "results", ".", "We", "include", "them", "below", "to", "complement", "the", "experiments", "in", "the", "main", "text", ".", "The", "new", "experiment", "is", "basically", "the", "same", "as", "the", "one", "in", "the", "main", "text", "except", "that", "we", "replace", "SYNTHIA", "with", "the", "GTA", "dataset", ".", "GTA", "is", "a", "synthetic", ",", "vehicle", "-", "egocentric", "image", "dataset", "collected", "from", "the", "open", "world", "in", "the", "realistically", "rendered", "computer", "game", "Grand", "Theft", "Auto", "V", "(", "GTA", ",", "or", "GTA5", ")", ".", "It", "contains", "24", ",", "996", "images", ",", "whose", "semantic", "segmentation", "annotations", "are", "fully", "compatible", "with", "the", "classes", "used", "in", "Cityscapes", ".", "Hence", "we", "use", "all", "the", "19", "official", "training", "classes", "in", "our", "experiment", ".", "The", "results", "are", "shown", "in", "Table", "[", "reference", "]", ".", "As", "in", "the", "main", "text", ",", "the", "same", "observations", "about", "our", "approach", "apply", "here", ".", "Additionally", ",", "we", "note", "that", "the", "results", "are", "overall", "better", "than", "those", "adapting", "from", "SYNTHIA", "to", "Cityscapes", ".", "This", "is", "not", "surprising", ",", "because", "the", "GTA", "images", "are", "more", "photo", "-", "realistic", "than", "SYNTHIA", "’s", "."], "entity": [[2, 5, ["Curriculum", "Domain", "Adaptation"], "Method"], [6, 11, ["Semantic", "Segmentation", "of", "Urban", "Scenes"], "Task"], [17, 20, ["convolutional", "neural", "networks"], "Method"], [21, 22, ["CNNs"], "Method"], [26, 28, ["semantic", "segmentation"], "Task"], [37, 39, ["industrial", "applications"], "Task"], [41, 43, ["autonomous", "driving"], "Task"], [44, 46, ["medical", "imaging"], "Task"], [51, 52, ["CNNs"], "Method"], [72, 74, ["computer", "graphics"], "Task"], [118, 123, ["curriculum", "-", "style", "learning", "approach"], "Method"], [126, 128, ["domain", "gap"], "Task"], [129, 131, ["semantic", "segmentation"], "Task"], [133, 136, ["curriculum", "domain", "adaptation"], "Method"], [210, 212, ["segmentation", "network"], "Method"], [270, 272, ["Computer", "Vision"], "Task"], [303, 305, ["domain", "adaptation"], "Task"], [306, 312, ["semantic", "image", "segmentation", "of", "urban", "scenes"], "Task"], [340, 344, ["vision", "of", "autonomous", "driving"], "Task"], [346, 353, ["understanding", "complex", "inner", "-", "city", "traffic", "scenes"], "Task"], [358, 360, ["semantic", "segmentation"], "Task"], [367, 370, ["Machine", "learning", "methods"], "Method"], [371, 374, ["automatic", "semantic", "segmentation"], "Task"], [388, 389, ["classifiers"], "Method"], [402, 404, ["semantic", "segmentation"], "Task"], [419, 420, ["annotation"], "Task"], [421, 423, ["quality", "control"], "Task"], [434, 436, ["Cityscapes", "dataset"], "Material"], [556, 559, ["Domain", "adaptation", "techniques"], "Method"], [563, 566, ["machine", "learning", "methods"], "Method"], [576, 578, ["computer", "vision"], "Task"], [579, 584, ["learning", "domain", "-", "invariant", "features"], "Task"], [601, 605, ["classification", "and", "regression", "problems"], "Task"], [675, 676, ["classifiers"], "Method"], [713, 715, ["linear", "transformation"], "Method"], [716, 718, ["kernel", "methods"], "Method"], [722, 725, ["deep", "learning", "methods"], "Method"], [744, 746, ["semantic", "segmentation"], "Task"], [754, 757, ["structured", "prediction", "problem"], "Task"], [760, 762, ["domain", "adaptation"], "Task"], [771, 773, ["structured", "prediction"], "Task"], [779, 781, ["domain", "adaptation"], "Task"], [807, 809, ["decision", "function"], "Method"], [810, 812, ["structured", "prediction"], "Task"], [816, 817, ["classification"], "Task"], [869, 871, ["feature", "representations"], "Method"], [909, 911, ["semantic", "segmentation"], "Task"], [919, 921, ["segmentation", "methods"], "Method"], [924, 927, ["deep", "neural", "networks"], "Method"], [938, 941, ["fully", "convolutional", "network"], "Method"], [942, 943, ["FCN"], "Method"], [1065, 1067, ["semantic", "segmentation"], "Task"], [1069, 1071, ["posterior", "regularization"], "Method"], [1091, 1097, ["curriculum", "-", "style", "domain", "adaptation", "approach"], "Method"], [1102, 1104, ["domain", "adaptation"], "Task"], [1126, 1129, ["curriculum", "domain", "adaptation"], "Task"], [1132, 1134, ["easy", "tasks"], "Task"], [1160, 1163, ["semantic", "segmentation", "network"], "Method"], [1165, 1167, ["hard", "task"], "Task"], [1307, 1309, ["label", "distributions"], "Task"], [1338, 1340, ["label", "distributions"], "Method"], [1343, 1348, ["domain", "adaptation", "of", "semantic", "segmentation"], "Task"], [1357, 1359, ["segmentation", "networks"], "Method"], [1404, 1406, ["segmentation", "network"], "Method"], [1474, 1479, ["curriculum", "-", "style", "domain", "adaptation"], "Method"], [1481, 1486, ["semantic", "segmentation", "of", "urban", "scenes"], "Task"], [1498, 1501, ["inferring", "label", "distributions"], "Task"], [1528, 1534, ["pixel", "-", "wise", "discriminative", "segmentation", "network"], "Method"], [1594, 1596, ["segmentation", "networks"], "Method"], [1603, 1605, ["network", "architecture"], "Method"], [1621, 1623, ["domain", "adaptation"], "Task"], [1624, 1626, ["semantic", "segmentation"], "Task"], [1632, 1634, ["transferring", "knowledge"], "Task"], [1643, 1645, ["Domain", "adaptation"], "Method"], [1647, 1650, ["machine", "learning", "algorithms"], "Method"], [1690, 1692, ["Domain", "adaptation"], "Method"], [1703, 1704, ["generalization"], "Task"], [1711, 1713, ["domain", "adaptation"], "Task"], [1716, 1720, ["classification", "and", "regression", "problems"], "Task"], [1730, 1734, ["real", "-", "world", "objects"], "Task"], [1746, 1749, ["deep", "neural", "networks"], "Method"], [1769, 1772, ["generic", "object", "detectors"], "Method"], [1788, 1790, ["pedestrian", "detections"], "Task"], [1821, 1823, ["Semantic", "segmentation"], "Task"], [1824, 1826, ["Semantic", "segmentation"], "Task"], [1861, 1864, ["convolutional", "neural", "network"], "Method"], [1865, 1866, ["CNN"], "Method"], [1868, 1870, ["semantic", "segmentation"], "Task"], [1880, 1881, ["CNNs"], "Method"], [1903, 1905, ["segmentation", "models"], "Method"], [1907, 1910, ["PASCAL", "VOC2012", "Challenge"], "Material"], [1919, 1921, ["segmentation", "competition"], "Task"], [1924, 1927, ["MS", "COCO", "Challenge"], "Material"], [1963, 1965, ["human", "workload"], "Metric"], [1992, 1997, ["high", "-", "quality", "segmentation", "networks"], "Task"], [2012, 2014, ["computer", "game"], "Method"], [2035, 2037, ["Domain", "adaptation"], "Task"], [2038, 2040, ["semantic", "segmentation"], "Task"], [2054, 2056, ["domain", "adaptation"], "Method"], [2059, 2060, ["segmentation"], "Task"], [2106, 2110, ["curriculum", "domain", "adaptation", "strategy"], "Method"], [2114, 2116, ["easy", "task"], "Task"], [2147, 2150, ["curriculum", "domain", "adaptation"], "Method"], [2151, 2157, ["semantic", "segmentation", "of", "urban", "scene", "images"], "Task"], [2193, 2195, ["structured", "prediction"], "Task"], [2198, 2200, ["semantic", "segmentation"], "Task"], [2204, 2212, ["cross", "-", "domain", "generalization", "of", "machine", "learning", "models"], "Method"], [2286, 2291, ["one", "-", "hot", "vector", "encoding"], "Method"], [2328, 2329, ["prediction"], "Task"], [2331, 2333, ["segmentation", "network"], "Method"], [2337, 2339, ["softmax", "function"], "Method"], [2417, 2419, ["network", "predictions"], "Method"], [2426, 2428, ["Domain", "adaptation"], "Task"], [2440, 2442, ["segmentation", "network"], "Method"], [2474, 2476, ["cross", "entropy"], "Metric"], [2477, 2478, ["training"], "Task"], [2521, 2523, ["objective", "function"], "Metric"], [2526, 2532, ["cross", "-", "domain", "generalizing", "segmentation", "network"], "Method"], [2537, 2544, ["pixel", "-", "wise", "cross", "-", "entropy", "loss"], "Metric"], [2614, 2617, ["domain", "adaptation", "context"], "Task"], [2675, 2677, ["objective", "function"], "Metric"], [2683, 2685, ["model", "compression"], "Task"], [2698, 2701, ["domain", "adaptation", "procedure"], "Method"], [2704, 2707, ["student", "”", "network"], "Method"], [2725, 2728, ["semantically", "segmenting", "images"], "Task"], [2751, 2753, ["image", "annotation"], "Task"], [2777, 2779, ["posterior", "regularization"], "Method"], [2784, 2786, ["posterior", "regularization"], "Method"], [2800, 2805, ["weakly", "supervised", "segmentation", "and", "detection"], "Task"], [2807, 2814, ["rule", "-", "regularized", "training", "of", "neural", "networks"], "Method"], [2819, 2822, ["domain", "adaptation", "setting"], "Task"], [2848, 2852, ["EM", "type", "of", "optimizations"], "Method"], [2857, 2859, ["segmentation", "network"], "Method"], [2866, 2869, ["deep", "learning", "tools"], "Method"], [2888, 2891, ["curriculum", "domain", "adaptation"], "Task"], [2941, 2946, ["Global", "label", "distributions", "of", "images"], "Task"], [2953, 2956, ["baseline", "segmentation", "network"], "Method"], [3066, 3069, ["machine", "learning", "models"], "Method"], [3077, 3079, ["label", "distribution"], "Task"], [3096, 3100, ["per", "-", "pixel", "predictions"], "Task"], [3129, 3134, ["Inception", "-", "Resnet", "-", "v2"], "Method"], [3153, 3154, ["classification"], "Task"], [3183, 3187, ["cross", "-", "entropy", "loss"], "Metric"], [3236, 3238, ["nonparametric", "method"], "Method"], [3242, 3244, ["nearest", "neighbors"], "Method"], [3245, 3246, ["NNs"], "Method"], [3273, 3275, ["NN", "retrieval"], "Task"], [3303, 3305, ["label", "distributions"], "Method"], [3330, 3332, ["landmark", "superpixels"], "Method"], [3409, 3411, ["segmentation", "network"], "Method"], [3457, 3461, ["label", "distributions", "of", "superpixels"], "Task"], [3464, 3465, ["superpixels"], "Task"], [3473, 3475, ["linear", "SVM"], "Method"], [3488, 3489, ["superpixels"], "Method"], [3490, 3493, ["linear", "spectral", "clustering"], "Method"], [3532, 3536, ["multi", "-", "class", "SVM"], "Method"], [3547, 3551, ["multi", "-", "class", "SVM"], "Method"], [3567, 3569, ["confidence", "score"], "Metric"], [3669, 3672, ["FCN", "-", "8s"], "Method"], [3677, 3680, ["PASCAL", "CONTEXT", "dataset"], "Material"], [3690, 3692, ["detection", "scores"], "Metric"], [3739, 3742, ["Curriculum", "domain", "adaptation"], "Task"], [3748, 3751, ["curriculum", "domain", "adaptation"], "Method"], [3772, 3774, ["domain", "adaptation"], "Method"], [3781, 3783, ["easy", "tasks"], "Task"], [3793, 3795, ["semantic", "segmentation"], "Task"], [3908, 3910, ["Segmentation", "network"], "Method"], [3911, 3912, ["optimization"], "Task"], [3918, 3921, ["FCN", "-", "8s"], "Method"], [3923, 3926, ["semantic", "segmentation", "network"], "Method"], [3930, 3932, ["convolutional", "layers"], "Method"], [3933, 3936, ["VGG", "-", "19"], "Method"], [3943, 3945, ["AdaDelta", "optimizer"], "Method"], [3970, 3972, ["baseline", "network"], "Method"], [3974, 3975, ["adaptation"], "Method"], [3999, 4000, ["Keras"], "Method"], [4013, 4016, ["Tesla", "K40", "GPU"], "Method"], [4020, 4024, ["deep", "domain", "adaptation", "methods"], "Method"], [4045, 4048, ["curriculum", "domain", "adaptation"], "Method"], [4054, 4056, ["segmentation", "networks"], "Method"], [4071, 4072, ["Cityscpaes"], "Material"], [4073, 4075, ["SYNTHIA", "datasets"], "Material"], [4079, 4080, ["Cityscapes"], "Material"], [4165, 4166, ["SYNTHIA"], "Material"], [4180, 4185, ["SYNTHIA", "-", "RAND", "-", "CITYSCAPES"], "Material"], [4189, 4190, ["Cityscapes"], "Material"], [4238, 4239, ["Cityscapes"], "Material"], [4242, 4247, ["SYNTHIA", "-", "RAND", "-", "CITYSCAPES"], "Material"], [4248, 4249, ["SYNTHIA"], "Material"], [4260, 4262, ["urban", "scenes"], "Material"], [4264, 4265, ["SYNTHIA"], "Material"], [4315, 4317, ["Cityscapes", "images"], "Material"], [4338, 4340, ["Cityscapes", "images"], "Material"], [4344, 4346, ["dash", "cameras"], "Method"], [4375, 4376, ["SYNTHIA"], "Material"], [4418, 4420, ["Cityscapes", "images"], "Material"], [4429, 4431, ["domain", "adaptation"], "Task"], [4432, 4433, ["SYNTHIA"], "Material"], [4434, 4435, ["Cityscapes"], "Material"], [4455, 4457, ["Cityscpaes", "image"], "Material"], [4462, 4463, ["SYNTHIA"], "Material"], [4488, 4490, ["query", "images"], "Material"], [4513, 4516, ["semantic", "segmentation", "problem"], "Task"], [4517, 4522, ["real", "images", "of", "urban", "scenes"], "Task"], [4525, 4526, ["Cityscapes"], "Material"], [4531, 4532, ["SYNTHIA"], "Material"], [4538, 4541, ["Cityscapes", "validation", "set"], "Material"], [4555, 4558, ["Cityscpaes", "training", "set"], "Material"], [4560, 4562, ["validation", "purpose"], "Task"], [4593, 4594, ["SYNTHIA"], "Material"], [4599, 4600, ["Cityscapes"], "Material"], [4663, 4664, ["SYNTHIA"], "Material"], [4678, 4680, ["Cityscapes", "dataset"], "Material"], [4688, 4695, ["PASCAL", "VOC", "intersection", "-", "over", "-", "union"], "Task"], [4700, 4701, ["TP"], "Metric"], [4702, 4703, ["FP"], "Metric"], [4705, 4706, ["FN"], "Metric"], [4742, 4744, ["segmentation", "network"], "Method"], [4770, 4774, ["inferring", "global", "label", "distributions"], "Task"], [4778, 4780, ["semantic", "segmentation"], "Task"], [4819, 4820, ["Cityscapes"], "Material"], [4864, 4866, ["baseline", "network"], "Method"], [4867, 4868, ["NoAdapt"], "Method"], [4880, 4882, ["adaptation", "methods"], "Method"], [4885, 4888, ["dumb", "uniform", "distribution"], "Method"], [4889, 4890, ["Uniform"], "Method"], [4901, 4903, ["baseline", "network"], "Method"], [4915, 4917, ["dumb", "prediction"], "Method"], [4918, 4920, ["Src", "mean"], "Method"], [4929, 4931, ["label", "distributions"], "Method"], [4964, 4970, ["semantic", "segmentation", "task", "of", "urban", "scenes"], "Task"], [4974, 4981, ["nearest", "neighbors", "(", "NN", ")", "based", "method"], "Method"], [4984, 4986, ["logistic", "regression"], "Method"], [4987, 4988, ["LR"], "Method"], [5006, 5007, ["LR"], "Method"], [5044, 5046, ["semantic", "segmentation"], "Task"], [5078, 5079, ["SYNTHIA"], "Material"], [5082, 5085, ["domain", "adaptation", "methods"], "Method"], [5102, 5106, ["multi", "-", "class", "SVM"], "Method"], [5145, 5147, ["segmentation", "network"], "Method"], [5165, 5166, ["classification"], "Task"], [5169, 5170, ["superpixels"], "Method"], [5193, 5194, ["IoU"], "Metric"], [5200, 5201, ["classification"], "Task"], [5206, 5207, ["accuracy"], "Metric"], [5212, 5214, ["classification", "accuracy"], "Metric"], [5236, 5238, ["classification", "accuracy"], "Metric"], [5272, 5277, ["pixel", "-", "level", "adversarial", "loss"], "Method"], [5309, 5310, ["Cityscapes"], "Material"], [5352, 5354, ["domain", "adaptation"], "Method"], [5361, 5362, ["adaptation"], "Method"], [5363, 5364, ["NoAdapt"], "Method"], [5414, 5419, ["pixel", "-", "wise", "discriminative", "network"], "Method"], [5474, 5479, ["superpixel", "classification", "-", "based", "segmentation"], "Method"], [5481, 5482, ["SP"], "Method"], [5483, 5485, ["SP", "Lndmk"], "Method"], [5491, 5493, ["baseline", "network"], "Method"], [5494, 5495, ["NoAdapt"], "Method"], [5503, 5505, ["landmark", "superpixels"], "Method"], [5507, 5509, ["segmentation", "network"], "Method"], [5510, 5514, ["Ours", "(", "SP", ")"], "Method"], [5517, 5519, ["mean", "IoU"], "Metric"], [5530, 5532, ["superpixel", "classification"], "Method"], [5534, 5536, ["baseline", "network"], "Method"], [5561, 5562, ["NoAdapt"], "Method"], [5592, 5596, ["superpixel", "-", "based", "methods"], "Method"], [5598, 5602, ["Ours", "(", "SP", ")"], "Method"], [5681, 5682, ["SP"], "Method"], [5726, 5728, ["semantic", "segmentation"], "Task"], [5769, 5774, ["superpixel", "-", "level", "label", "distributions"], "Method"], [5779, 5780, ["FCNs"], "Method"], [5789, 5791, ["segmentation", "network"], "Method"], [5792, 5795, ["FCN", "-", "8s"], "Method"], [5802, 5803, ["NoAdapt"], "Method"], [5878, 5881, ["domain", "adaptation", "problem"], "Task"], [5883, 5888, ["semantic", "segmentation", "of", "urban", "scenes"], "Task"], [5898, 5902, ["deep", "domain", "adaptation", "methods"], "Method"], [5903, 5904, ["classification"], "Task"], [5924, 5926, ["network", "layer"], "Method"], [5927, 5929, ["feature", "layer"], "Method"], [5938, 5941, ["FCN", "-", "8s"], "Method"], [5957, 5959, ["domain", "adaptation"], "Method"], [5965, 5967, ["domain", "classifier"], "Method"], [5998, 6000, ["domain", "adaptation"], "Task"], [6002, 6007, ["semantic", "segmentation", "of", "urban", "scenes"], "Task"], [6011, 6014, ["curriculum", "style", "approach"], "Method"], [6050, 6055, ["pixel", "-", "wise", "label", "assignment"], "Task"], [6069, 6072, ["semantic", "segmentation", "network"], "Method"], [6094, 6096, ["domain", "adaptation"], "Method"], [6125, 6129, ["curriculum", "domain", "adaptation", "framework"], "Method"], [6155, 6156, ["GPU"], "Method"], [6157, 6158, ["NVIDIA"], "Method"], [6200, 6202, ["GTA", "Cityscapes"], "Material"], [6214, 6216, ["Computer", "Vision"], "Task"], [6217, 6218, ["ICCV"], "Task"], [6271, 6272, ["SYNTHIA"], "Material"], [6277, 6278, ["GTA"], "Material"], [6298, 6302, ["Grand", "Theft", "Auto", "V"], "Material"], [6303, 6304, ["GTA"], "Material"], [6306, 6307, ["GTA5"], "Material"], [6328, 6329, ["Cityscapes"], "Material"], [6382, 6383, ["SYNTHIA"], "Material"], [6384, 6385, ["Cityscapes"], "Material"], [6401, 6403, ["SYNTHIA", "’s"], "Material"], [79, 80, ["CNN"], "Method"], [533, 534, ["SYNTHIA"], "Material"], [538, 539, ["Cityscapes"], "Material"], [2023, 2024, ["SYNTHIA"], "Material"], [3144, 3146, ["logistic", "regression"], "Method"], [3147, 3148, ["LR"], "Method"], [3223, 3224, ["LR"], "Method"], [4296, 4297, ["SYNTHIA"], "Material"], [4400, 4401, ["SYNTHIA"], "Material"], [5073, 5076, ["FCN", "-", "8s"], "Method"], [6274, 6275, ["GTA"], "Material"], [6393, 6394, ["GTA"], "Material"]]} {"doc_id": "3ca3993b1f3536b15112f759067f62e999c5d38f", "text": ["BB8", ":", "A", "Scalable", ",", "Accurate", ",", "Robust", "to", "Partial", "Occlusion", "Method", "for", "Predicting", "the", "3D", "Poses", "of", "Challenging", "Objects", "without", "Using", "Depth", "section", ":", "Abstract", "We", "introduce", "a", "novel", "method", "for", "3D", "object", "detection", "and", "pose", "estimation", "from", "color", "images", "only", ".", "We", "first", "use", "segmentation", "to", "detect", "the", "objects", "of", "interest", "in", "2D", "even", "in", "presence", "of", "partial", "occlusions", "and", "cluttered", "background", ".", "By", "contrast", "with", "recent", "patch", "-", "based", "methods", ",", "we", "rely", "on", "a", "\"", "holistic", "\"", "approach", ":", "We", "apply", "to", "the", "detected", "objects", "a", "Convolutional", "Neural", "Network", "(", "CNN", ")", "trained", "to", "predict", "their", "3D", "poses", "in", "the", "form", "of", "2D", "projections", "of", "the", "corners", "of", "their", "3D", "bounding", "boxes", ".", "This", ",", "however", ",", "is", "not", "sufficient", "for", "handling", "objects", "from", "the", "recent", "T", "-", "LESS", "dataset", ":", "These", "objects", "exhibit", "an", "axis", "of", "rotational", "symmetry", ",", "and", "the", "similarity", "of", "two", "images", "of", "such", "an", "object", "under", "two", "different", "poses", "makes", "training", "the", "CNN", "challenging", ".", "We", "solve", "this", "problem", "by", "restricting", "the", "range", "of", "poses", "used", "for", "training", ",", "and", "by", "introducing", "a", "classifier", "to", "identify", "the", "range", "of", "a", "pose", "at", "run", "-", "time", "before", "estimating", "it", ".", "We", "also", "use", "an", "optional", "additional", "step", "that", "refines", "the", "predicted", "poses", ".", "We", "improve", "the", "state", "-", "of", "-", "the", "-", "art", "on", "the", "LINEMOD", "dataset", "from", "73.7", "%", "[", "2", "]", "to", "89.3", "%", "of", "correctly", "registered", "RGB", "frames", ".", "We", "are", "also", "the", "first", "to", "report", "results", "on", "the", "Occlusion", "dataset", "[", "1", "]", "using", "color", "images", "only", ".", "We", "obtain", "54", "%", "of", "frames", "passing", "the", "Pose", "6D", "criterion", "on", "average", "on", "several", "sequences", "of", "the", "T", "-", "LESS", "dataset", ",", "compared", "to", "the", "67", "%", "of", "the", "state", "-", "of", "-", "the", "-", "art", "[", "10", "]", "on", "the", "same", "sequences", "which", "uses", "both", "color", "and", "depth", ".", "The", "full", "approach", "is", "also", "scalable", ",", "as", "a", "single", "network", "can", "be", "trained", "for", "multiple", "objects", "simultaneously", ".", "section", ":", "Introduction", "3D", "pose", "estimation", "of", "object", "instances", "has", "recently", "become", "a", "popular", "problem", "again", ",", "because", "of", "its", "application", "in", "robotics", ",", "virtual", "and", "augmented", "reality", ".", "Many", "recent", "approaches", "rely", "on", "depth", "maps", ",", "sometimes", "in", "conjunction", "with", "color", "images", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "However", ",", "it", "is", "not", "always", "possible", "to", "use", "depth", "cameras", ",", "as", "they", "fail", "(", "c", ")", "(", "d", ")", "Figure", "1", ".", "Zooms", "on", "estimated", "poses", "for", "(", "a", ")", "the", "Ape", "of", "the", "LINEMOD", "dataset", "[", "reference", "]", ",", "(", "b", ")", "the", "Driller", "of", "the", "Occlusion", "dataset", "[", "reference", "]", ",", "(", "c", ")", "and", "(", "d", ")", "three", "objects", "of", "the", "T", "-", "LESS", "[", "reference", "]", "dataset", ".", "The", "green", "bounding", "boxes", "correspond", "to", "the", "ground", "truth", "poses", ",", "and", "the", "blue", "bounding", "boxes", "to", "the", "poses", "estimated", "with", "our", "method", ".", "The", "two", "boxes", "often", "overlap", "almost", "perfectly", ",", "showing", "the", "accuracy", "of", "our", "estimated", "poses", ".", "The", "parts", "of", "the", "bounding", "boxes", "occluded", "by", "the", "object", "were", "removed", "using", "the", "object", "mask", "rendered", "from", "our", "estimated", "pose", ".", "In", "(", "b", ")", ",", "we", "can", "still", "obtain", "a", "good", "pose", "despite", "the", "large", "occlusion", "by", "the", "bench", "vise", ".", "In", "(", "c", ")", "and", "(", "d", ")", ",", "we", "also", "obtain", "very", "good", "estimates", "despite", "large", "occlusions", ",", "the", "similarities", "between", "the", "objects", ",", "and", "the", "fact", "that", "the", "symmetries", "challenge", "the", "learning", "algorithms", ".", "outdoor", "or", "on", "specular", "objects", ".", "In", "addition", ",", "they", "drain", "the", "batteries", "of", "mobile", "devices", ",", "being", "an", "active", "sensor", ".", "It", "is", "therefore", "desirable", "to", "rely", "only", "on", "color", "images", "for", "3D", "pose", "estimation", ",", "even", "if", "it", "is", "more", "challenging", ".", "Recent", "methods", "[", "reference", "][", "reference", "][", "reference", "]", "work", "by", "identifying", "the", "'", "object", "coordinates", "'", "of", "the", "pixels", ",", "which", "are", "the", "pixels", "'", "3D", "coordinates", "in", "a", "coordinate", "system", "related", "to", "the", "object", "[", "reference", "]", ".", "The", "object", "3D", "pose", "can", "then", "be", "estimated", "using", "a", "PnP", "algorithm", "from", "these", "2D", "-", "3D", "correspondences", ".", "[", "reference", "]", "obtain", "similar", "correspondences", "by", "associating", "some", "pixels", "in", "selected", "parts", "of", "the", "object", "with", "virtual", "3D", "points", ".", "However", ",", "obtaining", "these", "2D", "-", "3D", "correspondences", "from", "local", "patches", "is", "difficult", "and", "the", "output", "is", "typically", "very", "noisy", "for", "these", "methods", ".", "A", "robust", "optimization", "is", "then", "needed", "to", "estimate", "the", "pose", ".", "In", "this", "paper", ",", "we", "argue", "for", "a", "\"", "holistic", "\"", "approach", ",", "in", "the", "sense", "that", "we", "predict", "the", "pose", "of", "an", "object", "directly", "from", "its", "appearance", ",", "instead", "of", "identifying", "its", "individual", "surface", "points", ".", "As", "we", "will", "show", ",", "this", "approach", "provides", "significantly", "better", "results", ".", "We", "first", "detect", "the", "target", "objects", "in", "2D.", "We", "show", "that", "using", "object", "segmentation", "performs", "better", "for", "this", "task", "compared", "to", "a", "standard", "sliding", "window", "detector", ",", "in", "particular", "in", "presence", "of", "partial", "occlusion", ".", "We", "then", "apply", "a", "CNN", "to", "predict", "the", "3D", "pose", "of", "the", "detected", "objects", ".", "While", "the", "predicted", "3D", "pose", "can", "be", "represented", "directly", "by", "a", "translation", "and", "a", "rotation", ",", "we", "achieve", "better", "accuracy", "by", "using", "a", "representation", "similar", "to", "the", "one", "used", "in", "[", "reference", "]", "for", "object", "parts", ":", "We", "predict", "the", "2D", "projections", "of", "the", "corners", "of", "the", "object", "'s", "bounding", "box", ",", "and", "compute", "the", "3D", "pose", "from", "these", "2D", "-", "3D", "correspondences", "with", "a", "PnP", "algorithm", ".", "Compared", "to", "the", "object", "coordinate", "approaches", "the", "predictions", "are", "typically", "outlier", "-", "free", ",", "and", "no", "robust", "estimation", "is", "thus", "needed", ".", "Compared", "to", "the", "direct", "prediction", "of", "the", "pose", ",", "this", "also", "avoids", "the", "need", "for", "a", "meta", "-", "parameter", "to", "balance", "the", "translation", "and", "rotation", "terms", ".", "Unfortunately", ",", "this", "simple", "approach", "performs", "badly", "on", "the", "recent", "and", "challenging", "T", "-", "LESS", "dataset", ".", "This", "dataset", "is", "made", "of", "manufactured", "objects", "that", "are", "not", "only", "similar", "to", "each", "other", ",", "but", "also", "have", "one", "axis", "of", "rotational", "symmetry", ".", "For", "example", ",", "the", "squared", "box", "of", "Fig", ".", "1", "(", "c", ")", "has", "an", "angle", "of", "symmetry", "of", "90", "•", "and", "the", "other", "object", "has", "an", "angle", "of", "symmetry", "of", "0", "•", "since", "it", "is", "an", "object", "of", "revolution", ";", "Object", "#", "5", "in", "Fig", ".", "1", "(", "d", ")", "is", "not", "perfectly", "symmetrical", "but", "only", "because", "of", "the", "small", "screw", "on", "the", "top", "face", ".", "The", "approach", "described", "above", "fails", "on", "these", "objects", "because", "it", "tries", "to", "learn", "a", "mapping", "from", "the", "image", "space", "to", "the", "pose", "space", ".", "Since", "two", "images", "of", "a", "symmetrical", "object", "under", "two", "different", "poses", "look", "identical", ",", "the", "image", "-", "pose", "correspondence", "is", "in", "fact", "a", "one", "-", "to", "-", "many", "relationship", ".", "This", "issue", "is", "actually", "not", "restricted", "to", "our", "approach", ".", "For", "example", ",", "[", "reference", "]", ",", "which", "relies", "on", "object", "coordinates", ",", "does", "not", "provide", "results", "on", "the", "Bowl", "object", "of", "the", "LINEMOD", "dataset", ",", "an", "object", "with", "an", "axis", "of", "symmetry", ":", "It", "is", "not", "clear", "which", "coordinates", "should", "be", "assigned", "to", "the", "3D", "points", "of", "this", "object", ",", "as", "all", "the", "points", "on", "a", "circle", "orthogonal", "to", "the", "axis", "of", "symmetry", "have", "the", "same", "appearance", ".", "To", "solve", "this", "problem", ",", "we", "train", "the", "method", "described", "above", "using", "images", "of", "the", "object", "under", "rotation", "in", "a", "restricted", "range", ",", "such", "that", "the", "training", "set", "does", "not", "contain", "ambiguous", "images", ".", "In", "order", "to", "recover", "the", "object", "pose", "under", "a", "larger", "range", "of", "rotation", ",", "we", "train", "a", "classifer", "to", "tell", "under", "which", "range", "the", "object", "rotation", "is", ".", "Again", ",", "this", "is", "easy", "to", "do", "with", "a", "\"", "holistic", "\"", "approach", ",", "and", "this", "classifier", "takes", "an", "image", "of", "the", "entire", "object", "as", "input", ".", "As", "we", "will", "explain", "in", "more", "details", ",", "we", "can", "then", "always", "use", "the", "CNN", "trained", "on", "the", "restricted", "range", "to", "estimate", "any", "pose", ".", "In", "addition", ",", "we", "will", "show", "how", "to", "adapt", "this", "idea", "to", "handle", "\"", "approximatively", "symmetrical", "\"", "objects", "like", "Object", "#", "5", ".", "This", "approach", "allows", "us", "to", "obtain", "good", "performance", "on", "the", "T", "-", "LESS", "dataset", ".", "Finally", ",", "we", "show", "that", "we", "can", "add", "an", "optional", "last", "step", "to", "refine", "the", "pose", "estimates", "by", "using", "the", "\"", "feedback", "loop", "\"", "proposed", "in", "[", "reference", "]", "for", "hand", "detection", "in", "depth", "images", ":", "We", "train", "a", "network", "to", "improve", "the", "prediction", "of", "the", "2D", "projections", "by", "comparing", "the", "input", "image", "and", "a", "rendering", "of", "the", "object", "for", "the", "initial", "pose", "estimate", ".", "This", "allows", "us", "to", "improve", "even", "more", "our", "results", "on", "the", "LINEMOD", "and", "Occlusion", "datasets", ".", "Our", "full", "approach", ",", "which", "we", "call", "BB8", ",", "for", "the", "8", "corners", "of", "the", "bounding", "box", ",", "is", "also", "very", "fast", ",", "as", "it", "only", "requires", "to", "apply", "Deep", "Networks", "to", "the", "input", "image", "a", "few", "times", ".", "In", "the", "remainder", "of", "the", "paper", ",", "we", "first", "discuss", "related", "work", ",", "describe", "our", "approach", ",", "and", "compare", "it", "against", "the", "state", "-", "ofthe", "-", "art", "on", "the", "three", "available", "datasets", ".", "section", ":", "Related", "Work", "The", "literature", "on", "3D", "object", "detection", "is", "very", "large", ",", "thus", "we", "will", "focus", "only", "on", "recent", "works", ".", "Keypoint", "-", "based", "methods", "[", "reference", "][", "reference", "]", "were", "popular", "for", "a", "long", "time", "and", "perform", "well", "but", "only", "on", "very", "textured", "objects", ".", "The", "apparition", "of", "inexpensive", "3D", "cameras", "favored", "the", "development", "of", "methods", "suitable", "for", "untextured", "objects", ":", "[", "reference", "][", "reference", "]", "rely", "on", "depth", "data", "only", "and", "use", "votes", "from", "pairs", "of", "3D", "points", "and", "their", "normals", "to", "detect", "3D", "objects", ".", "[", "reference", "]", "uses", "a", "decision", "tree", "applied", "to", "RGB", "-", "D", "images", "to", "simultaneously", "recognize", "the", "objects", "and", "predict", "their", "poses", ".", "[", "reference", "][", "reference", "]", "consider", "a", "template", "-", "based", "representation", "computed", "from", "RGB", "-", "D", "or", "RGB", "data", ",", "which", "allows", "for", "large", "scale", "detection", "[", "reference", "]", ".", "However", ",", "this", "template", "approach", "is", "sensitive", "to", "partial", "occlusions", ".", "To", "tackle", "clutter", "and", "partial", "occlusions", ",", "[", "reference", "]", "and", "[", "reference", "]", "rely", "on", "local", "patches", "recognition", "performed", "with", "Random", "Forests", ".", "In", "particular", ",", "[", "reference", "]", "considers", "'", "3D", "object", "coordinates", "'", ":", "A", "Random", "Forest", "is", "trained", "to", "predict", "the", "3D", "location", "in", "the", "object", "coordinate", "system", "of", "each", "image", "location", ".", "The", "prediction", "of", "this", "forest", "is", "integrated", "in", "an", "energy", "function", "together", "with", "a", "term", "that", "compares", "the", "depth", "map", "with", "a", "rendering", "of", "the", "object", "and", "a", "term", "that", "penalizes", "pixels", "that", "lie", "on", "the", "object", "rendering", "but", "predicted", "by", "the", "forest", "to", "not", "be", "an", "object", "point", ".", "This", "energy", "function", "is", "optimized", "by", "a", "RANSAC", "procedure", ".", "[", "reference", "]", "replaces", "this", "energy", "function", "by", "an", "energy", "computed", "from", "the", "output", "of", "a", "CNN", "trained", "to", "compare", "observed", "image", "features", "and", "features", "computed", "from", "a", "3D", "rendering", "of", "the", "potentially", "detected", "object", ".", "This", "makes", "the", "approach", "very", "robust", "to", "partial", "occlusions", ".", "These", "works", ",", "however", ",", "are", "designed", "for", "RGB", "-", "D", "data", ".", "[", "reference", "]", "extends", "this", "work", "and", "relies", "on", "RGB", "data", "only", ",", "as", "we", "do", ".", "They", "use", "Auto", "-", "Context", "[", "reference", "]", "to", "obtain", "better", "predictions", "from", "the", "Random", "Forests", ",", "estimate", "a", "distribute", "over", "the", "object", "coordinates", "to", "handle", "the", "prediction", "uncertainties", "better", ",", "and", "propose", "a", "more", "sophisticated", "RANSAC", "-", "like", "method", "that", "scales", "with", "the", "number", "of", "objects", ".", "This", "results", "in", "an", "efficient", "and", "accurate", "method", ",", "however", ",", "robustness", "to", "partial", "occlusions", "are", "not", "demonstrated", ".", "[", "3", "]", "is", "related", "to", "[", "reference", "][", "reference", "][", "reference", "]", "but", "focuses", "on", "providing", "sparse", "2D", "-", "3D", "correspondences", "from", "reliable", "object", "parts", ".", "Unfortunately", ",", "it", "provides", "results", "on", "its", "own", "dataset", "only", ",", "not", "on", "more", "broadly", "available", "datasets", ".", "Like", "us", ",", "[", "reference", "]", "relies", "on", "a", "CNN", "to", "directly", "predict", "a", "3D", "pose", ",", "but", "in", "the", "form", "of", "a", "translation", "and", "a", "rotation", ".", "It", "considers", "camera", "relocalisation", "in", "urban", "environment", "rather", "than", "3D", "object", "detection", ",", "and", "uses", "the", "full", "image", "as", "input", "to", "the", "CNN", ".", "By", "predicting", "the", "2D", "projections", "of", "the", "corners", "of", "the", "bounding", "box", ",", "we", "avoid", "the", "need", "for", "a", "meta", "-", "parameter", "to", "balance", "the", "position", "and", "orientation", "errors", ".", "As", "shown", "in", "our", "experiments", ",", "the", "pose", "appears", "to", "be", "more", "accurate", "when", "predicted", "in", "this", "form", ".", "Intuitively", ",", "this", "should", "not", "be", "surprising", ",", "as", "predicting", "2D", "locations", "from", "a", "color", "images", "seems", "easier", "than", "predicting", "a", "3D", "translation", "and", "a", "quaternion", ",", "for", "example", ".", "[", "6", "]", "also", "uses", "a", "CNN", "to", "predict", "the", "3D", "pose", "of", "generic", "objects", "but", "from", "RGB", "-", "D", "data", ".", "It", "first", "segments", "the", "objects", "of", "interest", "to", "avoid", "the", "influence", "of", "clutter", ".", "We", "tried", "segmenting", "the", "objects", "before", "predicting", "the", "pose", "as", "well", ",", "however", ",", "this", "performed", "poorly", "on", "the", "LINEMOD", "dataset", ",", "because", "the", "segmented", "silhouttes", "were", "not", "very", "accurate", ",", "even", "with", "state", "-", "of", "-", "the", "-", "art", "segmentation", "methods", ".", "In", "summary", ",", "our", "method", "appears", "to", "be", "one", "of", "the", "first", "to", "deal", "with", "RGB", "data", "only", "to", "detect", "3D", "objects", "and", "estimate", "their", "poses", "on", "recent", "datasets", ".", "As", "we", "will", "show", "in", "the", "experiments", ",", "it", "outperforms", "the", "accuracy", "of", "the", "state", "-", "of", "-", "theart", "[", "reference", "]", "by", "a", "large", "margin", ".", "section", ":", "Proposed", "Approach", "In", "our", "approach", ",", "we", "first", "find", "the", "objects", "in", "2D", ",", "we", "obtain", "a", "first", "estimate", "of", "the", "3D", "poses", ",", "including", "objects", "with", "a", "rotational", "symmetry", ",", "and", "we", "finally", "refine", "the", "initial", "pose", "estimates", ".", "We", "describe", "each", "step", "in", "this", "section", ".", "section", ":", "Localizing", "the", "Objects", "in", "2D", "We", "first", "identify", "the", "2D", "centers", "of", "the", "objects", "of", "interest", "in", "the", "input", "images", ".", "We", "could", "use", "a", "standard", "2D", "object", "detector", ",", "but", "we", "developed", "an", "approach", "based", "on", "segmentation", "that", "resulted", "in", "better", "performance", "as", "it", "can", "provide", "accurate", "locations", "even", "under", "partial", "occlusions", ".", "Compared", "to", "our", "initial", "tests", "using", "a", "sliding", "window", ",", "this", "ap", "-", "proach", "improved", "our", "2D", "detection", "results", "from", "about", "75", "%", "to", "98.8", "%", "correct", "detection", "rate", "based", "on", "a", "IoU", "of", "0.5", ".", "We", "only", "need", "a", "low", "resolution", "segmentation", "and", "thus", "do", "not", "need", "a", "hourglass", "-", "shaped", "architecture", "[", "reference", "]", ",", "which", "makes", "our", "segmentation", "more", "efficient", ".", "As", "shown", "in", "Fig", ".", "2", ",", "our", "approach", "performs", "a", "two", "-", "level", "coarse", "-", "to", "-", "fine", "object", "segmentation", ".", "For", "each", "level", ",", "we", "train", "a", "single", "network", "for", "all", "the", "objects", ".", "The", "first", "network", "is", "obtained", "by", "replacing", "the", "last", "layer", "of", "VGG", "[", "reference", "]", "by", "a", "fully", "connected", "layer", "with", "the", "required", "number", "of", "output", "required", "by", "each", "step", ",", "and", "fine", "-", "tune", "it", ".", "The", "second", "network", "has", "a", "simple", ",", "ad", "hoc", "architecture", ".", "More", "exactly", ",", "the", "first", "network", "is", "trained", "to", "provide", "a", "very", "low", "resolution", "binary", "segmentation", "of", "the", "objects", "given", "an", "image", "region", "J", "of", "size", "128", "×", "128", "by", "minimizing", "the", "following", "objective", "function", ":", "where", "T", "s", "is", "a", "training", "set", "made", "of", "image", "regions", "J", ",", "and", "the", "corresponding", "segmentations", "S", "for", "object", "o", ",", "(", "f", "is", "the", "output", "of", "network", "f", "1", "φ", "for", "region", "J", "and", "object", "o.", "φ", "denotes", "the", "network", "'s", "parameters", ",", "optimized", "during", "training", ".", "For", "the", "LINEMOD", "and", "Occlusion", "datasets", ",", "there", "is", "at", "most", "one", "object", "for", "a", "given", "region", "J", ",", "but", "more", "objects", "can", "be", "present", "for", "the", "T", "-", "LESS", "dataset", ".", "At", "run", "-", "time", ",", "to", "get", "the", "segmentations", ",", "we", "compute", ":", "where", "s", "1", ",", "o", "is", "a", "8", "×", "8", "binary", "segmentation", "of", "J", "for", "object", "o", ",", "and", "τ", "1", "is", "a", "threshold", "used", "to", "binarize", "the", "network", "'s", "output", ".", "To", "obtain", "a", "binary", "segmentation", "for", "the", "full", "input", "image", ",", "we", "split", "this", "image", "into", "regions", "and", "compute", "the", "s", "1", ",", "o", "for", "each", "region", ".", "This", "gives", "us", "one", "binary", "segmentation", "S", "1", ",", "o", "for", "the", "full", "input", "image", ",", "and", "each", "possible", "object", ".", "This", "usually", "results", "in", "a", "single", "connected", "component", "per", "visible", "object", ";", "if", "several", "components", "are", "present", ",", "we", "keep", "only", "the", "largest", "one", "for", "each", "object", ".", "If", "the", "largest", "component", "in", "a", "segmentation", "S", "1", ",", "o", "is", "small", ",", "object", "o", "is", "likely", "not", "visible", ".", "For", "the", "remaining", "object", "(", "s", ")", ",", "we", "refine", "the", "shape", "of", "the", "largest", "component", "by", "applying", "a", "second", "network", "to", "each", "16", "×", "16", "image", "patch", "P", "that", "corresponds", "to", "an", "active", "location", "in", "S", "1", ":", "using", "notations", "similar", "to", "the", "ones", "in", "Eq", ".", "[", "reference", "]", ".", "Since", "the", "input", "to", "f", "2", "ψ", "(", "P", ")", "has", "a", "low", "resolution", ",", "we", "do", "not", "need", "a", "complex", "network", "such", "as", "VGG", "[", "reference", "]", ",", "and", "we", "use", "a", "much", "simpler", "architecture", "with", "2", "convolutional", "layers", "and", "2", "pooling", "layers", ".", "We", "finally", "obtain", "a", "segmentation", "S", "2", ",", "o", "with", "resolution", "64", "×", "48", "for", "the", "full", "input", "image", "and", "each", "visible", "object", "o.", "We", "therefore", "get", "the", "identities", "o", "of", "the", "visible", "object", "(", "s", ")", ",", "and", "for", "these", "objects", ",", "we", "use", "the", "segmentation", "centroids", "as", "their", "2D", "centers", ",", "to", "compute", "the", "3D", "poses", "of", "the", "objects", "as", "described", "below", ".", "section", ":", "Predicting", "the", "3D", "Pose", "We", "predict", "the", "3D", "pose", "of", "an", "object", "by", "applying", "a", "Deep", "Network", "to", "an", "image", "window", "W", "centered", "on", "the", "2D", "object", "center", "estimated", "as", "described", "in", "the", "previous", "section", ".", "As", "for", "the", "segmentation", ",", "we", "use", "VGG", "[", "reference", "]", "as", "a", "basis", "for", "this", "network", ".", "This", "allows", "us", "to", "handle", "all", "the", "objects", "of", "the", "target", "dataset", "with", "a", "single", "network", ".", "It", "is", "possible", "to", "directly", "predict", "the", "pose", "in", "the", "form", "of", "a", "3", "-", "vector", "and", "an", "exponential", "map", "for", "example", ",", "as", "in", "[", "reference", "]", ".", "However", ",", "a", "more", "accurate", "approach", "was", "proposed", "in", "[", "reference", "]", "for", "predicting", "the", "poses", "of", "object", "parts", ".", "To", "apply", "it", "here", ",", "we", "minimize", "the", "following", "cost", "function", "over", "the", "parameters", "Θ", "of", "network", "g", "Θ", ":", "where", "T", "is", "a", "training", "set", "made", "of", "image", "windows", "W", "containing", "object", "o", "under", "a", "pose", "defined", "by", "an", "exponential", "map", "e", "and", "a", "3", "-", "vector", "t.", "The", "M", "o", "i", "are", "the", "3D", "coordinates", "of", "the", "corners", "of", "the", "bounding", "box", "of", "object", "o", "in", "the", "object", "coordinate", "system", ".", "Proj", "e", ",", "t", "(", "M", ")", "projects", "the", "3D", "point", "M", "on", "the", "image", "from", "the", "pose", "defined", "by", "e", "and", "t.", "returns", "the", "two", "components", "of", "the", "output", "of", "g", "Θ", "corresponding", "to", "the", "predicted", "2D", "coordinates", "of", "the", "i", "-", "th", "corner", "for", "object", "o.", "•", "modulo", "180", "•", "(", "c", ")", ".", "Our", "solution", "is", "to", "restrict", "the", "range", "during", "training", "to", "be", "between", "0", "•", "and", "90", "•", ".", "We", "use", "a", "classifier", "to", "detect", "if", "the", "pose", "in", "an", "input", "image", "is", "between", "90", "•", "and", "180", "•", ".", "If", "this", "is", "the", "case", "(", "d", ")", ",", "we", "mirror", "the", "input", "image", "(", "e", ")", ",", "and", "mirror", "back", "the", "predicted", "projections", "for", "the", "corners", "(", "f", ")", ".", "At", "run", "-", "time", ",", "the", "segmentation", "gives", "the", "identity", "and", "the", "2D", "locations", "of", "the", "visible", "object", "(", "s", ")", "o.", "The", "3D", "pose", "can", "then", "be", "estimated", "for", "the", "correspondences", "between", "the", "3D", "points", "M", "o", "i", "and", "the", "predicted", "m", "i", "(", "(", "g", "Θ", "(", "W", ")", ")", "[", "o", "]", ")", "using", "a", "PnP", "algorithm", ".", "Other", "3D", "points", "could", "be", "used", "here", ",", "however", ",", "the", "corners", "of", "the", "bounding", "box", "are", "a", "natural", "choice", "as", "they", "frame", "the", "object", "and", "are", "well", "spread", "in", "space", "1", ".", "section", ":", "Handling", "Objects", "with", "an", "Axis", "of", "Symmetry", "If", "we", "apply", "the", "method", "described", "so", "far", "to", "the", "T", "-", "LESS", "dataset", ",", "the", "performances", "are", "significantly", "lower", "than", "the", "performances", "on", "the", "LINEMOD", "dataset", ".", "As", "mentioned", "in", "the", "introduction", ",", "this", "is", "because", "training", "images", "W", "in", "Eq", ".", "(", "4", ")", "for", "the", "objects", "of", "this", "dataset", "can", "be", "identical", "while", "having", "very", "different", "expected", "predictions", "Proj", "e", ",", "t", "(", "M", "o", "i", ")", ",", "because", "of", "the", "rotational", "symmetry", "of", "the", "objects", ".", "We", "first", "remark", "that", "for", "an", "object", "with", "an", "angle", "of", "symmetry", "α", ",", "its", "3D", "rotation", "around", "its", "axis", "of", "symmetry", "can", "be", "defined", "only", "modulo", "α", ",", "not", "2π", ".", "For", "an", "object", "with", "an", "angle", "of", "symmetry", "α", ",", "we", "can", "therefore", "restrict", "the", "poses", "used", "for", "training", "to", "the", "poses", "where", "the", "angle", "of", "rotation", "around", "the", "symmetry", "axis", "is", "within", "the", "range", "[", "0", ";", "α", "[", ",", "to", "avoid", "the", "ambiguity", "between", "images", ".", "However", ",", "this", "solves", "our", "problem", "only", "partially", ":", "Images", "at", "one", "extremity", "of", "this", "range", "of", "poses", "and", "the", "images", "at", "the", "other", "extremity", ",", "while", "not", "identical", ",", "still", "look", "very", "similar", ".", "As", "a", "result", ",", "for", "input", "images", "with", "an", "angle", "of", "rotation", "close", "to", "0", "modulo", "α", ",", "the", "pose", "prediction", "can", "still", "be", "very", "bad", ",", "as", "illustrated", "in", "Fig", ".", "3", ".", "To", "explain", "our", "solution", ",", "let", "us", "first", "denote", "by", "β", "the", "rotation", "angle", ",", "and", "introduce", "the", "intervals", "r", "1", "=", "[", "0", ";", "α", "/", "2", "[", "and", "r", "2", "=", "[", "α", "/", "2", ";", "α", "[", ".", "To", "avoid", "ambiguity", ",", "we", "restrict", "β", "to", "be", "in", "r", "1", "for", "the", "training", "images", "used", "in", "the", "optimization", "problem", "of", "Eq", ".", "(", "4", ")", ".", "The", "drawback", "is", "of", "course", "that", ",", "without", "doing", "anything", "else", ",", "we", "would", "not", "be", "able", "to", "estimate", "the", "poses", "when", "β", "is", "in", "r", "2", ".", "We", "therefore", "introduce", "a", "CNN", "classifier", "k", "(", "·", ")", "to", "predict", "at", "run", "-", "time", "if", "β", "is", "in", "r", "1", "or", "r", "2", ":", "If", "β", "is", "in", "r", "1", ",", "we", "can", "estimate", "the", "pose", "as", "before", ";", "If", "β", "is", "in", "r", "2", ",", "one", "option", "would", "be", "to", "apply", "another", "g", "Θ", "(", "·", ")", "network", "trained", "for", "this", "range", ".", "However", ",", "it", "is", "actually", "possible", "to", "use", "the", "same", "network", "g", "Θ", "(", "·", ")", "for", "both", "r", "1", "and", "r", "2", ",", "as", "follows", ".", "If", "the", "classifier", "predicts", "that", "β", "in", "in", "r", "2", ",", "we", "mirror", "the", "input", "image", "W", ":", "As", "illustrated", "in", "Fig", ".", "3", "(", "e", ")", ",", "the", "object", "appears", "in", "the", "mirror", "image", "with", "a", "rotation", "angle", "equal", "to", "α", "−", "β", ",", "which", "is", "in", "r", "1", ".", "Therefore", "we", "can", "apply", "g", "Θ", "(", "·", ")", "to", "the", "mirrored", "W", ".", "To", "obtain", "the", "correct", "pose", ",", "we", "finally", "mirror", "back", "the", "projections", "of", "the", "corners", "predicted", "by", "g", "Θ", "(", "·", ")", ".", "We", "currently", "consider", "the", "case", "where", "the", "axis", "of", "symmetry", "is", "more", "or", "less", "vertical", "in", "the", "image", ",", "and", "mirror", "the", "image", "from", "left", "to", "right", ".", "When", "the", "axis", "is", "closer", "to", "be", "horizontal", ",", "we", "should", "mirror", "the", "image", "from", "top", "to", "bottom", ".", "Objects", "of", "revolution", "are", "a", "special", "and", "simpler", "case", ":", "since", "their", "angle", "of", "symmetry", "is", "0", "•", ",", "we", "predict", "their", "poses", "under", "the", "same", "angle", "of", "rotation", ".", "For", "training", "the", "pose", "predictor", "g", "Θ", "(", "·", ")", ",", "we", "use", "the", "original", "training", "images", "with", "angles", "of", "rotation", "in", "r", "1", ",", "and", "mirror", "the", "training", "images", "with", "angles", "of", "rotation", "in", "r", "2", ".", "Handling", "Objects", "that", "are", "'", "Not", "Exactly", "Symmetrical", "'", "As", "mentioned", "in", "the", "introduction", ",", "some", "objects", "of", "the", "T", "-", "LESS", "dataset", "are", "only", "approximately", "symmetrical", ",", "such", "as", "Object", "#", "5", "in", "Fig", ".", "1", "(", "d", ")", ".", "The", "small", "details", "that", "make", "the", "object", "not", "perfectly", "symmetrical", ",", "however", ",", "do", "not", "help", "the", "optimization", "problem", "of", "Eq", ".", "(", "4", ")", ",", "but", "we", "would", "still", "like", "to", "predict", "the", "pose", "of", "this", "object", ".", "In", "the", "case", "of", "Object", "#", "5", ",", "we", "consider", "4", "regions", "instead", "of", "2", ":", "r", "1", "=", "[", "0", ";", "π", "/", "2", "[", ",", "r", "1", "=", "[", "π", "/", "2", ";", "π", "[", ",", "r", "3", "=", "[", "π", ";", "3π", "/", "2", "[", ",", "and", "r", "4", "=", "[", "3π", "/", "2", ";", "2π", "[", ",", "and", "we", "train", "the", "classifier", "k", "(", "·", ")", "to", "predict", "in", "which", "of", "these", "four", "regions", "the", "angle", "of", "rotation", "β", "is", ".", "If", "β", "∈", "r", "2", "or", "β", "∈", "r", "4", ",", "we", "mirror", "the", "image", "before", "computing", "the", "pose", "as", "before", ".", "Then", ",", "if", "β", "∈", "r", "3", "or", "β", "∈", "r", "4", ",", "we", "still", "have", "to", "add", "π", "to", "the", "angle", "of", "rotation", "of", "the", "recovered", "pose", "to", "get", "an", "angle", "between", "0", "and", "2π", ".", "section", ":", "Refining", "the", "Pose", "We", "also", "introduce", "an", "optional", "additional", "stage", "to", "improve", "the", "accuracy", "of", "the", "pose", "estimates", "inspired", "by", "[", "reference", "]", ".", "As", "illustrated", "in", "Fig", ".", "4", ",", "we", "train", "another", "CNN", "that", "predicts", "an", "update", "to", "improve", "the", "pose", ".", "Because", "this", "CNN", "takes", "4", "or", "6", "channels", "as", "input", ",", "it", "is", "not", "clear", "how", "we", "can", "use", "VGG", ",", "as", "we", "did", "for", "the", "previously", "introduced", "networks", ",", "and", "we", "use", "here", "one", "CNN", "per", "object", ".", "However", ",", "this", "stage", "is", "optional", ",", "and", "without", "it", ",", "we", "already", "outperform", "the", "-", "state", "-", "of", "-", "the", "-", "art", ".", "The", "first", "image", "is", "the", "image", "window", "W", "as", "for", "g", "Θ", "(", "·", ")", ".", "The", "second", "image", "depends", "on", "the", "current", "estimate", "of", "the", "pose", ":", "While", "[", "reference", "]", "generates", "a", "depth", "map", "with", "a", "deep", "network", ",", "we", "render", "(", "using", "OpenGL", ")", "either", "a", "binary", "mask", "or", "a", "color", "rendering", "of", "the", "target", "object", "as", "seen", "from", "this", "current", "estimate", ".", "More", "formally", "we", "train", "this", "CNN", "by", "minimizing", ":", "where", "h", "µ", "denotes", "the", "CNN", ",", "µ", "its", "parameters", ";", "N", "(", "e", ",", "t", ")", "is", "a", "set", "of", "poses", "sampled", "around", "pose", "(", "e", ",", "t", ")", ",", "and", "Render", "(", "e", ",", "t", ")", "a", "function", "that", "returns", "a", "binary", "mask", ",", "or", "a", "color", "rendering", ",", "of", "the", "target", "object", "seen", "from", "pose", "(", "e", ",", "t", ")", ".", "At", "run", "-", "time", ",", "given", "a", "current", "estimate", "of", "the", "object", "pose", "represented", "by", "the", "projections", "of", "the", "cornersv", "=", "[", ".", ".", ".m", "i", ".", ".", ".", "]", ",", "and", "the", "corresponding", "parameterisation", "(", "ê", ",", "t", ")", ",", "we", "can", "update", "this", "estimate", "by", "invoking", "h", "µ", "(", "·", ")", ":", "section", ":", "Generating", "Training", "Images", "In", "Section", "4", ",", "we", "will", "compare", "our", "method", "to", "the", "stateof", "-", "the", "art", "for", "3D", "object", "detection", "in", "color", "images", "[", "reference", "]", ",", "and", "like", "them", ",", "for", "each", "of", "15", "objects", "of", "the", "LINEMOD", "dataset", ",", "we", "use", "15", "%", "of", "the", "images", "for", "training", "and", "use", "the", "rest", "for", "testing", ".", "The", "training", "images", "are", "selected", "as", "in", "[", "reference", "]", ",", "such", "that", "relative", "orientation", "between", "them", "should", "be", "larger", "than", "a", "threshold", ".", "We", "also", "tried", "a", "random", "selection", ",", "and", "there", "was", "only", "a", "slight", "drop", "in", "performance", ",", "for", "some", "objects", "only", ".", "The", "selection", "method", "thus", "does", "not", "seem", "critical", ".", "The", "T", "-", "LESS", "dataset", "provides", "regularly", "sampled", "training", "images", ".", "As", "shown", "in", "Fig", ".", "5", ",", "we", "also", "use", "a", "similar", "method", "as", "[", "reference", "]", "to", "augment", "the", "training", "set", ":", "We", "extract", "the", "objects", "'", "silhouettes", "from", "these", "images", ",", "which", "can", "be", "done", "as", "the", "ground", "Figure", "5", ".", "Two", "generated", "training", "images", "for", "different", "objects", "from", "the", "LINEMOD", "dataset", "[", "reference", "]", ".", "The", "object", "is", "shifted", "from", "the", "center", "to", "handle", "the", "inaccuracy", "of", "the", "detection", "method", ",", "and", "the", "background", "is", "random", "to", "make", "sure", "that", "the", "network", "gΘ", "can", "not", "exploit", "the", "context", "specific", "to", "the", "dataset", ".", "truth", "poses", "and", "the", "objects", "'", "3D", "models", "are", "available", ".", "Note", "that", "this", "means", "the", "results", "are", "not", "influenced", "by", "the", "scene", "context", ",", "which", "makes", "the", "pose", "estimation", "more", "difficult", ".", "To", "be", "robust", "to", "clutter", "and", "scale", "changes", ",", "we", "scale", "the", "segmented", "objects", "by", "a", "factor", "of", "s", "∈", "[", "0.8", ",", "1.2", "]", ",", "and", "change", "the", "background", "by", "a", "patch", "extracted", "from", "a", "randomly", "picked", "image", "from", "the", "ImageNet", "dataset", "[", "reference", "]", ".", "Moreover", ",", "the", "object", "is", "shifted", "by", "some", "pixels", "from", "the", "center", "of", "the", "image", "window", "in", "both", "x", "and", "y", "directions", ".", "This", "helps", "us", "to", "handle", "small", "object", "localization", "errors", "made", "during", "detection", ".", "section", ":", "Experiments", "In", "this", "section", ",", "we", "present", "and", "discuss", "the", "results", "of", "our", "evaluation", ".", "We", "first", "describe", "the", "three", "evaluation", "metrics", "used", "in", "the", "literature", "and", "in", "this", "paper", ".", "We", "evaluate", "our", "method", "on", "all", "the", "possible", "datasets", "with", "color", "images", "for", "instance", "3D", "detection", "and", "pose", "estimation", "we", "are", "aware", "of", ":", "the", "LINEMOD", "[", "reference", "]", ",", "Occlusion", "[", "reference", "]", ",", "and", "T", "-", "LESS", "[", "reference", "]", "datasets", ".", "section", ":", "Evaluation", "Metrics", "As", "in", "[", "reference", "]", ",", "we", "use", "the", "percentage", "of", "correctly", "predicted", "poses", "for", "each", "sequence", "and", "each", "object", ",", "where", "a", "pose", "is", "considered", "correct", "if", "it", "passes", "the", "tests", "presented", "below", ".", "section", ":", "2D", "Projections", "[", "2", "]", "This", "is", "a", "metric", "suited", "for", "applications", "such", "as", "augmented", "reality", ".", "A", "pose", "is", "considered", "correct", "if", "the", "average", "of", "the", "2D", "distances", "between", "the", "projections", "of", "the", "object", "'s", "vertices", "from", "the", "estimated", "pose", "and", "the", "ground", "truth", "pose", "is", "less", "than", "5", "pixels", ".", "6D", "Pose", "[", "reference", "]", "With", "this", "metric", ",", "a", "pose", "is", "considered", "correct", "if", "the", "average", "of", "the", "3D", "distances", "between", "the", "transformed", "of", "the", "object", "'s", "vertices", "Table", "1", ".", "Evaluation", "using", "the", "2D", "Projections", "metric", "of", "using", "the", "2D", "projections", "of", "the", "bounding", "box", "(", "'", "BB", "'", ")", ",", "compared", "to", "the", "direct", "prediction", "of", "the", "pose", "(", "'", "Direct", "'", ")", ",", "and", "of", "the", "refinement", "methods", ".", "For", "this", "evaluation", ",", "we", "used", "the", "ground", "truth", "2D", "object", "center", "to", "avoid", "the", "influence", "of", "the", "detection", ".", "For", "the", "objects", "marked", "with", "a", "(", "*", ")", ",", "we", "optimize", "the", "value", "of", "the", "weight", "balancing", "the", "rotation", "and", "translation", "terms", "on", "the", "test", "set", ",", "giving", "an", "advantage", "to", "the", "'", "Direct", "'", "pose", "method", ".", "For", "the", "other", "objects", ",", "we", "used", "the", "value", "that", "is", "optimal", "for", "both", "the", "Ape", "and", "the", "Driller", ".", "is", "less", "than", "10", "%", "of", "the", "object", "'s", "diameter", ".", "V", "is", "the", "set", "of", "the", "object", "'s", "vertices", ",", "(", "ê", ",", "t", ")", "the", "estimated", "pose", "and", "(", "ē", ",", "t", ")", "the", "ground", "truth", "pose", ",", "and", "Tr", "e", ",", "t", "(", "·", ")", "a", "rigid", "transformation", "by", "rotation", "e", ",", "translation", "t.", "For", "the", "objects", "with", "ambigious", "poses", "due", "to", "symmetries", ",", "[", "reference", "]", "replaces", "this", "measure", "by", ":", "5", "cm", "5", "•", "Metric", "[", "reference", "]", "With", "this", "metric", ",", "a", "pose", "is", "considered", "correct", "if", "the", "translation", "and", "rotation", "errors", "are", "below", "5", "cm", "and", "5", "•", "respectively", ".", "section", ":", "Contributions", "of", "the", "Different", "Steps", "The", "columns", "'", "BB", "'", ",", "'", "Mask", "Ref", ".", "[", "reference", "]", ",", "and", "'", "RGB", "Ref", ".", "'", "of", "Table", "1", "compare", "the", "results", "of", "our", "method", "before", "and", "after", "two", "iterations", "of", "refinement", ",", "using", "either", "a", "binary", "mask", "or", "a", "color", "rendering", ".", "For", "this", "evaluation", ",", "we", "used", "the", "ground", "truth", "2D", "object", "center", "to", "avoid", "the", "influence", "of", "the", "detection", ".", "Using", "refinement", "improves", "the", "results", "on", "average", "by", "4.5", "%", "and", "6.3", "%", "for", "the", "mask", "and", "color", "rendering", "respectively", ".", "Using", "a", "color", "rendering", "systematically", "yields", "the", "best", "results", ",", "but", "using", "the", "binary", "mask", "yields", "already", "a", "significant", "improvement", ",", "showing", "that", "an", "untextured", "model", "can", "be", "used", ".", "[", "reference", "]", "and", "our", "method", "without", "and", "with", "RGB", "Refinement", "using", "our", "segmentation", "-", "based", "method", "to", "obtain", "the", "2D", "object", "centers", "on", "the", "LINEMOD", "dataset", ".", "[", "reference", "]", "does", "not", "provide", "results", "for", "the", "Bowl", "and", "the", "Cup", ",", "hence", "for", "the", "sake", "of", "comparison", "the", "average", "is", "taken", "over", "the", "first", "13", "objects", ".", "section", ":", "The", "LINEMOD", "Dataset", ":", "Comparison", "with", "[", "2", "]", "Table", "2", "compares", "our", "BB8", "method", "with", "and", "without", "RGB", "refinement", "against", "the", "one", "presented", "in", "[", "reference", "]", "on", "the", "LINEMOD", "dataset", ".", "Because", "of", "lack", "of", "space", ",", "we", "provide", "the", "results", "without", "refinement", "only", "for", "the", "2D", "Projection", "metric", ",", "however", ",", "the", "results", "for", "the", "other", "metrics", "are", "comparable", ".", "For", "this", "evaluation", ",", "we", "used", "the", "results", "of", "our", "detection", "method", "presented", "in", "Section", "3.1", ",", "not", "the", "ground", "truth", "2D", "object", "center", ".", "Our", "method", "outperforms", "[", "reference", "]", "by", "a", "large", "margin", ":", "15.6", "%", "for", "2D", "Projection", ",", "12.6", "%", "for", "6D", "Pose", "and", "28.4", "%", "for", "the", "5", "cm", "5", "•", "metric", ".", "Fig", ".", "7", "shows", "qualitative", "results", "for", "our", "method", "on", "this", "dataset", ".", "For", "most", "of", "the", "images", ",", "the", "two", "bounding", "boxes", ",", "for", "the", "ground", "truth", "pose", "and", "for", "the", "pose", "we", "estimate", ",", "overlap", "almost", "perfectly", ".", "section", ":", "The", "Occlusion", "Dataset", ":", "Robustness", "to", "Partial", "Occlusions", "The", "Occlusion", "dataset", "was", "created", "by", "[", "reference", "]", "from", "the", "LINEMOD", "dataset", ".", "The", "partial", "occlusions", "make", "it", "significantly", "more", "difficult", ",", "and", "to", "the", "best", "of", "our", "knowledge", ",", "the", "only", "published", "results", "use", "both", "color", "and", "depth", "data", ".", "[", "reference", "]", "provide", "results", "using", "only", "color", "images", ",", "but", "limited", "to", "2D", "detection", ",", "not", "3D", "pose", "estimation", ".", "We", "only", "use", "images", "from", "the", "LINEMOD", "dataset", "to", "generate", "our", "training", "images", "by", "using", "the", "approach", "explained", "in", "Section", "3.5", ",", "except", "that", "we", "also", "randomly", "superimpose", "objects", "extracted", "from", "the", "other", "sequences", "to", "the", "target", "ob", "-", "ject", "to", "be", "robust", "to", "occlusions", ".", "We", "do", "not", "use", "any", "image", "of", "the", "test", "sequence", "to", "avoid", "having", "occlusions", "similar", "to", "the", "ones", "presented", "in", "the", "test", "sequence", ".", "Although", "all", "the", "poses", "in", "the", "test", "sets", "are", "not", "visible", "in", "the", "training", "sequences", ",", "we", "can", "estimate", "accurate", "poses", "with", "a", "2D", "Projection", "error", "lower", "than", "15px", "for", "about", "80", "%", "of", "the", "frames", "for", "these", "seven", "objects", ".", "We", "do", "not", "report", "the", "performance", "of", "our", "method", "for", "the", "Eggbox", ",", "as", "more", "than", "70", "%", "of", "close", "poses", "are", "not", "seen", "in", "the", "training", "sequence", ".", "Some", "qualitative", "results", "are", "shown", "in", "the", "second", "row", "of", "Fig", ".", "7", ".", "To", "the", "best", "of", "our", "knowledge", ",", "we", "are", "the", "first", "to", "present", "results", "on", "this", "dataset", "using", "color", "images", "only", ".", "section", ":", "The", "T", "-", "LESS", "Dataset", ":", "Handling", "Objects", "with", "an", "Axis", "of", "Symmetry", "The", "test", "sequences", "of", "the", "T", "-", "LESS", "dataset", "are", "very", "challenging", ",", "with", "sometimes", "multiple", "instances", "of", "the", "same", "objects", "and", "a", "high", "amount", "of", "clutter", "and", "occlusion", ".", "We", "considered", "only", "Scenes", "#", "1", ",", "#", "2", ",", "#", "4", ",", "#", "5", ",", "and", "#", "7", "in", "our", "experiments", ".", "It", "is", "also", "difficult", "to", "compare", "against", "the", "only", "published", "work", "on", "T", "-", "LESS", "[", "reference", "]", ",", "as", "it", "provides", "the", "6D", "pose", "metric", "averaged", "per", "object", "or", "per", "scene", ",", "computed", "using", "RGB", "-", "D", "data", ",", "while", ",", "to", "the", "best", "of", "our", "knowledge", ",", "we", "are", "the", "first", "to", "report", "results", "on", "the", "T", "-", "LESS", "dataset", "using", "RGB", "images", "Figure", "7", ".", "Some", "qualitative", "results", ".", "First", "row", ":", "LINEMOD", "dataset", ";", "Second", "row", ":", "Occlusion", "dataset", ";", "Third", "row", ":", "T", "-", "LESS", "dataset", "(", "for", "objects", "of", "revolution", ",", "we", "represent", "the", "pose", "with", "a", "cylinder", "rather", "than", "a", "box", ")", ";", "Last", "row", ":", "Some", "failure", "cases", ".", "From", "left", "to", "right", ":", "An", "example", "of", "a", "pose", "rejected", "by", "the", "2D", "Projections", "metric", ",", "a", "failure", "due", "to", "the", "lack", "of", "corresponding", "poses", "in", "the", "training", "set", ",", "two", "examples", "from", "T", "-", "LESS", "rejected", "by", "the", "6D", "pose", "metric", ",", "and", "one", "failure", "due", "to", "the", "fact", "that", "some", "objects", "are", "made", "of", "several", "instances", "of", "another", "object", ".", "Table", "3", ".", "Our", "quantitative", "results", "on", "T", "-", "LESS", "[", "reference", "]", ".", "Most", "of", "the", "errors", "are", "along", "the", "z", "axis", "of", "the", "camera", ",", "as", "we", "rely", "on", "color", "images", ".", "only", ".", "Similarly", "to", "[", "reference", "]", ",", "we", "evaluate", "the", "poses", "with", "more", "than", "10", "%", "of", "the", "object", "surface", "visible", "in", "the", "ground", "truth", "poses", ".", "As", "shown", "in", "Table", "3", ",", "the", "6D", "Pose", "average", "per", "scene", "with", "our", "method", "is", "54", "%", ".", "The", "object", "3D", "orientation", "and", "translation", "along", "the", "x", "and", "y", "axes", "of", "the", "camera", "are", "typically", "very", "well", "estimated", ",", "and", "most", "of", "the", "error", "is", "along", "the", "z", "axis", ",", "which", "should", "not", "be", "surprising", "for", "a", "method", "using", "color", "images", "only", ".", "section", ":", "Computation", "Times", "Our", "implementation", "takes", "140", "ms", "for", "the", "segmentation", ",", "130", "ms", "for", "the", "pose", "prediction", ",", "and", "21", "ms", "for", "each", "refinement", "iteration", ",", "on", "an", "Intel", "Core", "i7", "-", "5820", "K", "3.30", "GHz", "desktop", "with", "a", "GeForce", "TITAN", "X.", "If", "there", "is", "only", "one", "object", "of", "interest", ",", "we", "can", "replace", "VGG", "by", "a", "specific", "network", "with", "a", "simpler", "architecture", ",", "the", "computation", "times", "then", "become", "20", "ms", "for", "the", "segmentation", "and", "12", "ms", "for", "the", "pose", "prediction", ",", "with", "similar", "accuracy", ".", "section", ":", "Conclusion", "Our", "\"", "holistic", "\"", "approach", ",", "made", "possible", "by", "the", "remarkable", "abilities", "of", "Deep", "Networks", "for", "regression", ",", "allowed", "us", "to", "significantly", "advance", "the", "state", "-", "of", "-", "the", "-", "art", "on", "3D", "pose", "estimation", "from", "color", "images", ",", "even", "on", "challenging", "objects", "from", "the", "T", "-", "LESS", "dataset", ".", "section", ":", "section", ":", "Acknowledgment", ":", "This", "work", "was", "funded", "by", "the", "Christian", "Doppler", "Laboratory", "for", "Semantic", "3D", "Computer", "Vision", ".", "section", ":"], "entity": [[0, 1, ["BB8"], "Method"], [9, 12, ["Partial", "Occlusion", "Method"], "Method"], [13, 20, ["Predicting", "the", "3D", "Poses", "of", "Challenging", "Objects"], "Task"], [32, 35, ["3D", "object", "detection"], "Task"], [36, 38, ["pose", "estimation"], "Task"], [46, 47, ["segmentation"], "Method"], [50, 56, ["objects", "of", "interest", "in", "2D", "even"], "Task"], [69, 73, ["patch", "-", "based", "methods"], "Method"], [79, 82, ["holistic", "\"", "approach"], "Method"], [90, 93, ["Convolutional", "Neural", "Network"], "Method"], [94, 95, ["CNN"], "Method"], [161, 162, ["CNN"], "Method"], [182, 183, ["classifier"], "Method"], [223, 225, ["LINEMOD", "dataset"], "Material"], [237, 239, ["RGB", "frames"], "Material"], [250, 252, ["Occlusion", "dataset"], "Material"], [268, 271, ["Pose", "6D", "criterion"], "Metric"], [352, 353, ["robotics"], "Task"], [354, 358, ["virtual", "and", "augmented", "reality"], "Task"], [431, 433, ["LINEMOD", "dataset"], "Material"], [444, 446, ["Occlusion", "dataset"], "Material"], [503, 504, ["accuracy"], "Metric"], [585, 587, ["learning", "algorithms"], "Method"], [607, 609, ["active", "sensor"], "Method"], [621, 624, ["3D", "pose", "estimation"], "Task"], [674, 676, ["3D", "pose"], "Task"], [682, 684, ["PnP", "algorithm"], "Method"], [737, 739, ["robust", "optimization"], "Method"], [756, 759, ["holistic", "\"", "approach"], "Method"], [808, 810, ["object", "segmentation"], "Method"], [819, 822, ["sliding", "window", "detector"], "Method"], [835, 836, ["CNN"], "Method"], [839, 845, ["3D", "pose", "of", "the", "detected", "objects"], "Task"], [865, 866, ["accuracy"], "Metric"], [880, 882, ["object", "parts"], "Task"], [911, 913, ["PnP", "algorithm"], "Method"], [917, 920, ["object", "coordinate", "approaches"], "Method"], [930, 932, ["robust", "estimation"], "Method"], [939, 944, ["direct", "prediction", "of", "the", "pose"], "Task"], [1155, 1157, ["Bowl", "object"], "Material"], [1159, 1161, ["LINEMOD", "dataset"], "Material"], [1256, 1257, ["classifer"], "Method"], [1277, 1280, ["holistic", "\"", "approach"], "Method"], [1283, 1284, ["classifier"], "Method"], [1308, 1309, ["CNN"], "Method"], [1372, 1374, ["pose", "estimates"], "Task"], [1378, 1380, ["feedback", "loop"], "Method"], [1387, 1389, ["hand", "detection"], "Task"], [1419, 1421, ["pose", "estimate"], "Task"], [1433, 1434, ["LINEMOD"], "Material"], [1435, 1437, ["Occlusion", "datasets"], "Material"], [1445, 1446, ["BB8"], "Method"], [1467, 1469, ["Deep", "Networks"], "Method"], [1517, 1520, ["3D", "object", "detection"], "Task"], [1533, 1537, ["Keypoint", "-", "based", "methods"], "Method"], [1562, 1564, ["3D", "cameras"], "Method"], [1605, 1607, ["decision", "tree"], "Method"], [1630, 1634, ["template", "-", "based", "representation"], "Method"], [1636, 1639, ["RGB", "-", "D"], "Material"], [1640, 1642, ["RGB", "data"], "Material"], [1646, 1649, ["large", "scale", "detection"], "Task"], [1656, 1658, ["template", "approach"], "Method"], [1680, 1683, ["local", "patches", "recognition"], "Method"], [1685, 1687, ["Random", "Forests"], "Method"], [1702, 1704, ["Random", "Forest"], "Method"], [1730, 1732, ["energy", "function"], "Method"], [1772, 1774, ["energy", "function"], "Method"], [1778, 1780, ["RANSAC", "procedure"], "Method"], [1797, 1798, ["CNN"], "Method"], [1809, 1811, ["3D", "rendering"], "Method"], [1849, 1851, ["RGB", "data"], "Material"], [1871, 1873, ["Random", "Forests"], "Method"], [1893, 1894, ["RANSAC"], "Method"], [1916, 1917, ["robustness"], "Metric"], [1941, 1946, ["sparse", "2D", "-", "3D", "correspondences"], "Task"], [1978, 1979, ["CNN"], "Method"], [1999, 2004, ["camera", "relocalisation", "in", "urban", "environment"], "Task"], [2006, 2009, ["3D", "object", "detection"], "Task"], [2019, 2020, ["CNN"], "Method"], [2106, 2107, ["CNN"], "Method"], [2110, 2115, ["3D", "pose", "of", "generic", "objects"], "Task"], [2155, 2157, ["LINEMOD", "dataset"], "Material"], [2176, 2178, ["segmentation", "methods"], "Method"], [2194, 2196, ["RGB", "data"], "Material"], [2220, 2221, ["accuracy"], "Metric"], [2275, 2277, ["pose", "estimates"], "Task"], [2314, 2317, ["2D", "object", "detector"], "Method"], [2325, 2326, ["segmentation"], "Method"], [2349, 2351, ["sliding", "window"], "Method"], [2358, 2360, ["2D", "detection"], "Task"], [2368, 2371, ["correct", "detection", "rate"], "Metric"], [2374, 2377, ["IoU", "of", "0.5"], "Metric"], [2382, 2385, ["low", "resolution", "segmentation"], "Method"], [2391, 2395, ["hourglass", "-", "shaped", "architecture"], "Method"], [2402, 2403, ["segmentation"], "Task"], [2419, 2427, ["level", "coarse", "-", "to", "-", "fine", "object", "segmentation"], "Task"], [2450, 2452, ["last", "layer"], "Method"], [2453, 2454, ["VGG"], "Method"], [2459, 2462, ["fully", "connected", "layer"], "Method"], [2486, 2489, ["ad", "hoc", "architecture"], "Method"], [2502, 2506, ["low", "resolution", "binary", "segmentation"], "Task"], [2523, 2525, ["objective", "function"], "Metric"], [2554, 2555, ["network"], "Method"], [2577, 2578, ["LINEMOD"], "Material"], [2579, 2581, ["Occlusion", "datasets"], "Material"], [2615, 2616, ["segmentations"], "Task"], [2630, 2632, ["binary", "segmentation"], "Method"], [2655, 2657, ["binary", "segmentation"], "Task"], [2684, 2690, ["binary", "segmentation", "S", "1", ",", "o"], "Method"], [2707, 2709, ["connected", "component"], "Method"], [2826, 2827, ["VGG"], "Method"], [2840, 2842, ["convolutional", "layers"], "Method"], [2844, 2846, ["pooling", "layers"], "Method"], [2851, 2856, ["segmentation", "S", "2", ",", "o"], "Method"], [2916, 2918, ["3D", "Pose"], "Task"], [2929, 2931, ["Deep", "Network"], "Method"], [2953, 2954, ["segmentation"], "Task"], [2957, 2958, ["VGG"], "Method"], [3027, 3033, ["predicting", "the", "poses", "of", "object", "parts"], "Task"], [3043, 3045, ["cost", "function"], "Metric"], [3050, 3053, ["network", "g", "Θ"], "Method"], [3184, 3185, ["classifier"], "Method"], [3239, 3240, ["segmentation"], "Method"], [3291, 3293, ["PnP", "algorithm"], "Method"], [3361, 3363, ["LINEMOD", "dataset"], "Material"], [3550, 3552, ["pose", "prediction"], "Task"], [3625, 3629, ["optimization", "problem", "of", "Eq"], "Task"], [3666, 3669, ["CNN", "classifier", "k"], "Method"], [3717, 3723, ["g", "Θ", "(", "·", ")", "network"], "Method"], [3738, 3741, ["network", "g", "Θ"], "Method"], [3757, 3758, ["classifier"], "Method"], [3810, 3812, ["g", "Θ"], "Method"], [3923, 3927, ["pose", "predictor", "g", "Θ"], "Task"], [4016, 4020, ["optimization", "problem", "of", "Eq"], "Task"], [4103, 4105, ["classifier", "k"], "Method"], [4197, 4198, ["accuracy"], "Metric"], [4200, 4202, ["pose", "estimates"], "Task"], [4218, 4219, ["CNN"], "Method"], [4230, 4231, ["CNN"], "Method"], [4247, 4248, ["VGG"], "Method"], [4263, 4264, ["CNN"], "Method"], [4329, 4331, ["deep", "network"], "Method"], [4336, 4337, ["OpenGL"], "Method"], [4362, 4363, ["CNN"], "Method"], [4371, 4372, ["CNN"], "Method"], [4487, 4489, ["Training", "Images"], "Material"], [4505, 4508, ["3D", "object", "detection"], "Task"], [4526, 4528, ["LINEMOD", "dataset"], "Material"], [4573, 4575, ["random", "selection"], "Method"], [4592, 4594, ["selection", "method"], "Method"], [4663, 4665, ["LINEMOD", "dataset"], "Material"], [4682, 4684, ["detection", "method"], "Method"], [4695, 4697, ["network", "gΘ"], "Method"], [4713, 4715, ["3D", "models"], "Method"], [4735, 4737, ["pose", "estimation"], "Task"], [4781, 4783, ["ImageNet", "dataset"], "Material"], [4815, 4819, ["small", "object", "localization", "errors"], "Task"], [4821, 4822, ["detection"], "Task"], [4845, 4847, ["evaluation", "metrics"], "Metric"], [4870, 4872, ["3D", "detection"], "Task"], [4873, 4875, ["pose", "estimation"], "Task"], [4881, 4882, ["LINEMOD"], "Material"], [4886, 4887, ["Occlusion"], "Material"], [4892, 4895, ["T", "-", "LESS"], "Material"], [4902, 4904, ["Evaluation", "Metrics"], "Metric"], [4955, 4957, ["augmented", "reality"], "Task"], [5028, 5031, ["2D", "Projections", "metric"], "Metric"], [5063, 5065, ["refinement", "methods"], "Method"], [5084, 5085, ["detection"], "Task"], [5120, 5124, ["Direct", "'", "pose", "method"], "Method"], [5140, 5141, ["Ape"], "Method"], [5143, 5144, ["Driller"], "Method"], [5303, 5305, ["color", "rendering"], "Method"], [5324, 5325, ["detection"], "Task"], [5327, 5328, ["refinement"], "Method"], [5341, 5345, ["mask", "and", "color", "rendering"], "Task"], [5349, 5351, ["color", "rendering"], "Method"], [5371, 5373, ["untextured", "model"], "Method"], [5386, 5388, ["RGB", "Refinement"], "Method"], [5390, 5394, ["segmentation", "-", "based", "method"], "Method"], [5402, 5404, ["LINEMOD", "dataset"], "Material"], [5455, 5457, ["RGB", "refinement"], "Method"], [5467, 5469, ["LINEMOD", "dataset"], "Material"], [5485, 5488, ["2D", "Projection", "metric"], "Metric"], [5510, 5512, ["detection", "method"], "Method"], [5539, 5541, ["2D", "Projection"], "Task"], [5556, 5557, ["metric"], "Metric"], [5601, 5603, ["Occlusion", "Dataset"], "Material"], [5604, 5605, ["Robustness"], "Metric"], [5609, 5611, ["Occlusion", "dataset"], "Material"], [5619, 5621, ["LINEMOD", "dataset"], "Material"], [5663, 5665, ["2D", "detection"], "Task"], [5667, 5670, ["3D", "pose", "estimation"], "Task"], [5677, 5679, ["LINEMOD", "dataset"], "Material"], [5764, 5767, ["2D", "Projection", "error"], "Metric"], [5793, 5794, ["Eggbox"], "Task"], [5927, 5930, ["T", "-", "LESS"], "Material"], [5938, 5941, ["6D", "pose", "metric"], "Metric"], [5978, 5980, ["RGB", "images"], "Material"], [5990, 5992, ["LINEMOD", "dataset"], "Material"], [5996, 5998, ["Occlusion", "dataset"], "Material"], [6045, 6048, ["2D", "Projections", "metric"], "Method"], [6066, 6069, ["T", "-", "LESS"], "Material"], [6072, 6075, ["6D", "pose", "metric"], "Metric"], [6102, 6105, ["T", "-", "LESS"], "Material"], [6164, 6169, ["6D", "Pose", "average", "per", "scene"], "Metric"], [6223, 6225, ["Computation", "Times"], "Metric"], [6232, 6233, ["segmentation"], "Task"], [6238, 6240, ["pose", "prediction"], "Task"], [6246, 6248, ["refinement", "iteration"], "Method"], [6277, 6278, ["VGG"], "Method"], [6288, 6290, ["computation", "times"], "Metric"], [6296, 6297, ["segmentation"], "Task"], [6302, 6304, ["pose", "prediction"], "Task"], [6307, 6308, ["accuracy"], "Metric"], [6314, 6317, ["holistic", "\"", "approach"], "Method"], [6325, 6327, ["Deep", "Networks"], "Method"], [6328, 6329, ["regression"], "Task"], [6344, 6347, ["3D", "pose", "estimation"], "Task"], [6378, 6382, ["Semantic", "3D", "Computer", "Vision"], "Task"], [130, 133, ["T", "-", "LESS"], "Material"], [278, 281, ["T", "-", "LESS"], "Material"], [334, 336, ["pose", "estimation"], "Task"], [461, 464, ["T", "-", "LESS"], "Material"], [975, 978, ["T", "-", "LESS"], "Material"], [1352, 1355, ["T", "-", "LESS"], "Material"], [1609, 1612, ["RGB", "-", "D"], "Material"], [1835, 1838, ["RGB", "-", "D"], "Material"], [2117, 2120, ["RGB", "-", "D"], "Material"], [2602, 2605, ["T", "-", "LESS"], "Material"], [3346, 3349, ["T", "-", "LESS"], "Material"], [3977, 3980, ["T", "-", "LESS"], "Material"], [4601, 4604, ["T", "-", "LESS"], "Material"], [5438, 5439, ["LINEMOD"], "Material"], [5450, 5451, ["BB8"], "Method"], [5850, 5853, ["T", "-", "LESS"], "Material"], [5867, 5870, ["T", "-", "LESS"], "Material"], [5950, 5953, ["RGB", "-", "D"], "Material"], [5973, 5976, ["T", "-", "LESS"], "Material"], [6002, 6005, ["T", "-", "LESS"], "Material"], [6357, 6360, ["T", "-", "LESS"], "Material"]]} {"doc_id": "3cf31ecb2724b5088783d7c96a5fc0d5604cbf41", "text": ["document", ":", "Simple", "and", "Accurate", "Dependency", "Parsing", "Using", "Bidirectional", "LSTM", "Feature", "Representations", "We", "present", "a", "simple", "and", "effective", "scheme", "for", "dependency", "parsing", "which", "is", "based", "on", "bidirectional", "-", "LSTMs", "(", "BiLSTMs", ")", ".", "Each", "sentence", "token", "is", "associated", "with", "a", "BiLSTM", "vector", "representing", "the", "token", "in", "its", "sentential", "context", ",", "and", "feature", "vectors", "are", "constructed", "by", "concatenating", "a", "few", "BiLSTM", "vectors", ".", "The", "BiLSTM", "is", "trained", "jointly", "with", "the", "parser", "objective", ",", "resulting", "in", "very", "effective", "feature", "extractors", "for", "parsing", ".", "We", "demonstrate", "the", "effectiveness", "of", "the", "approach", "by", "applying", "it", "to", "a", "greedy", "transition", "-", "based", "parser", "as", "well", "as", "to", "a", "globally", "optimized", "graph", "-", "based", "parser", ".", "The", "resulting", "parsers", "have", "very", "simple", "architectures", ",", "and", "match", "or", "surpass", "the", "state", "-", "of", "-", "the", "-", "art", "accuracies", "on", "English", "and", "Chinese", ".", "noitemsep", ",", "topsep=0pt", ",", "parsep=0pt", ",", "partopsep=0pt", "1.0em", "section", ":", "Introduction", "The", "focus", "of", "this", "paper", "is", "on", "feature", "representation", "for", "dependency", "parsing", ",", "using", "recent", "techniques", "from", "the", "neural", "-", "networks", "(", "‘", "‘", "deep", "learning", "’", "’", ")", "literature", ".", "Modern", "approaches", "to", "dependency", "parsing", "can", "be", "broadly", "categorized", "into", "graph", "-", "based", "and", "transition", "-", "based", "parsers", ".", "Graph", "-", "based", "parsers", "treat", "parsing", "as", "a", "search", "-", "based", "structured", "prediction", "problem", "in", "which", "the", "goal", "is", "learning", "a", "scoring", "function", "over", "dependency", "trees", "such", "that", "the", "correct", "tree", "is", "scored", "above", "all", "other", "trees", ".", "Transition", "-", "based", "parsers", "treat", "parsing", "as", "a", "sequence", "of", "actions", "that", "produce", "a", "parse", "tree", ",", "and", "a", "classifier", "is", "trained", "to", "score", "the", "possible", "actions", "at", "each", "stage", "of", "the", "process", "and", "guide", "the", "parsing", "process", ".", "Perhaps", "the", "simplest", "graph", "-", "based", "parsers", "are", "arc", "-", "factored", "(", "first", "order", ")", "models", ",", "in", "which", "the", "scoring", "function", "for", "a", "tree", "decomposes", "over", "the", "individual", "arcs", "of", "the", "tree", ".", "More", "elaborate", "models", "look", "at", "larger", "(", "overlapping", ")", "parts", ",", "requiring", "more", "sophisticated", "inference", "and", "training", "algorithms", ".", "The", "basic", "transition", "-", "based", "parsers", "work", "in", "a", "greedy", "manner", ",", "performing", "a", "series", "of", "locally", "-", "optimal", "decisions", ",", "and", "boast", "very", "fast", "parsing", "speeds", ".", "More", "advanced", "transition", "-", "based", "parsers", "introduce", "some", "search", "into", "the", "process", "using", "a", "beam", "or", "dynamic", "programming", ".", "Regardless", "of", "the", "details", "of", "the", "parsing", "framework", "being", "used", ",", "a", "crucial", "step", "in", "parser", "design", "is", "choosing", "the", "right", "feature", "function", "for", "the", "underlying", "statistical", "model", ".", "Recent", "work", "(", "see", "Section", "[", "reference", "]", "for", "an", "overview", ")", "attempt", "to", "alleviate", "parts", "of", "the", "feature", "function", "design", "problem", "by", "moving", "from", "linear", "to", "non", "-", "linear", "models", ",", "enabling", "the", "modeler", "to", "focus", "on", "a", "small", "set", "of", "‘", "‘", "core", "’", "’", "features", "and", "leaving", "it", "up", "to", "the", "machine", "-", "learning", "machinery", "to", "come", "up", "with", "good", "feature", "combinations", ".", "However", ",", "the", "need", "to", "carefully", "define", "a", "set", "of", "core", "features", "remains", ".", "For", "example", ",", "the", "work", "of", "chen2014fast", "uses", "18", "different", "elements", "in", "its", "feature", "function", ",", "while", "the", "work", "of", "pei2015effective", "uses", "21", "different", "elements", ".", "Other", "works", ",", "notably", "dyer2015transitionbased", "and", "le2014insideoutside", ",", "propose", "more", "sophisticated", "feature", "representations", ",", "in", "which", "the", "feature", "engineering", "is", "replaced", "with", "architecture", "engineering", ".", "In", "this", "work", ",", "we", "suggest", "an", "approach", "which", "is", "much", "simpler", "in", "terms", "of", "both", "feature", "engineering", "and", "architecture", "engineering", ".", "Our", "proposal", "(", "Section", "[", "reference", "]", ")", "is", "centered", "around", "BiRNNs", ",", "and", "more", "specifically", "BiLSTMs", ",", "which", "are", "strong", "and", "trainable", "sequence", "models", "(", "see", "Section", "[", "reference", "]", ")", ".", "The", "BiLSTM", "excels", "at", "representing", "elements", "in", "a", "sequence", "(", "i.e.", ",", "words", ")", "together", "with", "their", "contexts", ",", "capturing", "the", "element", "and", "an", "‘", "‘", "infinite", "’", "’", "window", "around", "it", ".", "We", "represent", "each", "word", "by", "its", "BiLSTM", "encoding", ",", "and", "use", "a", "concatenation", "of", "a", "minimal", "set", "of", "such", "BiLSTM", "encodings", "as", "our", "feature", "function", ",", "which", "is", "then", "passed", "to", "a", "non", "-", "linear", "scoring", "function", "(", "multi", "-", "layer", "perceptron", ")", ".", "Crucially", ",", "the", "BiLSTM", "is", "trained", "with", "the", "rest", "of", "the", "parser", "in", "order", "to", "learn", "a", "good", "feature", "representation", "for", "the", "parsing", "problem", ".", "If", "we", "set", "aside", "the", "inherent", "complexity", "of", "the", "BiLSTM", "itself", "and", "treat", "it", "as", "a", "black", "box", ",", "our", "proposal", "results", "in", "a", "pleasingly", "simple", "feature", "extractor", ".", "We", "demonstrate", "the", "effectiveness", "of", "the", "approach", "by", "using", "the", "BiLSTM", "feature", "extractor", "in", "two", "parsing", "architectures", ",", "transition", "-", "based", "(", "Section", "[", "reference", "]", ")", "as", "well", "as", "a", "graph", "-", "based", "(", "Section", "[", "reference", "]", ")", ".", "In", "the", "graph", "-", "based", "parser", ",", "we", "jointly", "train", "a", "structured", "-", "prediction", "model", "on", "top", "of", "a", "BiLSTM", ",", "propagating", "errors", "from", "the", "structured", "objective", "all", "the", "way", "back", "to", "the", "BiLSTM", "feature", "-", "encoder", ".", "To", "the", "best", "of", "our", "knowledge", ",", "we", "are", "the", "first", "to", "perform", "such", "end", "-", "to", "-", "end", "training", "of", "a", "structured", "prediction", "model", "and", "a", "recurrent", "feature", "extractor", "for", "non", "-", "sequential", "outputs", ".", "Aside", "from", "the", "novelty", "of", "the", "BiLSTM", "feature", "extractor", "and", "the", "end", "-", "to", "-", "end", "structured", "training", ",", "we", "rely", "on", "existing", "models", "and", "techniques", "from", "the", "parsing", "and", "structured", "prediction", "literature", ".", "We", "stick", "to", "the", "simplest", "parsers", "in", "each", "category", "–", "greedy", "inference", "for", "the", "transition", "-", "based", "architecture", ",", "and", "a", "first", "-", "order", ",", "arc", "-", "factored", "model", "for", "the", "graph", "-", "based", "architecture", ".", "Despite", "the", "simplicity", "of", "the", "parsing", "architectures", "and", "the", "feature", "functions", ",", "we", "achieve", "near", "state", "-", "of", "-", "the", "-", "art", "parsing", "accuracies", "in", "both", "English", "(", "93.1", "UAS", ")", "and", "Chinese", "(", "86.6", "UAS", ")", ",", "using", "a", "first", "-", "order", "parser", "with", "two", "features", "and", "while", "training", "solely", "on", "Treebank", "data", ",", "without", "relying", "on", "semi", "-", "supervised", "signals", "such", "as", "pre", "-", "trained", "word", "embeddings", ",", "word", "-", "clusters", ",", "or", "techniques", "such", "as", "tri", "-", "training", ".", "When", "also", "including", "pre", "-", "trained", "word", "embeddings", ",", "we", "obtain", "further", "improvements", ",", "with", "accuracies", "of", "93.9", "UAS", "(", "English", ")", "and", "87.6", "UAS", "(", "Chinese", ")", "for", "a", "greedy", "transition", "-", "based", "parser", "with", "11", "features", ",", "and", "93.6", "UAS", "(", "En", ")", "/", "87.4", "(", "Ch", ")", "for", "a", "greedy", "transition", "-", "based", "parser", "with", "4", "features", ".", "section", ":", "Background", "and", "Notation", "paragraph", ":", "Notation", "We", "use", "to", "denote", "a", "sequence", "of", "vectors", ".", "is", "a", "function", "parameterized", "with", "parameters", ".", "We", "write", "as", "shorthand", "for", "–", "an", "instantiation", "of", "with", "a", "specific", "set", "of", "parameters", ".", "We", "use", "to", "denote", "a", "vector", "concatenation", "operation", ",", "and", "to", "denote", "an", "indexing", "operation", "taking", "the", "th", "element", "of", "a", "vector", ".", "subsection", ":", "Feature", "Functions", "in", "Dependency", "Parsing", "Traditionally", ",", "state", "-", "of", "-", "the", "-", "art", "parsers", "rely", "on", "linear", "models", "over", "hand", "-", "crafted", "feature", "functions", ".", "The", "feature", "functions", "look", "at", "core", "components", "(", "e.g.", "‘", "‘", "word", "on", "top", "of", "stack", "’", "’", ",", "‘", "‘", "leftmost", "child", "of", "the", "second", "-", "to", "-", "top", "word", "on", "the", "stack", "’", "’", ",", "‘", "‘", "distance", "between", "the", "head", "and", "the", "modifier", "words", "’", "’", ")", ",", "and", "are", "comprised", "of", "several", "templates", ",", "where", "each", "template", "instantiates", "a", "binary", "indicator", "function", "over", "a", "conjunction", "of", "core", "elements", "(", "resulting", "in", "features", "of", "the", "form", "‘", "‘", "word", "on", "top", "of", "stack", "is", "X", "and", "leftmost", "child", "is", "Y", "and", "…", "’’", ")", ".", "The", "design", "of", "the", "feature", "function", "–", "which", "components", "to", "consider", "and", "which", "combinations", "of", "components", "to", "include", "–", "is", "a", "major", "challenge", "in", "parser", "design", ".", "Once", "a", "good", "feature", "function", "is", "proposed", "in", "a", "paper", "it", "is", "usually", "adopted", "in", "later", "works", ",", "and", "sometimes", "tweaked", "to", "improve", "performance", ".", "Examples", "of", "good", "feature", "functions", "are", "the", "feature", "-", "set", "proposed", "by", "zhang11acl", "for", "transition", "-", "based", "parsing", "(", "including", "roughly", "20", "core", "components", "and", "72", "feature", "templates", ")", ",", "and", "the", "feature", "-", "set", "proposed", "by", "mst", "for", "graph", "-", "based", "parsing", ",", "with", "the", "paper", "listing", "18", "templates", "for", "a", "first", "-", "order", "parser", ",", "while", "the", "first", "order", "feature", "-", "extractor", "in", "the", "actual", "implementation", "’s", "code", "(", "MSTParser", ")", "includes", "roughly", "a", "hundred", "feature", "templates", ".", "The", "core", "features", "in", "a", "transition", "-", "based", "parser", "usually", "look", "at", "information", "such", "as", "the", "word", "-", "identity", "and", "part", "-", "of", "-", "speech", "(", "POS", ")", "tags", "of", "a", "fixed", "number", "of", "words", "on", "top", "of", "the", "stack", ",", "a", "fixed", "number", "of", "words", "on", "the", "top", "of", "the", "buffer", ",", "the", "modifiers", "(", "usually", "left", "-", "most", "and", "right", "-", "most", ")", "of", "items", "on", "the", "stack", "and", "on", "the", "buffer", ",", "the", "number", "of", "modifiers", "of", "these", "elements", ",", "parents", "of", "words", "on", "the", "stack", ",", "and", "the", "length", "of", "the", "spans", "spanned", "by", "the", "words", "on", "the", "stack", ".", "The", "core", "features", "of", "a", "first", "-", "order", "graph", "-", "based", "parser", "usually", "take", "into", "account", "the", "word", "and", "POS", "of", "the", "head", "and", "modifier", "items", ",", "as", "well", "as", "POS", "-", "tags", "of", "the", "items", "around", "the", "head", "and", "modifier", ",", "POS", "tags", "of", "items", "between", "the", "head", "and", "modifier", ",", "and", "the", "distance", "and", "direction", "between", "the", "head", "and", "modifier", ".", "subsection", ":", "Related", "Research", "Efforts", "Coming", "up", "with", "a", "good", "feature", "-", "set", "for", "a", "parser", "is", "a", "hard", "and", "time", "consuming", "task", ",", "and", "many", "researchers", "attempt", "to", "reduce", "the", "required", "manual", "effort", ".", "The", "work", "of", "lei", "-", "EtAl:2014:P14", "-", "1", "suggests", "a", "low", "-", "rank", "tensor", "representation", "to", "automatically", "find", "good", "feature", "combinations", ".", "taubtabib2015template", "suggest", "a", "kernel", "-", "based", "approach", "to", "implicitly", "consider", "all", "possible", "feature", "combinations", "over", "sets", "of", "core", "-", "features", ".", "The", "recent", "popularity", "of", "neural", "networks", "prompted", "a", "move", "from", "templates", "of", "sparse", ",", "binary", "indicator", "features", "to", "dense", "core", "feature", "encodings", "fed", "into", "non", "-", "linear", "classifiers", ".", "chen2014fast", "encode", "each", "core", "feature", "of", "a", "greedy", "transition", "-", "based", "parser", "as", "a", "dense", "low", "-", "dimensional", "vector", ",", "and", "the", "vectors", "are", "then", "concatenated", "and", "fed", "into", "a", "non", "-", "linear", "classifier", "(", "multi", "-", "layer", "perceptron", ")", "which", "can", "potentially", "capture", "arbitrary", "feature", "combinations", ".", "weiss2015structured", "showed", "further", "gains", "using", "the", "same", "approach", "coupled", "with", "a", "somewhat", "improved", "set", "of", "core", "features", ",", "a", "more", "involved", "network", "architecture", "with", "skip", "-", "layers", ",", "beam", "search", "-", "decoding", ",", "and", "careful", "hyper", "-", "parameter", "tuning", ".", "pei2015effective", "apply", "a", "similar", "methodology", "to", "graph", "-", "based", "parsing", ".", "While", "the", "move", "to", "neural", "-", "network", "classifiers", "alleviates", "the", "need", "for", "hand", "-", "crafting", "feature", "-", "combinations", ",", "the", "need", "to", "carefully", "define", "a", "set", "of", "core", "features", "remain", ".", "For", "example", ",", "the", "feature", "representation", "in", "chen2014fast", "is", "a", "concatenation", "of", "18", "word", "vectors", ",", "18", "POS", "vectors", "and", "12", "dependency", "-", "label", "vectors", ".", "The", "above", "works", "tackle", "the", "effort", "in", "hand", "-", "crafting", "effective", "feature", "combinations", ".", "A", "different", "line", "of", "work", "attacks", "the", "feature", "-", "engineering", "problem", "by", "suggesting", "novel", "neural", "-", "network", "architectures", "for", "encoding", "the", "parser", "state", ",", "including", "intermediately", "-", "built", "subtrees", ",", "as", "vectors", "which", "are", "then", "fed", "to", "non", "-", "linear", "classifiers", ".", "Titov", "and", "Henderson", "encode", "the", "parser", "state", "using", "incremental", "sigmoid", "-", "belief", "networks", "titov", "-", "henderson:2007:IWPT2007", ".", "In", "the", "work", "of", "dyer2015transitionbased", ",", "the", "entire", "stack", "and", "buffer", "of", "a", "transition", "-", "based", "parser", "are", "encoded", "as", "a", "stack", "-", "LSTMs", ",", "where", "each", "stack", "element", "is", "itself", "based", "on", "a", "compositional", "representation", "of", "parse", "trees", ".", "le2014insideoutside", "encode", "each", "tree", "node", "as", "two", "compositional", "representations", "capturing", "the", "inside", "and", "outside", "structures", "around", "the", "node", ",", "and", "feed", "the", "representations", "into", "a", "reranker", ".", "A", "similar", "reranking", "approach", ",", "this", "time", "based", "on", "convolutional", "neural", "networks", ",", "is", "taken", "by", "zhu2015reranking", ".", "Finally", ",", "in", "kiperwasser2016ef", "we", "present", "an", "Easy", "-", "First", "parser", "based", "on", "a", "novel", "hierarchical", "-", "LSTM", "tree", "encoding", ".", "In", "contrast", "to", "these", ",", "the", "approach", "we", "present", "in", "this", "work", "results", "in", "much", "simpler", "feature", "functions", ",", "without", "resorting", "to", "elaborate", "network", "architectures", "or", "compositional", "tree", "representations", ".", "Work", "by", "vinlays2014grammar", "employs", "a", "sequence", "-", "to", "-", "sequence", "with", "attention", "architecture", "for", "constituency", "parsing", ".", "Each", "token", "in", "the", "input", "sentence", "is", "encoded", "in", "a", "deep", "-", "BiLSTM", "representation", ",", "and", "then", "the", "tokens", "are", "fed", "as", "input", "to", "a", "deep", "-", "LSTM", "that", "predicts", "a", "sequence", "of", "bracketing", "actions", "based", "on", "the", "already", "predicted", "bracketing", "as", "well", "as", "the", "encoded", "BiLSTM", "vectors", ".", "A", "trainable", "attention", "mechanism", "is", "used", "to", "guide", "the", "parser", "to", "relevant", "BiLSTM", "vectors", "at", "each", "stage", ".", "This", "architecture", "shares", "with", "ours", "the", "use", "of", "BiLSTM", "encoding", "and", "end", "-", "to", "-", "end", "training", ".", "The", "sequence", "of", "bracketing", "actions", "can", "be", "interpreted", "as", "a", "sequence", "of", "Shift", "and", "Reduce", "operations", "of", "a", "transition", "-", "based", "parser", ".", "However", ",", "while", "the", "parser", "of", "Vinyals", "et", "al", ".", "relies", "on", "a", "trainable", "attention", "mechanism", "for", "focusing", "on", "specific", "BiLSTM", "vectors", ",", "parsers", "in", "the", "transition", "-", "based", "family", "we", "use", "in", "Section", "[", "reference", "]", "use", "a", "human", "designed", "stack", "and", "buffer", "mechanism", "to", "manually", "direct", "the", "parser", "’s", "attention", ".", "While", "the", "effectiveness", "of", "the", "trainable", "attention", "approach", "is", "impressive", ",", "the", "stack", "-", "and", "-", "buffer", "guidance", "of", "transition", "-", "based", "parsers", "results", "in", "more", "robust", "learning", ".", "Indeed", ",", "work", "by", "cross2016incremental", ",", "published", "while", "working", "on", "the", "camera", "-", "ready", "version", "of", "this", "paper", ",", "show", "that", "the", "same", "methodology", "as", "ours", "is", "highly", "effective", "also", "for", "greedy", ",", "transition", "-", "based", "constituency", "parsing", ",", "surpassing", "the", "beam", "-", "based", "architecture", "of", "Vinyals", "et", "al", ".", "(", "88.3F", "vs.", "89.8F", "points", ")", "when", "trained", "on", "the", "Penn", "Treebank", "dataset", "and", "without", "using", "orthogonal", "methods", "such", "as", "ensembling", "and", "up", "-", "training", ".", "subsection", ":", "Bidirectional", "Recurrent", "Neural", "Networks", "Recurrent", "neural", "networks", "(", "RNNs", ")", "are", "statistical", "learners", "for", "modeling", "sequential", "data", ".", "An", "RNN", "allows", "one", "to", "model", "the", "th", "element", "in", "the", "sequence", "based", "on", "the", "past", "–", "the", "elements", "up", "to", "and", "including", "it", ".", "The", "RNN", "model", "provides", "a", "framework", "for", "conditioning", "on", "the", "entire", "history", "without", "resorting", "to", "the", "Markov", "assumption", "which", "is", "traditionally", "used", "for", "modeling", "sequences", ".", "RNNs", "were", "shown", "to", "be", "capable", "of", "learning", "to", "count", ",", "as", "well", "as", "to", "model", "line", "lengths", "and", "complex", "phenomena", "such", "as", "bracketing", "and", "code", "indentation", ".", "Our", "proposed", "feature", "extractors", "are", "based", "on", "a", "bidirectional", "recurrent", "neural", "network", "(", "BiRNN", ")", ",", "an", "extension", "of", "RNNs", "that", "take", "into", "account", "both", "the", "past", "and", "the", "future", ".", "We", "use", "a", "specific", "flavor", "of", "RNN", "called", "a", "long", "short", "-", "term", "memory", "network", "(", "LSTM", ")", ".", "For", "brevity", ",", "we", "treat", "RNN", "as", "an", "abstraction", ",", "without", "getting", "into", "the", "mathematical", "details", "of", "the", "implementation", "of", "the", "RNNs", "and", "LSTMs", ".", "For", "further", "details", "on", "RNNs", "and", "LSTMs", ",", "the", "reader", "is", "referred", "to", "goldberg", "-", "primer", "and", "cho", "-", "primer", ".", "The", "recurrent", "neural", "network", "(", "RNN", ")", "abstraction", "is", "a", "parameterized", "function", "mapping", "a", "sequence", "of", "input", "vectors", ",", "to", "a", "sequence", "of", "output", "vectors", ".", "Each", "output", "vector", "is", "conditioned", "on", "all", "the", "input", "vectors", ",", "and", "can", "be", "thought", "of", "as", "a", "summary", "of", "the", "prefix", "of", ".", "In", "our", "notation", ",", "we", "ignore", "the", "intermediate", "vectors", "and", "take", "the", "output", "of", "to", "be", "the", "vector", ".", "A", "bidirectional", "RNN", "is", "composed", "of", "two", "RNNs", ",", "and", ",", "one", "reading", "the", "sequence", "in", "its", "regular", "order", ",", "and", "the", "other", "reading", "it", "in", "reverse", ".", "Concretely", ",", "given", "a", "sequence", "of", "vectors", "and", "a", "desired", "index", ",", "the", "function", "is", "defined", "as", ":", "The", "vector", "is", "then", "a", "representation", "of", "the", "th", "item", "in", ",", "taking", "into", "account", "both", "the", "entire", "history", "and", "the", "entire", "future", "by", "concatenating", "the", "matching", "Rnn", "s.", "We", "can", "view", "the", "BiRNN", "encoding", "of", "an", "item", "as", "representing", "the", "item", "together", "with", "a", "context", "of", "an", "infinite", "window", "around", "it", ".", "paragraph", ":", "Computational", "Complexity", "Computing", "the", "BiRNN", "vectors", "encoding", "of", "the", "th", "element", "of", "a", "sequence", "requires", "time", "for", "computing", "the", "two", "RNNs", "and", "concatenating", "their", "outputs", ".", "A", "naive", "approach", "of", "computing", "the", "bidirectional", "representation", "of", "all", "elements", "result", "in", "computation", ".", "However", ",", "it", "is", "trivial", "to", "compute", "the", "BiRNN", "encoding", "of", "all", "sequence", "items", "in", "linear", "time", "by", "pre", "-", "computing", "and", ",", "keeping", "the", "intermediate", "representations", ",", "and", "concatenating", "the", "required", "elements", "as", "needed", ".", "paragraph", ":", "BiRNN", "Training", "Initially", ",", "the", "BiRNN", "encodings", "do", "not", "capture", "any", "particular", "information", ".", "During", "training", ",", "the", "encoded", "vectors", "are", "fed", "into", "further", "network", "layers", ",", "until", "at", "some", "point", "a", "prediction", "is", "made", ",", "and", "a", "loss", "is", "incurred", ".", "The", "back", "-", "propagation", "algorithm", "is", "used", "to", "compute", "the", "gradients", "of", "all", "the", "parameters", "in", "the", "network", "(", "including", "the", "BiRNN", "parameters", ")", "with", "respect", "to", "the", "loss", ",", "and", "an", "optimizer", "is", "used", "to", "update", "the", "parameters", "according", "to", "the", "gradients", ".", "The", "training", "procedure", "causes", "the", "BiRNN", "function", "to", "extract", "from", "the", "input", "sequence", "the", "relevant", "information", "for", "the", "task", "task", "at", "hand", ".", "paragraph", ":", "Going", "deeper", "We", "use", "a", "variant", "of", "deep", "bidirectional", "RNN", "(", "or", "-", "layer", "BiRNN", ")", "which", "is", "composed", "of", "BiRNN", "functions", "that", "feed", "into", "each", "other", ":", "the", "output", "of", "becomes", "the", "input", "of", ".", "Stacking", "BiRNNs", "in", "this", "way", "has", "been", "empirically", "shown", "to", "be", "effective", ".", "In", "this", "work", ",", "we", "use", "BiRNNs", "and", "deep", "-", "BiRNNs", "interchangeably", ",", "specifying", "the", "number", "of", "layers", "when", "needed", ".", "paragraph", ":", "Historical", "Notes", "RNNs", "were", "introduced", "by", "elman1990finding", ",", "and", "extended", "to", "BiRNNs", "by", "schuster1997bidirectional", ".", "The", "LSTM", "variant", "of", "RNNs", "is", "due", "to", "hochreiter1997long", ".", "BiLSTMs", "were", "recently", "popularized", "by", "graves2008supervised", ",", "and", "deep", "BiRNNs", "were", "introduced", "to", "NLP", "by", "irsoy2014opinion", ",", "who", "used", "them", "for", "sequence", "tagging", ".", "In", "the", "context", "of", "parsing", ",", "lewis2016lstm", "and", "Vaswani:2016:NAACL", "use", "a", "BiLSTM", "sequence", "tagging", "model", "to", "assign", "a", "CCG", "supertag", "for", "each", "token", "in", "the", "sentence", ".", "lewis2016lstm", "feeds", "the", "resulting", "supertags", "sequence", "into", "an", "A", "*", "CCG", "parser", ".", "Vaswani:2016:NAACL", "adds", "an", "additional", "layer", "of", "LSTM", "which", "receives", "the", "BiLSTM", "representation", "together", "with", "the", "k", "-", "best", "supertags", "for", "each", "word", "and", "outputs", "the", "most", "likely", "supertag", "given", "previous", "tags", ",", "and", "then", "feeds", "the", "predicted", "supertags", "to", "a", "discriminitively", "trained", "parser", ".", "In", "both", "works", ",", "the", "BiLSTM", "is", "trained", "to", "produce", "accurate", "CCG", "supertags", ",", "and", "is", "not", "aware", "of", "the", "global", "parsing", "objective", ".", "section", ":", "Our", "Approach", "We", "propose", "to", "replace", "the", "hand", "-", "crafted", "feature", "functions", "in", "favor", "of", "minimally", "-", "defined", "feature", "functions", "which", "make", "use", "of", "automatically", "learned", "Bidirectional", "LSTM", "representations", ".", "Given", "-", "words", "input", "sentence", "with", "words", "together", "with", "the", "corresponding", "POS", "tags", ",", "we", "associate", "each", "word", "and", "POS", "with", "embedding", "vectors", "and", ",", "and", "create", "a", "sequence", "of", "input", "vectors", "in", "which", "each", "is", "a", "concatenation", "of", "the", "corresponding", "word", "and", "POS", "vectors", ":", "The", "embeddings", "are", "trained", "together", "with", "the", "model", ".", "This", "encodes", "each", "word", "in", "isolation", ",", "disregarding", "its", "context", ".", "We", "introduce", "context", "by", "representing", "each", "input", "element", "as", "its", "(", "deep", ")", "BiLSTM", "vector", ",", ":", "Our", "feature", "function", "is", "then", "a", "concatenation", "of", "a", "small", "number", "of", "BiLSTM", "vectors", ".", "The", "exact", "feature", "function", "is", "parser", "dependent", "and", "will", "be", "discussed", "when", "discussing", "the", "corresponding", "parsers", ".", "The", "resulting", "feature", "vectors", "are", "then", "scored", "using", "a", "non", "-", "linear", "function", ",", "namely", "a", "multi", "-", "layer", "perceptron", "with", "one", "hidden", "layer", "(", "MLP", ")", ":", "where", "are", "the", "model", "parameters", ".", "Beside", "using", "the", "BiLSTM", "-", "based", "feature", "functions", ",", "we", "make", "use", "of", "standard", "parsing", "techniques", ".", "Crucially", ",", "the", "BiLSTM", "is", "trained", "jointly", "with", "the", "rest", "of", "the", "parsing", "objective", ".", "This", "allows", "it", "to", "learn", "representations", "which", "are", "suitable", "for", "the", "parsing", "task", ".", "Consider", "a", "concatenation", "of", "two", "BiLSTM", "vectors", "(", ")", "scored", "using", "an", "MLP", ".", "The", "scoring", "function", "has", "access", "to", "the", "words", "and", "POS", "-", "tags", "of", "and", ",", "as", "well", "as", "the", "words", "and", "POS", "-", "tags", "of", "the", "words", "in", "an", "infinite", "window", "surrounding", "them", ".", "As", "LSTMs", "are", "known", "to", "capture", "length", "and", "sequence", "position", "information", ",", "it", "is", "very", "plausible", "that", "the", "scoring", "function", "can", "be", "sensitive", "also", "to", "the", "distance", "between", "and", ",", "their", "ordering", ",", "and", "the", "sequential", "material", "between", "them", ".", "paragraph", ":", "Parsing", "-", "time", "Complexity", "Once", "the", "BiLSTM", "is", "trained", ",", "parsing", "is", "performed", "by", "first", "computing", "the", "BiLSTM", "encoding", "for", "each", "word", "in", "the", "sentence", "(", "a", "linear", "time", "operation", ")", ".", "Then", ",", "parsing", "proceeds", "as", "usual", ",", "where", "the", "feature", "extraction", "involves", "a", "concatenation", "of", "a", "small", "number", "of", "the", "pre", "-", "computed", "vectors", ".", "section", ":", "Transition", "-", "based", "Parser", "We", "begin", "by", "integrating", "the", "feature", "extractor", "in", "a", "transition", "-", "based", "parser", ".", "We", "follow", "the", "notation", "in", "tacl2013dynamic", ".", "The", "transition", "-", "based", "parsing", "framework", "assumes", "a", "transition", "system", ",", "an", "abstract", "machine", "that", "processes", "sentences", "and", "produces", "parse", "trees", ".", "The", "transition", "system", "has", "a", "set", "of", "configurations", "and", "a", "set", "of", "transitions", "which", "are", "applied", "to", "configurations", ".", "When", "parsing", "a", "sentence", ",", "the", "system", "is", "initialized", "to", "an", "initial", "configuration", "based", "on", "the", "input", "sentence", ",", "and", "transitions", "are", "repeatedly", "applied", "to", "this", "configuration", ".", "After", "a", "finite", "number", "of", "transitions", ",", "the", "system", "arrives", "at", "a", "terminal", "configuration", ",", "and", "a", "parse", "tree", "is", "read", "off", "the", "terminal", "configuration", ".", "In", "a", "greedy", "parser", ",", "a", "classifier", "is", "used", "to", "choose", "the", "transition", "to", "take", "in", "each", "configuration", ",", "based", "on", "features", "extracted", "from", "the", "configuration", "itself", ".", "The", "parsing", "algorithm", "is", "presented", "in", "Algorithm", "[", "reference", "]", "below", ".", "[", "h", "]", "Greedy", "transition", "-", "based", "parsing", "[", "1", "]", "Input", ":", "sentence", ",", "parameterized", "function", "with", "parameters", ".", "not", "Given", "a", "sentence", ",", "the", "parser", "is", "initialized", "with", "the", "configuration", "(", "line", "[", "reference", "]", ")", ".", "Then", ",", "a", "feature", "function", "represents", "the", "configuration", "as", "a", "vector", ",", "which", "is", "fed", "to", "a", "scoring", "function", "Score", "assigning", "scores", "to", "(", "configuration", ",", "transition", ")", "pairs", ".", "Score", "scores", "the", "possible", "transitions", ",", "and", "the", "highest", "scoring", "transition", "is", "chosen", "(", "line", "[", "reference", "]", ")", ".", "The", "transition", "is", "applied", "to", "the", "configuration", ",", "resulting", "in", "a", "new", "parser", "configuration", ".", "The", "process", "ends", "when", "reaching", "a", "final", "configuration", ",", "from", "which", "the", "resulting", "parse", "tree", "is", "read", "and", "returned", "(", "line", "[", "reference", "]", ")", ".", "Transition", "systems", "differ", "by", "the", "way", "they", "define", "configurations", ",", "and", "by", "the", "particular", "set", "of", "transitions", "available", "to", "them", ".", "A", "parser", "is", "determined", "by", "the", "choice", "of", "a", "transition", "system", ",", "a", "feature", "function", "and", "a", "scoring", "function", "Score", ".", "Our", "choices", "are", "detailed", "below", ".", "paragraph", ":", "The", "Arc", "-", "Hybrid", "System", "Many", "transition", "systems", "exist", "in", "the", "literature", ".", "In", "this", "work", ",", "we", "use", "the", "arc", "-", "hybrid", "transition", "system", ",", "which", "is", "similar", "to", "the", "more", "popular", "arc", "-", "standard", "system", ",", "but", "for", "which", "an", "efficient", "dynamic", "oracle", "is", "available", ".", "In", "the", "arc", "-", "hybrid", "system", ",", "a", "configuration", "consists", "of", "a", "stack", ",", "a", "buffer", ",", "and", "a", "set", "of", "dependency", "arcs", ".", "Both", "the", "stack", "and", "the", "buffer", "hold", "integer", "indices", "pointing", "to", "sentence", "elements", ".", "Given", "a", "sentence", ",", "the", "system", "is", "initialized", "with", "an", "empty", "stack", ",", "an", "empty", "arc", "set", ",", "and", ",", "where", "is", "the", "special", "root", "index", ".", "Any", "configuration", "with", "an", "empty", "stack", "and", "a", "buffer", "containing", "only", "is", "terminal", ",", "and", "the", "parse", "tree", "is", "given", "by", "the", "arc", "set", "of", ".", "The", "arc", "-", "hybrid", "system", "allows", "3", "possible", "transitions", ",", "Shift", ",", "and", ",", "defined", "as", ":", "The", "Shift", "transition", "moves", "the", "first", "item", "of", "the", "buffer", "(", ")", "to", "the", "stack", ".", "The", "Leftℓ", "transition", "removes", "the", "first", "item", "on", "top", "of", "the", "stack", "(", ")", "and", "attaches", "it", "as", "a", "modifier", "to", "with", "label", ",", "adding", "the", "arc", ".", "The", "Rightℓ", "transition", "removes", "from", "the", "stack", "and", "attaches", "it", "as", "a", "modifier", "to", "the", "next", "item", "on", "the", "stack", "(", ")", ",", "adding", "the", "arc", ".", "paragraph", ":", "Scoring", "Function", "Traditionally", ",", "the", "scoring", "function", "is", "a", "discriminative", "linear", "model", "of", "the", "form", ".", "The", "linearity", "of", "Score", "required", "the", "feature", "function", "to", "encode", "non", "-", "linearities", "in", "the", "form", "of", "combination", "features", ".", "We", "follow", "Chen", "and", "Manning", "chen2014fast", "and", "replace", "the", "linear", "scoring", "model", "with", "an", "MLP", ".", "paragraph", ":", "Simple", "Feature", "Function", "The", "feature", "function", "is", "typically", "complex", "(", "see", "Section", "[", "reference", "]", ")", ".", "Our", "feature", "function", "is", "the", "concatenated", "BiLSTM", "vectors", "of", "the", "top", "3", "items", "on", "the", "stack", "and", "the", "first", "item", "on", "the", "buffer", ".", "I.e.", ",", "for", "a", "configuration", "the", "feature", "extractor", "is", "defined", "as", ":", "This", "feature", "function", "is", "rather", "minimal", ":", "it", "takes", "into", "account", "the", "BiLSTM", "representations", "of", "and", ",", "which", "are", "the", "items", "affected", "by", "the", "possible", "transitions", "being", "scored", ",", "as", "well", "as", "one", "extra", "stack", "context", ".", "Figure", "1", "depicts", "transition", "scoring", "with", "our", "architecture", "and", "this", "feature", "function", ".", "Note", "that", ",", "unlike", "previous", "work", ",", "this", "feature", "function", "does", "not", "take", "into", "account", ",", "the", "already", "built", "structure", ".", "The", "high", "parsing", "accuracies", "in", "the", "experimental", "sections", "suggest", "that", "the", "BiLSTM", "encoding", "is", "capable", "of", "estimating", "a", "lot", "of", "the", "missing", "information", "based", "on", "the", "provided", "stack", "and", "buffer", "elements", "and", "the", "sequential", "content", "between", "them", ".", "While", "not", "explored", "in", "this", "work", ",", "relying", "on", "only", "four", "word", "indices", "for", "scoring", "an", "action", "results", "in", "very", "compact", "state", "signatures", ",", "making", "our", "proposed", "feature", "representation", "very", "appealing", "for", "use", "in", "transition", "-", "based", "parsers", "that", "employ", "dynamic", "-", "programming", "search", ".", "paragraph", ":", "Extended", "Feature", "Function", "One", "of", "the", "benefits", "of", "the", "greedy", "transition", "-", "based", "parsing", "framework", "is", "precisely", "its", "ability", "to", "look", "at", "arbitrary", "features", "from", "the", "already", "built", "tree", ".", "If", "we", "allow", "somewhat", "less", "minimal", "feature", "function", ",", "we", "could", "add", "the", "BiLSTM", "vectors", "corresponding", "to", "the", "right", "-", "most", "and", "left", "-", "most", "modifiers", "of", ",", "and", ",", "as", "well", "as", "the", "left", "-", "most", "modifier", "of", ",", "reaching", "a", "total", "of", "11", "BiLSTM", "vectors", ".", "We", "refer", "to", "this", "as", "the", "extended", "feature", "set", ".", "As", "we", "’ll", "see", "in", "Section", "[", "reference", "]", ",", "using", "the", "extended", "set", "does", "indeed", "improve", "parsing", "accuracies", "when", "using", "pre", "-", "trained", "word", "embeddings", ",", "but", "has", "a", "minimal", "effect", "in", "the", "fully", "-", "supervised", "case", ".", "subsection", ":", "Details", "of", "the", "Training", "Algorithm", "The", "training", "objective", "is", "to", "set", "the", "score", "of", "correct", "transitions", "above", "the", "scores", "of", "incorrect", "transitions", ".", "We", "use", "a", "margin", "-", "based", "objective", ",", "aiming", "to", "maximize", "the", "margin", "between", "the", "highest", "scoring", "correct", "action", "and", "the", "highest", "scoring", "incorrect", "action", ".", "The", "hinge", "loss", "at", "each", "parsing", "configuration", "is", "defined", "as", ":", "where", "is", "the", "set", "of", "possible", "transitions", "and", "is", "the", "set", "of", "correct", "(", "gold", ")", "transitions", "at", "the", "current", "stage", ".", "At", "each", "stage", "of", "the", "training", "process", "the", "parser", "scores", "the", "possible", "transitions", ",", "incurs", "a", "loss", ",", "selects", "a", "transition", "to", "follow", ",", "and", "moves", "to", "the", "next", "configuration", "based", "on", "it", ".", "The", "local", "losses", "are", "summed", "throughout", "the", "parsing", "process", "of", "a", "sentence", ",", "and", "the", "parameters", "are", "updated", "with", "respect", "to", "the", "sum", "of", "the", "losses", "at", "sentence", "boundaries", ".", "The", "gradients", "of", "the", "entire", "network", "(", "including", "the", "MLP", "and", "the", "BiLSTM", ")", "with", "respect", "to", "the", "sum", "of", "the", "losses", "are", "calculated", "using", "the", "backpropagation", "algorithm", ".", "As", "usual", ",", "we", "perform", "several", "training", "iterations", "over", "the", "training", "corpus", ",", "shuffling", "the", "order", "of", "sentences", "in", "each", "iteration", ".", "paragraph", ":", "Error", "-", "Exploration", "and", "Dynamic", "Oracle", "Training", "We", "follow", "tacl2013dynamic", ";", "coling2012dynamic", "in", "using", "error", "exploration", "training", "with", "a", "dynamic", "-", "oracle", ",", "which", "we", "briefly", "describe", "below", ".", "At", "each", "stage", "in", "the", "training", "process", ",", "the", "parser", "assigns", "scores", "to", "all", "the", "possible", "transitions", ".", "It", "then", "selects", "a", "transition", ",", "applies", "it", ",", "and", "moves", "to", "the", "next", "step", ".", "Which", "transition", "should", "be", "followed", "?", "A", "common", "approach", "follows", "the", "highest", "scoring", "transition", "that", "can", "lead", "to", "the", "gold", "tree", ".", "However", ",", "when", "training", "in", "this", "way", "the", "parser", "sees", "only", "configurations", "that", "result", "from", "following", "correct", "actions", ",", "and", "as", "a", "result", "tends", "to", "suffer", "from", "error", "propagation", "at", "test", "time", ".", "Instead", ",", "in", "error", "-", "exploration", "training", "the", "parser", "follows", "the", "highest", "scoring", "action", "in", "during", "training", "even", "if", "this", "action", "is", "incorrect", ",", "exposing", "it", "to", "configurations", "that", "result", "from", "erroneous", "decisions", ".", "This", "strategy", "requires", "defining", "the", "set", "such", "that", "the", "correct", "actions", "to", "take", "are", "well", "-", "defined", "also", "for", "states", "that", "can", "not", "lead", "to", "the", "gold", "tree", ".", "Such", "a", "set", "is", "called", "a", "dynamic", "oracle", ".", "We", "perform", "error", "-", "exploration", "training", "using", "the", "dynamic", "-", "oracle", "defined", "by", "tacl2013dynamic", ".", "paragraph", ":", "Aggressive", "Exploration", "We", "found", "that", "even", "when", "using", "error", "-", "exploration", ",", "after", "one", "iteration", "the", "model", "remembers", "the", "training", "set", "quite", "well", ",", "and", "does", "not", "make", "enough", "errors", "to", "make", "error", "-", "exploration", "effective", ".", "In", "order", "to", "expose", "the", "parser", "to", "more", "errors", ",", "we", "follow", "an", "aggressive", "-", "exploration", "scheme", ":", "we", "sometimes", "follow", "incorrect", "transitions", "also", "if", "they", "score", "below", "correct", "transitions", ".", "Specifically", ",", "when", "the", "score", "of", "the", "correct", "transition", "is", "greater", "than", "that", "of", "the", "wrong", "transition", "but", "the", "difference", "is", "smaller", "than", "a", "margin", "constant", ",", "we", "chose", "to", "follow", "the", "incorrect", "action", "with", "probability", "(", "we", "use", "in", "our", "experiments", ")", ".", "paragraph", ":", "Summary", "The", "greedy", "transition", "-", "based", "parser", "follows", "standard", "techniques", "from", "the", "literature", "(", "margin", "-", "based", "objective", ",", "dynamic", "oracle", "training", ",", "error", "exploration", ",", "MLP", "-", "based", "non", "-", "linear", "scoring", "function", ")", ".", "We", "depart", "from", "the", "literature", "by", "replacing", "the", "hand", "-", "crafted", "feature", "function", "over", "carefully", "selected", "components", "of", "the", "configuration", "with", "a", "concatenation", "of", "BiLSTM", "representations", "of", "a", "few", "prominent", "items", "on", "the", "stack", "and", "the", "buffer", ",", "and", "training", "the", "BiLSTM", "encoder", "jointly", "with", "the", "rest", "of", "the", "network", ".", "section", ":", "Graph", "-", "based", "Parser", "Graph", "-", "based", "parsing", "follows", "the", "common", "structured", "prediction", "paradigm", ":", "Given", "an", "input", "sentence", "(", "and", "the", "corresponding", "sequence", "of", "vectors", ")", "we", "look", "for", "the", "highest", "-", "scoring", "parse", "tree", "in", "the", "space", "of", "valid", "dependency", "trees", "over", ".", "In", "order", "to", "make", "the", "search", "tractable", ",", "the", "scoring", "function", "is", "decomposed", "to", "the", "sum", "of", "local", "scores", "for", "each", "part", "independently", ".", "In", "this", "work", ",", "we", "focus", "on", "arc", "-", "factored", "graph", "based", "approach", "presented", "in", "mst", ".", "Arc", "-", "factored", "parsing", "decomposes", "the", "score", "of", "a", "tree", "to", "the", "sum", "of", "the", "score", "of", "its", "head", "-", "modifier", "arcs", ":", "Given", "the", "scores", "of", "the", "arcs", "the", "highest", "scoring", "projective", "tree", "can", "be", "efficiently", "found", "using", "Eisner", "’s", "decoding", "algorithm", "eisner1996dep", ".", "McDonald", "et", "al", ".", "and", "most", "subsequent", "work", "estimate", "the", "local", "score", "of", "an", "arc", "by", "a", "linear", "model", "parameterized", "by", "a", "weight", "vector", ",", "and", "a", "feature", "function", "assigning", "a", "sparse", "feature", "vector", "for", "an", "arc", "linking", "modifier", "to", "head", ".", "We", "follow", "pei2015effective", "and", "replace", "the", "linear", "scoring", "function", "with", "an", "MLP", ".", "The", "feature", "extractor", "is", "usually", "complex", ",", "involving", "many", "elements", "(", "see", "Section", "[", "reference", "]", ")", ".", "In", "contrast", ",", "our", "feature", "extractor", "uses", "merely", "the", "BiLSTM", "encoding", "of", "the", "head", "word", "and", "the", "modifier", "word", ":", "The", "final", "model", "is", ":", "The", "architecture", "is", "illustrated", "in", "Figure", "[", "reference", "]", ".", "paragraph", ":", "Training", "The", "training", "objective", "is", "to", "set", "the", "score", "function", "such", "that", "correct", "tree", "is", "scored", "above", "incorrect", "ones", ".", "We", "use", "a", "margin", "-", "based", "objective", ",", "aiming", "to", "maximize", "the", "margin", "between", "the", "score", "of", "the", "gold", "tree", "and", "the", "highest", "scoring", "incorrect", "tree", ".", "We", "define", "a", "hinge", "loss", "with", "respect", "to", "a", "gold", "tree", "as", ":", "Each", "of", "the", "tree", "scores", "is", "then", "calculated", "by", "activating", "the", "MLP", "on", "the", "arc", "representations", ".", "The", "entire", "loss", "can", "viewed", "as", "the", "sum", "of", "multiple", "neural", "networks", ",", "which", "is", "sub", "-", "differentiable", ".", "We", "calculate", "the", "gradients", "of", "the", "entire", "network", "(", "including", "to", "the", "BiLSTM", "encoder", "and", "word", "embeddings", ")", ".", "paragraph", ":", "Labeled", "Parsing", "Up", "to", "now", ",", "we", "described", "unlabeled", "parsing", ".", "A", "possible", "approach", "for", "adding", "labels", "is", "to", "score", "the", "combination", "of", "an", "unlabeled", "arc", "and", "its", "label", "by", "considering", "the", "label", "as", "part", "of", "the", "arc", ".", "This", "results", "in", "parts", "that", "need", "to", "be", "scored", ",", "leading", "to", "slow", "parsing", "speeds", "and", "arguably", "a", "harder", "learning", "problem", ".", "Instead", ",", "we", "chose", "to", "first", "predict", "the", "unlabeled", "structure", "using", "the", "model", "given", "above", ",", "and", "then", "predict", "the", "label", "of", "each", "resulting", "arc", ".", "Using", "this", "approach", ",", "the", "number", "of", "parts", "stays", "small", ",", "enabling", "fast", "parsing", ".", "The", "labeling", "of", "an", "arc", "is", "performed", "using", "the", "same", "feature", "representation", "fed", "into", "a", "different", "MLP", "predictor", ":", "As", "before", "we", "use", "a", "margin", "based", "hinge", "loss", ".", "The", "labeler", "is", "trained", "on", "the", "gold", "trees", ".", "The", "BiLSTM", "encoder", "responsible", "for", "producing", "and", "is", "shared", "with", "the", "arc", "-", "factored", "parser", ":", "the", "same", "BiLSTM", "encoder", "is", "used", "in", "the", "parer", "and", "the", "labeler", ".", "This", "sharing", "of", "parameters", "can", "be", "seen", "as", "an", "instance", "of", "multi", "-", "task", "learning", ".", "As", "we", "show", "in", "Section", "[", "reference", "]", ",", "the", "sharing", "is", "effective", ":", "training", "the", "BiLSTM", "feature", "encoder", "to", "be", "good", "at", "predicting", "arc", "-", "labels", "significantly", "improves", "the", "parser", "’s", "unlabeled", "accuracy", ".", "paragraph", ":", "Loss", "augmented", "inference", "In", "initial", "experiments", ",", "the", "network", "learned", "quickly", "and", "overfit", "the", "data", ".", "In", "order", "to", "remedy", "this", ",", "we", "found", "it", "useful", "to", "use", "loss", "augmented", "inference", ".", "The", "intuition", "behind", "loss", "augmented", "inference", "is", "to", "update", "against", "trees", "which", "have", "high", "model", "scores", "and", "are", "also", "very", "wrong", ".", "This", "is", "done", "by", "augmenting", "the", "score", "of", "each", "part", "not", "belonging", "to", "the", "gold", "tree", "by", "adding", "a", "constant", "to", "its", "score", ".", "Formally", ",", "the", "loss", "transforms", "as", "follows", ":", "paragraph", ":", "Speed", "improvements", "The", "arc", "-", "factored", "model", "requires", "the", "scoring", "of", "arcs", ".", "Scoring", "is", "performed", "using", "an", "MLP", "with", "one", "hidden", "layer", ",", "resulting", "in", "matrix", "-", "vector", "multiplications", "from", "the", "input", "to", "the", "hidden", "layer", ",", "and", "multiplications", "from", "the", "hidden", "to", "the", "output", "layer", ".", "The", "first", "multiplications", "involve", "larger", "dimensional", "input", "and", "output", "vectors", ",", "and", "are", "the", "most", "time", "consuming", ".", "Fortunately", ",", "these", "can", "be", "reduced", "to", "multiplications", "and", "vector", "additions", ",", "by", "observing", "that", "the", "multiplication", "can", "be", "written", "as", "where", "and", "are", "are", "the", "first", "and", "second", "half", "of", "the", "matrix", "and", "reusing", "the", "products", "across", "different", "pairs", ".", "Summary", "The", "graph", "-", "based", "parser", "is", "straight", "-", "forward", "first", "-", "order", "parser", ",", "trained", "with", "a", "margin", "-", "based", "hinge", "-", "loss", "and", "loss", "-", "augmented", "inference", ".", "We", "depart", "from", "the", "literature", "by", "replacing", "the", "hand", "-", "crafted", "feature", "function", "with", "a", "concatenation", "of", "BiLSTM", "representations", "of", "the", "head", "and", "modifier", "words", ",", "and", "training", "the", "BiLSTM", "encoder", "jointly", "with", "the", "structured", "objective", ".", "We", "also", "introduce", "a", "novel", "multi", "-", "task", "learning", "approach", "for", "labeled", "parsing", "by", "training", "a", "second", "-", "stage", "arc", "-", "labeler", "sharing", "the", "same", "BiLSTM", "encoder", "with", "the", "unlabeled", "parser", ".", "section", ":", "Experiments", "and", "Results", "We", "evaluated", "our", "parsing", "model", "on", "English", "and", "Chinese", "data", ".", "For", "comparison", "purposes", "we", "follow", "the", "setup", "of", "dyer2015transitionbased", ".", "paragraph", ":", "Data", "For", "English", ",", "we", "used", "the", "Stanford", "Dependency", "(", "SD", ")", "conversion", "of", "the", "Penn", "Treebank", ",", "using", "the", "standard", "train", "/", "dev", "/", "test", "splits", "with", "the", "same", "predicted", "POS", "-", "tags", "as", "used", "in", "dyer2015transitionbased", ";", "chen2014fast", ".", "This", "dataset", "contains", "a", "few", "non", "-", "projective", "trees", ".", "Punctuation", "symbols", "are", "excluded", "from", "the", "evaluation", ".", "For", "Chinese", ",", "we", "use", "the", "Penn", "Chinese", "Treebank", "5.1", "(", "CTB5", ")", ",", "using", "the", "train", "/", "test", "/", "dev", "splits", "of", "with", "gold", "part", "-", "of", "-", "speech", "tags", ",", "also", "following", ".", "When", "using", "external", "word", "embeddings", ",", "we", "also", "use", "the", "same", "data", "as", "dyer2015transitionbased", ".", "paragraph", ":", "Implementation", "Details", "The", "parsers", "are", "implemented", "in", "python", ",", "using", "the", "PyCNN", "toolkit", "for", "neural", "network", "training", ".", "The", "code", "is", "available", "at", "the", "github", "repository", ".", "We", "use", "the", "LSTM", "variant", "implemented", "in", "PyCNN", ",", "and", "optimize", "using", "the", "Adam", "optimizer", ".", "Unless", "otherwise", "noted", ",", "we", "use", "the", "default", "values", "provided", "by", "PyCNN", "(", "e.g.", "for", "random", "initialization", ",", "learning", "rates", "etc", ")", ".", "The", "word", "and", "POS", "embeddings", "and", "are", "initialized", "to", "random", "values", "and", "trained", "together", "with", "the", "rest", "of", "the", "parsers", "’", "networks", ".", "In", "some", "experiments", ",", "we", "introduce", "also", "pre", "-", "trained", "word", "embeddings", ".", "In", "those", "cases", ",", "the", "vector", "representation", "of", "a", "word", "is", "a", "concatenation", "of", "its", "randomly", "-", "initialized", "vector", "embedding", "with", "its", "pre", "-", "trained", "word", "vector", ".", "Both", "are", "tuned", "during", "training", ".", "We", "use", "the", "same", "word", "vectors", "as", "in", "dyer2015transitionbased", "During", "training", ",", "we", "employ", "a", "variant", "of", "word", "dropout", ",", "and", "replace", "a", "word", "with", "the", "unknown", "-", "word", "symbol", "with", "probability", "that", "is", "inversely", "proportional", "to", "the", "frequency", "of", "the", "word", ".", "A", "word", "appearing", "times", "in", "the", "training", "corpus", "is", "replaced", "with", "the", "unknown", "symbol", "with", "probability", ".", "If", "a", "word", "was", "dropped", "the", "external", "embedding", "of", "the", "word", "is", "also", "dropped", "with", "probability", ".", "We", "train", "the", "parsers", "for", "up", "to", "30", "iterations", ",", "and", "choose", "the", "best", "model", "according", "to", "the", "UAS", "accuracy", "on", "the", "development", "set", ".", "paragraph", ":", "Hyperparameter", "Tuning", "We", "performed", "a", "very", "minimal", "hyper", "-", "parameter", "search", "with", "the", "graph", "-", "based", "parser", ",", "and", "use", "the", "same", "hyper", "-", "parameters", "for", "both", "parsers", ".", "The", "hyper", "-", "parameters", "of", "the", "final", "networks", "used", "for", "all", "the", "reported", "experiments", "are", "detailed", "in", "Table", "[", "reference", "]", ".", "Main", "Results", "Table", "[", "reference", "]", "lists", "the", "test", "-", "set", "accuracies", "of", "our", "best", "parsing", "models", ",", "compared", "to", "other", "state", "-", "of", "-", "the", "-", "art", "parsers", "from", "the", "literature", ".", "It", "is", "clear", "that", "our", "parsers", "are", "very", "competitive", ",", "despite", "using", "very", "simple", "parsing", "architectures", "and", "minimal", "feature", "extractors", ".", "When", "not", "using", "external", "embeddings", ",", "the", "first", "-", "order", "graph", "-", "based", "parser", "with", "2", "features", "outperforms", "all", "other", "systems", "that", "are", "not", "using", "external", "resources", ",", "including", "the", "third", "-", "order", "TurboParser", ".", "The", "greedy", "transition", "based", "parser", "with", "4", "features", "also", "matches", "or", "outperforms", "most", "other", "parsers", ",", "including", "the", "beam", "-", "based", "transition", "parser", "with", "heavily", "engineered", "features", "of", "Zhang", "and", "Nivre", "(", "2011", ")", "and", "the", "Stack", "-", "LSTM", "parser", "of", "dyer2015transitionbased", ",", "as", "well", "as", "the", "same", "parser", "when", "trained", "using", "a", "dynamic", "oracle", ".", "Moving", "from", "the", "simple", "(", "4", "features", ")", "to", "the", "extended", "(", "11", "features", ")", "feature", "set", "leads", "to", "some", "gains", "in", "accuracy", "for", "both", "English", "and", "Chinese", ".", "Interestingly", ",", "when", "adding", "external", "word", "embeddings", "the", "accuracy", "of", "the", "graph", "-", "based", "parser", "degrades", ".", "We", "are", "not", "sure", "why", "this", "happens", ",", "and", "leave", "the", "exploration", "of", "effective", "semi", "-", "supervised", "parsing", "with", "the", "graph", "-", "based", "model", "for", "future", "work", ".", "The", "greedy", "parser", "does", "manage", "to", "benefit", "from", "the", "external", "embeddings", ",", "and", "using", "them", "we", "also", "see", "gains", "from", "moving", "from", "the", "simple", "to", "the", "extended", "feature", "set", ".", "Both", "feature", "sets", "result", "in", "very", "competitive", "results", ",", "with", "the", "extended", "feature", "set", "yielding", "the", "best", "reported", "results", "for", "Chinese", ",", "and", "ranked", "second", "for", "English", ",", "after", "the", "heavily", "-", "tuned", "beam", "-", "based", "parser", "of", "weiss2015structured", ".", "paragraph", ":", "Additional", "Results", "We", "perform", "some", "ablation", "experiments", "in", "order", "to", "quantify", "the", "effect", "of", "the", "different", "components", "on", "our", "best", "models", "(", "Table", "[", "reference", "]", ")", ".", "Loss", "augmented", "inference", "is", "crucial", "for", "the", "success", "of", "the", "graph", "-", "based", "parser", ",", "and", "the", "multi", "-", "task", "learning", "scheme", "for", "the", "arc", "-", "labeler", "contributes", "nicely", "to", "the", "unlabeled", "scores", ".", "Dynamic", "oracle", "training", "yields", "nice", "gains", "for", "both", "English", "and", "Chinese", ".", "section", ":", "Conclusion", "We", "presented", "a", "pleasingly", "effective", "approach", "for", "feature", "extraction", "for", "dependency", "parsing", "based", "on", "a", "BiLSTM", "encoder", "that", "is", "trained", "jointly", "with", "the", "parser", ",", "and", "demonstrated", "its", "effectiveness", "by", "integrating", "it", "into", "two", "simple", "parsing", "models", ":", "a", "greedy", "transition", "-", "based", "parser", "and", "a", "globally", "optimized", "first", "-", "order", "graph", "-", "based", "parser", ",", "yielding", "very", "competitive", "parsing", "accuracies", "in", "both", "cases", ".", "paragraph", ":", "Acknowledgements", "This", "research", "is", "supported", "by", "the", "Intel", "Collaborative", "Research", "Institute", "for", "Computational", "Intelligence", "(", "ICRI", "-", "CI", ")", "and", "the", "Israeli", "Science", "Foundation", "(", "grant", "number", "1555", "/", "15", ")", ".", "We", "thank", "Lillian", "Lee", "for", "her", "important", "feedback", "and", "efforts", "invested", "in", "editing", "this", "paper", ".", "We", "also", "thank", "the", "reviewers", "for", "their", "valuable", "comments", ".", "bibliography", ":", "References"], "entity": [[5, 7, ["Dependency", "Parsing"], "Task"], [8, 12, ["Bidirectional", "LSTM", "Feature", "Representations"], "Method"], [20, 22, ["dependency", "parsing"], "Task"], [26, 29, ["bidirectional", "-", "LSTMs"], "Method"], [30, 31, ["BiLSTMs"], "Method"], [63, 64, ["BiLSTM"], "Method"], [69, 71, ["parser", "objective"], "Metric"], [76, 78, ["feature", "extractors"], "Method"], [79, 80, ["parsing"], "Task"], [93, 98, ["greedy", "transition", "-", "based", "parser"], "Method"], [103, 109, ["globally", "optimized", "graph", "-", "based", "parser"], "Method"], [112, 113, ["parsers"], "Method"], [130, 131, ["accuracies"], "Metric"], [132, 133, ["English"], "Material"], [134, 135, ["Chinese"], "Material"], [154, 156, ["feature", "representation"], "Task"], [157, 159, ["dependency", "parsing"], "Task"], [165, 168, ["neural", "-", "networks"], "Method"], [171, 173, ["deep", "learning"], "Method"], [181, 183, ["dependency", "parsing"], "Task"], [188, 191, ["graph", "-", "based"], "Method"], [192, 196, ["transition", "-", "based", "parsers"], "Method"], [197, 201, ["Graph", "-", "based", "parsers"], "Method"], [202, 203, ["parsing"], "Task"], [205, 211, ["search", "-", "based", "structured", "prediction", "problem"], "Task"], [235, 239, ["Transition", "-", "based", "parsers"], "Method"], [240, 241, ["parsing"], "Task"], [254, 255, ["classifier"], "Method"], [271, 273, ["parsing", "process"], "Task"], [282, 290, ["arc", "-", "factored", "(", "first", "order", ")", "models"], "Method"], [294, 296, ["scoring", "function"], "Method"], [322, 326, ["inference", "and", "training", "algorithms"], "Method"], [329, 333, ["transition", "-", "based", "parsers"], "Method"], [352, 354, ["parsing", "speeds"], "Metric"], [357, 361, ["transition", "-", "based", "parsers"], "Method"], [369, 373, ["beam", "or", "dynamic", "programming"], "Method"], [380, 382, ["parsing", "framework"], "Method"], [389, 391, ["parser", "design"], "Task"], [400, 402, ["statistical", "model"], "Method"], [421, 425, ["feature", "function", "design", "problem"], "Task"], [428, 434, ["linear", "to", "non", "-", "linear", "models"], "Method"], [457, 461, ["machine", "-", "learning", "machinery"], "Method"], [520, 522, ["feature", "representations"], "Method"], [526, 528, ["feature", "engineering"], "Method"], [531, 533, ["architecture", "engineering"], "Method"], [550, 552, ["feature", "engineering"], "Task"], [553, 555, ["architecture", "engineering"], "Task"], [567, 568, ["BiRNNs"], "Method"], [572, 573, ["BiLSTMs"], "Method"], [576, 581, ["strong", "and", "trainable", "sequence", "models"], "Method"], [590, 591, ["BiLSTM"], "Method"], [628, 630, ["BiLSTM", "encoding"], "Method"], [641, 643, ["BiLSTM", "encodings"], "Method"], [654, 659, ["non", "-", "linear", "scoring", "function"], "Method"], [660, 664, ["multi", "-", "layer", "perceptron"], "Method"], [669, 670, ["BiLSTM"], "Method"], [677, 678, ["parser"], "Method"], [684, 686, ["feature", "representation"], "Method"], [688, 690, ["parsing", "problem"], "Task"], [700, 702, ["BiLSTM", "itself"], "Method"], [717, 719, ["feature", "extractor"], "Method"], [730, 733, ["BiLSTM", "feature", "extractor"], "Method"], [735, 737, ["parsing", "architectures"], "Method"], [738, 741, ["transition", "-", "based"], "Method"], [772, 776, ["structured", "-", "prediction", "model"], "Method"], [780, 781, ["BiLSTM"], "Method"], [794, 798, ["BiLSTM", "feature", "-", "encoder"], "Method"], [817, 819, ["end", "training"], "Task"], [821, 824, ["structured", "prediction", "model"], "Method"], [826, 829, ["recurrent", "feature", "extractor"], "Method"], [841, 844, ["BiLSTM", "feature", "extractor"], "Method"], [846, 853, ["end", "-", "to", "-", "end", "structured", "training"], "Task"], [863, 868, ["parsing", "and", "structured", "prediction", "literature"], "Task"], [874, 875, ["parsers"], "Method"], [879, 881, ["greedy", "inference"], "Method"], [883, 887, ["transition", "-", "based", "architecture"], "Method"], [890, 898, ["first", "-", "order", ",", "arc", "-", "factored", "model"], "Method"], [910, 912, ["parsing", "architectures"], "Method"], [927, 929, ["parsing", "accuracies"], "Metric"], [931, 932, ["English"], "Material"], [934, 935, ["UAS"], "Metric"], [937, 938, ["Chinese"], "Material"], [940, 941, ["UAS"], "Metric"], [945, 949, ["first", "-", "order", "parser"], "Method"], [957, 959, ["Treebank", "data"], "Material"], [963, 967, ["semi", "-", "supervised", "signals"], "Method"], [975, 978, ["word", "-", "clusters"], "Method"], [983, 986, ["tri", "-", "training"], "Method"], [1002, 1003, ["accuracies"], "Metric"], [1005, 1006, ["UAS"], "Metric"], [1007, 1008, ["English"], "Material"], [1011, 1012, ["UAS"], "Metric"], [1013, 1014, ["Chinese"], "Material"], [1017, 1022, ["greedy", "transition", "-", "based", "parser"], "Method"], [1028, 1029, ["UAS"], "Metric"], [1030, 1031, ["En"], "Material"], [1035, 1036, ["Ch"], "Material"], [1039, 1044, ["greedy", "transition", "-", "based", "parser"], "Method"], [1116, 1118, ["Dependency", "Parsing"], "Task"], [1127, 1128, ["parsers"], "Method"], [1130, 1132, ["linear", "models"], "Method"], [1261, 1263, ["parser", "design"], "Task"], [1296, 1299, ["feature", "-", "set"], "Method"], [1303, 1307, ["transition", "-", "based", "parsing"], "Task"], [1326, 1327, ["mst"], "Method"], [1341, 1345, ["first", "-", "order", "parser"], "Method"], [1348, 1353, ["first", "order", "feature", "-", "extractor"], "Method"], [1360, 1361, ["MSTParser"], "Method"], [1374, 1378, ["transition", "-", "based", "parser"], "Method"], [1395, 1396, ["POS"], "Metric"], [1492, 1493, ["POS"], "Metric"], [1503, 1504, ["POS"], "Metric"], [1515, 1516, ["POS"], "Metric"], [1551, 1552, ["parser"], "Method"], [1581, 1586, ["low", "-", "rank", "tensor", "representation"], "Method"], [1596, 1600, ["kernel", "-", "based", "approach"], "Method"], [1618, 1620, ["neural", "networks"], "Method"], [1624, 1631, ["templates", "of", "sparse", ",", "binary", "indicator", "features"], "Method"], [1632, 1636, ["dense", "core", "feature", "encodings"], "Method"], [1638, 1642, ["non", "-", "linear", "classifiers"], "Method"], [1650, 1655, ["greedy", "transition", "-", "based", "parser"], "Method"], [1673, 1677, ["non", "-", "linear", "classifier"], "Method"], [1678, 1682, ["multi", "-", "layer", "perceptron"], "Method"], [1712, 1714, ["network", "architecture"], "Method"], [1715, 1718, ["skip", "-", "layers"], "Method"], [1719, 1723, ["beam", "search", "-", "decoding"], "Method"], [1726, 1730, ["hyper", "-", "parameter", "tuning"], "Method"], [1746, 1750, ["neural", "-", "network", "classifiers"], "Method"], [1777, 1779, ["feature", "representation"], "Method"], [1790, 1791, ["POS"], "Metric"], [1810, 1812, ["feature", "combinations"], "Task"], [1820, 1824, ["feature", "-", "engineering", "problem"], "Task"], [1827, 1831, ["neural", "-", "network", "architectures"], "Method"], [1850, 1854, ["non", "-", "linear", "classifiers"], "Method"], [1863, 1868, ["incremental", "sigmoid", "-", "belief", "networks"], "Method"], [1885, 1889, ["transition", "-", "based", "parser"], "Method"], [1893, 1896, ["stack", "-", "LSTMs"], "Method"], [1906, 1911, ["compositional", "representation", "of", "parse", "trees"], "Method"], [1919, 1921, ["compositional", "representations"], "Method"], [1937, 1938, ["reranker"], "Method"], [1941, 1943, ["reranking", "approach"], "Method"], [1948, 1951, ["convolutional", "neural", "networks"], "Method"], [1964, 1968, ["Easy", "-", "First", "parser"], "Method"], [1972, 1977, ["hierarchical", "-", "LSTM", "tree", "encoding"], "Method"], [2001, 2003, ["network", "architectures"], "Method"], [2004, 2007, ["compositional", "tree", "representations"], "Method"], [2013, 2021, ["sequence", "-", "to", "-", "sequence", "with", "attention", "architecture"], "Method"], [2022, 2024, ["constituency", "parsing"], "Task"], [2035, 2039, ["deep", "-", "BiLSTM", "representation"], "Method"], [2050, 2053, ["deep", "-", "LSTM"], "Method"], [2075, 2078, ["trainable", "attention", "mechanism"], "Method"], [2083, 2084, ["parser"], "Method"], [2100, 2102, ["BiLSTM", "encoding"], "Method"], [2103, 2109, ["end", "-", "to", "-", "end", "training"], "Task"], [2122, 2126, ["Shift", "and", "Reduce", "operations"], "Method"], [2128, 2132, ["transition", "-", "based", "parser"], "Method"], [2137, 2138, ["parser"], "Method"], [2146, 2149, ["trainable", "attention", "mechanism"], "Method"], [2156, 2157, ["parsers"], "Method"], [2159, 2163, ["transition", "-", "based", "family"], "Method"], [2172, 2178, ["human", "designed", "stack", "and", "buffer", "mechanism"], "Method"], [2191, 2194, ["trainable", "attention", "approach"], "Method"], [2213, 2214, ["learning"], "Task"], [2246, 2253, ["greedy", ",", "transition", "-", "based", "constituency", "parsing"], "Task"], [2256, 2260, ["beam", "-", "based", "architecture"], "Method"], [2275, 2278, ["Penn", "Treebank", "dataset"], "Material"], [2281, 2283, ["orthogonal", "methods"], "Method"], [2285, 2286, ["ensembling"], "Method"], [2287, 2290, ["up", "-", "training"], "Method"], [2293, 2297, ["Bidirectional", "Recurrent", "Neural", "Networks"], "Method"], [2297, 2300, ["Recurrent", "neural", "networks"], "Method"], [2301, 2302, ["RNNs"], "Method"], [2304, 2306, ["statistical", "learners"], "Method"], [2312, 2313, ["RNN"], "Method"], [2359, 2361, ["modeling", "sequences"], "Task"], [2362, 2363, ["RNNs"], "Method"], [2387, 2389, ["code", "indentation"], "Task"], [2392, 2394, ["feature", "extractors"], "Method"], [2398, 2402, ["bidirectional", "recurrent", "neural", "network"], "Method"], [2403, 2404, ["BiRNN"], "Method"], [2409, 2410, ["RNNs"], "Method"], [2427, 2428, ["RNN"], "Method"], [2430, 2436, ["long", "short", "-", "term", "memory", "network"], "Method"], [2437, 2439, ["LSTM", ")"], "Method"], [2445, 2446, ["RNN"], "Method"], [2461, 2462, ["RNNs"], "Method"], [2463, 2464, ["LSTMs"], "Method"], [2469, 2470, ["RNNs"], "Method"], [2471, 2472, ["LSTMs"], "Method"], [2487, 2490, ["recurrent", "neural", "network"], "Method"], [2491, 2492, ["RNN"], "Method"], [2496, 2498, ["parameterized", "function"], "Method"], [2562, 2563, ["RNNs"], "Method"], [2627, 2629, ["matching", "Rnn"], "Method"], [2656, 2658, ["Computational", "Complexity"], "Metric"], [2676, 2677, ["RNNs"], "Method"], [2688, 2693, ["bidirectional", "representation", "of", "all", "elements"], "Method"], [2695, 2696, ["computation"], "Task"], [2722, 2724, ["intermediate", "representations"], "Method"], [2759, 2761, ["network", "layers"], "Method"], [2778, 2782, ["back", "-", "propagation", "algorithm"], "Method"], [2809, 2810, ["optimizer"], "Method"], [2920, 2921, ["RNNs"], "Method"], [2929, 2930, ["BiRNNs"], "Method"], [2943, 2944, ["BiLSTMs"], "Method"], [2956, 2957, ["NLP"], "Task"], [2964, 2966, ["sequence", "tagging"], "Task"], [2971, 2972, ["parsing"], "Task"], [2978, 2982, ["BiLSTM", "sequence", "tagging", "model"], "Method"], [3002, 3006, ["A", "*", "CCG", "parser"], "Method"], [3013, 3014, ["LSTM"], "Method"], [3017, 3019, ["BiLSTM", "representation"], "Method"], [3047, 3050, ["discriminitively", "trained", "parser"], "Method"], [3056, 3057, ["BiLSTM"], "Method"], [3062, 3064, ["CCG", "supertags"], "Task"], [3071, 3074, ["global", "parsing", "objective"], "Metric"], [3101, 3106, ["automatically", "learned", "Bidirectional", "LSTM", "representations"], "Method"], [3118, 3119, ["POS"], "Metric"], [3126, 3127, ["POS"], "Metric"], [3150, 3151, ["POS"], "Metric"], [3220, 3221, ["parsers"], "Method"], [3231, 3235, ["non", "-", "linear", "function"], "Method"], [3238, 3242, ["multi", "-", "layer", "perceptron"], "Method"], [3243, 3246, ["one", "hidden", "layer"], "Method"], [3247, 3248, ["MLP"], "Method"], [3259, 3264, ["BiLSTM", "-", "based", "feature", "functions"], "Method"], [3270, 3272, ["parsing", "techniques"], "Method"], [3276, 3277, ["BiLSTM"], "Method"], [3285, 3287, ["parsing", "objective"], "Task"], [3299, 3301, ["parsing", "task"], "Task"], [3314, 3315, ["MLP"], "Method"], [3317, 3319, ["scoring", "function"], "Method"], [3325, 3326, ["POS"], "Metric"], [3337, 3338, ["POS"], "Metric"], [3351, 3352, ["LSTMs"], "Method"], [3368, 3370, ["scoring", "function"], "Method"], [3392, 3396, ["Parsing", "-", "time", "Complexity"], "Metric"], [3398, 3399, ["BiLSTM"], "Method"], [3402, 3403, ["parsing"], "Task"], [3409, 3411, ["BiLSTM", "encoding"], "Method"], [3426, 3427, ["parsing"], "Task"], [3433, 3435, ["feature", "extraction"], "Task"], [3451, 3455, ["Transition", "-", "based", "Parser"], "Method"], [3460, 3462, ["feature", "extractor"], "Method"], [3464, 3468, ["transition", "-", "based", "parser"], "Method"], [3477, 3482, ["transition", "-", "based", "parsing", "framework"], "Method"], [3484, 3486, ["transition", "system"], "Method"], [3488, 3490, ["abstract", "machine"], "Method"], [3499, 3501, ["transition", "system"], "Method"], [3573, 3575, ["greedy", "parser"], "Method"], [3577, 3578, ["classifier"], "Method"], [3600, 3602, ["parsing", "algorithm"], "Method"], [3614, 3619, ["Greedy", "transition", "-", "based", "parsing"], "Method"], [3637, 3638, ["parser"], "Method"], [3667, 3669, ["scoring", "function"], "Method"], [3741, 3743, ["Transition", "systems"], "Method"], [3763, 3764, ["parser"], "Method"], [3771, 3773, ["transition", "system"], "Method"], [3779, 3782, ["scoring", "function", "Score"], "Metric"], [3792, 3796, ["Arc", "-", "Hybrid", "System"], "Method"], [3797, 3799, ["transition", "systems"], "Method"], [3811, 3816, ["arc", "-", "hybrid", "transition", "system"], "Method"], [3824, 3828, ["arc", "-", "standard", "system"], "Method"], [3834, 3836, ["dynamic", "oracle"], "Method"], [3841, 3845, ["arc", "-", "hybrid", "system"], "Method"], [3931, 3935, ["arc", "-", "hybrid", "system"], "Method"], [3948, 3950, ["Shift", "transition"], "Method"], [4020, 4022, ["Scoring", "Function"], "Method"], [4025, 4027, ["scoring", "function"], "Method"], [4029, 4032, ["discriminative", "linear", "model"], "Method"], [4065, 4068, ["linear", "scoring", "model"], "Method"], [4070, 4071, ["MLP"], "Method"], [4092, 4094, ["feature", "function"], "Method"], [4121, 4123, ["feature", "extractor"], "Method"], [4128, 4130, ["feature", "function"], "Method"], [4139, 4141, ["BiLSTM", "representations"], "Method"], [4185, 4187, ["feature", "function"], "Method"], [4200, 4202, ["parsing", "accuracies"], "Metric"], [4209, 4211, ["BiLSTM", "encoding"], "Method"], [4263, 4265, ["feature", "representation"], "Method"], [4270, 4274, ["transition", "-", "based", "parsers"], "Method"], [4276, 4280, ["dynamic", "-", "programming", "search"], "Method"], [4292, 4298, ["greedy", "transition", "-", "based", "parsing", "framework"], "Method"], [4388, 4390, ["parsing", "accuracies"], "Metric"], [4405, 4409, ["fully", "-", "supervised", "case"], "Task"], [4415, 4417, ["Training", "Algorithm"], "Method"], [4418, 4420, ["training", "objective"], "Metric"], [4438, 4442, ["margin", "-", "based", "objective"], "Method"], [4502, 4503, ["parser"], "Method"], [4535, 4537, ["parsing", "process"], "Task"], [4567, 4568, ["MLP"], "Method"], [4570, 4571, ["BiLSTM"], "Method"], [4584, 4586, ["backpropagation", "algorithm"], "Method"], [4611, 4614, ["Error", "-", "Exploration"], "Method"], [4615, 4618, ["Dynamic", "Oracle", "Training"], "Task"], [4625, 4628, ["error", "exploration", "training"], "Method"], [4630, 4633, ["dynamic", "-", "oracle"], "Method"], [4649, 4650, ["parser"], "Method"], [4704, 4705, ["parser"], "Method"], [4723, 4725, ["error", "propagation"], "Task"], [4732, 4736, ["error", "-", "exploration", "training"], "Task"], [4737, 4738, ["parser"], "Method"], [4803, 4807, ["error", "-", "exploration", "training"], "Method"], [4818, 4820, ["Aggressive", "Exploration"], "Method"], [4826, 4829, ["error", "-", "exploration"], "Method"], [4850, 4853, ["error", "-", "exploration"], "Method"], [4860, 4861, ["parser"], "Method"], [4870, 4872, ["exploration", "scheme"], "Method"], [4934, 4939, ["greedy", "transition", "-", "based", "parser"], "Method"], [4946, 4950, ["margin", "-", "based", "objective"], "Metric"], [4951, 4954, ["dynamic", "oracle", "training"], "Method"], [4955, 4957, ["error", "exploration"], "Method"], [4958, 4966, ["MLP", "-", "based", "non", "-", "linear", "scoring", "function"], "Method"], [4990, 4994, ["concatenation", "of", "BiLSTM", "representations"], "Method"], [5009, 5011, ["BiLSTM", "encoder"], "Method"], [5021, 5025, ["Graph", "-", "based", "Parser"], "Method"], [5025, 5029, ["Graph", "-", "based", "parsing"], "Method"], [5032, 5035, ["structured", "prediction", "paradigm"], "Method"], [5097, 5103, ["arc", "-", "factored", "graph", "based", "approach"], "Method"], [5105, 5106, ["mst"], "Method"], [5107, 5111, ["Arc", "-", "factored", "parsing"], "Method"], [5146, 5150, ["Eisner", "’s", "decoding", "algorithm"], "Method"], [5162, 5164, ["local", "score"], "Metric"], [5169, 5171, ["linear", "model"], "Method"], [5179, 5181, ["feature", "function"], "Method"], [5200, 5203, ["linear", "scoring", "function"], "Method"], [5205, 5206, ["MLP"], "Method"], [5208, 5210, ["feature", "extractor"], "Method"], [5229, 5231, ["feature", "extractor"], "Method"], [5234, 5236, ["BiLSTM", "encoding"], "Method"], [5285, 5289, ["margin", "-", "based", "objective"], "Metric"], [5333, 5334, ["MLP"], "Method"], [5336, 5338, ["arc", "representations"], "Method"], [5341, 5342, ["loss"], "Metric"], [5346, 5351, ["sum", "of", "multiple", "neural", "networks"], "Method"], [5370, 5372, ["BiLSTM", "encoder"], "Method"], [5379, 5381, ["Labeled", "Parsing"], "Task"], [5387, 5389, ["unlabeled", "parsing"], "Task"], [5394, 5396, ["adding", "labels"], "Task"], [5430, 5433, ["slow", "parsing", "speeds"], "Task"], [5437, 5439, ["learning", "problem"], "Task"], [5479, 5480, ["parsing"], "Task"], [5491, 5493, ["feature", "representation"], "Method"], [5497, 5499, ["MLP", "predictor"], "Method"], [5505, 5509, ["margin", "based", "hinge", "loss"], "Method"], [5520, 5522, ["BiLSTM", "encoder"], "Method"], [5530, 5534, ["arc", "-", "factored", "parser"], "Method"], [5537, 5539, ["BiLSTM", "encoder"], "Method"], [5543, 5544, ["parer"], "Method"], [5546, 5547, ["labeler"], "Method"], [5559, 5563, ["multi", "-", "task", "learning"], "Task"], [5580, 5583, ["BiLSTM", "feature", "encoder"], "Method"], [5594, 5595, ["parser"], "Method"], [5596, 5598, ["unlabeled", "accuracy"], "Metric"], [5601, 5604, ["Loss", "augmented", "inference"], "Method"], [5629, 5632, ["loss", "augmented", "inference"], "Method"], [5636, 5639, ["loss", "augmented", "inference"], "Method"], [5682, 5684, ["loss", "transforms"], "Method"], [5692, 5696, ["arc", "-", "factored", "model"], "Method"], [5698, 5701, ["scoring", "of", "arcs"], "Task"], [5702, 5703, ["Scoring"], "Task"], [5707, 5708, ["MLP"], "Method"], [5710, 5712, ["hidden", "layer"], "Method"], [5715, 5719, ["matrix", "-", "vector", "multiplications"], "Method"], [5806, 5810, ["first", "-", "order", "parser"], "Method"], [5814, 5820, ["margin", "-", "based", "hinge", "-", "loss"], "Method"], [5821, 5825, ["loss", "-", "augmented", "inference"], "Method"], [5841, 5845, ["concatenation", "of", "BiLSTM", "representations"], "Method"], [5855, 5857, ["BiLSTM", "encoder"], "Method"], [5868, 5873, ["multi", "-", "task", "learning", "approach"], "Method"], [5874, 5876, ["labeled", "parsing"], "Task"], [5882, 5885, ["arc", "-", "labeler"], "Method"], [5888, 5890, ["BiLSTM", "encoder"], "Method"], [5892, 5894, ["unlabeled", "parser"], "Method"], [5903, 5905, ["parsing", "model"], "Method"], [5925, 5926, ["English"], "Material"], [5930, 5932, ["Stanford", "Dependency"], "Material"], [5933, 5934, ["SD"], "Material"], [5938, 5940, ["Penn", "Treebank"], "Material"], [5954, 5957, ["POS", "-", "tags"], "Metric"], [5983, 5984, ["Chinese"], "Material"], [5993, 5994, ["CTB5"], "Material"], [6037, 6038, ["parsers"], "Method"], [6041, 6042, ["python"], "Method"], [6045, 6047, ["PyCNN", "toolkit"], "Method"], [6048, 6051, ["neural", "network", "training"], "Method"], [6064, 6066, ["LSTM", "variant"], "Method"], [6068, 6069, ["PyCNN"], "Method"], [6074, 6076, ["Adam", "optimizer"], "Method"], [6088, 6089, ["PyCNN"], "Method"], [6095, 6097, ["learning", "rates"], "Metric"], [6103, 6104, ["POS"], "Metric"], [6119, 6122, ["parsers", "’", "networks"], "Method"], [6133, 6135, ["word", "embeddings"], "Method"], [6141, 6143, ["vector", "representation"], "Method"], [6151, 6156, ["randomly", "-", "initialized", "vector", "embedding"], "Method"], [6180, 6181, ["training"], "Task"], [6187, 6189, ["word", "dropout"], "Method"], [6250, 6251, ["parsers"], "Method"], [6274, 6276, ["Hyperparameter", "Tuning"], "Method"], [6281, 6285, ["hyper", "-", "parameter", "search"], "Method"], [6336, 6337, ["accuracies"], "Metric"], [6340, 6342, ["parsing", "models"], "Method"], [6353, 6354, ["parsers"], "Method"], [6363, 6364, ["parsers"], "Method"], [6372, 6374, ["parsing", "architectures"], "Method"], [6375, 6378, ["minimal", "feature", "extractors"], "Method"], [6409, 6413, ["third", "-", "order", "TurboParser"], "Method"], [6415, 6419, ["greedy", "transition", "based", "parser"], "Method"], [6428, 6429, ["parsers"], "Method"], [6432, 6437, ["beam", "-", "based", "transition", "parser"], "Method"], [6450, 6454, ["Stack", "-", "LSTM", "parser"], "Method"], [6462, 6463, ["parser"], "Method"], [6467, 6469, ["dynamic", "oracle"], "Method"], [6492, 6493, ["accuracy"], "Metric"], [6495, 6496, ["English"], "Material"], [6497, 6498, ["Chinese"], "Material"], [6507, 6508, ["accuracy"], "Metric"], [6530, 6534, ["semi", "-", "supervised", "parsing"], "Task"], [6545, 6547, ["greedy", "parser"], "Method"], [6594, 6595, ["Chinese"], "Material"], [6600, 6601, ["English"], "Material"], [6607, 6611, ["beam", "-", "based", "parser"], "Method"], [6644, 6647, ["Loss", "augmented", "inference"], "Task"], [6661, 6666, ["multi", "-", "task", "learning", "scheme"], "Method"], [6668, 6671, ["arc", "-", "labeler"], "Task"], [6678, 6681, ["Dynamic", "oracle", "training"], "Method"], [6686, 6687, ["English"], "Material"], [6688, 6689, ["Chinese"], "Material"], [6700, 6702, ["feature", "extraction"], "Task"], [6703, 6705, ["dependency", "parsing"], "Task"], [6708, 6710, ["BiLSTM", "encoder"], "Method"], [6716, 6717, ["parser"], "Method"], [6728, 6730, ["parsing", "models"], "Method"], [6732, 6737, ["greedy", "transition", "-", "based", "parser"], "Method"], [6752, 6754, ["parsing", "accuracies"], "Metric"], [277, 280, ["graph", "-", "based"], "Method"], [751, 754, ["graph", "-", "based"], "Method"], [763, 766, ["graph", "-", "based"], "Method"], [900, 903, ["graph", "-", "based"], "Method"], [1328, 1331, ["graph", "-", "based"], "Method"], [1481, 1484, ["graph", "-", "based"], "Method"], [1737, 1740, ["graph", "-", "based"], "Method"], [2205, 2209, ["transition", "-", "based", "parsers"], "Method"], [2337, 2338, ["RNN"], "Method"], [2557, 2558, ["RNN"], "Method"], [2634, 2635, ["BiRNN"], "Method"], [2660, 2661, ["BiRNN"], "Method"], [2705, 2706, ["BiRNN"], "Method"], [2735, 2736, ["BiRNN"], "Method"], [2740, 2741, ["BiRNN"], "Method"], [2798, 2799, ["BiRNN"], "Method"], [2826, 2827, ["BiRNN"], "Method"], [2855, 2856, ["RNN"], "Method"], [2860, 2861, ["BiRNN"], "Method"], [2866, 2867, ["BiRNN"], "Method"], [2883, 2884, ["BiRNNs"], "Method"], [2901, 2902, ["BiRNNs"], "Method"], [2905, 2906, ["BiRNNs"], "Method"], [2937, 2938, ["RNNs"], "Method"], [2952, 2953, ["BiRNNs"], "Method"], [5798, 5801, ["graph", "-", "based"], "Method"], [5906, 5907, ["English"], "Material"], [5908, 5909, ["Chinese"], "Material"], [5989, 5990, ["Chinese"], "Material"], [6265, 6266, ["UAS"], "Metric"], [6287, 6290, ["graph", "-", "based"], "Method"], [6389, 6392, ["graph", "-", "based"], "Method"], [6510, 6513, ["graph", "-", "based"], "Method"], [6536, 6539, ["graph", "-", "based"], "Method"], [6654, 6657, ["graph", "-", "based"], "Method"], [6744, 6747, ["graph", "-", "based"], "Method"]]} {"doc_id": "3daa086acd367dc971a2dc1382caba2031294233", "text": ["Holistic", ",", "Instance", "-", "level", "Human", "Parsing", "section", ":", "Abstract", "Object", "parsing", "-", "the", "task", "of", "decomposing", "an", "object", "into", "its", "semantic", "parts", "-", "has", "traditionally", "been", "formulated", "as", "a", "category", "-", "level", "segmentation", "problem", ".", "Consequently", ",", "when", "there", "are", "multiple", "objects", "in", "an", "image", ",", "current", "methods", "can", "not", "count", "the", "number", "of", "objects", "in", "the", "scene", ",", "nor", "can", "they", "determine", "which", "part", "belongs", "to", "which", "object", ".", "We", "address", "this", "problem", "by", "segmenting", "the", "parts", "of", "objects", "at", "an", "instance", "-", "level", ",", "such", "that", "each", "pixel", "in", "the", "image", "is", "assigned", "a", "part", "label", ",", "as", "well", "as", "the", "identity", "of", "the", "object", "it", "belongs", "to", ".", "Moreover", ",", "we", "show", "how", "this", "approach", "benefits", "us", "in", "obtaining", "segmentations", "at", "coarser", "granularities", "as", "well", ".", "Our", "proposed", "network", "is", "trained", "end", "-", "to", "-", "end", "given", "detections", ",", "and", "begins", "with", "a", "category", "-", "level", "segmentation", "module", ".", "Thereafter", ",", "a", "differentiable", "Conditional", "Random", "Field", ",", "defined", "over", "a", "variable", "number", "of", "instances", "for", "every", "input", "image", ",", "reasons", "about", "the", "identity", "of", "each", "part", "by", "associating", "it", "with", "a", "human", "detection", ".", "In", "contrast", "to", "other", "approaches", ",", "our", "method", "can", "handle", "the", "varying", "number", "of", "people", "in", "each", "image", "and", "our", "holistic", "network", "produces", "state", "-", "of", "-", "the", "-", "art", "results", "in", "instance", "-", "level", "part", "and", "human", "segmentation", ",", "together", "with", "competitive", "results", "in", "category", "-", "level", "part", "segmentation", ",", "all", "achieved", "by", "a", "single", "forward", "-", "pass", "through", "our", "neural", "network", ".", "section", ":", "Introduction", "Object", "parsing", ",", "the", "segmentation", "of", "an", "object", "into", "semantic", "parts", ",", "is", "naturally", "performed", "by", "humans", "to", "obtain", "a", "more", "detailed", "understanding", "of", "the", "scene", ".", "When", "performed", "automatically", "by", "computers", ",", "it", "has", "many", "practical", "applications", ",", "such", "as", "in", "human", "-", "robot", "interaction", ",", "human", "behaviour", "analysis", "and", "image", "descriptions", "for", "the", "visually", "impaired", ".", "Furthermore", ",", "detailed", "part", "information", "has", "been", "shown", "to", "be", "beneficial", "in", "other", "visual", "recognition", "tasks", "such", "as", "fine", "-", "grained", "recognition", "[", "reference", "]", ",", "human", "pose", "estimation", "[", "reference", "]", "and", "object", "detection", "[", "reference", "]", ".", "In", "this", "paper", ",", "we", "focus", "on", "the", "application", "of", "parsing", "humans", "as", "it", "is", "more", "commonly", "studied", ",", "although", "our", "method", "makes", "no", "assumptions", "on", "the", "type", "of", "object", "it", "is", "segmenting", ".", "In", "contrast", "to", "existing", "human", "parsing", "approaches", "[", "reference", "][", "reference", "][", "reference", "]", ",", "we", "operate", "at", "an", "instance", "level", "(", "to", "our", "knowledge", ",", "we", "are", "the", "first", "work", "to", "do", "so", ")", ".", "As", "shown", "in", "Fig", ".", "1", ",", "not", "only", "do", "we", "segment", "the", "various", "body", "parts", "of", "humans", "(", "Fig", ".", "1b", ")", ",", "but", "we", "associate", "each", "of", "these", "parts", "to", "one", "of", "the", "humans", "in", "the", "scene", "(", "Fig", ".", "1c", ")", ",", "which", "is", "particularly", "important", "for", "understanding", "scenes", "with", "multiple", "people", ".", "In", "contrast", "to", "existing", "instance", "segmentation", "work", "[", "reference", "]", "Part", "Segmentation", "Human", "Segmentation", "Figure", "1", ":", "Our", "proposed", "approach", "segments", "human", "parts", "at", "an", "instance", "level", "(", "c", ")", "(", "which", "to", "our", "knowledge", "is", "the", "first", "work", "to", "do", "so", ")", "from", "category", "-", "level", "part", "segmentations", "produced", "earlier", "in", "the", "network", "(", "b", ")", ".", "Moreover", ",", "we", "can", "easily", "obtain", "human", "instance", "segmentations", "(", "d", ")", "by", "taking", "the", "union", "of", "all", "pixels", "associated", "to", "a", "particular", "person", ".", "Therefore", ",", "our", "proposed", "end", "-", "to", "-", "end", "trained", "neural", "network", "parses", "humans", "into", "semantic", "parts", "at", "both", "category", "and", "instance", "level", "in", "a", "single", "forward", "-", "pass", ".", "Best", "viewed", "in", "colour", ".", "[", "reference", "]", ",", "we", "operate", "at", "a", "more", "detailed", "part", "level", ",", "enabling", "us", "to", "extract", "more", "comprehensive", "information", "of", "the", "scene", ".", "Furthermore", ",", "with", "our", "part", "-", "level", "instance", "segmentation", "of", "humans", ",", "we", "can", "easily", "recover", "human", "-", "level", "instance", "segmentation", "(", "by", "taking", "the", "union", "of", "all", "parts", "assigned", "to", "a", "particular", "instance", "as", "shown", "in", "Fig", ".", "1d", ")", ",", "and", "we", "show", "significant", "improvement", "over", "previous", "state", "-", "of", "-", "the", "-", "art", "in", "human", "instance", "-", "segmentation", "when", "doing", "so", ".", "Our", "approach", "is", "based", "on", "a", "deep", "Convolutional", "Neural", "Network", "(", "CNN", ")", ",", "which", "consists", "of", "an", "initial", "category", "-", "level", "part", "segmentation", "module", ".", "Using", "the", "output", "of", "a", "human", "detector", ",", "we", "are", "then", "able", "to", "associate", "segmented", "parts", "with", "detected", "humans", "in", "the", "image", "using", "a", "differentiable", "Conditional", "Random", "Field", "(", "CRF", ")", ",", "producing", "a", "part", "-", "level", "instance", "segmentation", "of", "the", "image", ".", "Our", "formulation", "is", "robust", "to", "false", "-", "positive", "detections", "as", "well", "as", "imperfect", "bounding", "boxes", "which", "do", "not", "cover", "the", "entire", "human", ",", "in", "contrast", "to", "other", "instance", "segmentation", "methods", "based", "on", "object", "detectors", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "Given", "object", "detections", ",", "our", "network", "is", "trained", "end", "-", "to", "-", "end", ",", "given", "detections", ",", "with", "a", "novel", "loss", "function", "which", "allows", "us", "to", "handle", "a", "variable", "number", "of", "human", "instances", "on", "every", "image", ".", "We", "evaluate", "our", "approach", "on", "the", "Pascal", "Person", "-", "Parts", "[", "reference", "]", "dataset", ",", "which", "contains", "humans", "in", "a", "diverse", "set", "of", "poses", "and", "occlusions", ".", "We", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", "on", "instancelevel", "segmentation", "of", "both", "body", "parts", "and", "humans", ".", "Moreover", ",", "our", "results", "on", "semantic", "part", "segmentation", "(", "which", "is", "not", "-", "instance", "aware", ")", "is", "also", "competitive", "with", "current", "state", "-", "of", "-", "theart", ".", "All", "of", "these", "results", "are", "achieved", "with", "a", "holistic", ",", "end", "-", "to", "-", "end", "trained", "model", "which", "parses", "humans", "at", "both", "an", "instance", "and", "category", "level", ",", "and", "outputs", "a", "dynamic", "number", "of", "instances", "per", "image", ",", "all", "in", "a", "single", "forward", "-", "pass", "through", "the", "network", ".", "section", ":", "Related", "Work", "The", "problem", "of", "object", "parsing", ",", "which", "aims", "to", "decompose", "objects", "into", "their", "semantic", "parts", ",", "has", "been", "addressed", "by", "numerous", "works", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ",", "most", "of", "which", "have", "concentrated", "on", "parsing", "humans", ".", "However", ",", "none", "of", "the", "aforementioned", "works", "have", "parsed", "objects", "at", "an", "instance", "level", "as", "shown", "in", "Fig", ".", "1", ",", "but", "rather", "category", "level", ".", "In", "fact", ",", "a", "lot", "of", "work", "on", "human", "parsing", "has", "focussed", "on", "datasets", "such", "as", "Fashionista", "[", "reference", "]", ",", "ATR", "[", "reference", "]", "and", "Deep", "Fashion", "[", "reference", "]", "where", "images", "typically", "contain", "only", "one", ",", "centred", "person", ".", "The", "notion", "of", "instance", "-", "level", "segmentation", "only", "matters", "when", "more", "than", "one", "person", "is", "present", "in", "an", "image", ",", "motivating", "us", "to", "evaluate", "our", "method", "on", "the", "Pascal", "Person", "-", "Parts", "dataset", "[", "reference", "]", "where", "multiple", "people", "can", "appear", "in", "unconstrained", "environments", ".", "Recent", "human", "parsing", "approaches", "have", "typically", "been", "similar", "to", "semantic", "segmentation", "works", "using", "fully", "convolutional", "networks", "(", "FCNs", ")", "[", "reference", "]", ",", "but", "trained", "to", "label", "parts", "[", "reference", "][", "reference", "][", "reference", "]", "instead", "of", "object", "classes", ".", "However", ",", "methods", "using", "only", "FCNs", "do", "not", "explicitly", "model", "the", "structure", "of", "a", "human", "body", ",", "and", "typically", "do", "not", "perform", "as", "well", "as", "methods", "which", "do", "[", "reference", "]", ".", "Structural", "priors", "of", "the", "human", "body", "have", "been", "encoded", "using", "pictorial", "structures", "[", "reference", "][", "reference", "]", ",", "Conditional", "Random", "Fields", "(", "CRFs", ")", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", "and", "more", "recently", ",", "with", "LSTMs", "[", "reference", "][", "reference", "]", ".", "The", "HAZN", "approach", "of", "[", "reference", "]", "addressed", "the", "problem", "that", "some", "parts", "are", "often", "very", "small", "compared", "to", "other", "parts", "and", "difficult", "to", "segment", "with", "scale", "-", "variant", "CNNs", ".", "This", "scale", "variation", "was", "handled", "by", "a", "cascade", "of", "three", "separatelytrained", "FCNs", ",", "each", "parsing", "different", "regions", "of", "the", "image", "at", "different", "scales", ".", "An", "early", "instance", "segmentation", "work", "by", "Winn", "et", "al", ".", "[", "reference", "]", "predicted", "the", "parts", "of", "an", "object", ",", "and", "then", "encouraged", "these", "parts", "to", "maintain", "a", "spatial", "ordering", ",", "characteristic", "of", "an", "instance", ",", "using", "asymmetric", "pairwise", "potentials", "in", "a", "CRF", ".", "However", ",", "subsequent", "work", "has", "not", "operated", "at", "a", "part", "level", ".", "Zhang", "et", "al", ".", "[", "reference", "][", "reference", "]", "performed", "instance", "segmentation", "of", "vehicles", "using", "an", "MRF", ".", "However", ",", "this", "graphical", "model", "was", "not", "trained", "end", "-", "to", "-", "end", "as", "done", "by", "[", "reference", "][", "reference", "][", "reference", "]", "and", "our", "approach", ".", "Furthermore", ",", "they", "assumed", "a", "maximum", "of", "9", "cars", "per", "image", ".", "Approaches", "using", "recurrent", "neural", "networks", "[", "reference", "][", "reference", "]", "can", "handle", "a", "variable", "number", "of", "instances", "per", "image", "by", "segmenting", "an", "instance", "per", "time", "-", "step", ",", "but", "are", "currently", "restricted", "to", "only", "one", "object", "category", ".", "Our", "method", ",", "on", "the", "other", "hand", ",", "is", "able", "to", "handle", "both", "an", "arbitrary", "number", "of", "objects", ",", "and", "multiple", "object", "categories", "in", "the", "image", "with", "a", "single", "forward", "-", "pass", "through", "the", "network", ".", "Various", "methods", "of", "instance", "segmentation", "have", "also", "involved", "modifying", "object", "detection", "systems", "to", "output", "segments", "instead", "of", "bounding", "boxes", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "However", ",", "these", "methods", "can", "not", "produce", "a", "segmentation", "map", "of", "the", "image", ",", "as", "shown", "in", "Fig", ".", "1", ",", "without", "postprocessing", "as", "they", "consider", "each", "detection", "independently", ".", "Although", "our", "method", "also", "uses", "an", "object", "detector", ",", "it", "considers", "all", "detections", "in", "the", "image", "jointly", "with", "an", "initial", "category", "-", "level", "segmentation", ",", "and", "produces", "segmentation", "maps", "naturally", "where", "one", "pixel", "can", "not", "belong", "to", "multiple", "instances", "in", "contrast", "to", "the", "aforementioned", "approaches", ".", "The", "idea", "of", "combining", "the", "outputs", "of", "a", "category", "-", "level", "segmentation", "network", "and", "an", "object", "detector", "to", "reason", "about", "different", "instances", "was", "also", "presented", "by", "[", "reference", "]", ".", "However", ",", "that", "system", "was", "not", "trained", "end", "-", "toend", ",", "could", "not", "segment", "instances", "outside", "the", "detector", "'s", "bounding", "box", ",", "and", "did", "not", "operate", "at", "a", "part", "level", ".", "section", ":", "Proposed", "Approach", "Our", "network", "(", "Fig", ".", "2", ")", "consists", "of", "two", "components", ":", "a", "category", "-", "level", "part", "segmentation", "module", ",", "and", "an", "instance", "segmentation", "module", ".", "As", "both", "of", "these", "modules", "are", "differentiable", ",", "they", "can", "be", "integrated", "into", "a", "single", "network", "and", "trained", "jointly", ".", "The", "instance", "segmentation", "module", "(", "Sec", ".", "3.2", ")", "uses", "the", "output", "of", "the", "first", "category", "-", "level", "segmentation", "module", "(", "Sec", ".", "3.1", ")", "as", "well", "as", "the", "outputs", "of", "an", "object", "detector", "as", "its", "input", ".", "It", "associates", "each", "pixel", "in", "the", "categorylevel", "segmentation", "with", "an", "object", "detection", ",", "resulting", "in", "an", "instance", "-", "level", "segmentation", "of", "the", "image", ".", "Given", "a", "H", "×W", "×", "3", "input", "image", ",", "I", ",", "the", "category", "-", "level", "part", "segmentation", "module", "produces", "a", "H", "×W", "×", "(", "P", "+", "1", ")", "dimensional", "output", "Q", "where", "P", "is", "the", "number", "of", "part", "classes", "in", "the", "dataset", "and", "one", "background", "class", ".", "There", "can", "be", "a", "variable", "number", ",", "D", ",", "of", "human", "detections", "per", "image", ",", "and", "the", "output", "of", "the", "instance", "segmentation", "module", "is", "an", "H", "×W", "×", "(", "PD", "+", "1", ")", "tensor", "denoting", "the", "probabilities", ",", "at", "each", "pixel", "in", "the", "image", ",", "of", "each", "of", "the", "P", "part", "classes", "belonging", "to", "one", "of", "the", "D", "detections", ".", "Two", "challenges", "of", "instance", "segmentation", "are", "the", "variable", "number", "of", "instances", "in", "every", "image", ",", "and", "the", "fact", "that", "permutations", "of", "instance", "labels", "lead", "to", "identical", "results", "(", "in", "Fig", ".", "1", ",", "how", "we", "order", "the", "different", "people", "does", "not", "matter", ")", ".", "Zhang", "et", "al", ".", "[", "reference", "][", "reference", "]", "resolve", "these", "issues", "by", "assuming", "a", "maximum", "number", "of", "instances", "and", "using", "the", "ground", "-", "truth", "depth", "ordering", "of", "instances", "respectively", ".", "Others", "have", "bypassed", "both", "of", "these", "issues", "by", "predicting", "each", "instance", "independently", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ",", "but", "this", "also", "allows", "a", "pixel", "to", "belong", "to", "multiple", "instances", ".", "Instead", ",", "we", "use", "a", "loss", "function", "(", "Sec", "3.3", ")", "that", "is", "based", "on", "\"", "matching", "\"", "the", "prediction", "to", "the", "ground", "-", "truth", ",", "allowing", "us", "to", "handle", "permutations", "of", "the", "ground", "truth", ".", "Furthermore", ",", "weight", "-", "sharing", "in", "our", "instance", "segmentation", "module", "allows", "us", "to", "segment", "a", "variable", "number", "of", "instances", "per", "image", ".", "As", "a", "result", ",", "we", "do", "not", "assume", "a", "maximum", "number", "of", "instances", ",", "consider", "all", "instances", "jointly", ",", "and", "train", "our", "network", "end", "-", "to", "-", "end", ",", "given", "object", "detections", ".", "section", ":", "Category", "-", "level", "part", "segmentation", "module", "The", "part", "segmentation", "module", "is", "a", "fully", "convolutional", "network", "[", "reference", "]", "based", "on", "ResNet", "-", "101", "[", "reference", "]", ".", "A", "common", "technique", ",", "presented", "in", "[", "reference", "][", "reference", "]", ",", "is", "to", "predict", "the", "image", "at", "three", "different", "scales", "(", "with", "network", "weights", "shared", "among", "all", "the", "scales", ")", ",", "and", "combine", "predictions", "together", "with", "learned", ",", "image", "-", "dependent", "weights", ".", "We", "take", "a", "different", "approach", "of", "fusing", "information", "at", "multiple", "scales", "-", "we", "pool", "the", "features", "after", "res5c", "[", "reference", "]", "at", "five", "different", "resolutions", "(", "by", "varying", "the", "pooling", "stride", ")", ",", "upsample", "the", "features", "to", "the", "resolution", "before", "pooling", ",", "and", "then", "concatenate", "these", "features", "before", "passing", "them", "to", "the", "final", "convolutional", "classifier", ",", "as", "proposed", "in", "[", "reference", "]", ".", "As", "we", "show", "in", "Sec", "4.4", ",", "this", "approach", "achieves", "better", "semantic", "segmentation", "results", "than", "[", "reference", "][", "reference", "]", ".", "We", "denote", "the", "output", "of", "this", "module", "by", "the", "tensor", ",", "Q", ",", "where", "Q", "i", "(", "l", ")", "is", "the", "probability", "of", "pixel", "i", "being", "assigned", "label", "l", "∈", "{", "0", ",", "1", ",", "2", ",", "...", ",", "P}.", "Further", "details", "of", "this", "module", "are", "included", "in", "the", "appendix", ".", "section", ":", "Instance", "-", "level", "segmentation", "module", "This", "module", "creates", "an", "instance", "-", "level", "segmentation", "of", "the", "image", "by", "associating", "each", "pixel", "in", "the", "input", "category", "-", "level", "segmentation", ",", "Q", ",", "with", "one", "of", "the", "D", "input", "human", "-", "detections", "or", "the", "background", "label", ".", "Let", "there", "be", "D", "input", "human", "-", "detections", "for", "the", "image", ",", "where", "the", "i", "-", "th", "detection", "is", "represented", "by", "B", "i", ",", "the", "set", "of", "pixels", "lying", "within", "the", "four", "corners", "of", "its", "bounding", "box", ",", "and", "s", "i", "∈", "[", "0", ",", "1", "]", ",", "the", "detection", "score", ".", "We", "assume", "that", "the", "0", "-", "th", "detection", "refers", "to", "the", "background", "label", ".", "Furthermore", ",", "we", "define", "a", "multinomial", "random", "variable", ",", "V", "i", ",", "at", "each", "of", "the", "N", "pixels", "in", "the", "image", ",", "and", "let", ".", "This", "variable", "can", "take", "on", "a", "label", "from", "the", "set", "{", "1", ",", "2", ",", "...", ",", "D", "}", "×", "{", "1", ",", "2", ",", "...", ",", "P", "}", "∪", "{", "(", "0", ",", "0", ")", "}", "since", "each", "of", "the", "P", "part", "labels", "can", "be", "associated", "with", "one", "of", "the", "D", "human", "detections", ",", "or", "that", "pixel", "could", "belong", "to", "the", "background", "label", ",", "(", "0", ",", "0", ")", ".", "We", "formulate", "a", "Conditional", "Random", "Field", "over", "these", "V", "variables", ",", "where", "the", "energy", "of", "the", "assignment", "v", "to", "all", "of", "the", "instance", "variables", "V", "consists", "of", "two", "unary", "terms", ",", "and", "one", "pairwise", "term", "(", "whose", "weighting", "co", "-", "efficients", "are", "all", "learned", "via", "backpropagation", ")", ":", "(", "The", "unary", "and", "pairwise", "potentials", "are", "computed", "within", "our", "neural", "network", ",", "differentiable", "with", "respect", "to", "their", "input", "and", "parameters", ",", "and", "described", "in", "Sec", ".", "3.2.1", "through", "3.2.3", ".", "The", "Maximum", "-", "a", "-", "Posteriori", "(", "MAP", ")", "estimate", "of", "our", "CRF", "(", "since", "the", "energy", "in", "Eq", ".", "1", "characterises", "a", "Gibbs", "distribution", ")", "is", "computed", "as", "the", "final", "labelling", "produced", "by", "our", "network", ".", "We", "perform", "the", "iterative", "mean", "-", "field", "inference", "algorithm", "to", "approximately", "compute", "the", "MAP", "solution", "by", "minimising", "Eq", ".", "1", ".", "As", "shown", "by", "Zheng", "et", "al", ".", "[", "reference", "]", ",", "this", "can", "be", "formulated", "as", "a", "Recurrent", "Neural", "Network", "(", "RNN", ")", ",", "allowing", "it", "to", "be", "trained", "end", "-", "to", "-", "end", "as", "part", "of", "a", "larger", "network", ".", "However", ",", "as", "our", "network", "is", "input", "a", "variable", "number", "of", "detections", "per", "image", ",", "D", ",", "the", "label", "space", "of", "the", "CRF", "is", "dynamic", ".", "Therefore", ",", "unlike", "[", "reference", "]", ",", "the", "parameters", "of", "our", "CRF", "are", "not", "class", "-", "specific", "to", "allow", "for", "this", "variable", "number", "of", "\"", "channels", "\"", ".", "section", ":", "Box", "Consistency", "Term", "We", "observe", "that", "in", "most", "cases", ",", "a", "body", "part", "belonging", "to", "a", "person", "is", "located", "inside", "the", "bounding", "box", "of", "the", "person", ".", "Based", "on", "this", "observation", ",", "the", "box", "consistency", "term", "is", "employed", "to", "encourage", "pixel", "locations", "inside", "a", "human", "bounding", "box", "B", "i", "to", "be", "associated", "with", "the", "i", "-", "th", "human", "detection", ".", "The", "box", "term", "potential", "at", "spatial", "location", "k", "for", "body", "part", "j", "of", "a", "human", "i", "is", "assigned", "either", "0", "for", "k", "/", "∈", "B", "i", ",", "or", "the", "product", "of", "the", "detection", "score", ",", "s", "i", ",", "and", "the", "category", "-", "level", "part", "segmentation", "confidence", ",", "Note", "that", "this", "potential", "may", "be", "robust", "to", "false", "-", "positive", "detections", "when", "the", "category", "-", "level", "segmentation", "and", "human", "detection", "do", "not", "agree", "with", "each", "other", ",", "since", "Q", "k", "(", "l", ")", ",", "the", "probability", "of", "a", "pixel", "k", "taking", "on", "body", "-", "part", "label", "l", ",", "is", "low", ".", "Furthermore", ",", "note", "that", "we", "use", "one", "humandetection", "to", "reason", "about", "the", "identity", "of", "all", "parts", "which", "constitute", "that", "human", ".", "section", ":", "Global", "Term", "A", "possible", "shortcoming", "for", "the", "box", "consistency", "potential", "is", "that", "if", "some", "pixels", "belonging", "to", "a", "human", "instance", "fall", "outside", "the", "bounding", "box", "and", "are", "consequently", "assigned", "0", "for", "the", "box", "consistency", "term", "potential", ",", "they", "would", "be", "lost", "in", "the", "final", "instance", "segmentation", "prediction", ".", "Visually", ",", "the", "generated", "instance", "masks", "would", "appear", "truncated", "along", "the", "bounding", "box", "boundaries", "-", "a", "problem", "suffered", "by", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "To", "overcome", "this", "undesirable", "effect", ",", "we", "introduce", "the", "global", "potential", ":", "it", "complements", "the", "box", "consistency", "term", "by", "assuming", "that", "a", "pixel", "is", "equally", "likely", "to", "belong", "to", "any", "one", "of", "the", "detected", "humans", ".", "It", "is", "expressed", "as", "Prediction", ",", "P", "Original", "ground", "-", "truth", ",", "Y", "\"", "Matched", "\"", "ground", "-", "truth", ",", "Y", "*", "Figure", "3", ":", "As", "different", "permutations", "of", "the", "ground", "-", "truth", "are", "equivalent", "in", "the", "case", "of", "instance", "segmentation", ",", "we", "\"", "match", "\"", "the", "original", "ground", "-", "truth", ",", "Y", ",", "to", "our", "network", "'s", "prediction", ",", "P", ",", "to", "obtain", "the", "\"", "matched", "\"", "ground", "-", "truth", "which", "we", "use", "to", "compute", "our", "loss", "during", "training", ".", "section", ":", "Pairwise", "Term", "Our", "pairwise", "term", "is", "composed", "of", "densely", "-", "connected", "Gaussian", "kernels", "[", "reference", "]", "which", "are", "commonly", "used", "in", "segmentation", "literature", "[", "reference", "][", "reference", "]", ".", "This", "pairwise", "potential", "encourages", "both", "spatial", "and", "appearance", "consistency", ",", "and", "we", "find", "these", "priors", "to", "be", "suitable", "in", "the", "case", "of", "instancelevel", "segmentation", "as", "well", ".", "As", "in", "[", "reference", "]", ",", "the", "weighting", "parameters", "of", "these", "potentials", "are", "learned", "via", "backpropagation", ",", "though", "in", "our", "case", ",", "the", "weights", "are", "shared", "among", "all", "classes", ".", "section", ":", "Loss", "function", "and", "network", "training", "We", "first", "pre", "-", "train", "the", "category", "-", "level", "segmentation", "part", "of", "our", "network", ",", "as", "described", "in", "the", "appendix", ".", "Thereafter", ",", "we", "add", "the", "instance", "segmentation", "module", ",", "and", "train", "with", "a", "permutationinvariant", "loss", "function", "which", "is", "backpropagated", "through", "both", "our", "instance", "-", "and", "categorylevel", "segmentation", "networks", ".", "Since", "all", "permutations", "of", "an", "instance", "segmentation", "have", "the", "same", "qualitative", "result", ",", "we", "\"", "match", "\"", "the", "original", "ground", "-", "truth", "to", "our", "prediction", "before", "computing", "the", "loss", ",", "as", "shown", "in", "Fig", ".", "3", ".", "This", "matching", "is", "based", "on", "the", "Intersection", "over", "Union", "(", "IoU", ")", "[", "reference", "]", "of", "a", "predicted", "and", "ground", "-", "truth", "instance", ",", "similar", "to", "[", "reference", "]", ".", "Let", "Y", "=", "{", "y", "1", ",", "y", "2", ",", "...", ",", "y", "m", "}", ",", "a", "set", "of", "m", "segments", ",", "denote", "the", "ground", "-", "truth", "labelling", "of", "an", "image", ",", "where", "each", "segment", "is", "an", "instance", "and", "has", "a", "part", "label", "assigned", "to", "it", ".", "Similarly", ",", "let", "P", "=", "{", "p", "1", ",", "p", "2", ",", "...", ",", "p", "n", "}", "denote", "our", "n", "predicted", "instances", ",", "each", "with", "an", "associated", "part", "label", ".", "Note", "that", "m", "and", "n", "need", "not", "be", "the", "same", "as", "we", "may", "predict", "greater", "or", "fewer", "instances", "than", "there", "actually", "are", "in", "the", "image", ".", "The", "\"", "matched", "\"", "ground", "truth", ",", "Y", "*", "is", "the", "permutation", "of", "the", "original", "ground", "-", "truth", "labelling", "which", "maximises", "the", "IoU", "between", "our", "prediction", ",", "P", "and", "ground", "-", "truth", "where", "π", "(", "Y", ")", "denotes", "the", "set", "of", "all", "permutations", "of", "Y.", "Note", "that", "we", "define", "the", "IoU", "between", "all", "segments", "of", "different", "labels", "to", "be", "0", ".", "Eq", ".", "4", "can", "be", "solved", "efficiently", "using", "the", "Hungarian", "algorithm", "as", "it", "can", "be", "formulated", "as", "a", "bipartite", "graph", "matching", "problem", ",", "and", "once", "we", "have", "the", "\"", "matched", "\"", "ground", "-", "truth", ",", "Y", "*", ",", "we", "can", "apply", "any", "loss", "function", "to", "it", "and", "train", "our", "network", "for", "segmentation", ".", "In", "our", "case", ",", "we", "use", "the", "standard", "cross", "-", "entropy", "loss", "function", "on", "the", "\"", "matched", "\"", "ground", "truth", ".", "In", "addition", ",", "we", "employ", "Online", "Hard", "Example", "Mining", "(", "OHEM", ")", ",", "and", "only", "compute", "our", "loss", "over", "the", "top", "K", "pixels", "with", "the", "highest", "loss", "in", "the", "training", "mini", "-", "batch", ".", "We", "found", "that", "during", "training", ",", "many", "pixels", "already", "had", "a", "high", "probability", "of", "being", "assigned", "to", "the", "correct", "class", ".", "By", "only", "selecting", "the", "top", "K", "pixels", "with", "the", "highest", "loss", ",", "we", "are", "able", "to", "encourage", "our", "network", "to", "improve", "on", "the", "pixels", "it", "is", "currently", "misclassifying", ",", "as", "opposed", "to", "increasing", "the", "probability", "of", "a", "pixel", "it", "is", "already", "classifying", "correctly", ".", "This", "approach", "was", "inspired", "by", "\"", "bootstrapping", "\"", "[", "reference", "][", "reference", "]", "or", "\"", "hard", "-", "negative", "mining", "\"", "[", "reference", "]", "commonly", "used", "in", "training", "object", "detectors", ".", "However", ",", "these", "methods", "mined", "hard", "examples", "from", "the", "entire", "dataset", ".", "Our", "approach", "is", "most", "similar", "to", "[", "reference", "]", ",", "who", "mined", "hard", "examples", "online", "from", "each", "mini", "-", "batch", "in", "the", "context", "of", "detection", ".", "Similar", "to", "the", "aforementioned", "works", ",", "we", "found", "OHEM", "to", "improve", "our", "overall", "results", ",", "as", "shown", "in", "Sec", ".", "4.2", ".", "section", ":", "Obtaining", "segmentations", "at", "other", "granularities", "Given", "the", "part", "instance", "prediction", "produced", "by", "our", "proposed", "network", ",", "we", "are", "able", "to", "easily", "obtain", "human", "instance", "segmentation", "and", "semantic", "part", "segmentation", ".", "In", "order", "to", "achieve", "human", "instance", "segmentation", ",", "we", "map", "the", "predicted", "part", "instance", "labels", "(", "i", ",", "j", ")", ",", "i.e.", "part", "j", "of", "person", "i", ",", "to", "i.", "Whereas", "to", "obtain", "semantic", "part", "segmentation", ",", "we", "map", "predicted", "part", "instance", "labels", "(", "i", ",", "j", ")", "to", "j", "instead", ".", "section", ":", "Experiments", "We", "describe", "our", "dataset", "and", "experimental", "set", "-", "up", "in", "Sec", ".", "4.1", ",", "before", "presenting", "results", "on", "instance", "-", "level", "part", "segmentation", "(", "Fig", ".", "1c", ")", ",", "instance", "-", "level", "human", "segmentation", "(", "Fig", ".", "1d", ")", "and", "semantic", "part", "segmentation", "(", "Fig", ".", "1b", ")", ".", "Additional", "quantitative", "and", "qualitative", "results", ",", "failure", "cases", "and", "experimental", "details", "are", "included", "in", "the", "appendix", ".", "section", ":", "Experimental", "Set", "-", "up", "We", "evaluate", "our", "proposed", "method", "on", "the", "Pascal", "Person", "-", "Part", "dataset", "[", "reference", "]", "which", "contains", "1716", "training", "images", ",", "and", "1817", "test", "images", ".", "This", "dataset", "contains", "multiple", "people", "per", "image", "in", "unconstrained", "poses", "and", "environments", ",", "and", "contains", "six", "human", "body", "part", "classes", "(", "Fig", ".", "1b", ")", ",", "as", "well", "as", "the", "background", "label", ".", "As", "described", "in", "Sec", ".", "3.3", ",", "we", "initially", "pre", "-", "train", "our", "categorylevel", "segmentation", "module", "before", "training", "for", "instance", "-", "level", "segmentation", ".", "This", "module", "is", "first", "trained", "on", "the", "21", "classes", "of", "the", "Pascal", "VOC", "dataset", "[", "reference", "]", ",", "and", "then", "finetuned", "on", "the", "seven", "classes", "of", "the", "Pascal", "Part", "training", "set", "using", "category", "-", "level", "annotations", ".", "Finally", ",", "we", "train", "for", "instance", "segmentation", "with", "instance", "-", "level", "ground", "truth", ".", "Full", "details", "of", "our", "training", "process", ",", "including", "all", "hyperparameters", "such", "as", "learning", "rate", ",", "are", "in", "the", "appendix", ".", "To", "clarify", "these", "details", ",", "we", "will", "also", "release", "our", "code", ".", "We", "use", "the", "standard", "AP", "r", "metric", "[", "reference", "]", "for", "evaluating", "instance", "-", "level", "segmentation", ":", "the", "mean", "Average", "Precision", "of", "our", "predictions", "is", "computed", "where", "a", "prediction", "is", "considered", "correct", "if", "its", "IoU", "with", "a", "ground", "-", "truth", "instance", "is", "above", "a", "certain", "threshold", ".", "This", "is", "similar", "to", "the", "AP", "metric", "used", "in", "object", "detection", ".", "However", ",", "in", "detection", ",", "the", "IoU", "between", "groundtruth", "and", "predicted", "bounding", "boxes", "is", "computed", ",", "whereas", "here", ",", "the", "IoU", "between", "regions", "is", "computed", ".", "Furthermore", ",", "in", "detection", ",", "an", "overlap", "threshold", "of", "0.5", "is", "used", ",", "whereas", "we", "vary", "this", "threshold", ".", "Finally", ",", "we", "define", "the", "AP", "r", "vol", "which", "is", "the", "mean", "of", "the", "AP", "r", "score", "for", "overlap", "thresholds", "varying", "from", "0.1", "to", "0.9", "in", "increments", "of", "0.1", ".", "We", "use", "the", "publicly", "available", "R", "-", "FCN", "detection", "framework", "[", "reference", "]", ",", "and", "train", "a", "new", "model", "with", "data", "from", "VOC", "2012", "[", "reference", "]", "that", "do", "not", "overlap", "with", "any", "of", "our", "test", "sets", ".", "We", "train", "with", "all", "object", "classes", "of", "VOC", ",", "and", "only", "use", "the", "output", "for", "the", "human", "class", ".", "Non", "-", "maximal", "suppression", "is", "performed", "on", "all", "detections", "before", "being", "fed", "into", "our", "network", ".", "Table", "1", "shows", "our", "results", "on", "part", "-", "level", "instance", "segmentation", "on", "the", "Pascal", "Person", "-", "Part", "dataset", ".", "To", "our", "knowledge", ",", "we", "are", "the", "first", "work", "to", "do", "this", ",", "and", "hence", "we", "study", "the", "effects", "of", "various", "design", "choices", "on", "overall", "performance", ".", "We", "also", "use", "the", "publicly", "available", "code", "for", "MNC", "[", "reference", "]", ",", "which", "won", "the", "MS", "-", "COCO", "2016", "instance", "segmentation", "challenge", ",", "and", "finetune", "their", "public", "model", "trained", "on", "VOC", "2011", "[", "reference", "]", "on", "Person", "-", "Part", "instances", "as", "a", "baseline", ".", "section", ":", "Results", "on", "Instance", "-", "level", "Part", "Segmentation", "We", "first", "train", "our", "model", "in", "a", "piecewise", "manner", ",", "by", "first", "optimising", "the", "parameters", "of", "the", "category", "-", "level", "segmentation", "module", ",", "and", "then", "\"", "freezing", "\"", "the", "weights", "of", "this", "module", "and", "only", "training", "the", "instance", "network", ".", "Initially", ",", "we", "only", "use", "the", "box", "consistency", "term", "(", "Sec", ".", "3.2.1", ")", "in", "the", "Instance", "CRF", ",", "resulting", "in", "an", "AP", "r", "at", "0.5", "of", "38.0", "%", ".", "Note", "that", "this", "model", "is", "equivalent", "to", "our", "reimplementation", "of", "[", "reference", "]", ".", "Adding", "in", "the", "global", "potential", "(", "Sec", ".", "3.2.2", ")", "helps", "us", "cope", "with", "bounding", "boxes", "which", "do", "not", "cover", "the", "whole", "human", ",", "and", "we", "see", "an", "improvement", "at", "all", "IoU", "thresholds", ".", "Training", "our", "entire", "network", "end", "-", "to", "-", "end", "gives", "further", "benefits", ".", "We", "then", "train", "all", "variants", "of", "our", "model", "with", "OHEM", ",", "and", "observe", "consistent", "improvements", "across", "all", "IoU", "thresholds", "with", "respect", "to", "the", "corresponding", "baseline", ".", "Here", ",", "we", "set", "K", "=", "2", "[", "reference", "]", ",", "meaning", "that", "we", "computed", "our", "loss", "over", "2", "[", "reference", "]", "or", "approximately", "12", "%", "of", "the", "hardest", "pixels", "in", "each", "training", "image", "(", "since", "we", "train", "at", "full", "resolution", ")", ".", "We", "also", "employ", "OHEM", "when", "pre", "-", "training", "the", "category", "-", "level", "segmentation", "module", "of", "our", "network", ",", "and", "observe", "minimal", "difference", "in", "the", "final", "result", "if", "we", "use", "OHEM", "when", "training", "the", "category", "-", "level", "segmentation", "module", "but", "not", "the", "instance", "segmentation", "module", ".", "Training", "end", "-", "to", "-", "end", "with", "OHEM", "achieves", "2.6", "%", "higher", "in", "AP", "r", "at", "0.5", ",", "and", "1.8", "%", "higher", "AP", "r", "vol", "over", "a", "piecewise", "-", "trained", "baseline", "model", "without", "OHEM", "and", "only", "the", "box", "term", "(", "second", "row", ")", ",", "which", "is", "equivalent", "to", "the", "model", "of", "[", "reference", "]", ".", "Furthermore", ",", "our", "AP", "r", "vol", "is", "1.7", "%", "greater", "than", "the", "strong", "MNC", "[", "reference", "]", "baseline", ".", "Note", "that", "although", "[", "reference", "]", "also", "performed", "instance", "-", "level", "segmentation", "on", "the", "same", "dataset", ",", "their", "evaluation", "was", "only", "done", "using", "human", "instance", "labels", ",", "which", "is", "similar", "to", "our", "following", "experiment", "on", "human", "instance", "segmentation", ".", "section", ":", "Results", "on", "Human", "Instance", "Segmentation", "We", "can", "trivially", "obtain", "instance", "-", "level", "segmentations", "of", "humans", "(", "Fig", "1d", ")", ",", "as", "mentioned", "in", "Sec", ".", "3.4", ".", "Table", "2", "shows", "our", "state", "-", "of", "-", "the", "-", "art", "instance", "segmentation", "results", "for", "humans", "on", "the", "VOC", "2012", "validation", "set", "[", "reference", "]", ".", "We", "use", "the", "best", "model", "from", "the", "previous", "section", "as", "there", "is", "DeepLab", "*", "[", "reference", "]", "53.0", "Attention", "[", "reference", "]", "56.4", "HAZN", "[", "reference", "]", "57.5", "LG", "-", "LSTM", "[", "reference", "]", "58.0", "Graph", "LSTM", "[", "reference", "]", "60.2", "DeepLab", "v2", "[", "reference", "]", "64.9", "RefineNet", "[", "reference", "]", "68.6", "Ours", ",", "pre", "-", "trained", "65.9", "Ours", ",", "final", "network", "66.3", "*", "Result", "reported", "in", "[", "reference", "]", "no", "overlap", "between", "the", "Pascal", "Person", "-", "Part", "training", "set", ",", "and", "the", "VOC", "2012", "validation", "set", ".", "As", "Tab", ".", "2", "shows", ",", "our", "proposed", "approach", "outperforms", "previous", "state", "-", "of", "-", "the", "-", "art", "by", "a", "significant", "margin", ",", "particularly", "at", "high", "IoU", "thresholds", ".", "Our", "model", "receives", "extra", "supervision", "in", "its", "part", "labels", ",", "but", "the", "fact", "that", "our", "network", "can", "implicitly", "infer", "relationships", "between", "different", "parts", "whilst", "training", "may", "help", "it", "handle", "occluding", "instances", "better", "than", "other", "approaches", ",", "leading", "to", "better", "instance", "segmentation", "performance", ".", "The", "fact", "that", "our", "network", "is", "trained", "with", "part", "-", "level", "annotations", "may", "also", "help", "it", "identify", "small", "features", "of", "humans", "better", ",", "leading", "to", "more", "precise", "segmentations", "and", "thus", "improvements", "at", "high", "AP", "r", "thresholds", ".", "Our", "AP", "r", "at", "each", "IoU", "threshold", "for", "human", "instance", "segmentation", "is", "higher", "than", "that", "for", "part", "instance", "segmentation", "(", "Tab", ".", "1", ")", ".", "This", "is", "because", "parts", "are", "smaller", "than", "entire", "humans", ",", "and", "thus", "more", "difficult", "to", "localise", "accurately", ".", "An", "alternate", "method", "of", "performing", "instance", "-", "level", "part", "segmentation", "may", "be", "to", "first", "obtain", "an", "instance", "-", "level", "human", "segmentation", "using", "another", "method", "from", "Tab", ".", "2", ",", "and", "then", "partition", "it", "into", "the", "various", "body", "parts", "of", "a", "human", ".", "However", ",", "our", "approach", ",", "which", "groups", "parts", "into", "instances", ",", "is", "validated", "by", "the", "fact", "that", "it", "achieves", "state", "-", "of", "-", "the", "-", "art", "instance", "-", "level", "human", "segmentation", "performance", ".", "section", ":", "Results", "on", "Category", "-", "level", "Part", "Segmentation", "Finally", ",", "our", "model", "is", "also", "able", "to", "produce", "category", "-", "level", "segmentations", "(", "as", "shown", "in", "Fig", ".", "1b", ")", ".", "This", "can", "be", "obtained", "from", "the", "output", "of", "the", "category", "-", "level", "segmentation", "module", ",", "or", "from", "our", "instance", "module", "as", "described", "in", "Sec", ".", "3.4", ".", "As", "shown", "in", "Tab", ".", "3", ",", "our", "semantic", "segmentation", "results", "are", "competitive", "with", "current", "state", "-", "of", "-", "the", "-", "art", ".", "By", "training", "our", "entire", "network", "consisting", "of", "the", "category", "-", "level", "and", "instance", "-", "level", "segmentation", "modules", "jointly", ",", "and", "then", "obtaining", "the", "semantic", "segmentation", "from", "the", "final", "instance", "segmentation", "output", "by", "our", "network", ",", "we", "are", "able", "to", "obtain", "a", "small", "improvement", "of", "0.4", "%", "in", "mean", "IoU", "over", "the", "output", "of", "the", "initial", "semantic", "segmentation", "module", ".", "section", ":", "Conclusion", "Our", "proposed", ",", "end", "-", "to", "-", "end", "trained", "network", "outputs", "instance", "-", "level", "body", "part", "and", "human", "segmentations", ",", "as", "well", "as", "category", "-", "level", "part", "segmentations", "in", "a", "single", "forward", "-", "pass", ".", "Moreover", ",", "section", ":", "Input", "Semantic", "Segmentation", "Instance", "Segmentation", "Ground", "Truth", "Figure", "4", ":", "Some", "results", "of", "our", "system", ".", "The", "first", "column", "shows", "the", "input", "image", "and", "the", "input", "detections", "we", "obtained", "from", "training", "the", "R", "-", "FCN", "detector", "[", "reference", "]", ".", "The", "second", "and", "third", "columns", "show", "our", "final", "semantic", "segmentation", "(", "Sec", ".", "3.4", ")", "and", "instance", "-", "level", "part", "segmentation", ".", "First", "row", ":", "our", "network", "can", "deal", "with", "poor", "bounding", "box", "localisation", ",", "as", "it", "manages", "to", "segment", "the", "third", "person", "from", "the", "left", "although", "the", "bounding", "box", "only", "partially", "covers", "her", ".", "Second", "row", ":", "our", "method", "is", "robust", "against", "false", "positive", "detections", "because", "of", "the", "box", "term", ".", "Observe", "that", "the", "bowl", "of", "the", "rightmost", "person", "in", "the", "bottom", "row", "is", "falsely", "detected", "as", "a", "person", ",", "but", "rejected", "in", "the", "final", "prediction", ".", "Following", "rows", ":", "we", "are", "able", "to", "handle", "overlapping", "bounding", "boxes", "by", "reasoning", "globally", "using", "the", "Instance", "CRF", ".", "we", "have", "shown", "how", "segmenting", "objects", "into", "their", "constituent", "parts", "helps", "us", "segment", "the", "object", "as", "a", "whole", "with", "our", "state", "-", "of", "-", "the", "-", "art", "results", "on", "instance", "-", "level", "segmentation", "of", "both", "body", "parts", "and", "entire", "humans", ".", "Furthermore", ",", "our", "category", "-", "level", "segmentations", "improve", "after", "training", "for", "instance", "-", "level", "segmentation", ".", "Our", "future", "work", "is", "to", "train", "the", "object", "detector", "end", "-", "to", "-", "end", "as", "well", ".", "Moreover", ",", "the", "improvement", "that", "we", "obtained", "in", "instance", "segmentation", "of", "humans", "as", "a", "result", "of", "first", "segmenting", "parts", "motivates", "us", "to", "explore", "weakly", "-", "supervised", "methods", "which", "do", "not", "require", "explicit", "object", "part", "annotations", ".", "In", "our", "main", "paper", ",", "we", "reported", "our", "AP", "r", "results", "averaged", "over", "all", "classes", ".", "Fig", ".", "5", "visualises", "the", "perclass", "results", "of", "our", "best", "model", "at", "different", "IoU", "thresholds", ".", "Fig", ".", "6", "displays", "the", "success", "cases", "of", "our", "method", ",", "while", "Fig", ".", "7", "shows", "examples", "of", "failure", "cases", ".", "Furthermore", ",", "we", "illustrate", "the", "strengths", "and", "weaknesses", "of", "our", "part", "instance", "segmentation", "method", "in", "comparison", "to", "MNC", "[", "reference", "]", "in", "Fig", ".", "8", ",", "and", "compare", "our", "instance", "-", "level", "human", "segmentation", "results", ",", "which", "we", "obtain", "by", "the", "simple", "mapping", "described", "in", "Sec", ".", "3.4", "of", "our", "main", "paper", ",", "to", "MNC", "in", "Fig", ".", "9", ".", "Finally", ",", "we", "attach", "an", "additional", "video", ".", "We", "run", "our", "system", "offline", ",", "on", "a", "frame", "-", "by", "-", "frame", "basis", "on", "the", "entire", "music", "video", ",", "and", "show", "how", "our", "method", "is", "able", "to", "accurately", "parse", "humans", "at", "both", "category", "and", "instance", "level", "on", "internet", "data", "outside", "the", "Pascal", "dataset", ".", "Instance", "-", "level", "segmentation", "of", "videos", "requires", "data", "association", ".", "We", "use", "a", "simple", ",", "greedy", "method", "which", "operates", "on", "a", "frame", "-", "by", "-", "frame", "basis", ".", "Segments", "from", "one", "frame", "are", "associated", "to", "segments", "in", "the", "next", "frame", "based", "on", "the", "IoU", ",", "using", "the", "same", "method", "we", "use", "for", "our", "loss", "function", "as", "described", "in", "Sec", ".", "3.3", "of", "the", "main", "paper", ".", "It", "shows", "that", "our", "method", "achieves", "best", "instance", "accuracy", "for", "the", "head", "category", ",", "and", "finds", "lower", "arms", "and", "lower", "legs", "most", "challenging", "to", "segment", "correctly", ".", "This", "is", "likely", "because", "of", "the", "thin", "shape", "of", "the", "lower", "limbs", "which", "is", "known", "to", "pose", "difficulty", "for", "semantic", "segmentation", ".", "section", ":", "Input", "Semantic", "Segmentation", "Instance", "Segmentation", "Ground", "Truth", "Figure", "6", ":", "Success", "cases", "of", "our", "method", ".", "The", "first", "column", "shows", "the", "input", "image", "and", "the", "input", "detections", "we", "obtained", "from", "training", "the", "R", "-", "FCN", "detector", "[", "reference", "]", ".", "The", "second", "column", "shows", "our", "final", "semantic", "segmentation", "(", "as", "described", "in", "Sec", ".", "3.4", "of", "the", "main", "paper", ")", ".", "Our", "proposed", "method", "is", "able", "to", "leverage", "an", "initial", "category", "-", "level", "segmentation", "network", "and", "human", "detections", "to", "produce", "accurate", "instance", "-", "level", "part", "segmentation", "as", "shown", "in", "the", "third", "column", ".", "First", "row", ":", "unlike", "MNC", "which", "predicts", "for", "each", "part", "instance", "independently", ",", "our", "method", "reasons", "globally", "and", "jointly", ".", "As", "a", "result", ",", "MNC", "predicts", "two", "instances", "of", "lower", "legs", "for", "the", "same", "lower", "leg", "of", "the", "second", "and", "third", "person", "from", "the", "left", ".", "Furthermore", ",", "with", "a", "dedicated", "category", "-", "level", "segmentation", "module", ",", "we", "are", "less", "prone", "to", "false", "negatives", ",", "whereas", "MNC", "misses", "the", "legs", "of", "the", "rightmost", "person", ",", "and", "the", "lower", "arm", "of", "the", "second", "person", "from", "the", "right", ".", "Second", "row", ":", "while", "we", "can", "handle", "poor", "bounding", "box", "localisation", "because", "of", "our", "global", "potential", "term", ",", "MNC", "is", "unable", "to", "segment", "regions", "outside", "the", "bounding", "boxes", "it", "generates", ".", "Consequently", ",", "only", "one", "lower", "arm", "of", "the", "person", "on", "the", "left", "is", "segmented", "as", "the", "other", "one", "is", "outside", "the", "bounding", "box", ".", "The", "square", "corners", "of", "the", "segmented", "lower", "arm", "correspond", "to", "the", "limits", "imposed", "by", "the", "bounding", "box", "which", "MNC", "internally", "uses", "(", "box", "generation", "is", "the", "first", "stage", "of", "the", "cascade", "[", "reference", "]", ")", ".", "Third", "row", ":", "By", "analysing", "an", "image", "globally", "and", "employing", "a", "differentiable", "CRF", ",", "our", "method", "can", "produce", "more", "precise", "boundaries", ".", "As", "MNC", "does", "not", "perform", "category", "-", "level", "segmentation", "over", "the", "entire", "image", ",", "it", "has", "no", "incentive", "to", "produce", "a", "coherent", "and", "continuous", "prediction", ".", "Visually", ",", "this", "is", "reflected", "in", "the", "gaps", "of", "\"", "background", "\"", "between", "body", "parts", "of", "the", "same", "person", ".", "Fourth", "row", ":", "MNC", "predicts", "two", "instances", "of", "lower", "leg", "for", "the", "second", "person", "from", "the", "right", ",", "and", "fails", "to", "segment", "any", "lower", "arms", "for", "all", "four", "people", "due", "to", "the", "aforementioned", "problems", ".", "[", "reference", "]", "using", "the", "default", "parameters", "and", "extract", "only", "its", "human", "instance", "predictions", ".", "In", "contrast", "with", "proposal", "-", "driven", "methods", "such", "as", "MNC", ",", "our", "approach", "assigns", "each", "pixel", "to", "only", "one", "instance", ",", "is", "robust", "against", "non", "-", "ideal", "bounding", "boxes", ",", "and", "often", "produces", "better", "boundaries", "due", "to", "the", "Instance", "CRF", "which", "is", "trained", "endto", "-", "end", ".", "First", "and", "second", "row", ":", "since", "MNC", "predicts", "instances", "independently", ",", "it", "is", "prone", "to", "predicting", "multiple", "instances", "for", "a", "single", "person", ".", "Third", "row", ":", "due", "to", "the", "global", "potential", "term", ",", "we", "can", "segment", "regions", "outside", "of", "a", "detection", "bounding", "box", "which", "fails", "to", "cover", "the", "entire", "person", ",", "whereas", "MNC", "is", "unable", "to", "recover", "from", "such", "imperfect", "bounding", "boxes", ",", "leading", "to", "its", "frequent", "occurrences", "of", "truncated", "instance", "predictions", ".", "Fourth", "row", ":", "a", "case", "where", "MNC", "and", "our", "method", "show", "different", "failure", "modes", ".", "MNC", "predicts", "three", "people", "where", "there", "are", "only", "two", ",", "and", "our", "method", "can", "only", "predict", "one", "instance", "due", "to", "a", "missing", "detection", ".", "MNC", "is", "unable", "to", "recover", "from", "a", "false", "positive", "detection", "and", "predicts", "two", "people", ".", "Second", "row", ":", "while", "both", "MNC", "and", "our", "method", "start", "off", "with", "poor", "bounding", "box", "localisation", "that", "does", "not", "cover", "the", "whole", "instance", ",", "we", "are", "able", "to", "segment", "the", "entire", "person", ",", "whereas", "MNC", "is", "bounded", "by", "its", "flawed", "region", "proposal", ".", "Third", "row", ":", "MNC", "performs", "better", "in", "this", "case", "as", "it", "is", "able", "to", "segment", "the", "infant", ",", "whereas", "we", "miss", "her", "completely", "due", "to", "a", "false", "negative", "person", "detection", ".", "section", ":", "B", "Additional", "information", "We", "detail", "our", "initial", "category", "-", "level", "segmentation", "module", "and", "compare", "it", "to", "DeepLab", "-", "v2", "[", "reference", "]", "in", "Sec", ".", "B.1", ",", "present", "our", "network", "training", "details", "in", "Sec", ".", "B.2", ",", "and", "finally", "describe", "how", "we", "train", "the", "MNC", "model", "which", "serves", "as", "our", "baseline", "in", "Sec", ".", "B.3", ".", "section", ":", "B.1", "Details", "of", "the", "category", "-", "level", "segmentation", "module", "As", "shown", "in", "Fig", "10b", ",", "the", "structure", "of", "our", "category", "-", "level", "segmentation", "module", "consists", "of", "a", "ResNet", "-", "101", "backbone", ",", "and", "a", "classifier", "that", "extracts", "multi", "-", "scale", "features", "from", "the", "ResNet", "-", "101", "output", "by", "using", "average", "pooling", "with", "different", "kernel", "sizes", ".", "While", "our", "category", "-", "level", "segmentation", "module", "and", "the", "Deeplab", "-", "v2", "network", "(", "Fig", ".", "10a", ")", "of", "Chen", "et", "al", ".", "[", "reference", "]", "both", "attempt", "to", "exploit", "multi", "-", "scale", "information", "in", "the", "image", ",", "the", "approach", "of", "[", "reference", "]", "entails", "executing", "three", "forward", "passes", "for", "each", "image", ",", "whereas", "we", "only", "need", "a", "single", "forward", "pass", ".", "In", "comparison", "to", "Deeplab", "-", "v2", ",", "our", "network", "saves", "both", "memory", "and", "time", ",", "and", "achieves", "better", "performance", ".", "To", "carry", "out", "a", "single", "forward", "pass", ",", "our", "network", "uses", "4.3", "GB", "of", "memory", "while", "Deeplab", "-", "v2", "[", "reference", "]", "needs", "9.5", "GB", ",", "120", "%", "more", "than", "ours", ".", "Speed", "-", "wise", ",", "our", "network", "runs", "forward", "passes", "at", "0.255", "seconds", "per", "image", "(", "3.9", "fps", ")", ",", "whereas", "Deeplab", "-", "v2", "takes", "55", "%", "longer", ",", "at", "0.396", "seconds", "per", "image", "(", "2.5", "fps", ")", "on", "average", ".", "When", "Deeplab", "-", "v2", "adds", "a", "CRF", "with", "10", "mean", "-", "field", "iterations", "to", "post", "-", "process", "the", "network", "output", ",", "it", "gains", "a", "small", "improvement", "in", "mean", "IoU", "by", "0.54", "%", "[", "reference", "]", ",", "but", "it", "requires", "11.2", "GB", "of", "memory", "to", "make", "a", "forward", "pass", "(", "140", "%", "of", "the", "total", "amount", "used", "by", "our", "full", "network", "including", "the", "instance", "-", "level", "segmentation", "module", ")", ",", "and", "takes", "0.960", "seconds", "per", "image", "(", "1.0", "fps", ")", ",", "almost", "a", "quater", "of", "our", "frame", "rate", ".", "Tests", "are", "done", "on", "a", "single", "GeForce", "GTX", "Titan", "X", "(", "Maxwell", ")", "card", ".", "Overall", ",", "we", "are", "able", "to", "achieve", "better", "segmentation", "accuracy", "(", "as", "shown", "in", "Tab", ".", "3", "of", "our", "main", "paper", ")", "and", "is", "more", "memory", "-", "and", "time", "-", "efficient", "than", "Deeplab", "-", "v2", ".", "section", ":", "B.2", "Training", "our", "proposed", "network", "B.2.1", "Training", "the", "category", "-", "level", "segmentation", "module", "We", "initialise", "our", "semantic", "segmentation", "network", "with", "the", "COCO", "pre", "-", "trained", "ResNet", "-", "101", "weights", "provided", "by", "[", "reference", "]", ".", "Training", "is", "first", "performed", "on", "the", "Pascal", "VOC", "2012", "training", "set", "using", "the", "extra", "annotations", "from", "[", "reference", "]", ",", "which", "combine", "to", "a", "total", "of", "9012", "training", "images", ".", "Care", "is", "taken", "to", "ensure", "that", "all", "images", "from", "the", "Pascal", "Person", "-", "Parts", "test", "set", "is", "excluded", "from", "this", "training", "set", ".", "A", "polynomial", "learning", "rate", "policy", "is", "adopted", "such", "that", "the", "effective", "learning", "rate", "at", "iteration", "i", "is", "given", "by", "l", "i", "=", "l", "0", "(", "1", "−", "i", "i", "max", ")", "p", ",", "where", "the", "base", "learning", "rate", ",", "l", "0", ",", "is", "set", "to", "6.25", "×", "10", "−4", ",", "the", "total", "number", "of", "iterations", ",", "i", "max", ",", "is", "set", "to", "30k", ",", "and", "the", "power", ",", "p", ",", "is", "set", "to", "0.9", ".", "A", "batch", "size", "of", "16", "is", "used", ".", "However", ",", "due", "to", "memory", "constraints", ",", "we", "simulate", "this", "batch", "size", "by", "\"", "accumulating", "gradients", "\"", ":", "We", "carry", "out", "16", "forward", "and", "backward", "passes", "with", "one", "image", "per", "iteration", ",", "and", "only", "perform", "the", "weight", "update", "after", "completing", "all", "16", "passes", ".", "We", "use", "a", "momentum", "of", "0.9", "and", "weight", "decay", "of", "1", "×", "10", "−4", "for", "these", "experiments", ".", "After", "30k", "of", "iterations", "are", "completed", ",", "we", "take", "the", "best", "performing", "model", "and", "finetune", "on", "the", "Pascal", "Person", "-", "Parts", "training", "set", "using", "the", "same", "training", "scheme", "as", "described", "above", ".", "Note", "that", "the", "parameters", "of", "the", "batch", "normalisation", "modules", "are", "kept", "unchanged", "in", "the", "whole", "learning", "process", ".", "Online", "data", "-", "augmentation", "is", "performed", "during", "training", "to", "regularise", "the", "model", ".", "The", "training", "images", "are", "randomly", "mirrored", ",", "scaled", "by", "a", "ratio", "between", "0.5", "and", "2", ",", "rotated", "by", "an", "angle", "between", "-", "10", "and", "10", "degrees", ",", "translated", "by", "a", "random", "amount", "in", "the", "HSV", "colour", "space", ",", "and", "blurred", "with", "a", "randomly", "-", "sized", "Gaussian", "kernel", ",", "all", "on", "-", "the", "-", "fly", ".", "We", "observe", "that", "these", "techniques", "are", "effective", "at", "reducing", "the", "accuracy", "gap", "between", "training", "and", "testing", ",", "leading", "to", "overall", "higher", "test", "accuracies", ".", "[", "reference", "]", "and", "our", "network", "structure", ".", "The", "numbers", "following", "the", "layer", "type", "denote", "the", "kernel", "size", "and", "number", "of", "filters", ".", "For", "pooling", "layers", ",", "only", "their", "kernel", "sizes", "are", "shown", "as", "the", "number", "of", "filters", "is", "not", "applicable", ".", "The", "upsampling", "ratios", "can", "be", "inferred", "from", "the", "context", ".", "Fig", ".", "10a", ":", "in", "the", "Deeplab", "-", "v2", "architecture", ",", "a", "513×513×3", "input", "image", "is", "downsampled", "by", "two", "different", "ratios", "(", "0.75", "and", "0.5", ")", "to", "produce", "multi", "-", "scale", "input", "at", "three", "different", "resolutions", ".", "The", "three", "resolutions", "are", "independently", "processed", "by", "a", "ResNet", "-", "101", "-", "based", "network", "using", "shared", "weights", "(", "shown", "by", "the", "individually", "coloured", "paths", ")", ".", "The", "output", "feature", "maps", "are", "then", "upsampled", "where", "appropriate", ",", "combined", "by", "taking", "the", "elementwise", "maximum", ",", "and", "finally", "upsampled", "back", "to", "513×513", ".", "Fig", ".", "10b", ":", "the", "category", "-", "level", "segmentation", "module", "proposed", "in", "this", "paper", "forwards", "an", "input", "image", "of", "size", "521×521×3", "through", "a", "ResNet", "-", "101", "-", "based", "CNN", ",", "producing", "a", "feature", "map", "of", "resolution", "66×66×2048", ".", "This", "feature", "map", "is", "average", "-", "pooled", "with", "four", "different", "kernel", "sizes", ",", "giving", "us", "four", "feature", "maps", "with", "spatial", "resolutions", "1×1", ",", "2×2", ",", "3×3", ",", "and", "6×6", "respectively", ".", "Each", "feature", "map", "undergoes", "convolution", "and", "upsampling", ",", "before", "being", "concatenated", "together", "with", "each", "other", "and", "the", "66×66×2048", "ResNet", "-", "101", "output", ".", "This", "is", "followed", "by", "a", "convolution", "layer", "that", "reduces", "the", "dimension", "of", "the", "concatenated", "features", "to", "512", ",", "and", "a", "convolutional", "classifier", "that", "maps", "the", "512", "channels", "to", "the", "size", "of", "label", "space", "in", "the", "dataset", ".", "Finally", ",", "the", "prediction", "is", "upsampled", "back", "to", "521×521", ".", "In", "both", "Fig", ".", "10a", "and", "10b", ",", "the", "ResNet", "-", "101", "backbone", "uses", "dilated", "convolution", "such", "that", "its", "output", "at", "res5c", "is", "at", "1", "/", "8", "of", "the", "input", "resolution", ",", "instead", "of", "1", "/", "32", "for", "the", "original", "ResNet", "-", "101", "[", "reference", "]", ".", "The", "convolutional", "classifiers", "(", "coloured", "in", "purple", ")", "output", "C", "channels", ",", "corresponding", "to", "the", "number", "of", "classes", "in", "the", "dataset", "including", "a", "background", "class", ".", "For", "the", "Pascal", "Person", "-", "Parts", "Dataset", ",", "C", "is", "7", ".", "Best", "viewed", "in", "colour", ".", "section", ":", "B.2.2", "Training", "the", "instance", "-", "level", "segmentation", "module", "In", "our", "model", ",", "the", "pairwise", "term", "of", "the", "fully", "-", "connected", "CRF", "takes", "the", "following", "form", ":", "where", "µ", "(", "·", ",", "·", ")", "is", "a", "compatibility", "function", ",", "k", "(", "·", ",", "·", ")", "is", "a", "kernel", "function", ",", "and", "f", "i", "is", "a", "feature", "vector", "at", "spatial", "location", "i", "containing", "the", "3", "-", "dimensional", "colour", "vector", "I", "i", "and", "the", "2", "-", "dimensional", "position", "vector", "p", "i", "[", "reference", "]", ".", "We", "further", "define", "the", "kernel", "as", "follows", ":", "where", "w", "[", "reference", "]", "and", "w", "[", "reference", "]", "are", "the", "linear", "combination", "weights", "for", "the", "bilateral", "term", "and", "the", "Gaussian", "term", "respectively", ".", "In", "order", "to", "determine", "the", "initial", "values", "for", "the", "parameters", "in", "the", "Instance", "CRF", "to", "train", "from", ",", "we", "carry", "out", "a", "random", "search", ".", "According", "to", "the", "search", "results", ",", "the", "best", "prediction", "accuracy", "is", "obtained", "by", "initialising", "w", "(", "1", ")", "=", "8", ",", "w", "(", "2", ")", "=", "2", ",", "θ", "α", "=", "2", ",", "θ", "β", "=", "8", ",", "θ", "γ", "=", "2", ".", "Furthermore", ",", "we", "use", "a", "fixed", "learning", "rate", "of", "1", "×", "10", "−6", ",", "momentum", "of", "0.9", ",", "and", "weight", "decay", "of", "1", "×", "10", "−4", "for", "training", "both", "the", "instance", "-", "level", "and", "category", "-", "level", "segmentation", "modules", "jointly", ".", "Although", "we", "previously", "use", "the", "polynomial", "learning", "rate", "policy", ",", "we", "find", "that", "for", "training", "the", "instance", "-", "level", "segmentation", "module", ",", "a", "fixed", "learning", "rate", "leads", "to", "better", "results", ".", "Furthermore", ",", "our", "experiments", "show", "that", "a", "batch", "size", "of", "one", "works", "best", "at", "this", "training", "stage", ".", "Using", "this", "scheme", ",", "we", "train", "for", "175k", "iterations", ",", "or", "approximately", "100", "epochs", ".", "section", ":", "B.3", "Training", "Multi", "-", "task", "Network", "Cascades", "(", "MNC", ")", "We", "use", "the", "publicly", "available", "Multi", "-", "task", "Network", "Cascades", "(", "MNC", ")", "framework", "[", "reference", "]", ",", "and", "train", "a", "new", "model", "for", "instance", "-", "level", "part", "segmentation", "using", "the", "Pascal", "Person", "-", "Parts", "dataset", ".", "The", "weights", "are", "initialised", "with", "the", "officially", "released", "MNC", "model", "1", "which", "has", "been", "trained", "on", "Pascal", "VOC", "2011", "/", "SBD", "[", "reference", "]", ".", "The", "base", "learning", "rate", "is", "set", "to", "1", "×", "10", "−3", ",", "which", "is", "reduced", "by", "10", "times", "after", "20k", "iterations", ".", "A", "total", "of", "25k", "training", "iterations", "are", "carried", "out", ".", "A", "batch", "size", "of", "8", ",", "momentum", "of", "0.9", "and", "weight", "decay", "of", "5", "×", "10", "−4", "are", "used", ".", "These", "settings", "are", "identical", "to", "the", "ones", "used", "in", "training", "the", "original", "MNC", "and", "provided", "in", "their", "public", "source", "code", ".", "Using", "these", "settings", ",", "we", "are", "also", "able", "to", "reproduce", "the", "experimental", "results", "obtained", "in", "the", "original", "MNC", "paper", "[", "reference", "]", ",", "and", "hence", "we", "believe", "that", "the", "MNC", "model", "we", "have", "trained", "acts", "as", "a", "strong", "baseline", "for", "our", "proposed", "approach", ".", "section", ":", "section", ":", "section", ":", "section", ":", "Appendix", "In", "this", "appendix", ",", "we", "present", "additional", "results", "of", "our", "proposed", "approach", "in", "Sec", ".", "A", ",", "and", "provide", "additional", "training", "and", "implementation", "details", "in", "Sec", ".", "B", "(", "both", "for", "our", "model", ",", "and", "the", "strong", "MNC", "baseline", "[", "reference", "]", ")", ".", "section", ":"], "entity": [[0, 7, ["Holistic", ",", "Instance", "-", "level", "Human", "Parsing"], "Task"], [10, 12, ["Object", "parsing"], "Task"], [16, 19, ["decomposing", "an", "object"], "Task"], [30, 35, ["category", "-", "level", "segmentation", "problem"], "Task"], [123, 124, ["segmentations"], "Task"], [147, 152, ["category", "-", "level", "segmentation", "module"], "Method"], [185, 187, ["human", "detection"], "Task"], [208, 210, ["holistic", "network"], "Method"], [220, 224, ["instance", "-", "level", "part"], "Method"], [225, 227, ["human", "segmentation"], "Method"], [233, 238, ["category", "-", "level", "part", "segmentation"], "Task"], [244, 247, ["forward", "-", "pass"], "Method"], [249, 251, ["neural", "network"], "Method"], [255, 257, ["Object", "parsing"], "Task"], [259, 260, ["segmentation"], "Task"], [297, 301, ["human", "-", "robot", "interaction"], "Task"], [302, 305, ["human", "behaviour", "analysis"], "Task"], [306, 308, ["image", "descriptions"], "Task"], [310, 312, ["visually", "impaired"], "Task"], [326, 329, ["visual", "recognition", "tasks"], "Task"], [331, 335, ["fine", "-", "grained", "recognition"], "Task"], [339, 342, ["human", "pose", "estimation"], "Task"], [346, 348, ["object", "detection"], "Task"], [362, 364, ["parsing", "humans"], "Task"], [390, 393, ["human", "parsing", "approaches"], "Method"], [482, 484, ["instance", "segmentation"], "Method"], [488, 492, ["Part", "Segmentation", "Human", "Segmentation"], "Task"], [522, 527, ["category", "-", "level", "part", "segmentations"], "Method"], [569, 573, ["end", "trained", "neural", "network"], "Method"], [624, 631, ["part", "-", "level", "instance", "segmentation", "of", "humans"], "Method"], [636, 641, ["human", "-", "level", "instance", "segmentation"], "Task"], [677, 681, ["human", "instance", "-", "segmentation"], "Task"], [692, 695, ["Convolutional", "Neural", "Network"], "Method"], [696, 697, ["CNN"], "Method"], [704, 710, ["category", "-", "level", "part", "segmentation", "module"], "Method"], [716, 718, ["human", "detector"], "Method"], [736, 739, ["Conditional", "Random", "Field"], "Method"], [740, 741, ["CRF"], "Method"], [745, 750, ["part", "-", "level", "instance", "segmentation"], "Method"], [781, 784, ["instance", "segmentation", "methods"], "Method"], [786, 788, ["object", "detectors"], "Method"], [820, 822, ["loss", "function"], "Method"], [843, 847, ["Pascal", "Person", "-", "Parts"], "Material"], [875, 881, ["instancelevel", "segmentation", "of", "both", "body", "parts"], "Task"], [889, 892, ["semantic", "part", "segmentation"], "Method"], [897, 899, ["instance", "aware"], "Task"], [921, 928, ["end", "-", "to", "-", "end", "trained", "model"], "Method"], [967, 969, ["object", "parsing"], "Task"], [1004, 1006, ["parsing", "humans"], "Task"], [1041, 1043, ["human", "parsing"], "Task"], [1049, 1050, ["Fashionista"], "Material"], [1054, 1055, ["ATR"], "Material"], [1059, 1061, ["Deep", "Fashion"], "Material"], [1077, 1081, ["instance", "-", "level", "segmentation"], "Method"], [1120, 1123, ["human", "parsing", "approaches"], "Method"], [1128, 1130, ["semantic", "segmentation"], "Task"], [1132, 1135, ["fully", "convolutional", "networks"], "Method"], [1136, 1137, ["FCNs"], "Method"], [1164, 1165, ["FCNs"], "Method"], [1209, 1212, ["Conditional", "Random", "Fields"], "Method"], [1213, 1214, ["CRFs"], "Method"], [1229, 1230, ["LSTMs"], "Method"], [1237, 1239, ["HAZN", "approach"], "Method"], [1262, 1266, ["scale", "-", "variant", "CNNs"], "Method"], [1274, 1275, ["cascade"], "Method"], [1277, 1279, ["separatelytrained", "FCNs"], "Method"], [1293, 1295, ["instance", "segmentation"], "Task"], [1333, 1334, ["CRF"], "Method"], [1357, 1361, ["instance", "segmentation", "of", "vehicles"], "Task"], [1363, 1364, ["MRF"], "Method"], [1368, 1370, ["graphical", "model"], "Method"], [1406, 1409, ["recurrent", "neural", "networks"], "Method"], [1481, 1483, ["instance", "segmentation"], "Task"], [1487, 1490, ["object", "detection", "systems"], "Method"], [1515, 1520, ["segmentation", "map", "of", "the", "image"], "Task"], [1543, 1545, ["object", "detector"], "Method"], [1557, 1561, ["category", "-", "level", "segmentation"], "Method"], [1564, 1566, ["segmentation", "maps"], "Task"], [1591, 1596, ["category", "-", "level", "segmentation", "network"], "Method"], [1598, 1600, ["object", "detector"], "Method"], [1661, 1667, ["category", "-", "level", "part", "segmentation", "module"], "Method"], [1670, 1673, ["instance", "segmentation", "module"], "Method"], [1695, 1698, ["instance", "segmentation", "module"], "Method"], [1708, 1714, ["first", "category", "-", "level", "segmentation", "module"], "Method"], [1726, 1728, ["object", "detector"], "Method"], [1738, 1740, ["categorylevel", "segmentation"], "Method"], [1742, 1744, ["object", "detection"], "Task"], [1748, 1752, ["instance", "-", "level", "segmentation"], "Method"], [1768, 1774, ["category", "-", "level", "part", "segmentation", "module"], "Method"], [1823, 1826, ["instance", "segmentation", "module"], "Method"], [1866, 1868, ["instance", "segmentation"], "Task"], [1980, 1982, ["Sec", "3.3"], "Method"], [2010, 2013, ["weight", "-", "sharing"], "Method"], [2015, 2018, ["instance", "segmentation", "module"], "Method"], [2060, 2062, ["object", "detections"], "Task"], [2065, 2071, ["Category", "-", "level", "part", "segmentation", "module"], "Method"], [2072, 2075, ["part", "segmentation", "module"], "Method"], [2077, 2080, ["fully", "convolutional", "network"], "Method"], [2085, 2088, ["ResNet", "-", "101"], "Method"], [2189, 2191, ["convolutional", "classifier"], "Method"], [2210, 2212, ["semantic", "segmentation"], "Task"], [2273, 2278, ["Instance", "-", "level", "segmentation", "module"], "Method"], [2282, 2286, ["instance", "-", "level", "segmentation"], "Method"], [2296, 2300, ["category", "-", "level", "segmentation"], "Method"], [2366, 2368, ["detection", "score"], "Metric"], [2373, 2377, ["0", "-", "th", "detection"], "Task"], [2524, 2525, ["backpropagation"], "Method"], [2537, 2539, ["neural", "network"], "Method"], [2559, 2568, ["Maximum", "-", "a", "-", "Posteriori", "(", "MAP", ")", "estimate"], "Method"], [2570, 2571, ["CRF"], "Method"], [2589, 2590, ["labelling"], "Task"], [2598, 2604, ["iterative", "mean", "-", "field", "inference", "algorithm"], "Method"], [2608, 2610, ["MAP", "solution"], "Method"], [2633, 2636, ["Recurrent", "Neural", "Network"], "Method"], [2637, 2639, ["RNN", ")"], "Method"], [2679, 2680, ["CRF"], "Method"], [2694, 2695, ["CRF"], "Method"], [2746, 2749, ["box", "consistency", "term"], "Method"], [2770, 2772, ["human", "detection"], "Task"], [2805, 2807, ["detection", "score"], "Metric"], [2828, 2832, ["false", "-", "positive", "detections"], "Task"], [2834, 2838, ["category", "-", "level", "segmentation"], "Task"], [2839, 2841, ["human", "detection"], "Task"], [2939, 2942, ["instance", "segmentation", "prediction"], "Task"], [3012, 3013, ["Prediction"], "Task"], [3047, 3049, ["instance", "segmentation"], "Task"], [3066, 3067, ["prediction"], "Method"], [3085, 3086, ["loss"], "Metric"], [3087, 3088, ["training"], "Task"], [3091, 3093, ["Pairwise", "Term"], "Method"], [3099, 3104, ["densely", "-", "connected", "Gaussian", "kernels"], "Method"], [3112, 3113, ["segmentation"], "Task"], [3142, 3144, ["instancelevel", "segmentation"], "Task"], [3162, 3163, ["backpropagation"], "Method"], [3182, 3184, ["network", "training"], "Method"], [3190, 3195, ["category", "-", "level", "segmentation", "part"], "Method"], [3210, 3213, ["instance", "segmentation", "module"], "Method"], [3218, 3221, ["permutationinvariant", "loss", "function"], "Method"], [3227, 3233, ["instance", "-", "and", "categorylevel", "segmentation", "networks"], "Method"], [3262, 3263, ["loss"], "Metric"], [3277, 3283, ["Intersection", "over", "Union", "(", "IoU", ")"], "Method"], [3419, 3423, ["ground", "-", "truth", "labelling"], "Method"], [3426, 3427, ["IoU"], "Metric"], [3474, 3476, ["Hungarian", "algorithm"], "Method"], [3483, 3487, ["bipartite", "graph", "matching", "problem"], "Task"], [3507, 3509, ["loss", "function"], "Method"], [3516, 3517, ["segmentation"], "Task"], [3526, 3531, ["cross", "-", "entropy", "loss", "function"], "Method"], [3544, 3548, ["Online", "Hard", "Example", "Mining"], "Task"], [3549, 3551, ["OHEM", ")"], "Method"], [3556, 3557, ["loss"], "Metric"], [3644, 3645, ["bootstrapping"], "Method"], [3653, 3657, ["hard", "-", "negative", "mining"], "Method"], [3665, 3667, ["object", "detectors"], "Method"], [3704, 3705, ["detection"], "Task"], [3714, 3715, ["OHEM"], "Method"], [3731, 3732, ["segmentations"], "Task"], [3737, 3740, ["part", "instance", "prediction"], "Method"], [3752, 3755, ["human", "instance", "segmentation"], "Method"], [3756, 3759, ["semantic", "part", "segmentation"], "Method"], [3764, 3767, ["human", "instance", "segmentation"], "Method"], [3793, 3796, ["semantic", "part", "segmentation"], "Method"], [3833, 3838, ["instance", "-", "level", "part", "segmentation"], "Method"], [3844, 3849, ["instance", "-", "level", "human", "segmentation"], "Method"], [3855, 3858, ["semantic", "part", "segmentation"], "Method"], [3894, 3899, ["Pascal", "Person", "-", "Part", "dataset"], "Material"], [3959, 3962, ["categorylevel", "segmentation", "module"], "Method"], [3965, 3969, ["instance", "-", "level", "segmentation"], "Method"], [3981, 3984, ["Pascal", "VOC", "dataset"], "Material"], [3997, 4001, ["Pascal", "Part", "training", "set"], "Material"], [4012, 4014, ["instance", "segmentation"], "Task"], [4030, 4031, ["hyperparameters"], "Method"], [4033, 4035, ["learning", "rate"], "Metric"], [4057, 4060, ["AP", "r", "metric"], "Metric"], [4065, 4069, ["instance", "-", "level", "segmentation"], "Method"], [4071, 4074, ["mean", "Average", "Precision"], "Metric"], [4081, 4082, ["prediction"], "Task"], [4105, 4107, ["AP", "metric"], "Metric"], [4109, 4111, ["object", "detection"], "Task"], [4115, 4116, ["detection"], "Task"], [4141, 4142, ["detection"], "Task"], [4192, 4197, ["R", "-", "FCN", "detection", "framework"], "Method"], [4232, 4233, ["VOC"], "Material"], [4244, 4248, ["Non", "-", "maximal", "suppression"], "Method"], [4266, 4271, ["part", "-", "level", "instance", "segmentation"], "Task"], [4273, 4278, ["Pascal", "Person", "-", "Part", "dataset"], "Material"], [4314, 4315, ["MNC"], "Method"], [4322, 4329, ["MS", "-", "COCO", "2016", "instance", "segmentation", "challenge"], "Material"], [4343, 4347, ["Person", "-", "Part", "instances"], "Material"], [4355, 4360, ["Instance", "-", "level", "Part", "Segmentation"], "Method"], [4377, 4382, ["category", "-", "level", "segmentation", "module"], "Method"], [4397, 4399, ["instance", "network"], "Method"], [4422, 4424, ["AP", "r"], "Metric"], [4500, 4501, ["OHEM"], "Method"], [4508, 4510, ["IoU", "thresholds"], "Metric"], [4563, 4564, ["OHEM"], "Method"], [4569, 4574, ["category", "-", "level", "segmentation", "module"], "Method"], [4589, 4590, ["OHEM"], "Method"], [4593, 4598, ["category", "-", "level", "segmentation", "module"], "Method"], [4601, 4604, ["instance", "segmentation", "module"], "Method"], [4612, 4613, ["OHEM"], "Method"], [4618, 4620, ["AP", "r"], "Metric"], [4632, 4637, ["piecewise", "-", "trained", "baseline", "model"], "Method"], [4638, 4639, ["OHEM"], "Method"], [4642, 4644, ["box", "term"], "Method"], [4673, 4674, ["MNC"], "Method"], [4687, 4691, ["instance", "-", "level", "segmentation"], "Method"], [4714, 4717, ["human", "instance", "segmentation"], "Method"], [4722, 4725, ["Human", "Instance", "Segmentation"], "Task"], [4729, 4733, ["instance", "-", "level", "segmentations"], "Method"], [4758, 4760, ["instance", "segmentation"], "Task"], [4762, 4763, ["humans"], "Method"], [4796, 4797, ["HAZN"], "Method"], [4801, 4804, ["LG", "-", "LSTM"], "Method"], [4808, 4810, ["Graph", "LSTM"], "Method"], [4847, 4853, ["Pascal", "Person", "-", "Part", "training", "set"], "Material"], [4929, 4931, ["instance", "segmentation"], "Task"], [4960, 4961, ["segmentations"], "Task"], [4971, 4973, ["AP", "r"], "Metric"], [4978, 4981, ["human", "instance", "segmentation"], "Method"], [4986, 4989, ["part", "instance", "segmentation"], "Task"], [5018, 5023, ["instance", "-", "level", "part", "segmentation"], "Method"], [5029, 5034, ["instance", "-", "level", "human", "segmentation"], "Method"], [5092, 5097, ["Category", "-", "level", "Part", "Segmentation"], "Task"], [5106, 5110, ["category", "-", "level", "segmentations"], "Task"], [5128, 5133, ["category", "-", "level", "segmentation", "module"], "Method"], [5137, 5139, ["instance", "module"], "Method"], [5154, 5156, ["semantic", "segmentation"], "Task"], [5192, 5194, ["semantic", "segmentation"], "Method"], [5197, 5199, ["instance", "segmentation"], "Method"], [5216, 5218, ["mean", "IoU"], "Metric"], [5224, 5227, ["semantic", "segmentation", "module"], "Method"], [5238, 5241, ["end", "trained", "network"], "Method"], [5242, 5250, ["instance", "-", "level", "body", "part", "and", "human", "segmentations"], "Task"], [5254, 5259, ["category", "-", "level", "part", "segmentations"], "Method"], [5271, 5277, ["Semantic", "Segmentation", "Instance", "Segmentation", "Ground", "Truth"], "Task"], [5302, 5306, ["R", "-", "FCN", "detector"], "Method"], [5318, 5320, ["semantic", "segmentation"], "Task"], [5326, 5331, ["instance", "-", "level", "part", "segmentation"], "Method"], [5341, 5344, ["bounding", "box", "localisation"], "Task"], [5373, 5376, ["false", "positive", "detections"], "Metric"], [5456, 5460, ["instance", "-", "level", "segmentation"], "Method"], [5462, 5464, ["body", "parts"], "Task"], [5471, 5475, ["category", "-", "level", "segmentations"], "Method"], [5479, 5483, ["instance", "-", "level", "segmentation"], "Method"], [5491, 5493, ["object", "detector"], "Method"], [5509, 5513, ["instance", "segmentation", "of", "humans"], "Task"], [5524, 5528, ["weakly", "-", "supervised", "methods"], "Method"], [5545, 5547, ["AP", "r"], "Metric"], [5600, 5604, ["part", "instance", "segmentation", "method"], "Method"], [5607, 5608, ["MNC"], "Method"], [5622, 5624, ["human", "segmentation"], "Method"], [5632, 5633, ["mapping"], "Method"], [5693, 5695, ["instance", "level"], "Metric"], [5700, 5702, ["Pascal", "dataset"], "Material"], [5703, 5709, ["Instance", "-", "level", "segmentation", "of", "videos"], "Task"], [5710, 5712, ["data", "association"], "Task"], [5718, 5720, ["greedy", "method"], "Method"], [5756, 5758, ["loss", "function"], "Method"], [5776, 5778, ["instance", "accuracy"], "Metric"], [5780, 5782, ["head", "category"], "Task"], [5815, 5817, ["semantic", "segmentation"], "Task"], [5821, 5827, ["Semantic", "Segmentation", "Instance", "Segmentation", "Ground", "Truth"], "Task"], [5852, 5856, ["R", "-", "FCN", "detector"], "Method"], [5866, 5868, ["semantic", "segmentation"], "Task"], [5890, 5895, ["category", "-", "level", "segmentation", "network"], "Method"], [5901, 5906, ["instance", "-", "level", "part", "segmentation"], "Method"], [5917, 5918, ["MNC"], "Method"], [5937, 5938, ["MNC"], "Method"], [5964, 5969, ["category", "-", "level", "segmentation", "module"], "Method"], [5979, 5980, ["MNC"], "Method"], [6008, 6011, ["bounding", "box", "localisation"], "Task"], [6018, 6019, ["MNC"], "Method"], [6073, 6074, ["MNC"], "Method"], [6077, 6079, ["box", "generation"], "Method"], [6114, 6115, ["MNC"], "Method"], [6118, 6122, ["category", "-", "level", "segmentation"], "Method"], [6136, 6138, ["continuous", "prediction"], "Task"], [6162, 6163, ["MNC"], "Method"], [6212, 6216, ["proposal", "-", "driven", "methods"], "Method"], [6218, 6219, ["MNC"], "Method"], [6262, 6263, ["MNC"], "Method"], [6308, 6309, ["MNC"], "Method"], [6325, 6328, ["truncated", "instance", "predictions"], "Task"], [6335, 6336, ["MNC"], "Method"], [6344, 6345, ["MNC"], "Method"], [6365, 6367, ["missing", "detection"], "Task"], [6368, 6369, ["MNC"], "Method"], [6375, 6378, ["false", "positive", "detection"], "Task"], [6388, 6389, ["MNC"], "Method"], [6396, 6399, ["bounding", "box", "localisation"], "Method"], [6417, 6418, ["MNC"], "Method"], [6423, 6425, ["region", "proposal"], "Method"], [6429, 6430, ["MNC"], "Method"], [6452, 6456, ["false", "negative", "person", "detection"], "Task"], [6466, 6471, ["category", "-", "level", "segmentation", "module"], "Method"], [6475, 6478, ["DeepLab", "-", "v2"], "Method"], [6521, 6526, ["category", "-", "level", "segmentation", "module"], "Method"], [6536, 6541, ["category", "-", "level", "segmentation", "module"], "Method"], [6544, 6548, ["ResNet", "-", "101", "backbone"], "Method"], [6551, 6552, ["classifier"], "Method"], [6566, 6568, ["average", "pooling"], "Method"], [6575, 6580, ["category", "-", "level", "segmentation", "module"], "Method"], [6582, 6586, ["Deeplab", "-", "v2", "network"], "Method"], [6638, 6641, ["Deeplab", "-", "v2"], "Method"], [6660, 6662, ["forward", "pass"], "Method"], [6671, 6674, ["Deeplab", "-", "v2"], "Method"], [6707, 6710, ["Deeplab", "-", "v2"], "Method"], [6722, 6723, ["fps"], "Metric"], [6728, 6731, ["Deeplab", "-", "v2"], "Method"], [6733, 6734, ["CRF"], "Method"], [6736, 6740, ["mean", "-", "field", "iterations"], "Method"], [6754, 6756, ["mean", "IoU"], "Metric"], [6785, 6787, ["full", "network"], "Method"], [6812, 6814, ["frame", "rate"], "Metric"], [6821, 6829, ["GeForce", "GTX", "Titan", "X", "(", "Maxwell", ")", "card"], "Method"], [6838, 6840, ["segmentation", "accuracy"], "Metric"], [6862, 6865, ["Deeplab", "-", "v2"], "Method"], [6876, 6881, ["category", "-", "level", "segmentation", "module"], "Method"], [6884, 6887, ["semantic", "segmentation", "network"], "Method"], [6909, 6914, ["Pascal", "VOC", "2012", "training", "set"], "Material"], [6957, 6961, ["polynomial", "learning", "rate", "policy"], "Method"], [6966, 6969, ["effective", "learning", "rate"], "Metric"], [6991, 6994, ["base", "learning", "rate"], "Metric"], [7075, 7077, ["weight", "update"], "Method"], [7139, 7142, ["batch", "normalisation", "modules"], "Method"], [7148, 7150, ["learning", "process"], "Task"], [7151, 7155, ["Online", "data", "-", "augmentation"], "Task"], [7158, 7159, ["training"], "Task"], [7209, 7211, ["Gaussian", "kernel"], "Method"], [7229, 7231, ["accuracy", "gap"], "Metric"], [7232, 7233, ["training"], "Metric"], [7234, 7235, ["testing"], "Task"], [7240, 7242, ["test", "accuracies"], "Metric"], [7301, 7305, ["Deeplab", "-", "v2", "architecture"], "Method"], [7340, 7346, ["ResNet", "-", "101", "-", "based", "network"], "Method"], [7387, 7392, ["category", "-", "level", "segmentation", "module"], "Method"], [7414, 7416, ["feature", "map"], "Method"], [7421, 7423, ["feature", "map"], "Method"], [7452, 7454, ["feature", "map"], "Method"], [7455, 7456, ["convolution"], "Method"], [7457, 7458, ["upsampling"], "Method"], [7479, 7481, ["convolution", "layer"], "Method"], [7494, 7496, ["convolutional", "classifier"], "Method"], [7530, 7534, ["ResNet", "-", "101", "backbone"], "Method"], [7535, 7537, ["dilated", "convolution"], "Method"], [7569, 7571, ["convolutional", "classifiers"], "Method"], [7750, 7752, ["random", "search"], "Method"], [7761, 7763, ["prediction", "accuracy"], "Metric"], [7802, 7804, ["learning", "rate"], "Metric"], [7826, 7835, ["instance", "-", "level", "and", "category", "-", "level", "segmentation", "modules"], "Method"], [7842, 7846, ["polynomial", "learning", "rate", "policy"], "Method"], [7861, 7863, ["learning", "rate"], "Metric"], [7905, 7910, ["Multi", "-", "task", "Network", "Cascades"], "Method"], [7911, 7912, ["MNC"], "Method"], [7918, 7923, ["Multi", "-", "task", "Network", "Cascades"], "Method"], [7924, 7925, ["MNC"], "Method"], [7937, 7942, ["instance", "-", "level", "part", "segmentation"], "Method"], [7976, 7979, ["base", "learning", "rate"], "Metric"], [8039, 8040, ["MNC"], "Method"], [157, 160, ["Conditional", "Random", "Field"], "Method"], [1102, 1106, ["Pascal", "Person", "-", "Parts"], "Material"], [2482, 2485, ["Conditional", "Random", "Field"], "Method"], [4162, 4164, ["AP", "r"], "Metric"], [4171, 4173, ["AP", "r"], "Metric"], [4209, 4210, ["VOC"], "Material"], [4337, 4338, ["VOC"], "Material"], [4417, 4418, ["CRF"], "Method"], [4627, 4629, ["AP", "r"], "Metric"], [4663, 4665, ["AP", "r"], "Metric"], [4765, 4766, ["VOC"], "Material"], [4856, 4857, ["VOC"], "Material"], [4966, 4968, ["AP", "r"], "Metric"], [5084, 5086, ["human", "segmentation"], "Method"], [5181, 5185, ["instance", "-", "level", "segmentation"], "Method"], [5425, 5426, ["CRF"], "Method"], [5644, 5645, ["MNC"], "Method"], [6103, 6104, ["CRF"], "Method"], [6248, 6249, ["CRF"], "Method"], [6503, 6504, ["MNC"], "Method"], [6789, 6793, ["instance", "-", "level", "segmentation"], "Method"], [6943, 6947, ["Pascal", "Person", "-", "Parts"], "Material"], [7118, 7122, ["Pascal", "Person", "-", "Parts"], "Material"], [7410, 7411, ["CNN"], "Method"], [7596, 7600, ["Pascal", "Person", "-", "Parts"], "Material"], [7616, 7620, ["instance", "-", "level", "segmentation"], "Method"], [7633, 7634, ["CRF"], "Method"], [7741, 7742, ["CRF"], "Method"], [7853, 7857, ["instance", "-", "level", "segmentation"], "Method"], [7944, 7948, ["Pascal", "Person", "-", "Parts"], "Material"], [7958, 7959, ["MNC"], "Method"], [7967, 7968, ["VOC"], "Material"], [8065, 8066, ["MNC"], "Method"], [8077, 8078, ["MNC"], "Method"], [8138, 8139, ["MNC"], "Method"]]} {"doc_id": "408e8eecc14c5cc60bbdfc486ba7a7fc97031788", "text": ["Discriminative", "Unsupervised", "Feature", "Learning", "with", "Convolutional", "Neural", "Networks", "Current", "methods", "for", "training", "convolutional", "neural", "networks", "depend", "on", "large", "amounts", "of", "labeled", "samples", "for", "supervised", "training", ".", "In", "this", "paper", "we", "present", "an", "approach", "for", "training", "a", "convolutional", "neural", "network", "using", "only", "unlabeled", "data", ".", "We", "train", "the", "network", "to", "discriminate", "between", "a", "set", "of", "surrogate", "classes", ".", "Each", "surrogate", "class", "is", "formed", "by", "applying", "a", "variety", "of", "transformations", "to", "a", "randomly", "sampled", "’", "seed", "’", "image", "patch", ".", "We", "find", "that", "this", "simple", "feature", "learning", "algorithm", "is", "surprisingly", "successful", "when", "applied", "to", "visual", "object", "recognition", ".", "The", "feature", "representation", "learned", "by", "our", "algorithm", "achieves", "classification", "results", "matching", "or", "outperforming", "the", "current", "state", "-", "of", "-", "the", "-", "art", "for", "unsupervised", "learning", "on", "several", "popular", "datasets", "(", "STL", "-", "10", ",", "CIFAR", "-", "10", ",", "Caltech", "-", "101", ")", ".", "1", "Introduction", "Convolutional", "neural", "networks", "(", "CNNs", ")", "trained", "via", "backpropagation", "were", "recently", "shown", "to", "perform", "well", "on", "image", "classification", "tasks", "with", "millions", "of", "training", "images", "and", "thousands", "of", "categories", "[", "1", ",", "2", "]", ".", "The", "feature", "representation", "learned", "by", "these", "networks", "achieves", "state", "-", "of", "-", "the", "-", "art", "performance", "not", "only", "on", "the", "classification", "task", "for", "which", "the", "network", "was", "trained", ",", "but", "also", "on", "various", "other", "visual", "recognition", "tasks", ",", "for", "example", ":", "classification", "on", "Caltech", "-", "101", "[", "2", ",", "3", "]", ",", "Caltech", "-", "256", "[", "2", "]", "and", "the", "CaltechUCSD", "birds", "dataset", "[", "3", "]", ";", "scene", "recognition", "on", "the", "SUN", "-", "397", "database", "[", "3", "]", ";", "detection", "on", "the", "PASCAL", "VOC", "dataset", "[", "4", "]", ".", "This", "capability", "to", "generalize", "to", "new", "datasets", "makes", "supervised", "CNN", "training", "an", "attractive", "approach", "for", "generic", "visual", "feature", "learning", ".", "The", "downside", "of", "supervised", "training", "is", "the", "need", "for", "expensive", "labeling", ",", "as", "the", "amount", "of", "required", "labeled", "samples", "grows", "quickly", "the", "larger", "the", "model", "gets", ".", "The", "large", "performance", "increase", "achieved", "by", "methods", "based", "on", "the", "work", "of", "Krizhevsky", "et", "al", ".", "[", "1", "]", "was", ",", "for", "example", ",", "only", "possible", "due", "to", "massive", "efforts", "on", "manually", "annotating", "millions", "of", "images", ".", "For", "this", "reason", ",", "unsupervised", "learning", "–", "although", "currently", "underperforming", "–", "remains", "an", "appealing", "paradigm", ",", "since", "it", "can", "make", "use", "of", "raw", "unlabeled", "images", "and", "videos", ".", "Furthermore", ",", "on", "vision", "tasks", "outside", "classification", "it", "is", "not", "even", "certain", "whether", "training", "based", "on", "object", "class", "labels", "is", "advantageous", ".", "For", "example", ",", "unsupervised", "feature", "learning", "is", "known", "to", "be", "beneficial", "for", "image", "restoration", "[", "5", "]", "and", "recent", "results", "show", "that", "it", "outperforms", "supervised", "feature", "learning", "also", "on", "descriptor", "matching", "[", "6", "]", ".", "In", "this", "work", "we", "combine", "the", "power", "of", "a", "discriminative", "objective", "with", "the", "major", "advantage", "of", "unsupervised", "feature", "learning", ":", "cheap", "data", "acquisition", ".", "We", "introduce", "a", "novel", "training", "procedure", "for", "convolutional", "neural", "networks", "that", "does", "not", "require", "any", "labeled", "data", ".", "It", "rather", "relies", "on", "an", "automatically", "generated", "surrogate", "task", ".", "The", "task", "is", "created", "by", "taking", "the", "idea", "of", "data", "augmentation", "–", "which", "is", "commonly", "used", "in", "supervised", "learning", "–", "to", "the", "extreme", ".", "Starting", "with", "trivial", "surrogate", "classes", "consisting", "of", "one", "random", "image", "patch", "each", ",", "we", "augment", "the", "data", "by", "applying", "a", "random", "set", "of", "transformations", "to", "each", "patch", ".", "Then", "we", "train", "a", "CNN", "to", "classify", "these", "surrogate", "classes", ".", "We", "refer", "to", "this", "method", "as", "exemplar", "training", "of", "convolutional", "neural", "networks", "(", "Exemplar", "-", "CNN", ")", ".", "The", "feature", "representation", "learned", "by", "Exemplar", "-", "CNN", "is", ",", "by", "construction", ",", "discriminative", "and", "invariant", "to", "typical", "transformations", ".", "We", "confirm", "this", "both", "theoretically", "and", "empirically", ",", "showing", "that", "this", "approach", "matches", "or", "outperforms", "all", "previous", "unsupervised", "feature", "learning", "methods", "on", "the", "standard", "image", "classification", "benchmarks", "STL", "-", "10", ",", "CIFAR", "-", "10", ",", "and", "Caltech", "-", "101", ".", "1.1", "Related", "Work", "Our", "approach", "is", "related", "to", "a", "large", "body", "of", "work", "on", "unsupervised", "learning", "of", "invariant", "features", "and", "training", "of", "convolutional", "neural", "networks", ".", "Convolutional", "training", "is", "commonly", "used", "in", "both", "supervised", "and", "unsupervised", "methods", "to", "utilize", "the", "invariance", "of", "image", "statistics", "to", "translations", "(", "e.g.", "LeCun", "et", "al", ".", "[", "7", "]", ",", "Kavukcuoglu", "et", "al", ".", "[", "8", "]", ",", "Krizhevsky", "et", "al", ".", "[", "1", "]", ")", ".", "Similar", "to", "our", "approach", "the", "current", "surge", "of", "successful", "methods", "employing", "convolutional", "neural", "networks", "for", "object", "recognition", "often", "rely", "on", "data", "augmentation", "to", "generate", "additional", "training", "samples", "for", "their", "classification", "objective", "(", "e.g.", "Krizhevsky", "et", "al", ".", "[", "1", "]", ",", "Zeiler", "and", "Fergus", "[", "2", "]", ")", ".", "While", "we", "share", "the", "architecture", "(", "a", "convolutional", "neural", "network", ")", "with", "these", "approaches", ",", "our", "method", "does", "not", "rely", "on", "any", "labeled", "training", "data", ".", "In", "unsupervised", "learning", ",", "several", "studies", "on", "learning", "invariant", "representations", "exist", ".", "Denoising", "autoencoders", "[", "9", "]", ",", "for", "example", ",", "learn", "features", "that", "are", "robust", "to", "noise", "by", "trying", "to", "reconstruct", "data", "from", "randomly", "perturbed", "input", "samples", ".", "Zou", "et", "al", ".", "[", "10", "]", "learn", "invariant", "features", "from", "video", "by", "enforcing", "a", "temporal", "slowness", "constraint", "on", "the", "feature", "representation", "learned", "by", "a", "linear", "autoencoder", ".", "Sohn", "and", "Lee", "[", "11", "]", "and", "Hui", "[", "12", "]", "learn", "features", "invariant", "to", "local", "image", "transformations", ".", "In", "contrast", "to", "our", "discriminative", "approach", ",", "all", "these", "methods", "rely", "on", "directly", "modeling", "the", "input", "distribution", "and", "are", "typically", "hard", "to", "use", "for", "jointly", "training", "multiple", "layers", "of", "a", "CNN", ".", "The", "idea", "of", "learning", "features", "that", "are", "invariant", "to", "transformations", "has", "also", "been", "explored", "for", "supervised", "training", "of", "neural", "networks", ".", "The", "research", "most", "similar", "to", "ours", "is", "early", "work", "on", "tangent", "propagation", "[", "13", "]", "(", "and", "the", "related", "double", "backpropagation", "[", "14", "]", ")", "which", "aims", "to", "learn", "invariance", "to", "small", "predefined", "transformations", "in", "a", "neural", "network", "by", "directly", "penalizing", "the", "derivative", "of", "the", "output", "with", "respect", "to", "the", "magnitude", "of", "the", "transformations", ".", "In", "contrast", ",", "our", "algorithm", "does", "not", "regularize", "the", "derivative", "explicitly", ".", "Thus", "it", "is", "less", "sensitive", "to", "the", "magnitude", "of", "the", "applied", "transformation", ".", "This", "work", "is", "also", "loosely", "related", "to", "the", "use", "of", "unlabeled", "data", "for", "regularizing", "supervised", "algorithms", ",", "for", "example", "self", "-", "training", "[", "15", "]", "or", "entropy", "regularization", "[", "16", "]", ".", "In", "contrast", "to", "these", "semi", "-", "supervised", "methods", ",", "Exemplar", "-", "CNN", "training", "does", "not", "require", "any", "labeled", "data", ".", "Finally", ",", "the", "idea", "of", "creating", "an", "auxiliary", "task", "in", "order", "to", "learn", "a", "good", "data", "representation", "was", "used", "by", "Ahmed", "et", "al", ".", "[", "17", "]", ",", "Collobert", "et", "al", ".", "[", "18", "]", ".", "2", "Creating", "Surrogate", "Training", "Data", "The", "input", "to", "the", "training", "procedure", "is", "a", "set", "of", "unlabeled", "images", ",", "which", "come", "from", "roughly", "the", "same", "distribution", "as", "the", "images", "to", "which", "we", "later", "aim", "to", "apply", "the", "learned", "features", ".", "We", "randomly", "sample", "N", "∈", "[", "50", ",", "32000", "]", "patches", "of", "size", "32×32", "pixels", "from", "different", "images", "at", "varying", "positions", "and", "scales", "forming", "the", "initial", "training", "set", "X", "=", "{", "x1", ",", ".", ".", ".xN}.", "We", "are", "interested", "in", "patches", "containing", "objects", "or", "parts", "of", "objects", ",", "hence", "we", "sample", "only", "from", "regions", "containing", "considerable", "gradients", ".", "We", "define", "a", "family", "of", "transformations", "{", "Tα|α", "∈", "A", "}", "parameterized", "by", "vectors", "α", "∈", "A", ",", "where", "A", "is", "the", "set", "of", "all", "possible", "parameter", "vectors", ".", "Each", "transformation", "Tα", "is", "a", "composition", "of", "elementary", "transformations", "from", "the", "following", "list", ":", "•", "translation", ":", "vertical", "or", "horizontal", "translation", "by", "a", "distance", "within", "0.2", "of", "the", "patch", "size", ";", "•", "scaling", ":", "multiplication", "of", "the", "patch", "scale", "by", "a", "factor", "between", "0.7", "and", "1.4", ";", "•", "rotation", ":", "rotation", "of", "the", "image", "by", "an", "angle", "up", "to", "20", "degrees", ";", "•", "contrast", "1", ":", "multiply", "the", "projection", "of", "each", "patch", "pixel", "onto", "the", "principal", "components", "of", "the", "set", "of", "all", "pixels", "by", "a", "factor", "between", "0.5", "and", "2", "(", "factors", "are", "independent", "for", "each", "principal", "component", "and", "the", "same", "for", "all", "pixels", "within", "a", "patch", ")", ";", "•", "contrast", "2", ":", "raise", "saturation", "and", "value", "(", "S", "and", "V", "components", "of", "the", "HSV", "color", "representation", ")", "of", "all", "pixels", "to", "a", "power", "between", "0.25", "and", "4", "(", "same", "for", "all", "pixels", "within", "a", "patch", ")", ",", "multiply", "these", "values", "by", "a", "factor", "between", "0.7", "and", "1.4", ",", "add", "to", "them", "a", "value", "between", "−0.1", "and", "0.1", ";", "•", "color", ":", "add", "a", "value", "between", "−0.1", "and", "0.1", "to", "the", "hue", "(", "H", "component", "of", "the", "HSV", "color", "representation", ")", "of", "all", "pixels", "in", "the", "patch", "(", "the", "same", "value", "is", "used", "for", "all", "pixels", "within", "a", "patch", ")", ".", "All", "numerical", "parameters", "of", "elementary", "transformations", ",", "when", "concatenated", "together", ",", "form", "a", "single", "parameter", "vector", "α", ".", "For", "each", "initial", "patch", "xi", "∈", "X", "we", "sample", "K", "∈", "[", "1", ",", "300", "]", "random", "parameter", "vectors", "{", "α1i", ",", ".", ".", ".", ",", "αKi", "}", "and", "apply", "the", "corresponding", "transformations", "Ti", "=", "{", "Tα1i", ",", ".", ".", ".", ",", "TαKi", "}", "to", "the", "patch", "xi", ".", "This", "yields", "the", "set", "of", "its", "transformed", "versions", "Sxi", "=", "Tixi", "=", "{", "Txi|T", "∈", "Ti}.", "Afterwards", "we", "subtract", "the", "mean", "of", "each", "pixel", "over", "the", "whole", "resulting", "dataset", ".", "We", "do", "not", "apply", "any", "other", "preprocessing", ".", "Exemplary", "patches", "sampled", "from", "the", "STL", "-", "10", "unlabeled", "dataset", "are", "shown", "in", "Fig", ".", "1", ".", "Examples", "of", "transformed", "versions", "of", "one", "patch", "are", "shown", "in", "Fig", ".", "2", ".", "3", "Learning", "Algorithm", "Given", "the", "sets", "of", "transformed", "image", "patches", ",", "we", "declare", "each", "of", "these", "sets", "to", "be", "a", "class", "by", "assigning", "label", "i", "to", "the", "class", "Sxi", ".", "We", "next", "train", "a", "CNN", "to", "discriminate", "between", "these", "surrogate", "classes", ".", "Formally", ",", "we", "minimize", "the", "following", "loss", "function", ":", "L", "(", "X", ")", "=", "∑", "xi∈X", "∑", "T∈Ti", "l", "(", "i", ",", "Txi", ")", ",", "(", "1", ")", "where", "l", "(", "i", ",", "Txi", ")", "is", "the", "loss", "on", "the", "transformed", "sample", "Txi", "with", "(", "surrogate", ")", "true", "label", "i.", "We", "use", "a", "CNN", "with", "a", "softmax", "output", "layer", "and", "optimize", "the", "multinomial", "negative", "log", "likelihood", "of", "the", "network", "output", ",", "hence", "in", "our", "case", "l", "(", "i", ",", "Txi", ")", "=", "M", "(", "ei", ",", "f", "(", "Txi", ")", ")", ",", "M", "(", "y", ",", "f", ")", "=", "−〈y", ",", "log", "f", "〉", "=", "−", "∑", "k", "yk", "log", "fk", ",", "(", "2", ")", "where", "f", "(", "·", ")", "denotes", "the", "function", "computing", "the", "values", "of", "the", "output", "layer", "of", "the", "CNN", "given", "the", "input", "data", ",", "and", "ei", "is", "the", "ith", "standard", "basis", "vector", ".", "We", "note", "that", "in", "the", "limit", "of", "an", "infinite", "number", "of", "transformations", "per", "surrogate", "class", ",", "the", "objective", "function", "(", "1", ")", "takes", "the", "form", "L̂", "(", "X", ")", "=", "∑", "xi∈X", "Eα", "[", "l", "(", "i", ",", "Tαxi", ")", "]", ",", "(", "3", ")", "which", "we", "shall", "analyze", "in", "the", "next", "section", ".", "Intuitively", ",", "the", "classification", "problem", "described", "above", "serves", "to", "ensure", "that", "different", "input", "samples", "can", "be", "distinguished", ".", "At", "the", "same", "time", ",", "it", "enforces", "invariance", "to", "the", "specified", "transformations", ".", "In", "the", "following", "sections", "we", "provide", "a", "foundation", "for", "this", "intuition", ".", "We", "first", "present", "a", "formal", "analysis", "of", "the", "objective", ",", "separating", "it", "into", "a", "well", "defined", "classification", "problem", "and", "a", "regularizer", "that", "enforces", "invariance", "(", "resembling", "the", "analysis", "in", "Wager", "et", "al", ".", "[", "19", "]", ")", ".", "We", "then", "discuss", "the", "derived", "properties", "of", "this", "classification", "problem", "and", "compare", "it", "to", "common", "practices", "for", "unsupervised", "feature", "learning", ".", "3.1", "Formal", "Analysis", "We", "denote", "by", "α", "∈", "A", "the", "random", "vector", "of", "transformation", "parameters", ",", "by", "g", "(", "x", ")", "the", "vector", "of", "activations", "of", "the", "second", "-", "to", "-", "last", "layer", "of", "the", "network", "when", "presented", "the", "input", "patch", "x", ",", "by", "W", "the", "matrix", "of", "the", "weights", "of", "the", "last", "network", "layer", ",", "by", "h", "(", "x", ")", "=", "Wg", "(", "x", ")", "the", "last", "layer", "activations", "before", "applying", "the", "softmax", ",", "and", "by", "f", "(", "x", ")", "=", "softmax", "(", "h", "(", "x", ")", ")", "the", "output", "of", "the", "network", ".", "By", "plugging", "in", "the", "definition", "of", "the", "softmax", "activation", "function", "softmax", "(", "z", ")", "=", "exp", "(", "z", ")/", "‖", "exp", "(", "z", ")", "‖1", "(", "4", ")", "the", "objective", "function", "(", "3", ")", "with", "loss", "(", "2", ")", "takes", "the", "form∑", "xi∈X", "Eα", "[", "−〈ei", ",", "h", "(", "Tαxi", ")", "〉", "+", "log", "‖", "exp", "(", "h", "(", "Tαxi", "))", "‖1", "]", ".", "(", "5", ")", "With", "ĝi", "=", "Eα", "[", "g", "(", "Tαxi", ")", "]", "being", "the", "average", "feature", "representation", "of", "transformed", "versions", "of", "the", "image", "patch", "xi", "we", "can", "rewrite", "Eq", ".", "(", "5", ")", "as∑", "xi∈X", "[", "−〈ei", ",", "Wĝi〉", "+", "log", "‖", "exp", "(", "Wĝi", ")", "‖1", "]", "+", "∑", "xi∈X", "[", "Eα", "[", "log", "‖", "exp", "(", "h", "(", "Tαxi", "))", "‖1", "]", "−", "log", "‖", "exp", "(", "Wĝi", ")", "‖1", "]", ".", "(", "6", ")", "The", "first", "sum", "is", "the", "objective", "function", "of", "a", "multinomial", "logistic", "regression", "problem", "with", "input", "-", "target", "pairs", "(", "ĝi", ",", "ei", ")", ".", "This", "objective", "falls", "back", "to", "the", "transformation", "-", "free", "instance", "classification", "problem", "L", "(", "X", ")", "=", "∑", "xi∈X", "l", "(", "i", ",", "xi", ")", "if", "g", "(", "xi", ")", "=", "Eα", "[", "g", "(", "Tαx", ")]", ".", "In", "general", ",", "this", "equality", "does", "not", "hold", "and", "thus", "the", "first", "sum", "enforces", "correct", "classification", "of", "the", "average", "representation", "Eα", "[", "g", "(", "Tαxi", ")", "]", "for", "a", "given", "input", "sample", ".", "For", "a", "truly", "invariant", "representation", ",", "however", ",", "the", "equality", "is", "achieved", ".", "Similarly", ",", "if", "we", "suppose", "that", "Tαx", "=", "x", "for", "α", "=", "0", ",", "that", "for", "small", "values", "of", "α", "the", "feature", "representation", "g", "(", "Tαxi", ")", "is", "approximately", "linear", "with", "respect", "to", "α", "and", "that", "the", "random", "variable", "α", "is", "centered", ",", "i.e.", "Eα", "[", "α", "]", "=", "0", ",", "then", "ĝi", "=", "Eα", "[", "g", "(", "Tαxi", ")", "]", "≈", "Eα", "[", "g", "(", "xi", ")", "+", "∇α", "(", "g", "(", "Tαxi", "))", "|α=0", "α", "]", "=", "g", "(", "xi", ")", ".", "The", "second", "sum", "in", "Eq", ".", "(", "6", ")", "can", "be", "seen", "as", "a", "regularizer", "enforcing", "all", "h", "(", "Tαxi", ")", "to", "be", "close", "to", "their", "average", "value", ",", "i.e.", ",", "the", "feature", "representation", "is", "sought", "to", "be", "approximately", "invariant", "to", "the", "transformations", "Tα", ".", "To", "show", "this", "we", "use", "the", "convexity", "of", "the", "function", "log", "‖", "exp", "(", "·", ")", "‖1", "and", "Jensen", "’s", "inequality", ",", "which", "yields", "(", "proof", "in", "supplementary", "material", ")", "Eα", "[", "log", "‖", "exp", "(", "h", "(", "Tαxi", "))", "‖1", "]", "−", "log", "‖", "exp", "(", "Wĝi", ")", "‖1", "≥", "0", ".", "(", "7", ")", "If", "the", "feature", "representation", "is", "perfectly", "invariant", ",", "then", "h", "(", "Tαxi", ")", "=", "Wĝi", "and", "inequality", "(", "7", ")", "turns", "to", "equality", ",", "meaning", "that", "the", "regularizer", "reaches", "its", "global", "minimum", ".", "3.2", "Conceptual", "Comparison", "to", "Previous", "Unsupervised", "Learning", "Methods", "Suppose", "we", "want", "to", "unsupervisedly", "learn", "a", "feature", "representation", "useful", "for", "a", "recognition", "task", ",", "for", "example", "classification", ".", "The", "mapping", "from", "input", "images", "x", "to", "a", "feature", "representation", "g", "(", "x", ")", "should", "then", "satisfy", "two", "requirements", ":", "(", "1", ")", "there", "must", "be", "at", "least", "one", "feature", "that", "is", "similar", "for", "images", "of", "the", "same", "category", "y", "(", "invariance", ")", ";", "(", "2", ")", "there", "must", "be", "at", "least", "one", "feature", "that", "is", "sufficiently", "different", "for", "images", "of", "different", "categories", "(", "ability", "to", "discriminate", ")", ".", "Most", "unsupervised", "feature", "learning", "methods", "aim", "to", "learn", "such", "a", "representation", "by", "modeling", "the", "input", "distribution", "p", "(", "x", ")", ".", "This", "is", "based", "on", "the", "assumption", "that", "a", "good", "model", "of", "p", "(", "x", ")", "contains", "information", "about", "the", "category", "distribution", "p", "(", "y|x", ")", ".", "That", "is", ",", "if", "a", "representation", "is", "learned", ",", "from", "which", "a", "given", "sample", "can", "be", "reconstructed", "perfectly", ",", "then", "the", "representation", "is", "expected", "to", "also", "encode", "information", "about", "the", "category", "of", "the", "sample", "(", "ability", "to", "discriminate", ")", ".", "Additionally", ",", "the", "learned", "representation", "should", "be", "invariant", "to", "variations", "in", "the", "samples", "that", "are", "irrelevant", "for", "the", "classification", "task", ",", "i.e.", ",", "it", "should", "adhere", "to", "the", "manifold", "hypothesis", "(", "see", "e.g.", "Rifai", "et", "al", ".", "[", "20", "]", "for", "a", "recent", "discussion", ")", ".", "Invariance", "is", "classically", "achieved", "by", "regularization", "of", "the", "latent", "representation", ",", "e.g.", ",", "by", "enforcing", "sparsity", "[", "8", "]", "or", "robustness", "to", "noise", "[", "9", "]", ".", "In", "contrast", ",", "the", "discriminative", "objective", "in", "Eq", ".", "(", "1", ")", "does", "not", "directly", "model", "the", "input", "distribution", "p", "(", "x", ")", "but", "learns", "a", "representation", "that", "discriminates", "between", "input", "samples", ".", "The", "representation", "is", "not", "required", "to", "reconstruct", "the", "input", ",", "which", "is", "unnecessary", "in", "a", "recognition", "or", "matching", "task", ".", "This", "leaves", "more", "degrees", "of", "freedom", "to", "model", "the", "desired", "variability", "of", "a", "sample", ".", "As", "shown", "in", "our", "analysis", "(", "see", "Eq", ".", "(", "7", ")", ")", ",", "we", "achieve", "partial", "invariance", "to", "transformations", "applied", "during", "surrogate", "data", "creation", "by", "forcing", "the", "representation", "g", "(", "Tαxi", ")", "of", "the", "transformed", "image", "patch", "to", "be", "predictive", "of", "the", "surrogate", "label", "assigned", "to", "the", "original", "image", "patch", "xi", ".", "It", "should", "be", "noted", "that", "this", "approach", "assumes", "that", "the", "transformations", "Tα", "do", "not", "change", "the", "identity", "of", "the", "image", "content", ".", "If", "we", ",", "for", "example", ",", "use", "a", "color", "transformation", "we", "will", "force", "the", "network", "to", "be", "invariant", "to", "this", "change", "and", "can", "not", "expect", "the", "extracted", "features", "to", "perform", "well", "in", "a", "task", "relying", "on", "color", "information", "(", "such", "as", "differentiating", "black", "panthers", "from", "pumas", ")", "1", ".", "4", "Experiments", "To", "compare", "our", "discriminative", "approach", "to", "previous", "unsupervised", "feature", "learning", "methods", ",", "we", "report", "classification", "results", "on", "the", "STL", "-", "10", "[", "21", "]", ",", "CIFAR", "-", "10", "[", "22", "]", "and", "Caltech", "-", "101", "[", "23", "]", "datasets", ".", "Moreover", ",", "we", "assess", "the", "influence", "of", "the", "augmentation", "parameters", "on", "the", "classification", "performance", "and", "study", "the", "invariance", "properties", "of", "the", "network", ".", "4.1", "Experimental", "Setup", "The", "datasets", "we", "test", "on", "differ", "in", "the", "number", "of", "classes", "(", "10", "for", "CIFAR", "and", "STL", ",", "101", "for", "Caltech", ")", "and", "the", "number", "of", "samples", "per", "class", ".", "STL", "is", "especially", "well", "suited", "for", "unsupervised", "learning", "as", "it", "contains", "a", "large", "set", "of", "100", ",", "000", "unlabeled", "samples", ".", "In", "all", "experiments", "(", "except", "for", "the", "dataset", "transfer", "experiment", "in", "the", "supplementary", "material", ")", "we", "extracted", "surrogate", "training", "data", "from", "the", "unlabeled", "subset", "of", "STL", "-", "10", ".", "When", "testing", "on", "CIFAR", "-", "10", ",", "we", "resized", "the", "images", "from", "32×32", "pixels", "to", "64×64", "pixels", "so", "that", "the", "scale", "of", "depicted", "objects", "roughly", "matches", "the", "two", "other", "datasets", ".", "We", "worked", "with", "two", "network", "architectures", ".", "A", "“", "small", "”", "network", "was", "used", "to", "evaluate", "the", "influence", "of", "different", "components", "of", "the", "augmentation", "procedure", "on", "classification", "performance", ".", "It", "consists", "of", "two", "convolutional", "layers", "with", "64", "filters", "each", "followed", "by", "a", "fully", "connected", "layer", "with", "128", "neurons", ".", "This", "last", "layer", "is", "succeeded", "by", "a", "softmax", "layer", ",", "which", "serves", "as", "the", "network", "output", ".", "A", "“", "large", "”", "network", ",", "consisting", "of", "three", "convolutional", "layers", "with", "64", ",", "128", "and", "256", "filters", "respectively", "followed", "by", "a", "fully", "connected", "layer", "with", "512", "neurons", ",", "was", "trained", "to", "compare", "our", "method", "to", "the", "state", "-", "of", "-", "theart", ".", "In", "both", "models", "all", "convolutional", "filters", "are", "connected", "to", "a", "5×5", "region", "of", "their", "input", ".", "2×2", "maxpooling", "was", "performed", "after", "the", "first", "and", "second", "convolutional", "layers", ".", "Dropout", "[", "24", "]", "was", "applied", "to", "the", "fully", "connected", "layers", ".", "We", "trained", "the", "networks", "using", "an", "implementation", "based", "on", "Caffe", "[", "25", "]", ".", "Details", "on", "the", "training", ",", "the", "hyperparameter", "settings", ",", "and", "an", "analysis", "of", "the", "performance", "depending", "on", "the", "network", "architecture", "is", "provided", "in", "the", "supplementary", "material", ".", "Our", "code", "and", "training", "data", "are", "available", "at", "http:", "//", "lmb.informatik.uni", "-", "freiburg.de", "/", "resources", ".", "We", "applied", "the", "feature", "representation", "to", "images", "of", "arbitrary", "size", "by", "convolutionally", "computing", "the", "responses", "of", "all", "the", "network", "layers", "except", "the", "top", "softmax", ".", "To", "each", "feature", "map", ",", "we", "applied", "the", "pooling", "method", "that", "is", "commonly", "used", "for", "the", "respective", "dataset", ":", "1", ")", "4", "-", "quadrant", "max", "-", "pooling", ",", "resulting", "in", "4", "values", "per", "feature", "map", ",", "which", "is", "the", "standard", "procedure", "for", "STL", "-", "10", "and", "CIFAR", "-", "10", "[", "26", ",", "10", ",", "27", ",", "12", "]", ";", "2", ")", "3", "-", "layer", "spatial", "pyramid", ",", "i.e.", "max", "-", "pooling", "over", "the", "whole", "image", "as", "well", "as", "within", "4", "quadrants", "and", "within", "the", "cells", "of", "a", "4", "×", "4", "grid", ",", "resulting", "in", "1", "+", "4", "+", "16", "=", "21", "values", "per", "feature", "map", ",", "which", "is", "the", "standard", "for", "Caltech", "-", "101", "[", "28", ",", "10", ",", "29", "]", ".", "Finally", ",", "we", "trained", "a", "linear", "support", "vector", "machine", "(", "SVM", ")", "on", "the", "pooled", "features", ".", "On", "all", "datasets", "we", "used", "the", "standard", "training", "and", "test", "protocols", ".", "On", "STL", "-", "10", "the", "SVM", "was", "trained", "on", "10", "pre", "-", "defined", "folds", "of", "the", "training", "data", ".", "We", "report", "the", "mean", "and", "standard", "deviation", "achieved", "on", "the", "fixed", "test", "set", ".", "For", "CIFAR", "-", "10", "we", "report", "two", "results", ":", "(", "1", ")", "training", "the", "SVM", "on", "the", "whole", "CIFAR", "-", "10", "training", "set", "(", "’", "CIFAR", "-", "10", "’", ")", ";", "(", "2", ")", "the", "average", "over", "10", "random", "selections", "of", "400", "training", "samples", "per", "class", "(", "’", "CIFAR", "-", "10", "(", "400", ")", "’", ")", ".", "For", "Caltech", "-", "101", "we", "followed", "the", "usual", "protocol", "of", "selecting", "30", "random", "samples", "per", "class", "for", "training", "and", "not", "more", "than", "50", "samples", "per", "class", "for", "testing", ".", "This", "was", "repeated", "10", "times", ".", "4.2", "Classification", "Results", "In", "Table", "1", "we", "compare", "Exemplar", "-", "CNN", "to", "several", "unsupervised", "feature", "learning", "methods", ",", "including", "the", "current", "state", "-", "of", "-", "the", "-", "art", "on", "each", "dataset", ".", "We", "also", "list", "the", "state", "-", "of", "-", "the", "-", "art", "for", "supervised", "learning", "(", "which", "is", "not", "directly", "comparable", ")", ".", "Additionally", "we", "show", "the", "dimensionality", "of", "the", "feature", "vectors", "1Such", "cases", "could", "be", "covered", "either", "by", "careful", "selection", "of", "applied", "transformations", "or", "by", "combining", "features", "from", "multiple", "networks", "trained", "with", "different", "sets", "of", "transformations", "and", "letting", "the", "final", "classifier", "choose", "which", "features", "to", "use", ".", "produced", "by", "each", "method", "before", "final", "pooling", ".", "The", "small", "network", "was", "trained", "on", "8000", "surrogate", "classes", "containing", "150", "samples", "each", "and", "the", "large", "one", "on", "16000", "classes", "with", "100", "samples", "each", ".", "The", "features", "extracted", "from", "the", "larger", "network", "match", "or", "outperform", "the", "best", "prior", "result", "on", "all", "datasets", ".", "This", "is", "despite", "the", "fact", "that", "the", "dimensionality", "of", "the", "feature", "vector", "is", "smaller", "than", "that", "of", "most", "other", "approaches", "and", "that", "the", "networks", "are", "trained", "on", "the", "STL", "-", "10", "unlabeled", "dataset", "(", "i.e.", "they", "are", "used", "in", "a", "transfer", "learning", "manner", "when", "applied", "to", "CIFAR", "-", "10", "and", "Caltech", "101", ")", ".", "The", "increase", "in", "performance", "is", "especially", "pronounced", "when", "only", "few", "labeled", "samples", "are", "available", "for", "training", "the", "SVM", "(", "as", "is", "the", "case", "for", "all", "the", "datasets", "except", "full", "CIFAR", "-", "10", ")", ".", "This", "is", "in", "agreement", "with", "previous", "evidence", "that", "with", "increasing", "feature", "vector", "dimensionality", "and", "number", "of", "labeled", "samples", ",", "training", "an", "SVM", "becomes", "less", "dependent", "on", "the", "quality", "of", "the", "features", "[", "26", ",", "12", "]", ".", "Remarkably", ",", "on", "STL", "-", "10", "we", "achieve", "an", "accuracy", "of", "72.8", "%", ",", "which", "is", "a", "large", "improvement", "over", "all", "previously", "reported", "results", ".", "4.3", "Detailed", "Analysis", "We", "performed", "additional", "experiments", "(", "using", "the", "“", "small", "”", "network", ")", "to", "study", "the", "effect", "of", "three", "design", "choices", "in", "Exemplar", "-", "CNN", "training", "and", "validate", "the", "invariance", "properties", "of", "the", "learned", "features", ".", "Experiments", "on", "sampling", "’", "seed", "’", "patches", "from", "different", "datasets", "can", "be", "found", "in", "the", "supplementary", ".", "4.3.1", "Number", "of", "Surrogate", "Classes", "We", "varied", "the", "number", "N", "of", "surrogate", "classes", "between", "50", "and", "32000", ".", "As", "a", "sanity", "check", ",", "we", "also", "tried", "classification", "with", "random", "filters", ".", "The", "results", "are", "shown", "in", "Fig", ".", "3", ".", "Clearly", ",", "the", "classification", "accuracy", "increases", "with", "the", "number", "of", "surrogate", "classes", "until", "it", "reaches", "an", "optimum", "at", "about", "8000", "surrogate", "classes", "after", "which", "it", "did", "not", "change", "or", "even", "decreased", ".", "This", "is", "to", "be", "expected", ":", "the", "larger", "the", "number", "of", "surrogate", "classes", ",", "the", "more", "likely", "it", "is", "to", "draw", "very", "similar", "or", "even", "identical", "samples", ",", "which", "are", "hard", "or", "impossible", "to", "discriminate", ".", "Few", "such", "cases", "are", "not", "detrimental", "to", "the", "classification", "performance", ",", "but", "as", "soon", "as", "such", "collisions", "dominate", "the", "set", "of", "surrogate", "labels", ",", "the", "discriminative", "loss", "is", "no", "longer", "reasonable", "and", "training", "the", "network", "to", "the", "surrogate", "task", "no", "longer", "succeeds", ".", "To", "check", "the", "validity", "of", "this", "explanation", "we", "also", "plot", "in", "Fig", ".", "3", "the", "classification", "error", "on", "the", "validation", "set", "(", "taken", "from", "the", "surrogate", "data", ")", "computed", "after", "training", "the", "network", ".", "It", "rapidly", "grows", "as", "the", "number", "of", "surrogate", "classes", "increases", ".", "We", "also", "observed", "that", "the", "optimal", "number", "of", "surrogate", "classes", "increases", "with", "the", "size", "of", "the", "network", "(", "not", "shown", "in", "the", "figure", ")", ",", "but", "eventually", "saturates", ".", "This", "demonstrates", "the", "main", "limitation", "of", "our", "approach", "to", "randomly", "sample", "’", "seed", "’", "patches", ":", "it", "does", "not", "scale", "to", "arbitrarily", "large", "amounts", "of", "unlabeled", "data", ".", "However", ",", "we", "do", "not", "see", "this", "as", "a", "fundamental", "restriction", "and", "discuss", "possible", "solutions", "in", "Section", "5", ".", "4.3.2", "Number", "of", "Samples", "per", "Surrogate", "Class", "Fig", ".", "4", "shows", "the", "classification", "accuracy", "when", "the", "number", "K", "of", "training", "samples", "per", "surrogate", "class", "varies", "between", "1", "and", "300", ".", "The", "performance", "improves", "with", "more", "samples", "per", "surrogate", "class", "and", "2", "On", "Caltech", "-", "101", "one", "can", "either", "measure", "average", "accuracy", "over", "all", "samples", "(", "average", "overall", "accuracy", ")", "or", "calculate", "the", "accuracy", "for", "each", "class", "and", "then", "average", "these", "values", "(", "average", "per", "-", "class", "accuracy", ")", ".", "These", "differ", ",", "as", "some", "classes", "contain", "fewer", "than", "50", "test", "samples", ".", "Most", "researchers", "in", "ML", "use", "average", "overall", "accuracy", ".", "saturates", "at", "around", "100", "samples", ".", "This", "indicates", "that", "this", "amount", "is", "sufficient", "to", "approximate", "the", "formal", "objective", "from", "Eq", ".", "(", "3", ")", ",", "hence", "further", "increasing", "the", "number", "of", "samples", "does", "not", "significantly", "change", "the", "optimization", "problem", ".", "On", "the", "other", "hand", ",", "if", "the", "number", "of", "samples", "is", "too", "small", ",", "there", "is", "insufficient", "data", "to", "learn", "the", "desired", "invariance", "properties", ".", "4.3.3", "Types", "of", "Transformations", "We", "varied", "the", "transformations", "used", "for", "creating", "the", "surrogate", "data", "to", "analyze", "their", "influence", "on", "the", "final", "classification", "performance", ".", "The", "set", "of", "’", "seed", "’", "patches", "was", "fixed", ".", "The", "result", "is", "shown", "in", "Fig", ".", "5", ".", "The", "value", "’", "0", "’", "corresponds", "to", "applying", "random", "compositions", "of", "all", "elementary", "transformations", ":", "scaling", ",", "rotation", ",", "translation", ",", "color", "variation", ",", "and", "contrast", "variation", ".", "Different", "columns", "of", "the", "plot", "show", "the", "difference", "in", "classification", "accuracy", "as", "we", "discarded", "some", "types", "of", "elementary", "transformations", ".", "Several", "tendencies", "can", "be", "observed", ".", "First", ",", "rotation", "and", "scaling", "have", "only", "a", "minor", "impact", "on", "the", "performance", ",", "while", "translations", ",", "color", "variations", "and", "contrast", "variations", "are", "significantly", "more", "important", ".", "Secondly", ",", "the", "results", "on", "STL", "-", "10", "and", "CIFAR", "-", "10", "consistently", "show", "that", "spatial", "invariance", "and", "color", "-", "contrast", "invariance", "are", "approximately", "of", "equal", "importance", "for", "the", "classification", "performance", ".", "This", "indicates", "that", "variations", "in", "color", "and", "contrast", ",", "though", "often", "neglected", ",", "may", "also", "improve", "performance", "in", "a", "supervised", "learning", "scenario", ".", "Thirdly", ",", "on", "Caltech", "-", "101", "color", "and", "contrast", "transformations", "are", "much", "more", "important", "compared", "to", "spatial", "transformations", "than", "on", "the", "two", "other", "datasets", ".", "This", "is", "not", "surprising", ",", "since", "Caltech", "-", "101", "images", "are", "often", "well", "aligned", ",", "and", "this", "dataset", "bias", "makes", "spatial", "invariance", "less", "useful", ".", "4.3.4", "Invariance", "Properties", "of", "the", "Learned", "Representation", "In", "a", "final", "experiment", ",", "we", "analyzed", "to", "which", "extent", "the", "representation", "learned", "by", "the", "network", "is", "invariant", "to", "the", "transformations", "applied", "during", "training", ".", "We", "randomly", "sampled", "500", "images", "from", "the", "STL", "-", "10", "test", "set", "and", "applied", "a", "range", "of", "transformations", "(", "translation", ",", "rotation", ",", "contrast", ",", "color", ")", "to", "each", "image", ".", "To", "avoid", "empty", "regions", "beyond", "the", "image", "boundaries", "when", "applying", "spatial", "transformations", ",", "we", "cropped", "the", "central", "64×64", "pixel", "sub", "-", "patch", "from", "each", "96×96", "pixel", "image", ".", "We", "then", "applied", "two", "measures", "of", "invariance", "to", "these", "patches", ".", "First", ",", "as", "an", "explicit", "measure", "of", "invariance", ",", "we", "calculated", "the", "normalized", "Euclidean", "distance", "between", "normalized", "feature", "vectors", "of", "the", "original", "image", "patch", "and", "the", "transformed", "one", "[", "10", "]", "(", "see", "the", "supplementary", "material", "for", "details", ")", ".", "The", "downside", "of", "this", "approach", "is", "that", "the", "distance", "between", "extracted", "features", "does", "not", "take", "into", "account", "how", "informative", "and", "discriminative", "they", "are", ".", "We", "there", "-", "fore", "evaluated", "a", "second", "measure", "–", "classification", "performance", "depending", "on", "the", "magnitude", "of", "the", "transformation", "applied", "to", "the", "classified", "patches", "–", "which", "does", "not", "come", "with", "this", "problem", ".", "To", "compute", "the", "classification", "accuracy", ",", "we", "trained", "an", "SVM", "on", "the", "central", "64", "×", "64", "pixel", "patches", "from", "one", "fold", "of", "the", "STL", "-", "10", "training", "set", "and", "measured", "classification", "performance", "on", "all", "transformed", "versions", "of", "500", "samples", "from", "the", "test", "set", ".", "The", "results", "of", "both", "experiments", "are", "shown", "in", "Fig", ".", "6", ".", "Due", "to", "space", "restrictions", "we", "show", "only", "few", "representative", "plots", ".", "Overall", "the", "experiment", "empirically", "confirms", "that", "the", "Exemplar", "-", "CNN", "objective", "leads", "to", "learning", "invariant", "features", ".", "Features", "in", "the", "third", "layer", "and", "the", "final", "pooled", "feature", "representation", "compare", "favorably", "to", "a", "HOG", "baseline", "(", "Fig", ".", "6", "(", "a", ")", ")", ".", "Furthermore", ",", "adding", "stronger", "transformations", "in", "the", "surrogate", "training", "data", "leads", "to", "more", "invariant", "classification", "with", "respect", "to", "these", "transformations", "(", "Fig", ".", "6", "(", "b", ")-(", "d", ")", ")", ".", "However", ",", "adding", "too", "much", "contrast", "variation", "may", "deteriorate", "classification", "performance", "(", "Fig", ".", "6", "(", "d", ")", ")", ".", "One", "possible", "reason", "is", "that", "level", "of", "contrast", "can", "be", "a", "useful", "feature", ":", "for", "example", ",", "strong", "edges", "in", "an", "image", "are", "usually", "more", "important", "than", "weak", "ones", ".", "5", "Discussion", "We", "have", "proposed", "a", "discriminative", "objective", "for", "unsupervised", "feature", "learning", "by", "training", "a", "CNN", "without", "class", "labels", ".", "The", "core", "idea", "is", "to", "generate", "a", "set", "of", "surrogate", "labels", "via", "data", "augmentation", ".", "The", "features", "learned", "by", "the", "network", "yield", "a", "large", "improvement", "in", "classification", "accuracy", "compared", "to", "features", "obtained", "with", "previous", "unsupervised", "methods", ".", "These", "results", "strongly", "indicate", "that", "a", "discriminative", "objective", "is", "superior", "to", "objectives", "previously", "used", "for", "unsupervised", "feature", "learning", ".", "One", "potential", "shortcoming", "of", "the", "proposed", "method", "is", "that", "in", "its", "current", "state", "it", "does", "not", "scale", "to", "arbitrarily", "large", "datasets", ".", "Two", "probable", "reasons", "for", "this", "are", "that", "(", "1", ")", "as", "the", "number", "of", "surrogate", "classes", "grows", "larger", ",", "many", "of", "them", "become", "similar", ",", "which", "contradicts", "the", "discriminative", "objective", ",", "and", "(", "2", ")", "the", "surrogate", "task", "we", "use", "is", "relatively", "simple", "and", "does", "not", "allow", "the", "network", "to", "learn", "invariance", "to", "complex", "variations", ",", "such", "as", "3D", "viewpoint", "changes", "or", "inter", "-", "instance", "variation", ".", "We", "hypothesize", "that", "the", "presented", "approach", "could", "learn", "more", "powerful", "higher", "-", "level", "features", ",", "if", "the", "surrogate", "data", "were", "more", "diverse", ".", "This", "could", "be", "achieved", "by", "using", "additional", "weak", "supervision", ",", "for", "example", ",", "by", "means", "of", "video", "or", "a", "small", "number", "of", "labeled", "samples", ".", "Another", "possible", "way", "of", "obtaining", "richer", "surrogate", "training", "data", "and", "at", "the", "same", "time", "avoiding", "similar", "surrogate", "classes", "would", "be", "(", "unsupervised", ")", "merging", "of", "similar", "surrogate", "classes", ".", "We", "see", "these", "as", "interesting", "directions", "for", "future", "work", ".", "Acknowledgements", "We", "acknowledge", "funding", "by", "the", "ERC", "Starting", "Grant", "VideoLearn", "(", "279401", ")", ";", "the", "work", "was", "also", "partly", "supported", "by", "the", "BrainLinks", "-", "BrainTools", "Cluster", "of", "Excellence", "funded", "by", "the", "German", "Research", "Foundation", "(", "DFG", ",", "grant", "number", "EXC", "1086", ")", "."], "entity": [[0, 8, ["Discriminative", "Unsupervised", "Feature", "Learning", "with", "Convolutional", "Neural", "Networks"], "Method"], [12, 15, ["convolutional", "neural", "networks"], "Method"], [23, 25, ["supervised", "training"], "Task"], [83, 86, ["feature", "learning", "algorithm"], "Method"], [92, 95, ["visual", "object", "recognition"], "Task"], [97, 99, ["feature", "representation"], "Method"], [104, 105, ["classification"], "Task"], [119, 121, ["unsupervised", "learning"], "Method"], [126, 129, ["STL", "-", "10"], "Material"], [130, 133, ["CIFAR", "-", "10"], "Material"], [134, 137, ["Caltech", "-", "101"], "Material"], [141, 144, ["Convolutional", "neural", "networks"], "Method"], [145, 146, ["CNNs"], "Method"], [149, 150, ["backpropagation"], "Method"], [157, 160, ["image", "classification", "tasks"], "Task"], [176, 178, ["feature", "representation"], "Method"], [195, 197, ["classification", "task"], "Task"], [209, 212, ["visual", "recognition", "tasks"], "Task"], [216, 217, ["classification"], "Task"], [235, 238, ["CaltechUCSD", "birds", "dataset"], "Material"], [242, 244, ["scene", "recognition"], "Task"], [246, 250, ["SUN", "-", "397", "database"], "Material"], [254, 255, ["detection"], "Task"], [257, 260, ["PASCAL", "VOC", "dataset"], "Material"], [272, 275, ["supervised", "CNN", "training"], "Method"], [279, 283, ["generic", "visual", "feature", "learning"], "Task"], [287, 289, ["supervised", "training"], "Task"], [293, 295, ["expensive", "labeling"], "Task"], [352, 354, ["unsupervised", "learning"], "Method"], [379, 381, ["vision", "tasks"], "Task"], [382, 383, ["classification"], "Task"], [401, 404, ["unsupervised", "feature", "learning"], "Method"], [410, 412, ["image", "restoration"], "Task"], [422, 425, ["supervised", "feature", "learning"], "Method"], [427, 429, ["descriptor", "matching"], "Task"], [442, 444, ["discriminative", "objective"], "Method"], [449, 452, ["unsupervised", "feature", "learning"], "Method"], [453, 456, ["cheap", "data", "acquisition"], "Task"], [461, 463, ["training", "procedure"], "Method"], [464, 467, ["convolutional", "neural", "networks"], "Method"], [482, 484, ["surrogate", "task"], "Task"], [494, 496, ["data", "augmentation"], "Task"], [502, 504, ["supervised", "learning"], "Task"], [541, 542, ["CNN"], "Method"], [554, 560, ["exemplar", "training", "of", "convolutional", "neural", "networks"], "Method"], [561, 564, ["Exemplar", "-", "CNN"], "Method"], [567, 569, ["feature", "representation"], "Method"], [571, 574, ["Exemplar", "-", "CNN"], "Method"], [603, 607, ["unsupervised", "feature", "learning", "methods"], "Method"], [613, 616, ["STL", "-", "10"], "Material"], [617, 620, ["CIFAR", "-", "10"], "Material"], [622, 625, ["Caltech", "-", "101"], "Material"], [648, 651, ["convolutional", "neural", "networks"], "Method"], [652, 654, ["Convolutional", "training"], "Method"], [659, 663, ["supervised", "and", "unsupervised", "methods"], "Method"], [710, 713, ["convolutional", "neural", "networks"], "Method"], [714, 716, ["object", "recognition"], "Task"], [719, 721, ["data", "augmentation"], "Task"], [728, 730, ["classification", "objective"], "Metric"], [775, 777, ["unsupervised", "learning"], "Method"], [781, 784, ["learning", "invariant", "representations"], "Task"], [786, 788, ["Denoising", "autoencoders"], "Method"], [833, 835, ["feature", "representation"], "Method"], [838, 840, ["linear", "autoencoder"], "Method"], [864, 866, ["discriminative", "approach"], "Method"], [890, 891, ["CNN"], "Method"], [907, 912, ["supervised", "training", "of", "neural", "networks"], "Task"], [923, 925, ["tangent", "propagation"], "Method"], [932, 934, ["double", "backpropagation"], "Method"], [1006, 1009, ["regularizing", "supervised", "algorithms"], "Method"], [1012, 1015, ["self", "-", "training"], "Task"], [1019, 1021, ["entropy", "regularization"], "Method"], [1029, 1033, ["semi", "-", "supervised", "methods"], "Method"], [1034, 1038, ["Exemplar", "-", "CNN", "training"], "Method"], [1052, 1054, ["auxiliary", "task"], "Task"], [1060, 1062, ["data", "representation"], "Method"], [1083, 1086, ["Surrogate", "Training", "Data"], "Material"], [1209, 1210, ["Tα"], "Method"], [1239, 1240, ["scaling"], "Method"], [1331, 1334, ["HSV", "color", "representation"], "Method"], [1390, 1392, ["H", "component"], "Method"], [1394, 1397, ["HSV", "color", "representation"], "Method"], [1542, 1544, ["transformed", "versions"], "Method"], [1555, 1557, ["Learning", "Algorithm"], "Method"], [1588, 1589, ["CNN"], "Method"], [1602, 1604, ["loss", "function"], "Metric"], [1605, 1608, ["L", "(", "X"], "Method"], [1649, 1650, ["CNN"], "Method"], [1652, 1655, ["softmax", "output", "layer"], "Method"], [1728, 1729, ["CNN"], "Method"], [1760, 1762, ["objective", "function"], "Metric"], [1800, 1802, ["classification", "problem"], "Task"], [1856, 1858, ["classification", "problem"], "Task"], [1860, 1861, ["regularizer"], "Method"], [1886, 1888, ["classification", "problem"], "Task"], [1895, 1898, ["unsupervised", "feature", "learning"], "Method"], [1900, 1902, ["Formal", "Analysis"], "Task"], [2023, 2025, ["objective", "function"], "Metric"], [2073, 2079, ["average", "feature", "representation", "of", "transformed", "versions"], "Method"], [2081, 2083, ["image", "patch"], "Method"], [2141, 2143, ["objective", "function"], "Metric"], [2145, 2149, ["multinomial", "logistic", "regression", "problem"], "Task"], [2166, 2175, ["transformation", "-", "free", "instance", "classification", "problem", "L", "(", "X"], "Task"], [2212, 2214, ["correct", "classification"], "Metric"], [2234, 2236, ["invariant", "representation"], "Method"], [2265, 2270, ["feature", "representation", "g", "(", "Tαxi"], "Method"], [2342, 2343, ["regularizer"], "Method"], [2360, 2362, ["feature", "representation"], "Method"], [2431, 2433, ["feature", "representation"], "Method"], [2443, 2444, ["Wĝi"], "Method"], [2456, 2457, ["regularizer"], "Method"], [2467, 2470, ["Unsupervised", "Learning", "Methods"], "Method"], [2477, 2479, ["feature", "representation"], "Method"], [2482, 2484, ["recognition", "task"], "Task"], [2487, 2488, ["classification"], "Task"], [2497, 2502, ["feature", "representation", "g", "(", "x"], "Method"], [2559, 2563, ["unsupervised", "feature", "learning", "methods"], "Method"], [2573, 2578, ["distribution", "p", "(", "x", ")"], "Method"], [2663, 2665, ["classification", "task"], "Task"], [2691, 2692, ["Invariance"], "Task"], [2696, 2701, ["regularization", "of", "the", "latent", "representation"], "Method"], [2711, 2712, ["robustness"], "Metric"], [2722, 2724, ["discriminative", "objective"], "Metric"], [2766, 2770, ["recognition", "or", "matching", "task"], "Task"], [2808, 2811, ["surrogate", "data", "creation"], "Task"], [2814, 2818, ["representation", "g", "(", "Tαxi"], "Method"], [2915, 2917, ["discriminative", "approach"], "Method"], [2919, 2923, ["unsupervised", "feature", "learning", "methods"], "Method"], [2926, 2927, ["classification"], "Task"], [2930, 2933, ["STL", "-", "10"], "Material"], [2937, 2940, ["CIFAR", "-", "10"], "Material"], [2944, 2947, ["Caltech", "-", "101"], "Material"], [2964, 2966, ["classification", "performance"], "Metric"], [2969, 2971, ["invariance", "properties"], "Metric"], [2992, 2993, ["CIFAR"], "Material"], [2994, 2995, ["STL"], "Material"], [2998, 2999, ["Caltech"], "Material"], [3008, 3009, ["STL"], "Material"], [3014, 3016, ["unsupervised", "learning"], "Method"], [3036, 3038, ["dataset", "transfer"], "Task"], [3054, 3057, ["STL", "-", "10"], "Material"], [3061, 3064, ["CIFAR", "-", "10"], "Material"], [3100, 3101, ["network"], "Method"], [3112, 3114, ["augmentation", "procedure"], "Method"], [3115, 3116, ["classification"], "Task"], [3122, 3124, ["convolutional", "layers"], "Method"], [3125, 3127, ["64", "filters"], "Method"], [3131, 3134, ["fully", "connected", "layer"], "Method"], [3145, 3147, ["softmax", "layer"], "Method"], [3159, 3160, ["network"], "Method"], [3164, 3166, ["convolutional", "layers"], "Method"], [3177, 3180, ["fully", "connected", "layer"], "Method"], [3202, 3204, ["convolutional", "filters"], "Method"], [3214, 3216, ["2×2", "maxpooling"], "Method"], [3223, 3225, ["convolutional", "layers"], "Method"], [3226, 3227, ["Dropout"], "Method"], [3234, 3237, ["fully", "connected", "layers"], "Method"], [3247, 3248, ["Caffe"], "Method"], [3255, 3256, ["training"], "Task"], [3298, 3300, ["feature", "representation"], "Method"], [3306, 3307, ["convolutionally"], "Method"], [3328, 3330, ["pooling", "method"], "Method"], [3341, 3347, ["4", "-", "quadrant", "max", "-", "pooling"], "Method"], [3362, 3365, ["STL", "-", "10"], "Material"], [3366, 3369, ["CIFAR", "-", "10"], "Material"], [3381, 3386, ["3", "-", "layer", "spatial", "pyramid"], "Method"], [3388, 3391, ["max", "-", "pooling"], "Method"], [3447, 3451, ["linear", "support", "vector", "machine"], "Method"], [3452, 3453, ["SVM"], "Method"], [3466, 3470, ["training", "and", "test", "protocols"], "Metric"], [3472, 3475, ["STL", "-", "10"], "Material"], [3476, 3477, ["SVM"], "Method"], [3505, 3508, ["CIFAR", "-", "10"], "Material"], [3518, 3519, ["SVM"], "Method"], [3529, 3532, ["CIFAR", "-", "10"], "Material"], [3552, 3555, ["CIFAR", "-", "10"], "Material"], [3562, 3565, ["Caltech", "-", "101"], "Material"], [3597, 3598, ["Classification"], "Task"], [3604, 3607, ["Exemplar", "-", "CNN"], "Method"], [3609, 3613, ["unsupervised", "feature", "learning", "methods"], "Method"], [3640, 3642, ["supervised", "learning"], "Task"], [3688, 3689, ["classifier"], "Method"], [3701, 3702, ["pooling"], "Task"], [3704, 3706, ["small", "network"], "Method"], [3786, 3789, ["transfer", "learning", "manner"], "Method"], [3792, 3795, ["CIFAR", "-", "10"], "Material"], [3796, 3798, ["Caltech", "101"], "Material"], [3817, 3818, ["SVM"], "Method"], [3855, 3856, ["SVM"], "Method"], [3874, 3877, ["STL", "-", "10"], "Material"], [3880, 3881, ["accuracy"], "Metric"], [3909, 3910, ["network"], "Method"], [3920, 3924, ["Exemplar", "-", "CNN", "training"], "Method"], [3977, 3978, ["classification"], "Method"], [3979, 3981, ["random", "filters"], "Method"], [3994, 3996, ["classification", "accuracy"], "Metric"], [4067, 4068, ["classification"], "Task"], [4084, 4086, ["discriminative", "loss"], "Metric"], [4096, 4098, ["surrogate", "task"], "Task"], [4117, 4119, ["classification", "error"], "Metric"], [4235, 4237, ["classification", "accuracy"], "Metric"], [4272, 4274, ["average", "accuracy"], "Metric"], [4278, 4281, ["average", "overall", "accuracy"], "Metric"], [4285, 4286, ["accuracy"], "Metric"], [4295, 4300, ["average", "per", "-", "class", "accuracy"], "Metric"], [4318, 4319, ["ML"], "Task"], [4320, 4323, ["average", "overall", "accuracy"], "Metric"], [4361, 4363, ["optimization", "problem"], "Task"], [4410, 4412, ["classification", "performance"], "Metric"], [4469, 4471, ["classification", "accuracy"], "Metric"], [4518, 4521, ["STL", "-", "10"], "Material"], [4522, 4525, ["CIFAR", "-", "10"], "Material"], [4542, 4543, ["classification"], "Task"], [4564, 4567, ["supervised", "learning", "scenario"], "Task"], [4619, 4621, ["Invariance", "Properties"], "Metric"], [4623, 4625, ["Learned", "Representation"], "Method"], [4727, 4728, ["invariance"], "Metric"], [4793, 4795, ["classification", "performance"], "Metric"], [4819, 4821, ["classification", "accuracy"], "Metric"], [4825, 4826, ["SVM"], "Method"], [4846, 4847, ["classification"], "Task"], [4890, 4894, ["Exemplar", "-", "CNN", "objective"], "Method"], [4908, 4911, ["pooled", "feature", "representation"], "Method"], [4915, 4917, ["HOG", "baseline"], "Method"], [4940, 4941, ["classification"], "Task"], [4966, 4968, ["classification", "performance"], "Metric"], [5013, 5015, ["discriminative", "objective"], "Method"], [5016, 5019, ["unsupervised", "feature", "learning"], "Method"], [5022, 5026, ["CNN", "without", "class", "labels"], "Method"], [5039, 5041, ["data", "augmentation"], "Task"], [5053, 5055, ["classification", "accuracy"], "Metric"], [5061, 5063, ["unsupervised", "methods"], "Method"], [5070, 5072, ["discriminative", "objective"], "Method"], [5079, 5082, ["unsupervised", "feature", "learning"], "Method"], [5133, 5135, ["discriminative", "objective"], "Metric"], [5141, 5143, ["surrogate", "task"], "Task"], [38, 39, ["network"], "Method"], [47, 48, ["network"], "Method"], [200, 201, ["network"], "Method"], [218, 221, ["Caltech", "-", "101"], "Material"], [640, 642, ["unsupervised", "learning"], "Method"], [757, 758, ["network"], "Method"], [950, 951, ["network"], "Method"], [1528, 1529, ["STL"], "Material"], [1664, 1665, ["network"], "Method"], [1934, 1935, ["network"], "Method"], [1952, 1953, ["network"], "Method"], [1992, 1993, ["network"], "Method"], [2875, 2876, ["network"], "Method"], [2973, 2974, ["network"], "Method"], [3093, 3094, ["network"], "Method"], [3152, 3153, ["network"], "Method"], [3270, 3271, ["network"], "Method"], [3313, 3314, ["network"], "Method"], [3431, 3434, ["Caltech", "-", "101"], "Material"], [3522, 3525, ["CIFAR", "-", "10"], "Material"], [3734, 3735, ["network"], "Method"], [3774, 3775, ["STL"], "Material"], [3829, 3832, ["CIFAR", "-", "10"], "Material"], [4093, 4094, ["network"], "Method"], [4134, 4135, ["network"], "Method"], [4163, 4164, ["network"], "Method"], [4265, 4268, ["Caltech", "-", "101"], "Material"], [4447, 4448, ["scaling"], "Method"], [4490, 4491, ["scaling"], "Method"], [4571, 4574, ["Caltech", "-", "101"], "Material"], [4599, 4602, ["Caltech", "-", "101"], "Material"], [4640, 4641, ["network"], "Method"], [4657, 4658, ["STL"], "Material"], [4839, 4840, ["STL"], "Material"], [5047, 5048, ["network"], "Method"], [5153, 5154, ["network"], "Method"]]} {"doc_id": "4365eb43a635bc6431dfaf3af1f7bf7bf55522cc", "text": ["document", ":", "CoupleNet", ":", "Coupling", "Global", "Structure", "with", "Local", "Parts", "for", "Object", "Detection", "The", "region", "-", "based", "Convolutional", "Neural", "Network", "(", "CNN", ")", "detectors", "such", "as", "Faster", "R", "-", "CNN", "or", "R", "-", "FCN", "have", "already", "shown", "promising", "results", "for", "object", "detection", "by", "combining", "the", "region", "proposal", "subnetwork", "and", "the", "classification", "subnetwork", "together", ".", "Although", "R", "-", "FCN", "has", "achieved", "higher", "detection", "speed", "while", "keeping", "the", "detection", "performance", ",", "the", "global", "structure", "information", "is", "ignored", "by", "the", "position", "-", "sensitive", "score", "maps", ".", "To", "fully", "explore", "the", "local", "and", "global", "properties", ",", "in", "this", "paper", ",", "we", "propose", "a", "novel", "fully", "convolutional", "network", ",", "named", "as", "CoupleNet", ",", "to", "couple", "the", "global", "structure", "with", "local", "parts", "for", "object", "detection", ".", "Specifically", ",", "the", "object", "proposals", "obtained", "by", "the", "Region", "Proposal", "Network", "(", "RPN", ")", "are", "fed", "into", "the", "the", "coupling", "module", "which", "consists", "of", "two", "branches", ".", "One", "branch", "adopts", "the", "position", "-", "sensitive", "RoI", "(", "PSRoI", ")", "pooling", "to", "capture", "the", "local", "part", "information", "of", "the", "object", ",", "while", "the", "other", "employs", "the", "RoI", "pooling", "to", "encode", "the", "global", "and", "context", "information", ".", "Next", ",", "we", "design", "different", "coupling", "strategies", "and", "normalization", "ways", "to", "make", "full", "use", "of", "the", "complementary", "advantages", "between", "the", "global", "and", "local", "branches", ".", "Extensive", "experiments", "demonstrate", "the", "effectiveness", "of", "our", "approach", ".", "We", "achieve", "state", "-", "of", "-", "the", "-", "art", "results", "on", "all", "three", "challenging", "datasets", ",", "a", "mAP", "of", "on", "VOC07", ",", "on", "VOC12", ",", "and", "on", "COCO", ".", "Codes", "will", "be", "made", "publicly", "available", ".", "section", ":", "Introduction", "General", "object", "detection", "requires", "to", "accurately", "locate", "and", "classify", "all", "targets", "in", "the", "image", "or", "video", ".", "Compared", "to", "specific", "object", "detection", ",", "such", "as", "face", ",", "pedestrian", "and", "vehicle", "detection", ",", "general", "object", "detection", "often", "faces", "more", "challenges", "due", "to", "the", "large", "inter", "-", "class", "appearance", "differences", ".", "The", "variations", "arise", "not", "only", "from", "changes", "in", "a", "variety", "of", "non", "-", "rigid", "deformations", ",", "but", "also", "due", "to", "the", "truncations", ",", "occlusions", "and", "inter", "-", "class", "interference", ".", "However", ",", "no", "matter", "how", "complicated", "the", "objects", "are", ",", "when", "humans", "identify", "a", "target", ",", "the", "recognition", "of", "object", "categories", "is", "subserved", "by", "both", "a", "global", "process", "that", "retrieves", "structural", "information", "and", "a", "local", "process", "that", "is", "sensitive", "to", "individual", "parts", ".", "This", "motivates", "us", "to", "build", "a", "detection", "model", "that", "fused", "both", "global", "and", "local", "information", ".", "With", "the", "revival", "of", "Convolutional", "Neural", "Networks", "(", "CNN", ")", ",", "CNN", "-", "based", "object", "detection", "pipelines", "have", "been", "proposed", "consecutively", "and", "made", "impressive", "improvements", "in", "generic", "benchmarks", ",", "PASCAL", "VOC", "and", "MS", "COCO", ".", "As", "two", "representative", "region", "-", "based", "CNN", "approaches", ",", "Fast", "/", "Faster", "R", "-", "CNN", "uses", "a", "certain", "subnetwork", "to", "predict", "the", "category", "of", "each", "region", "proposal", "while", "R", "-", "FCN", "conducts", "the", "inference", "with", "the", "position", "-", "sensitive", "score", "maps", ".", "Through", "removing", "the", "RoI", "-", "wise", "subnetwork", ",", "R", "-", "FCN", "has", "achieved", "higher", "detection", "speed", "while", "keeping", "the", "detection", "performance", ".", "However", ",", "the", "global", "structure", "information", "is", "ignored", "by", "the", "PSRoI", "pooling", ".", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "using", "PSRoI", "pooling", "to", "extract", "local", "part", "information", "for", "final", "object", "category", "prediction", ",", "R", "-", "FCN", "leads", "to", "a", "low", "confidence", "score", "of", "0.08", "for", "the", "sofa", "detection", "since", "the", "local", "responses", "of", "sofa", "are", "disturbed", "by", "a", "women", "and", "a", "dog", "(", "they", "are", "also", "the", "categories", "that", "need", "to", "be", "detected", ")", ".", "Conversely", ",", "the", "global", "structure", "of", "sofa", "could", "be", "extracted", "by", "the", "RoI", "pooling", ",", "but", "the", "confidence", "score", "is", "0.45", ",", "which", "is", "also", "very", "low", "for", "the", "incomplete", "structure", "of", "sofa", ".", "By", "coupling", "the", "global", "confidence", "with", "the", "local", "part", "confidence", "together", ",", "we", "can", "obtain", "a", "more", "reliable", "prediction", "with", "the", "confidence", "score", "of", "0.78", ".", "In", "fact", ",", "the", "idea", "of", "fusing", "global", "and", "local", "information", "together", "is", "widely", "used", "in", "lots", "of", "visual", "tasks", ".", "In", "fingerprint", "recognition", ",", "Gu", "combined", "the", "global", "orientation", "field", "and", "local", "minutiae", "cue", "to", "largely", "improve", "the", "performance", ".", "In", "clique", "-", "graph", "matching", ",", "Nie", "proposed", "a", "clique", "-", "graph", "matching", "method", "by", "preserving", "global", "clique", "-", "to", "-", "clique", "correspondence", "and", "local", "unary", "and", "pairwise", "correspondences", ".", "In", "scene", "parsing", ",", "Zhao", "designed", "a", "pyramid", "pooling", "module", "to", "effectively", "extract", "hierarchical", "global", "contextual", "prior", ",", "and", "then", "concatenated", "it", "with", "the", "local", "FCN", "feature", "to", "improve", "the", "performance", ".", "In", "traditional", "object", "detection", ",", "Felzenszwalb", "incorporated", "a", "global", "root", "model", "and", "several", "finer", "local", "part", "models", "to", "represent", "highly", "variable", "objects", ".", "All", "of", "which", "show", "that", "effective", "combination", "of", "the", "global", "structural", "properties", "and", "local", "fine", "-", "grained", "details", "can", "achieve", "complementary", "advantages", ".", "Therefore", ",", "to", "fully", "explore", "the", "global", "and", "local", "clues", ",", "in", "this", "paper", ",", "we", "propose", "a", "novel", "full", "convolutional", "network", "named", "as", "CoupleNet", ",", "to", "couple", "the", "global", "structure", "and", "local", "parts", "to", "boost", "the", "detection", "accuracy", ".", "Specifically", ",", "the", "object", "proposals", "obtained", "by", "the", "RPN", "are", "fed", "into", "the", "coupling", "module", "which", "consists", "of", "two", "branches", ".", "One", "branch", "adopts", "the", "PSRoI", "pooling", "to", "capture", "the", "local", "part", "information", "of", "the", "object", ",", "while", "the", "other", "employs", "the", "RoI", "pooling", "to", "encode", "the", "global", "and", "context", "information", ".", "Moreover", ",", "we", "design", "different", "coupling", "strategies", "and", "normalization", "ways", "to", "make", "full", "use", "of", "the", "complementary", "advantages", "between", "the", "global", "and", "local", "branches", ".", "With", "the", "coupling", "structure", ",", "our", "network", "can", "jointly", "learn", "the", "local", ",", "global", "and", "context", "expression", "of", "the", "objects", ",", "which", "makes", "the", "model", "have", "a", "more", "powerful", "representation", "capacity", "and", "generalization", "ability", ".", "Extensive", "experiments", "demonstrate", "that", "CoupleNet", "can", "significantly", "improve", "the", "detection", "performance", ".", "Our", "detector", "shows", "competitive", "results", "on", "PASCAL", "VOC", "07", "/", "12", "and", "MS", "COCO", "compared", "to", "other", "state", "-", "of", "-", "the", "-", "art", "detectors", ",", "even", "with", "model", "ensemble", "approaches", ".", "In", "summary", ",", "our", "main", "contributions", "are", "as", "follows", ":", "1", ".", "We", "propose", "a", "unified", "fully", "convolutional", "network", "to", "jointly", "learn", "the", "local", ",", "global", "and", "context", "information", "for", "object", "detection", ".", "2", ".", "We", "design", "different", "normalization", "methods", "and", "coupling", "strategies", "to", "mine", "the", "compatibility", "and", "complementarity", "between", "the", "global", "and", "local", "branches", ".", "3", ".", "We", "achieve", "the", "state", "-", "of", "-", "the", "-", "art", "results", "on", "all", "three", "challenging", "datasets", ",", "a", "mAP", "of", "on", "VOC07", ",", "on", "VOC12", ",", "and", "on", "MS", "COCO", ".", "section", ":", "Related", "work", "Before", "the", "arrival", "of", "CNN", ",", "visual", "tasks", "have", "been", "dominated", "by", "traditional", "paradigms", ".", "As", "one", "of", "an", "outstanding", "framework", ",", "DPM", "described", "the", "object", "system", "using", "mixtures", "of", "multi", "-", "scale", "deformable", "part", "models", ",", "including", "a", "coarse", "global", "root", "model", "and", "several", "finer", "local", "part", "models", ".", "The", "root", "model", "extracts", "structural", "information", "of", "the", "objects", ",", "while", "the", "part", "models", "capture", "local", "appearance", "properties", "of", "an", "object", ".", "The", "sum", "of", "root", "response", "and", "weighted", "average", "response", "of", "each", "part", "is", "used", "as", "the", "final", "confidence", "of", "an", "object", ".", "Although", "DPM", "provides", "an", "elegant", "framework", "for", "object", "detection", ",", "the", "hand", "-", "crafted", "features", ",", "improved", "HOG", ",", "are", "not", "discriminative", "enough", "to", "express", "the", "diversity", "of", "object", "categories", ".", "This", "is", "also", "the", "main", "reason", "that", "CNN", "completely", "surpassed", "the", "traditional", "methods", "in", "a", "short", "period", "time", ".", "In", "order", "to", "leverage", "the", "great", "success", "of", "deep", "neural", "networks", "for", "image", "classification", ",", "considerable", "object", "detection", "methods", "based", "on", "deep", "learning", "have", "been", "proposed", ".", "Although", "there", "are", "end", "-", "to", "-", "end", "detection", "frameworks", ",", "like", "SSD", ",", "YOLO", "and", "DenseBox", ",", "region", "-", "based", "systems", "(", "Fast", "/", "Faster", "R", "-", "CNN", "and", "R", "-", "FCN", ")", "still", "dominate", "the", "detection", "accuracy", "on", "generic", "benchmarks", ".", "Compared", "to", "the", "end", "-", "to", "-", "end", "framework", ",", "the", "region", "-", "based", "systems", "have", "several", "advantages", ".", "Firstly", ",", "by", "exploiting", "a", "divide", "-", "and", "-", "conquer", "strategy", ",", "the", "two", "-", "step", "framework", "is", "more", "stable", "and", "easier", "to", "converge", ".", "Secondly", ",", "without", "the", "complicated", "data", "augmentation", "and", "training", "skills", ",", "you", "can", "still", "easily", "achieve", "state", "-", "of", "-", "the", "-", "art", "performance", ".", "The", "main", "reason", "for", "these", "advantages", "is", "that", "there", "is", "a", "certain", "structure", "to", "encode", "translation", "variance", "features", "for", "each", "proposal", ",", "since", "in", "deep", "networks", ",", "higher", "-", "layers", "contain", "more", "semantic", "meaning", "and", "less", "location", "information", ".", "As", "a", "consequence", ",", "a", "RoI", "-", "wise", "subnetwork", "or", "a", "position", "-", "sensitive", "RoI", "pooling", "layer", "is", "used", "to", "achieve", "the", "translation", "variance", "in", "region", "-", "based", "systems", ".", "However", ",", "all", "the", "existing", "region", "-", "based", "systems", "utilize", "either", "the", "region", "-", "level", "or", "part", "-", "level", "features", "to", "learn", "the", "variations", ",", "where", "each", "one", "alone", "is", "not", "representative", "enough", "for", "a", "variety", "of", "challenging", "situations", ".", "Therefore", ",", "this", "motivates", "us", "to", "design", "a", "certain", "structure", "to", "take", "advantages", "of", "both", "the", "global", "and", "local", "features", ".", "In", "addition", ",", "context", "is", "known", "to", "play", "an", "important", "role", "in", "visual", "recognition", ".", "Considerable", "works", "have", "been", "proposed", "for", "exploting", "context", "in", "object", "detection", ".", "Bell", "explored", "the", "use", "of", "recurrent", "neural", "networks", "to", "model", "the", "contextual", "information", ".", "Gidaris", "proposed", "to", "utilize", "multiple", "contextual", "regions", "around", "the", "object", ".", "Cai", "collected", "the", "context", "by", "padding", "the", "proposals", "for", "pedestrian", "and", "car", "detection", ".", "Similar", "to", "these", "works", ",", "we", "also", "absorb", "the", "context", "prior", "to", "enhance", "the", "global", "feature", "representation", ".", "section", ":", "CoupleNet", "In", "this", "section", ",", "we", "first", "introduce", "the", "architecture", "of", "the", "proposed", "CoupleNet", "for", "object", "detection", ".", "Then", "we", "explain", "in", "detail", "how", "we", "incorporate", "local", "representations", ",", "global", "appearance", "and", "contextual", "information", "for", "robust", "object", "detection", ".", "subsection", ":", "Network", "architecture", "The", "architecture", "of", "our", "proposed", "CoupleNet", "is", "illustrated", "in", "Figure", "[", "reference", "]", ".", "Our", "CoupleNet", "includes", "two", "different", "branches", ":", "a", ")", "a", "local", "part", "-", "sensitive", "fully", "convolutional", "network", "to", "learn", "the", "object", "-", "specific", "parts", ",", "denoted", "as", "local", "FCN", ";", "b", ")", "a", "global", "region", "-", "sensitive", "fully", "convolutional", "network", "to", "encode", "the", "whole", "appearance", "structure", "and", "context", "prior", "of", "the", "object", ",", "denoted", "as", "global", "FCN", ".", "We", "first", "use", "the", "ImageNet", "pre", "-", "trained", "ResNet", "-", "101", "released", "in", "to", "initialize", "our", "network", ".", "For", "our", "detection", "task", ",", "we", "remove", "the", "last", "average", "pooling", "layer", "and", "the", "fc", "layer", ".", "Given", "an", "input", "image", ",", "we", "extract", "candidate", "proposals", "by", "using", "the", "Region", "Proposal", "Network", "(", "RPN", ")", ",", "which", "also", "shares", "convolution", "features", "with", "CoupleNet", "following", ".", "Then", "each", "proposal", "flows", "to", "two", "different", "branches", ":", "the", "local", "FCN", "and", "the", "global", "FCN", ".", "Finally", ",", "the", "output", "of", "global", "and", "local", "FCN", "are", "coupled", "together", "as", "the", "final", "score", "of", "the", "object", ".", "We", "also", "perform", "class", "-", "agnostic", "bounding", "box", "regression", "in", "a", "similar", "way", ".", "subsection", ":", "Local", "FCN", "To", "effectively", "capture", "the", "specific", "fine", "-", "grained", "parts", "in", "local", "FCN", ",", "we", "construct", "a", "set", "of", "part", "-", "sensitive", "score", "maps", "by", "appending", "a", "1x1", "convolutional", "layer", "with", "channels", ",", "where", "means", "we", "divide", "the", "object", "into", "local", "parts", "(", "here", "is", "set", "to", "the", "default", "value", "7", ")", "and", "is", "the", "number", "of", "object", "categories", "plus", "background", ".", "For", "each", "category", ",", "there", "are", "totally", "channels", "and", "each", "channel", "is", "responsible", "for", "encoding", "a", "specific", "part", "of", "the", "object", ".", "The", "final", "score", "of", "a", "category", "is", "determined", "by", "voting", "the", "responses", ".", "Here", "we", "use", "position", "-", "sensitive", "RoI", "pooling", "layer", "in", "to", "extract", "object", "-", "specific", "parts", "and", "we", "simply", "perform", "average", "pooling", "for", "voting", ".", "Then", ",", "we", "obtain", "a", "-", "d", "vector", "which", "indicates", "the", "probability", "that", "the", "object", "belongs", "to", "each", "class", ".", "This", "procedure", "is", "equivalent", "to", "dividing", "a", "strong", "object", "category", "decision", "into", "the", "sum", "of", "multiple", "weak", "classifiers", ",", "which", "serves", "as", "the", "ensemble", "of", "several", "part", "models", ".", "Here", "we", "refer", "this", "part", "ensemble", "as", "local", "structure", "representation", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "(", "a", ")", ",", "for", "the", "truncated", "person", ",", "one", "can", "hardly", "get", "a", "strong", "response", "from", "the", "global", "description", "of", "the", "person", "due", "to", "truncation", ",", "on", "the", "contrary", ",", "our", "local", "FCN", "can", "effectively", "capture", "several", "specific", "parts", ",", "such", "as", "human", "nose", ",", "mouth", ",", ",", "which", "correspond", "to", "the", "regions", "with", "large", "responses", "in", "the", "feature", "map", ".", "We", "argue", "that", "the", "local", "FCN", "is", "much", "concerned", "with", "the", "internal", "structure", "and", "components", ",", "which", "can", "effectively", "reflect", "the", "local", "properties", "of", "visual", "object", ",", "especially", "when", "the", "object", "is", "occluded", "or", "the", "whole", "boundary", "is", "incomplete", ".", "However", ",", "for", "those", "having", "simple", "spatial", "structure", "and", "encompassing", "considerable", "background", "in", "the", "bounding", "box", ",", "dining", "table", ",", "the", "local", "FCN", "alone", "is", "difficult", "to", "make", "robust", "predictions", ".", "Thus", "it", "is", "necessary", "to", "add", "the", "global", "structure", "information", "to", "enhance", "the", "discrimination", ".", "subsection", ":", "Global", "FCN", "For", "the", "global", "FCN", ",", "we", "aim", "to", "describe", "the", "object", "by", "using", "the", "whole", "region", "-", "level", "features", ".", "Firstly", ",", "we", "attach", "a", "1024", "-", "d", "1x1", "convolutional", "layer", "after", "the", "last", "convolutional", "block", "in", "ResNet", "-", "101", "for", "reducing", "the", "dimension", ".", "Due", "to", "the", "diverse", "size", "of", "the", "object", ",", "we", "insert", "a", "RoI", "pooling", "layer", "in", "to", "extract", "a", "fixed", "-", "length", "feature", "vector", "as", "the", "global", "structure", "description", "of", "the", "object", ".", "Secondly", ",", "we", "use", "two", "convolutional", "layers", "with", "kernal", "size", "and", "respectively", "(", "is", "set", "to", "the", "default", "value", "7", ")", "to", "further", "abstract", "the", "global", "representation", "of", "RoI.", "Finally", ",", "the", "output", "of", "1x1", "convolution", "is", "fed", "into", "the", "classifier", "whose", "output", "is", "also", "a", "-", "d", "vector", ".", "In", "addition", ",", "context", "prior", "is", "the", "most", "basic", "and", "important", "factor", "for", "visual", "recognition", "tasks", ".", "For", "example", ",", "the", "boat", "usually", "travels", "in", "the", "water", "while", "is", "unlikely", "to", "fly", "in", "the", "sky", ".", "Despite", "the", "higher", "layers", "in", "deep", "neural", "network", "can", "involve", "the", "spatial", "context", "information", "around", "the", "objects", "due", "to", "the", "large", "receptive", "field", ",", "Zhou", "have", "shown", "that", "the", "practical", "receptive", "field", "is", "actually", "much", "smaller", "than", "the", "theoretical", "one", ".", "Therefore", ",", "it", "is", "necessary", "to", "explicitly", "collect", "the", "surrounding", "information", "to", "reduce", "the", "chance", "of", "misclassification", ".", "To", "enhance", "the", "feature", "representation", "ability", "of", "the", "global", "FCN", ",", "here", "we", "introduce", "the", "contextual", "information", "as", "an", "effective", "supplement", ".", "Specifically", ",", "we", "extend", "the", "context", "region", "by", "2", "times", "larger", "than", "the", "size", "of", "original", "proposal", ".", "Then", "the", "features", "RoI", "pooled", "from", "the", "original", "region", "and", "context", "region", "are", "concatenated", "together", "and", "fed", "into", "the", "latter", "RoI", "-", "wise", "subnetwork", ".", "As", "shown", "in", "Figure", "[", "reference", "]", ",", "the", "context", "region", "is", "embedded", "into", "the", "global", "branch", "to", "extract", "a", "more", "complete", "appearance", "structure", "and", "discriminative", "prior", "representation", ",", "which", "will", "help", "the", "classifier", "to", "better", "identity", "the", "object", "categories", ".", "Due", "to", "the", "RoI", "pooling", "operation", ",", "the", "global", "FCN", "describes", "the", "proposal", "as", "a", "whole", "with", "CNN", "features", ",", "which", "can", "be", "seen", "as", "a", "global", "structure", "description", "of", "the", "object", ".", "Therefore", ",", "it", "can", "easily", "deal", "with", "the", "objects", "with", "intact", "structure", "and", "finer", "scale", ".", "As", "shown", "in", "Figure", "[", "reference", "]", "(", "b", ")", ",", "our", "global", "FCN", "shows", "a", "large", "confidence", "for", "the", "dining", "table", ".", "However", ",", "in", "most", "cases", ",", "natural", "scenes", "consist", "of", "considerable", "objects", "with", "occlusions", "or", "truncations", ",", "making", "the", "detection", "more", "difficult", ".", "Figure", "[", "reference", "]", "(", "a", ")", "shows", "that", "using", "the", "global", "structure", "information", "alone", "can", "hardly", "make", "a", "confident", "prediction", "for", "the", "truncated", "person", ".", "By", "adding", "local", "part", "structural", "supports", ",", "the", "detection", "performance", "can", "be", "significantly", "boosted", ".", "Therefore", ",", "it", "is", "essential", "to", "combine", "both", "local", "and", "global", "descriptions", "for", "a", "robust", "detection", ".", "subsection", ":", "Coupling", "structure", "To", "match", "the", "same", "order", "of", "magnitude", ",", "we", "apply", "a", "normalization", "operation", "to", "the", "output", "of", "local", "and", "global", "FCN", "before", "they", "are", "combined", "together", ".", "We", "explored", "two", "different", "methods", "to", "perform", "normalization", ":", "an", "L2", "normalization", "layer", "or", "a", "1x1", "convolutional", "layer", "to", "model", "the", "scale", ".", "Meanwhile", ",", "how", "to", "couple", "the", "local", "and", "global", "output", "is", "also", "a", "problem", "that", "needs", "to", "be", "researched", ".", "Here", ",", "we", "investigated", "three", "different", "coupling", "methods", ":", "element", "-", "wise", "sum", ",", "element", "-", "wise", "product", "and", "element", "-", "wise", "maximum", ".", "Our", "experiments", "show", "that", "using", "1x1", "convolution", "along", "with", "element", "-", "wise", "sum", "achieves", "the", "best", "performance", "and", "we", "will", "discuss", "it", "in", "Section", "[", "reference", "]", ".", "With", "the", "coupling", "structure", ",", "CoupleNet", "simultaneously", "exploits", "the", "local", "parts", ",", "global", "structure", "and", "context", "prior", "for", "object", "detection", ".", "The", "whole", "network", "is", "fully", "convolutional", "and", "benefits", "from", "approximate", "joint", "training", "and", "multi", "-", "task", "learning", ".", "We", "also", "note", "that", "the", "global", "branch", "can", "be", "regarded", "as", "a", "lightweight", "Faster", "R", "-", "CNN", ",", "in", "which", "all", "learnable", "parameters", "are", "from", "convolutional", "layers", "and", "the", "depth", "of", "RoI", "-", "wise", "subnetwork", "is", "only", "two", ".", "Therefore", ",", "the", "computational", "complexity", "is", "far", "less", "than", "the", "subnetwork", "in", "ResNet", "-", "based", "Faster", "R", "-", "CNN", "system", "whose", "depth", "is", "ten", ".", "As", "a", "consequence", ",", "our", "CoupleNet", "can", "perform", "the", "inference", "efficiently", ",", "which", "runs", "slightly", "slower", "than", "R", "-", "FCN", "but", "much", "more", "faster", "than", "Faster", "R", "-", "CNN", ".", "section", ":", "Experiments", "We", "train", "and", "evaluate", "our", "method", "on", "three", "challenging", "object", "detection", "datasets", ":", "PASCAL", "VOC2007", ",", "VOC2012", "and", "MS", "COCO", ".", "Since", "all", "these", "three", "datasets", "contain", "a", "variety", "of", "circumstances", ",", "which", "can", "sufficiently", "verify", "the", "effectiveness", "of", "our", "method", ".", "We", "demonstrate", "state", "-", "of", "-", "the", "-", "art", "results", "on", "all", "three", "datasets", "without", "bells", "and", "whistles", ".", "subsection", ":", "Ablation", "studies", "on", "VOC2007", "We", "first", "perform", "experiments", "on", "PASCAL", "VOC", "2007", "with", "20", "object", "categories", "for", "detailed", "analysis", "of", "our", "proposed", "CoupleNet", "detector", ".", "We", "train", "the", "models", "on", "the", "union", "set", "of", "VOC", "2007", "trainval", "and", "VOC", "2012", "trainval", "(", "“", "07", "+", "12", "”", ")", "following", ",", "and", "evaluate", "on", "VOC", "2007", "test", "set", ".", "Object", "detection", "accuracy", "is", "measured", "by", "mean", "Average", "Precision", "(", "mAP", ")", ",", "all", "the", "ablation", "experiments", "use", "single", "-", "scale", "training", "and", "testing", ",", "and", "we", "did", "not", "add", "the", "context", "prior", ".", "Normalization", ".", "Since", "features", "extracted", "form", "different", "layers", "of", "CNN", "show", "various", "of", "scales", ",", "it", "is", "essential", "to", "normalize", "different", "features", "before", "coupling", "them", "together", ".", "Bell", "proposed", "to", "use", "L2", "normalization", "to", "each", "RoI", "-", "pooled", "feature", "and", "re", "-", "scale", "back", "up", "by", "a", "empirical", "scale", ",", "which", "shows", "a", "great", "gain", "on", "VOC", "dataset", ".", "In", "this", "paper", ",", "we", "also", "explore", "two", "different", "normalization", "ways", "to", "normalize", "the", "output", "of", "local", "and", "global", "FCN", ":", "an", "L2", "normalization", "layer", "or", "a", "1x1", "convolutional", "layer", "to", "learn", "the", "scale", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "we", "find", "that", "the", "use", "of", "L2", "normalization", "decreases", "the", "performance", "greatly", ",", "even", "worse", "than", "the", "direct", "addition", "(", "without", "any", "normalization", "ways", ")", ".", "To", "explain", "such", "a", "phenomenon", ",", "we", "measured", "the", "outputs", "of", "two", "branches", "before", "and", "after", "L2", "normalization", ".", "We", "further", "found", "that", "L2", "normalization", "reduces", "the", "output", "gap", "between", "different", "categories", ",", "which", "results", "in", "a", "smaller", "score", "gap", ".", "As", "we", "know", ",", "a", "small", "score", "gap", "between", "different", "categories", "always", "means", "the", "classifier", "can", "not", "make", "a", "confident", "prediction", ".", "Therefore", ",", "we", "assume", "that", "this", "is", "the", "reason", "for", "the", "performance", "degradation", ".", "Moreover", ",", "we", "also", "exploit", "a", "1x1", "convolution", "to", "adaptively", "learn", "the", "scales", "between", "the", "global", "and", "local", "branches", ".", "Table", "[", "reference", "]", "shows", "that", "using", "1x1", "convolution", "increases", "by", "points", "compared", "to", "the", "direct", "addition", "and", "points", "over", "R", "-", "FCN", ".", "Therefore", ",", "we", "use", "1x1", "convolution", "to", "replace", "the", "L2", "normalization", "in", "the", "following", "experiments", ".", "Coupling", "strategy", ".", "We", "explore", "three", "different", "response", "coupling", "strategies", ":", "element", "-", "wise", "sum", ",", "element", "-", "wise", "product", "and", "element", "-", "wise", "maximum", ".", "Table", "[", "reference", "]", "shows", "the", "comparison", "results", "for", "the", "above", "three", "different", "implementations", ".", "We", "can", "see", "that", "the", "element", "-", "wise", "sum", "always", "achieves", "the", "best", "performance", "even", "though", "in", "different", "normalization", "methods", ".", "Generally", ",", "current", "advanced", "residual", "networks", "also", "use", "element", "-", "wise", "sum", "as", "the", "effective", "way", "to", "integrate", "information", "from", "previous", "layers", ",", "which", "greatly", "facilitates", "the", "circulation", "of", "information", "and", "achieves", "the", "complementary", "advantages", ".", "For", "element", "-", "wise", "product", ",", "we", "argue", "that", "the", "system", "is", "relatively", "unstable", "and", "is", "susceptible", "to", "the", "weak", "side", ",", "which", "results", "in", "a", "large", "gradient", "to", "update", "the", "weak", "branch", "that", "makes", "it", "difficult", "to", "converge", ".", "For", "element", "-", "wise", "maximum", ",", "it", "equals", "to", "an", "ensemble", "model", "within", "the", "network", "to", "some", "extent", ",", "which", "losts", "the", "advantages", "of", "mutual", "support", "compared", "to", "element", "-", "wise", "sum", "when", "both", "two", "branches", "are", "failed", "to", "detect", "the", "object", ".", "Moreover", ",", "a", "better", "coupling", "strategy", "can", "be", "taken", "into", "consideration", "as", "the", "future", "work", "to", "further", "improve", "the", "accuracy", ",", "such", "as", "designing", "a", "more", "subtle", "nonlinear", "structure", "to", "learn", "the", "coupling", "relationship", ".", "Model", "ensemble", ".", "Model", "ensemble", "is", "commonly", "used", "to", "improve", "the", "final", "detection", "performance", ",", "since", "diverse", "initialization", "of", "parameters", "and", "the", "randomness", "of", "training", "samples", "both", "lead", "to", "different", "performance", "for", "the", "same", "model", ".", "Although", "the", "differences", "and", "complementarities", "will", "be", "more", "pronounced", "for", "different", "models", ",", "the", "promotion", "is", "often", "very", "limited", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "we", "also", "compare", "our", "CoupleNet", "with", "the", "model", "ensemble", ".", "For", "a", "fair", "comparison", ",", "we", "first", "re", "-", "implemented", "Faster", "R", "-", "CNN", "using", "ResNet", "-", "101", "and", "online", "hard", "example", "mining", "(", "OHEM", ")", ",", "which", "achieves", "a", "mAP", "of", "on", "VOC07", "(", "in", "original", "paper", "without", "OHEM", ")", ".", "We", "also", "re", "-", "implemented", "R", "-", "FCN", "with", "appropriate", "joint", "training", "using", "the", "public", "available", "code", "py", "-", "R", "-", "FCN", ",", "which", "achieves", "a", "slightly", "lower", "result", "compared", "to", "(", "vs.", ")", ".", "We", "use", "our", "reimplementation", "models", "to", "conduct", "the", "comparisons", "for", "consistency", ".", "We", "found", "that", "the", "promotion", "brought", "by", "model", "ensemble", "is", "less", "than", "1", "point", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "it", "is", "far", "less", "than", "our", "method", "(", ")", ".", "On", "the", "one", "hand", ",", "we", "argue", "that", "the", "naive", "model", "ensemble", "just", "combines", "the", "results", "together", "and", "does", "not", "essentially", "guide", "the", "learning", "process", "of", "the", "network", ",", "while", "our", "CoupleNet", "can", "simultaneously", "utilize", "the", "global", "and", "local", "information", "to", "update", "the", "network", "and", "to", "infer", "the", "final", "results", ".", "On", "the", "other", "hand", ",", "our", "method", "enjoys", "end", "-", "to", "-", "end", "training", "and", "there", "is", "no", "need", "to", "train", "multiple", "models", ",", "thus", "greatly", "reducing", "the", "training", "time", ".", "Amount", "of", "parameters", ".", "Since", "our", "CoupleNet", "introduces", "a", "few", "more", "parameters", "compared", "with", "the", "single", "branch", "detectors", ",", "to", "further", "verify", "effectiveness", "of", "the", "coupling", "structure", ",", "here", "we", "increase", "the", "parameters", "of", "the", "prediction", "head", "for", "each", "single", "branch", "implementation", "to", "maintain", "the", "same", "amount", "of", "parameters", "with", "CoupleNet", "for", "comparison", ".", "In", "detail", ",", "we", "add", "a", "new", "residual", "variant", "block", "with", "three", "convolution", "layers", ",", "where", "the", "kernel", "size", "is", "1x1x256", ",", "3x3x256", "and", "1x1x1024", "respectively", ",", "to", "the", "prediction", "sub", "-", "network", ".", "We", "found", "that", "the", "standard", "R", "-", "FCN", "with", "one", "or", "two", "extra", "heads", "got", "a", "mAP", "of", "and", "respectively", "in", "VOC07", ",", "which", "is", "slightly", "higher", "than", "our", "re", "-", "implemented", "version", "(", ")", "in", "as", "shown", "in", "Table", "[", "reference", "]", ".", "Meanwhile", ",", "our", "global", "FCN", ",", "which", "performs", "the", "ROI", "Pooling", "on", "top", "of", "conv5", ",", "got", "a", "relative", "higher", "gain", "(", "a", "mAP", "of", "for", "one", "head", ",", "for", "two", "heads", ")", ".", "The", "results", "indicate", "that", "simply", "adding", "more", "prediction", "layers", "obtains", "a", "very", "limited", "performance", "gain", ",", "while", "our", "coupling", "structure", "shows", "more", "discriminative", "power", "with", "the", "same", "amount", "of", "parameters", ".", "subsection", ":", "Results", "on", "VOC2007", "Using", "the", "public", "available", "ResNet", "-", "101", "as", "the", "initialization", "model", ",", "we", "note", "that", "our", "method", "is", "easy", "to", "follow", "and", "the", "hyper", "-", "parameters", "for", "training", "are", "the", "same", "as", "in", ".", "Similarly", ",", "we", "use", "the", "dilation", "strategy", "to", "reduce", "the", "effective", "stride", "of", "ResNet", "-", "101", ",", "just", "as", "shows", ",", "thus", "both", "the", "global", "and", "local", "branches", "have", "a", "stride", "of", "16", ".", "We", "also", "use", "a", "1", "-", "GPU", "implementation", ",", "and", "the", "effective", "mini", "-", "batch", "size", "is", "2", "images", "by", "setting", "the", "to", "2", ".", "The", "whole", "network", "is", "trained", "for", "80k", "iterations", "with", "a", "learning", "rate", "of", "0.001", "and", "then", "for", "30k", "iterations", "with", "0.0001", ".", "In", "addition", ",", "the", "context", "prior", "is", "proposed", "to", "further", "boost", "the", "performance", "while", "keeping", "the", "iterations", "unchanged", ".", "Finally", ",", "we", "also", "perform", "multi", "-", "scale", "training", "with", "the", "shorter", "sides", "of", "images", "are", "randomly", "resized", "from", "480", "to", "864", ".", "Table", "[", "reference", "]", "shows", "the", "detailed", "comparisons", "with", "Faster", "R", "-", "CNN", "and", "R", "-", "FCN", ".", "As", "we", "can", "see", "that", "our", "single", "model", "achieves", "a", "mAP", "of", ",", "which", "outperforms", "the", "R", "-", "FCN", "by", "2.2", "points", ".", "However", ",", "while", "embedding", "the", "context", "prior", "to", "the", "global", "branch", ",", "our", "mAP", "rises", "up", "to", ",", "which", "is", "the", "current", "best", "single", "model", "detector", "to", "our", "knowledge", ".", "Moreover", ",", "we", "also", "evaluate", "the", "inference", "time", "of", "our", "network", "using", "a", "NVIDIA", "TITAN", "X", "GPU", "(", "pascal", ")", "along", "with", "CUDA", "8.0", "and", "cuDNN", "-", "v5.1", ".", "As", "shown", "in", "the", "last", "column", "of", "Table", "[", "reference", "]", ",", "our", "method", "is", "slightly", "slower", "than", "R", "-", "FCN", ",", "which", "also", "reaches", "a", "real", "-", "time", "speed", "(", "8.2", "fps", "or", "9.8", "fps", "without", "context", ")", "and", "achieves", "the", "best", "trade", "-", "off", "between", "accuracy", "and", "speed", ".", "We", "argue", "that", "the", "sharing", "process", "of", "feature", "extraction", "between", "two", "branches", "and", "the", "design", "of", "lightweight", "RoI", "-", "wise", "subnetwork", "after", "RoI", "pooling", "both", "greatly", "reduce", "the", "model", "complexity", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "we", "also", "compared", "our", "method", "with", "other", "state", "-", "of", "-", "the", "-", "art", "single", "model", ".", "We", "found", "that", "our", "method", "outperforms", "the", "others", "with", "a", "large", "margin", ",", "including", "the", "advanced", "end", "-", "to", "-", "end", "SSD", "method", ",", "which", "requires", "complicated", "data", "augmentation", "and", "careful", "training", "skills", ".", "Just", "as", "discussed", "earlier", ",", "CoupleNet", "shows", "a", "large", "gain", "over", "the", "classes", "with", "occlusions", ",", "truncations", "and", "considerable", "background", "information", ",", "like", "sofa", ",", "person", ",", "table", "and", "chair", ",", "which", "verifies", "our", "analyses", ".", "We", "also", "observed", "a", "large", "improvement", "for", "airplane", ",", "bird", ",", "boat", "and", "pottedplant", ",", "which", "usually", "have", "class", "-", "specific", "backgrounds", ",", "the", "sky", "for", "airplane", "and", "bird", ",", "water", "for", "boat", "and", "so", "on", ".", "Therefore", ",", "the", "context", "surrounding", "the", "objects", "provides", "an", "extra", "auxiliary", "discrimination", ".", "subsection", ":", "Results", "on", "VOC2012", "We", "also", "evaluate", "our", "method", "on", "the", "more", "challenging", "VOC2012", "dataset", "by", "submitting", "results", "to", "the", "public", "evaluation", "server", ".", "We", "use", "VOC07", "trainval", ",", "VOC07", "test", "and", "VOC12", "trainval", "as", "the", "training", "set", ",", "which", "consists", "of", "21k", "images", "in", "total", ".", "We", "also", "follow", "the", "similar", "hyper", "-", "parameter", "settings", "in", "VOC07", "but", "change", "the", "iterations", ",", "since", "there", "are", "more", "training", "images", ".", "We", "train", "our", "models", "with", "4", "GPUs", ",", "and", "the", "effective", "mini", "-", "batch", "size", "thus", "becomes", "4", "(", "1", "per", "GPU", ")", ".", "As", "a", "result", ",", "the", "network", "is", "trained", "for", "60k", "iterations", "with", "a", "learning", "rate", "of", "0.001", "and", "0.0001", "for", "the", "following", "20k", "iterations", ".", "Table", "[", "reference", "]", "shows", "the", "results", "on", "the", "VOC2012", "test", "set", ".", "Our", "method", "obtains", "a", "top", "mAP", "of", ",", "which", "is", "2.8", "points", "higher", "than", "R", "-", "FCN", ".", "We", "note", "that", "without", "using", "the", "extra", "tricks", "in", "the", "testing", "phase", ",", "our", "detector", "is", "the", "first", "one", "with", "a", "mAP", "higher", "than", ".", "Similar", "promotions", "over", "the", "specific", "classes", "analysed", "in", "VOC07", "are", "also", "observed", ",", "which", "once", "again", "validates", "the", "effectiveness", "of", "our", "method", ".", "Figure", "[", "reference", "]", "shows", "some", "detection", "examples", "on", "VOC", "2012", "test", "set", ".", "subsection", ":", "Results", "on", "MS", "COCO", "Next", "we", "present", "more", "results", "on", "the", "Microsoft", "COCO", "object", "detection", "dataset", ".", "The", "dataset", "consists", "of", "80k", "training", "set", ",", "40k", "validation", "set", "and", "20k", "test", "-", "dev", "set", ",", "which", "involves", "80", "object", "categories", ".", "All", "our", "models", "are", "trained", "on", "the", "union", "set", "of", "80k", "training", "set", "and", "40k", "validation", "set", ",", "and", "evaluated", "on", "20k", "test", "-", "dev", "set", ".", "The", "COCO", "standard", "metric", "denotes", "as", "AP", ",", "which", "is", "evaluated", "at", ".", "Following", "the", "VOC2012", ",", "a", "4", "-", "GPU", "implementation", "is", "used", "to", "accelerate", "the", "training", "process", ".", "We", "use", "an", "initial", "learning", "rate", "of", "0.001", "for", "the", "first", "510k", "iterations", "and", "0.0001", "for", "the", "next", "70k", "iterations", ".", "In", "addition", ",", "we", "conduct", "multi", "-", "scale", "training", "with", "the", "scales", "are", "randomly", "sampled", "from", "while", "testing", "in", "a", "single", "scale", ".", "Table", "[", "reference", "]", "shows", "our", "results", ".", "Our", "single", "-", "scale", "trained", "detector", "has", "already", "achieved", "a", "result", "of", ",", "which", "outperforms", "the", "R", "-", "FCN", "by", "3.9", "points", ".", "In", "addition", ",", "the", "multi", "-", "scale", "training", "further", "improves", "the", "performance", "up", "to", ".", "Interestingly", ",", "we", "observed", "that", "the", "more", "challenging", "the", "dataset", ",", "the", "more", "the", "promotion", "(", ",", "for", "VOC07", ",", "for", "VOC12", "and", "for", "COCO", ",", "all", "in", "multi", "-", "scale", "training", ")", ",", "which", "directly", "proves", "that", "our", "approach", "can", "effectively", "cope", "with", "a", "variety", "of", "complex", "situations", ".", "section", ":", "Conclusion", "In", "this", "paper", ",", "we", "present", "the", "CoupleNet", ",", "a", "concise", "yet", "effective", "network", "that", "simultaneously", "couples", "global", ",", "local", "and", "context", "cues", "for", "accurate", "object", "detection", ".", "Our", "system", "naturally", "combines", "the", "advantages", "of", "different", "region", "-", "based", "approaches", "with", "the", "coupling", "structure", ".", "With", "the", "combination", "of", "local", "part", "representation", ",", "global", "structural", "information", "and", "the", "contextual", "assistance", ",", "our", "CoupleNet", "achieves", "state", "-", "of", "-", "the", "-", "art", "results", "on", "the", "challenging", "PASCAL", "VOC", "and", "COCO", "datasets", "without", "using", "any", "extra", "tricks", "in", "the", "testing", "phase", ",", "which", "validates", "the", "effectiveness", "of", "our", "method", ".", "bibliography", ":", "References"], "entity": [[2, 3, ["CoupleNet"], "Method"], [11, 13, ["Object", "Detection"], "Task"], [17, 20, ["Convolutional", "Neural", "Network"], "Method"], [21, 22, ["CNN"], "Method"], [31, 34, ["R", "-", "FCN"], "Method"], [40, 42, ["object", "detection"], "Task"], [45, 48, ["region", "proposal", "subnetwork"], "Method"], [50, 52, ["classification", "subnetwork"], "Method"], [55, 58, ["R", "-", "FCN"], "Method"], [61, 63, ["detection", "speed"], "Metric"], [66, 67, ["detection"], "Task"], [100, 103, ["fully", "convolutional", "network"], "Method"], [106, 107, ["CoupleNet"], "Method"], [117, 119, ["object", "detection"], "Task"], [128, 131, ["Region", "Proposal", "Network"], "Method"], [132, 133, ["RPN"], "Method"], [139, 141, ["coupling", "module"], "Method"], [151, 159, ["position", "-", "sensitive", "RoI", "(", "PSRoI", ")", "pooling"], "Method"], [174, 176, ["RoI", "pooling"], "Method"], [189, 191, ["coupling", "strategies"], "Method"], [235, 236, ["mAP"], "Metric"], [238, 239, ["VOC07"], "Material"], [241, 242, ["VOC12"], "Material"], [245, 246, ["COCO"], "Material"], [282, 288, ["face", ",", "pedestrian", "and", "vehicle", "detection"], "Task"], [353, 357, ["recognition", "of", "object", "categories"], "Task"], [362, 364, ["global", "process"], "Method"], [370, 372, ["local", "process"], "Method"], [385, 387, ["detection", "model"], "Method"], [399, 402, ["Convolutional", "Neural", "Networks"], "Method"], [403, 404, ["CNN"], "Method"], [424, 426, ["PASCAL", "VOC"], "Material"], [427, 429, ["MS", "COCO"], "Material"], [458, 461, ["R", "-", "FCN"], "Method"], [463, 464, ["inference"], "Task"], [475, 479, ["RoI", "-", "wise", "subnetwork"], "Method"], [480, 483, ["R", "-", "FCN"], "Method"], [486, 488, ["detection", "speed"], "Metric"], [491, 492, ["detection"], "Task"], [504, 506, ["PSRoI", "pooling"], "Method"], [516, 518, ["PSRoI", "pooling"], "Method"], [525, 528, ["object", "category", "prediction"], "Task"], [529, 532, ["R", "-", "FCN"], "Method"], [535, 538, ["low", "confidence", "score"], "Metric"], [542, 544, ["sofa", "detection"], "Task"], [583, 585, ["RoI", "pooling"], "Method"], [588, 590, ["confidence", "score"], "Metric"], [626, 628, ["confidence", "score"], "Metric"], [649, 651, ["visual", "tasks"], "Task"], [653, 655, ["fingerprint", "recognition"], "Task"], [673, 677, ["clique", "-", "graph", "matching"], "Task"], [681, 686, ["clique", "-", "graph", "matching", "method"], "Method"], [703, 705, ["scene", "parsing"], "Task"], [709, 712, ["pyramid", "pooling", "module"], "Method"], [736, 738, ["object", "detection"], "Task"], [742, 745, ["global", "root", "model"], "Method"], [747, 751, ["finer", "local", "part", "models"], "Method"], [799, 802, ["full", "convolutional", "network"], "Method"], [804, 805, ["CoupleNet"], "Method"], [817, 819, ["detection", "accuracy"], "Metric"], [828, 829, ["RPN"], "Method"], [833, 835, ["coupling", "module"], "Method"], [845, 847, ["PSRoI", "pooling"], "Method"], [862, 864, ["RoI", "pooling"], "Method"], [877, 879, ["coupling", "strategies"], "Method"], [926, 928, ["representation", "capacity"], "Metric"], [929, 931, ["generalization", "ability"], "Metric"], [936, 937, ["CoupleNet"], "Method"], [941, 942, ["detection"], "Task"], [950, 953, ["PASCAL", "VOC", "07"], "Material"], [954, 955, ["12"], "Material"], [956, 958, ["MS", "COCO"], "Material"], [972, 975, ["model", "ensemble", "approaches"], "Method"], [991, 995, ["unified", "fully", "convolutional", "network"], "Method"], [1006, 1008, ["object", "detection"], "Task"], [1014, 1016, ["normalization", "methods"], "Method"], [1017, 1019, ["coupling", "strategies"], "Method"], [1052, 1053, ["mAP"], "Metric"], [1055, 1056, ["VOC07"], "Material"], [1058, 1059, ["VOC12"], "Material"], [1062, 1064, ["MS", "COCO"], "Material"], [1073, 1074, ["CNN"], "Method"], [1075, 1077, ["visual", "tasks"], "Task"], [1091, 1092, ["DPM"], "Method"], [1094, 1096, ["object", "system"], "Method"], [1097, 1105, ["mixtures", "of", "multi", "-", "scale", "deformable", "part", "models"], "Method"], [1108, 1112, ["coarse", "global", "root", "model"], "Method"], [1114, 1118, ["finer", "local", "part", "models"], "Method"], [1120, 1122, ["root", "model"], "Method"], [1131, 1133, ["part", "models"], "Method"], [1164, 1165, ["DPM"], "Method"], [1170, 1172, ["object", "detection"], "Task"], [1180, 1181, ["HOG"], "Method"], [1201, 1202, ["CNN"], "Method"], [1221, 1224, ["deep", "neural", "networks"], "Method"], [1225, 1227, ["image", "classification"], "Task"], [1234, 1236, ["deep", "learning"], "Method"], [1247, 1250, ["end", "detection", "frameworks"], "Method"], [1252, 1253, ["SSD"], "Method"], [1254, 1255, ["YOLO"], "Method"], [1256, 1257, ["DenseBox"], "Method"], [1258, 1262, ["region", "-", "based", "systems"], "Method"], [1270, 1273, ["R", "-", "FCN"], "Method"], [1277, 1279, ["detection", "accuracy"], "Metric"], [1286, 1292, ["end", "-", "to", "-", "end", "framework"], "Method"], [1294, 1298, ["region", "-", "based", "systems"], "Method"], [1307, 1313, ["divide", "-", "and", "-", "conquer", "strategy"], "Method"], [1317, 1319, ["step", "framework"], "Method"], [1332, 1334, ["data", "augmentation"], "Method"], [1376, 1378, ["deep", "networks"], "Method"], [1396, 1400, ["RoI", "-", "wise", "subnetwork"], "Method"], [1402, 1408, ["position", "-", "sensitive", "RoI", "pooling", "layer"], "Method"], [1416, 1420, ["region", "-", "based", "systems"], "Method"], [1426, 1430, ["region", "-", "based", "systems"], "Method"], [1494, 1496, ["visual", "recognition"], "Task"], [1503, 1505, ["exploting", "context"], "Task"], [1506, 1508, ["object", "detection"], "Task"], [1514, 1517, ["recurrent", "neural", "networks"], "Method"], [1523, 1524, ["Gidaris"], "Method"], [1543, 1547, ["pedestrian", "and", "car", "detection"], "Task"], [1562, 1565, ["global", "feature", "representation"], "Method"], [1568, 1569, ["CoupleNet"], "Method"], [1581, 1582, ["CoupleNet"], "Method"], [1583, 1585, ["object", "detection"], "Task"], [1594, 1596, ["local", "representations"], "Method"], [1609, 1611, ["Network", "architecture"], "Method"], [1616, 1617, ["CoupleNet"], "Method"], [1626, 1627, ["CoupleNet"], "Method"], [1635, 1642, ["local", "part", "-", "sensitive", "fully", "convolutional", "network"], "Method"], [1652, 1654, ["local", "FCN"], "Method"], [1658, 1665, ["global", "region", "-", "sensitive", "fully", "convolutional", "network"], "Method"], [1680, 1682, ["global", "FCN"], "Method"], [1687, 1688, ["ImageNet"], "Material"], [1703, 1705, ["detection", "task"], "Task"], [1709, 1713, ["last", "average", "pooling", "layer"], "Method"], [1715, 1717, ["fc", "layer"], "Method"], [1730, 1733, ["Region", "Proposal", "Network"], "Method"], [1734, 1736, ["RPN", ")"], "Method"], [1756, 1758, ["local", "FCN"], "Method"], [1760, 1762, ["global", "FCN"], "Method"], [1768, 1772, ["global", "and", "local", "FCN"], "Method"], [1786, 1792, ["class", "-", "agnostic", "bounding", "box", "regression"], "Method"], [1799, 1801, ["Local", "FCN"], "Method"], [1811, 1813, ["local", "FCN"], "Method"], [1819, 1824, ["part", "-", "sensitive", "score", "maps"], "Method"], [1828, 1830, ["convolutional", "layer"], "Method"], [1900, 1906, ["position", "-", "sensitive", "RoI", "pooling", "layer"], "Method"], [1917, 1919, ["average", "pooling"], "Method"], [1920, 1921, ["voting"], "Task"], [1958, 1960, ["weak", "classifiers"], "Method"], [1968, 1970, ["part", "models"], "Method"], [1978, 1981, ["local", "structure", "representation"], "Method"], [2021, 2023, ["local", "FCN"], "Method"], [2055, 2057, ["local", "FCN"], "Method"], [2112, 2114, ["local", "FCN"], "Method"], [2135, 2136, ["discrimination"], "Task"], [2139, 2141, ["Global", "FCN"], "Method"], [2143, 2145, ["global", "FCN"], "Method"], [2166, 2172, ["1024", "-", "d", "1x1", "convolutional", "layer"], "Method"], [2198, 2201, ["RoI", "pooling", "layer"], "Method"], [2224, 2226, ["convolutional", "layers"], "Method"], [2244, 2248, ["global", "representation", "of", "RoI."], "Method"], [2253, 2255, ["1x1", "convolution"], "Method"], [2259, 2260, ["classifier"], "Method"], [2282, 2285, ["visual", "recognition", "tasks"], "Task"], [2310, 2313, ["deep", "neural", "network"], "Method"], [2362, 2363, ["misclassification"], "Task"], [2367, 2369, ["feature", "representation"], "Task"], [2372, 2374, ["global", "FCN"], "Method"], [2424, 2428, ["RoI", "-", "wise", "subnetwork"], "Method"], [2454, 2457, ["discriminative", "prior", "representation"], "Method"], [2462, 2463, ["classifier"], "Method"], [2473, 2476, ["RoI", "pooling", "operation"], "Method"], [2478, 2480, ["global", "FCN"], "Method"], [2531, 2533, ["global", "FCN"], "Method"], [2561, 2562, ["detection"], "Task"], [2599, 2600, ["detection"], "Task"], [2620, 2622, ["robust", "detection"], "Task"], [2638, 2640, ["normalization", "operation"], "Method"], [2644, 2648, ["local", "and", "global", "FCN"], "Method"], [2661, 2662, ["normalization"], "Task"], [2664, 2667, ["L2", "normalization", "layer"], "Method"], [2669, 2672, ["1x1", "convolutional", "layer"], "Method"], [2703, 2705, ["coupling", "methods"], "Method"], [2706, 2710, ["element", "-", "wise", "sum"], "Method"], [2711, 2715, ["element", "-", "wise", "product"], "Method"], [2716, 2720, ["element", "-", "wise", "maximum"], "Method"], [2726, 2728, ["1x1", "convolution"], "Method"], [2730, 2734, ["element", "-", "wise", "sum"], "Method"], [2754, 2755, ["CoupleNet"], "Method"], [2767, 2769, ["object", "detection"], "Task"], [2775, 2776, ["convolutional"], "Method"], [2779, 2782, ["approximate", "joint", "training"], "Method"], [2783, 2787, ["multi", "-", "task", "learning"], "Task"], [2813, 2815, ["convolutional", "layers"], "Method"], [2830, 2832, ["computational", "complexity"], "Metric"], [2857, 2858, ["CoupleNet"], "Method"], [2861, 2862, ["inference"], "Task"], [2869, 2872, ["R", "-", "FCN"], "Method"], [2894, 2896, ["object", "detection"], "Task"], [2901, 2902, ["VOC2012"], "Material"], [2903, 2905, ["MS", "COCO"], "Material"], [2948, 2950, ["Ablation", "studies"], "Task"], [2951, 2952, ["VOC2007"], "Material"], [2979, 2985, ["union", "set", "of", "VOC", "2007", "trainval"], "Material"], [2986, 2989, ["VOC", "2012", "trainval"], "Material"], [3001, 3005, ["VOC", "2007", "test", "set"], "Material"], [3006, 3009, ["Object", "detection", "accuracy"], "Metric"], [3012, 3015, ["mean", "Average", "Precision"], "Metric"], [3016, 3017, ["mAP"], "Metric"], [3040, 3041, ["Normalization"], "Task"], [3049, 3050, ["CNN"], "Method"], [3071, 3073, ["L2", "normalization"], "Method"], [3096, 3098, ["VOC", "dataset"], "Material"], [3108, 3110, ["normalization", "ways"], "Method"], [3115, 3119, ["local", "and", "global", "FCN"], "Method"], [3121, 3124, ["L2", "normalization", "layer"], "Method"], [3126, 3129, ["1x1", "convolutional", "layer"], "Method"], [3148, 3150, ["L2", "normalization"], "Method"], [3159, 3161, ["direct", "addition"], "Method"], [3184, 3186, ["L2", "normalization"], "Method"], [3191, 3193, ["L2", "normalization"], "Method"], [3206, 3208, ["score", "gap"], "Metric"], [3223, 3224, ["classifier"], "Method"], [3251, 3253, ["1x1", "convolution"], "Method"], [3272, 3274, ["1x1", "convolution"], "Method"], [3280, 3282, ["direct", "addition"], "Method"], [3285, 3288, ["R", "-", "FCN"], "Method"], [3293, 3295, ["1x1", "convolution"], "Method"], [3298, 3300, ["L2", "normalization"], "Method"], [3305, 3307, ["Coupling", "strategy"], "Method"], [3312, 3315, ["response", "coupling", "strategies"], "Method"], [3316, 3320, ["element", "-", "wise", "sum"], "Method"], [3321, 3325, ["element", "-", "wise", "product"], "Method"], [3326, 3330, ["element", "-", "wise", "maximum"], "Method"], [3351, 3355, ["element", "-", "wise", "sum"], "Method"], [3364, 3366, ["normalization", "methods"], "Method"], [3371, 3373, ["residual", "networks"], "Method"], [3375, 3379, ["element", "-", "wise", "sum"], "Method"], [3404, 3408, ["element", "-", "wise", "product"], "Task"], [3444, 3448, ["element", "-", "wise", "maximum"], "Method"], [3453, 3455, ["ensemble", "model"], "Method"], [3490, 3492, ["coupling", "strategy"], "Method"], [3505, 3506, ["accuracy"], "Metric"], [3521, 3523, ["Model", "ensemble"], "Method"], [3524, 3526, ["Model", "ensemble"], "Method"], [3533, 3534, ["detection"], "Task"], [3589, 3590, ["CoupleNet"], "Method"], [3592, 3594, ["model", "ensemble"], "Method"], [3610, 3613, ["ResNet", "-", "101"], "Method"], [3614, 3618, ["online", "hard", "example", "mining"], "Method"], [3619, 3620, ["OHEM"], "Method"], [3625, 3626, ["mAP"], "Metric"], [3628, 3629, ["VOC07"], "Material"], [3634, 3635, ["OHEM"], "Method"], [3642, 3645, ["R", "-", "FCN"], "Method"], [3647, 3649, ["joint", "training"], "Method"], [3651, 3659, ["public", "available", "code", "py", "-", "R", "-", "FCN"], "Method"], [3675, 3677, ["reimplementation", "models"], "Method"], [3691, 3693, ["model", "ensemble"], "Method"], [3726, 3729, ["naive", "model", "ensemble"], "Method"], [3740, 3742, ["learning", "process"], "Method"], [3748, 3749, ["CoupleNet"], "Method"], [3776, 3782, ["end", "-", "to", "-", "end", "training"], "Task"], [3796, 3798, ["training", "time"], "Metric"], [3805, 3806, ["CoupleNet"], "Method"], [3814, 3817, ["single", "branch", "detectors"], "Method"], [3824, 3826, ["coupling", "structure"], "Method"], [3849, 3850, ["CoupleNet"], "Method"], [3860, 3863, ["residual", "variant", "block"], "Method"], [3865, 3867, ["convolution", "layers"], "Method"], [3882, 3886, ["prediction", "sub", "-", "network"], "Method"], [3892, 3895, ["R", "-", "FCN"], "Method"], [3903, 3904, ["mAP"], "Metric"], [3908, 3909, ["VOC07"], "Material"], [3934, 3936, ["global", "FCN"], "Method"], [3940, 3942, ["ROI", "Pooling"], "Method"], [3945, 3946, ["conv5"], "Method"], [3954, 3955, ["mAP"], "Metric"], [3983, 3985, ["coupling", "structure"], "Method"], [4000, 4001, ["VOC2007"], "Material"], [4005, 4008, ["ResNet", "-", "101"], "Method"], [4010, 4012, ["initialization", "model"], "Method"], [4040, 4042, ["dilation", "strategy"], "Method"], [4073, 4077, ["1", "-", "GPU", "implementation"], "Method"], [4081, 4085, ["mini", "-", "batch", "size"], "Metric"], [4104, 4106, ["learning", "rate"], "Metric"], [4140, 4144, ["multi", "-", "scale", "training"], "Task"], [4172, 4175, ["R", "-", "FCN"], "Method"], [4192, 4195, ["R", "-", "FCN"], "Method"], [4212, 4213, ["mAP"], "Metric"], [4235, 4237, ["inference", "time"], "Metric"], [4242, 4246, ["NVIDIA", "TITAN", "X", "GPU"], "Method"], [4251, 4253, ["CUDA", "8.0"], "Method"], [4254, 4257, ["cuDNN", "-", "v5.1"], "Method"], [4276, 4279, ["R", "-", "FCN"], "Method"], [4284, 4288, ["real", "-", "time", "speed"], "Metric"], [4290, 4291, ["fps"], "Metric"], [4305, 4306, ["accuracy"], "Metric"], [4307, 4308, ["speed"], "Metric"], [4316, 4318, ["feature", "extraction"], "Method"], [4325, 4330, ["lightweight", "RoI", "-", "wise", "subnetwork"], "Method"], [4331, 4333, ["RoI", "pooling"], "Method"], [4337, 4339, ["model", "complexity"], "Metric"], [4362, 4364, ["single", "model"], "Method"], [4381, 4388, ["end", "-", "to", "-", "end", "SSD", "method"], "Method"], [4392, 4394, ["data", "augmentation"], "Task"], [4404, 4405, ["CoupleNet"], "Method"], [4489, 4490, ["VOC2012"], "Material"], [4506, 4509, ["public", "evaluation", "server"], "Metric"], [4518, 4520, ["VOC12", "trainval"], "Material"], [4543, 4544, ["VOC07"], "Material"], [4562, 4563, ["GPUs"], "Method"], [4593, 4595, ["learning", "rate"], "Metric"], [4632, 4635, ["R", "-", "FCN"], "Method"], [4646, 4648, ["testing", "phase"], "Task"], [4657, 4658, ["mAP"], "Metric"], [4669, 4670, ["VOC07"], "Material"], [4693, 4697, ["VOC", "2012", "test", "set"], "Material"], [4702, 4704, ["MS", "COCO"], "Material"], [4769, 4772, ["COCO", "standard", "metric"], "Metric"], [4774, 4775, ["AP"], "Method"], [4783, 4784, ["VOC2012"], "Material"], [4786, 4790, ["4", "-", "GPU", "implementation"], "Method"], [4795, 4797, ["training", "process"], "Task"], [4802, 4804, ["learning", "rate"], "Metric"], [4824, 4828, ["multi", "-", "scale", "training"], "Task"], [4851, 4856, ["single", "-", "scale", "trained", "detector"], "Method"], [4866, 4869, ["R", "-", "FCN"], "Method"], [4877, 4881, ["multi", "-", "scale", "training"], "Method"], [4906, 4907, ["VOC07"], "Material"], [4909, 4910, ["VOC12"], "Material"], [4912, 4913, ["COCO"], "Material"], [4916, 4920, ["multi", "-", "scale", "training"], "Task"], [4948, 4949, ["CoupleNet"], "Method"], [4977, 4981, ["region", "-", "based", "approaches"], "Method"], [4990, 4993, ["local", "part", "representation"], "Method"], [5003, 5004, ["CoupleNet"], "Method"], [5016, 5018, ["PASCAL", "VOC"], "Material"], [5019, 5021, ["COCO", "datasets"], "Material"], [29, 30, ["CNN"], "Method"], [258, 260, ["object", "detection"], "Task"], [277, 279, ["object", "detection"], "Task"], [290, 292, ["object", "detection"], "Task"], [406, 407, ["CNN"], "Method"], [409, 411, ["object", "detection"], "Task"], [436, 437, ["CNN"], "Method"], [444, 445, ["CNN"], "Method"], [1229, 1231, ["object", "detection"], "Task"], [1268, 1269, ["CNN"], "Method"], [1604, 1606, ["object", "detection"], "Task"], [1691, 1694, ["ResNet", "-", "101"], "Method"], [1743, 1744, ["CoupleNet"], "Method"], [2178, 2181, ["ResNet", "-", "101"], "Method"], [2487, 2488, ["CNN"], "Method"], [2804, 2805, ["CNN"], "Method"], [2845, 2846, ["CNN"], "Method"], [2880, 2881, ["CNN"], "Method"], [2899, 2900, ["VOC2007"], "Material"], [2957, 2959, ["PASCAL", "VOC"], "Material"], [2970, 2971, ["CoupleNet"], "Method"], [2993, 2994, ["12"], "Material"], [3608, 3609, ["CNN"], "Method"], [4048, 4051, ["ResNet", "-", "101"], "Method"], [4170, 4171, ["CNN"], "Method"], [4186, 4187, ["mAP"], "Metric"], [4499, 4500, ["VOC2012"], "Material"], [4512, 4513, ["VOC07"], "Material"], [4515, 4516, ["VOC07"], "Material"], [4614, 4615, ["VOC2012"], "Material"], [4623, 4624, ["mAP"], "Metric"], [4713, 4715, ["object", "detection"], "Task"], [4966, 4968, ["object", "detection"], "Task"]]} {"doc_id": "436b07bebaa1d1f05ef85415e10374048d25334d", "text": ["OUTRAGEOUSLY", "LARGE", "NEURAL", "NETWORKS", ":", "THE", "SPARSELY", "-", "GATED", "MIXTURE", "-", "OF", "-", "EXPERTS", "LAYER", "section", ":", "ABSTRACT", "The", "capacity", "of", "a", "neural", "network", "to", "absorb", "information", "is", "limited", "by", "its", "number", "of", "parameters", ".", "Conditional", "computation", ",", "where", "parts", "of", "the", "network", "are", "active", "on", "a", "per", "-", "example", "basis", ",", "has", "been", "proposed", "in", "theory", "as", "a", "way", "of", "dramatically", "increasing", "model", "capacity", "without", "a", "proportional", "increase", "in", "computation", ".", "In", "practice", ",", "however", ",", "there", "are", "significant", "algorithmic", "and", "performance", "challenges", ".", "In", "this", "work", ",", "we", "address", "these", "challenges", "and", "finally", "realize", "the", "promise", "of", "conditional", "computation", ",", "achieving", "greater", "than", "1000x", "improvements", "in", "model", "capacity", "with", "only", "minor", "losses", "in", "computational", "efficiency", "on", "modern", "GPU", "clusters", ".", "We", "introduce", "a", "Sparsely", "-", "Gated", "Mixture", "-", "of", "-", "Experts", "layer", "(", "MoE", ")", ",", "consisting", "of", "up", "to", "thousands", "of", "feed", "-", "forward", "sub", "-", "networks", ".", "A", "trainable", "gating", "network", "determines", "a", "sparse", "combination", "of", "these", "experts", "to", "use", "for", "each", "example", ".", "We", "apply", "the", "MoE", "to", "the", "tasks", "of", "language", "modeling", "and", "machine", "translation", ",", "where", "model", "capacity", "is", "critical", "for", "absorbing", "the", "vast", "quantities", "of", "knowledge", "available", "in", "the", "training", "corpora", ".", "We", "present", "model", "architectures", "in", "which", "a", "MoE", "with", "up", "to", "137", "billion", "parameters", "is", "applied", "convolutionally", "between", "stacked", "LSTM", "layers", ".", "On", "large", "language", "modeling", "and", "machine", "translation", "benchmarks", ",", "these", "models", "achieve", "significantly", "better", "results", "than", "state", "-", "of", "-", "the", "-", "art", "at", "lower", "computational", "cost", ".", "section", ":", "INTRODUCTION", "AND", "RELATED", "WORK", "section", ":", "CONDITIONAL", "COMPUTATION", "Exploiting", "scale", "in", "both", "training", "data", "and", "model", "size", "has", "been", "central", "to", "the", "success", "of", "deep", "learning", ".", "When", "datasets", "are", "sufficiently", "large", ",", "increasing", "the", "capacity", "(", "number", "of", "parameters", ")", "of", "neural", "networks", "can", "give", "much", "better", "prediction", "accuracy", ".", "This", "has", "been", "shown", "in", "domains", "such", "as", "text", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ",", "images", "[", "reference", "][", "reference", "]", ",", "and", "audio", "[", "reference", "][", "reference", "]", ".", "For", "typical", "deep", "learning", "models", ",", "where", "the", "entire", "model", "is", "activated", "for", "every", "example", ",", "this", "leads", "to", "a", "roughly", "quadratic", "blow", "-", "up", "in", "training", "costs", ",", "as", "both", "the", "model", "size", "and", "the", "number", "of", "training", "examples", "increase", ".", "Unfortunately", ",", "the", "advances", "in", "computing", "power", "and", "distributed", "computation", "fall", "short", "of", "meeting", "such", "demand", ".", "Various", "forms", "of", "conditional", "computation", "have", "been", "proposed", "as", "a", "way", "to", "increase", "model", "capacity", "without", "a", "proportional", "increase", "in", "computational", "costs", "[", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "][", "reference", "]", ".", "In", "these", "schemes", ",", "large", "parts", "of", "a", "network", "are", "active", "or", "inactive", "on", "a", "per", "-", "example", "basis", ".", "The", "gating", "decisions", "may", "be", "binary", "or", "sparse", "and", "continuous", ",", "stochastic", "or", "deterministic", ".", "Various", "forms", "of", "reinforcement", "learning", "and", "back", "-", "propagation", "are", "proposed", "for", "trarining", "the", "gating", "decisions", ".", "•", "Model", "capacity", "is", "most", "critical", "for", "very", "large", "data", "sets", ".", "The", "existing", "literature", "on", "conditional", "computation", "deals", "with", "relatively", "small", "image", "recognition", "data", "sets", "consisting", "of", "up", "to", "600", ",", "000", "images", ".", "It", "is", "hard", "to", "imagine", "that", "the", "labels", "of", "these", "images", "provide", "a", "sufficient", "signal", "to", "adequately", "train", "a", "model", "with", "millions", ",", "let", "alone", "billions", "of", "parameters", ".", "In", "this", "work", ",", "we", "for", "the", "first", "time", "address", "all", "of", "the", "above", "challenges", "and", "finally", "realize", "the", "promise", "of", "conditional", "computation", ".", "We", "obtain", "greater", "than", "1000x", "improvements", "in", "model", "capacity", "with", "only", "minor", "losses", "in", "computational", "efficiency", "and", "significantly", "advance", "the", "state", "-", "of", "-", "the", "-", "art", "results", "on", "public", "language", "modeling", "and", "translation", "data", "sets", ".", "section", ":", "OUR", "APPROACH", ":", "THE", "SPARSELY", "-", "GATED", "MIXTURE", "-", "OF", "-", "EXPERTS", "LAYER", "Our", "approach", "to", "conditional", "computation", "is", "to", "introduce", "a", "new", "type", "of", "general", "purpose", "neural", "network", "component", ":", "a", "Sparsely", "-", "Gated", "Mixture", "-", "of", "-", "Experts", "Layer", "(", "MoE", ")", ".", "The", "MoE", "consists", "of", "a", "number", "of", "experts", ",", "each", "a", "simple", "feed", "-", "forward", "neural", "network", ",", "and", "a", "trainable", "gating", "network", "which", "selects", "a", "sparse", "combination", "of", "the", "experts", "to", "process", "each", "input", "(", "see", "Figure", "1", ")", ".", "All", "parts", "of", "the", "network", "are", "trained", "jointly", "by", "back", "-", "propagation", ".", "While", "the", "introduced", "technique", "is", "generic", ",", "in", "this", "paper", "we", "focus", "on", "language", "modeling", "and", "machine", "translation", "tasks", ",", "which", "are", "known", "to", "benefit", "from", "very", "large", "models", ".", "In", "particular", ",", "we", "apply", "a", "MoE", "convolutionally", "between", "stacked", "LSTM", "layers", "[", "reference", "]", ",", "as", "in", "Figure", "1", ".", "The", "MoE", "is", "called", "once", "for", "each", "position", "in", "the", "text", ",", "selecting", "a", "potentially", "different", "combination", "of", "experts", "at", "each", "position", ".", "The", "different", "experts", "tend", "to", "become", "highly", "specialized", "based", "on", "syntax", "and", "semantics", "(", "see", "Appendix", "E", "Table", "9", ")", ".", "On", "both", "language", "modeling", "and", "machine", "translation", "benchmarks", ",", "we", "improve", "on", "best", "published", "results", "at", "a", "fraction", "of", "the", "computational", "cost", ".", "section", ":", "RELATED", "WORK", "ON", "MIXTURES", "OF", "EXPERTS", "Since", "its", "introduction", "more", "than", "two", "decades", "ago", "[", "reference", "][", "reference", "]", ",", "the", "mixture", "-", "of", "-", "experts", "approach", "has", "been", "the", "subject", "of", "much", "research", ".", "Different", "types", "of", "expert", "architectures", "hae", "been", "proposed", "such", "as", "SVMs", "[", "reference", "]", ",", "Gaussian", "Processes", "[", "reference", "][", "reference", "][", "reference", "]", ",", "Dirichlet", "Processes", "[", "reference", "]", ",", "and", "deep", "networks", ".", "Other", "work", "has", "focused", "on", "different", "expert", "configurations", "such", "as", "a", "hierarchical", "structure", "[", "reference", "]", ",", "infinite", "numbers", "of", "experts", "[", "reference", "]", ",", "and", "adding", "experts", "sequentially", "[", "reference", "]", ".", "[", "reference", "]", "suggest", "an", "ensemble", "model", "in", "the", "format", "of", "mixture", "of", "experts", "for", "machine", "translation", ".", "The", "gating", "network", "is", "trained", "on", "a", "pre", "-", "trained", "ensemble", "NMT", "model", ".", "The", "works", "above", "concern", "top", "-", "level", "mixtures", "of", "experts", ".", "The", "mixture", "of", "experts", "is", "the", "whole", "model", ".", "[", "reference", "]", "introduce", "the", "idea", "of", "using", "multiple", "MoEs", "with", "their", "own", "gating", "networks", "as", "parts", "of", "a", "deep", "model", ".", "It", "is", "intuitive", "that", "the", "latter", "approach", "is", "more", "powerful", ",", "since", "complex", "problems", "may", "contain", "many", "sub", "-", "problems", "each", "requiring", "different", "experts", ".", "They", "also", "allude", "in", "their", "conclusion", "to", "the", "potential", "to", "introduce", "sparsity", ",", "turning", "MoEs", "into", "a", "vehicle", "for", "computational", "computation", ".", "Our", "work", "builds", "on", "this", "use", "of", "MoEs", "as", "a", "general", "purpose", "neural", "network", "component", ".", "While", "[", "reference", "]", "uses", "two", "stacked", "MoEs", "allowing", "for", "two", "sets", "of", "gating", "decisions", ",", "our", "convolutional", "application", "of", "the", "MoE", "allows", "for", "different", "gating", "decisions", "at", "each", "position", "in", "the", "text", ".", "We", "also", "realize", "sparse", "gating", "and", "demonstrate", "its", "use", "as", "a", "practical", "way", "to", "massively", "increase", "model", "capacity", ".", "section", ":", "THE", "STRUCTURE", "OF", "THE", "MIXTURE", "-", "OF", "-", "EXPERTS", "LAYER", "The", "Mixture", "-", "of", "-", "Experts", "(", "MoE", ")", "layer", "consists", "of", "a", "set", "of", "n", "\"", "expert", "networks", "\"", "E", "1", ",", "·", "·", "·", ",", "E", "n", ",", "and", "a", "\"", "gating", "network", "\"", "G", "whose", "output", "is", "a", "sparse", "n", "-", "dimensional", "vector", ".", "Figure", "1", "shows", "an", "overview", "of", "the", "MoE", "module", ".", "The", "experts", "are", "themselves", "neural", "networks", ",", "each", "with", "their", "own", "parameters", ".", "Although", "in", "principle", "we", "only", "require", "that", "the", "experts", "accept", "the", "same", "sized", "inputs", "and", "produce", "the", "same", "-", "sized", "outputs", ",", "in", "our", "initial", "investigations", "in", "this", "paper", ",", "we", "restrict", "ourselves", "to", "the", "case", "where", "the", "models", "are", "feed", "-", "forward", "networks", "with", "identical", "architectures", ",", "but", "with", "separate", "parameters", ".", "Let", "us", "denote", "by", "G", "(", "x", ")", "and", "E", "i", "(", "x", ")", "the", "output", "of", "the", "gating", "network", "and", "the", "output", "of", "the", "i", "-", "th", "expert", "network", "for", "a", "given", "input", "x.", "The", "output", "y", "of", "the", "MoE", "module", "can", "be", "written", "as", "follows", ":", "We", "save", "computation", "based", "on", "the", "sparsity", "of", "the", "output", "of", "G", "(", "x", ")", ".", "Wherever", "G", "(", "x", ")", "i", "=", "0", ",", "we", "need", "not", "compute", "E", "i", "(", "x", ")", ".", "In", "our", "experiments", ",", "we", "have", "up", "to", "thousands", "of", "experts", ",", "but", "only", "need", "to", "evaluate", "a", "handful", "of", "them", "for", "every", "example", ".", "If", "the", "number", "of", "experts", "is", "very", "large", ",", "we", "can", "reduce", "the", "branching", "factor", "by", "using", "a", "two", "-", "level", "hierarchical", "MoE.", "In", "a", "hierarchical", "MoE", ",", "a", "primary", "gating", "network", "chooses", "a", "sparse", "weighted", "combination", "of", "\"", "experts", "\"", ",", "each", "of", "which", "is", "itself", "a", "secondary", "mixture", "-", "of", "-", "experts", "with", "its", "own", "gating", "network", ".", "In", "the", "following", "we", "focus", "on", "ordinary", "MoEs", ".", "We", "provide", "more", "details", "on", "hierarchical", "MoEs", "in", "Appendix", "B.", "Our", "implementation", "is", "related", "to", "other", "models", "of", "conditional", "computation", ".", "A", "MoE", "whose", "experts", "are", "simple", "weight", "matrices", "is", "similar", "to", "the", "parameterized", "weight", "matrix", "proposed", "in", ".", "A", "MoE", "whose", "experts", "have", "one", "hidden", "layer", "is", "similar", "to", "the", "block", "-", "wise", "dropout", "described", "in", "[", "reference", "]", ",", "where", "the", "dropped", "-", "out", "layer", "is", "sandwiched", "between", "fully", "-", "activated", "layers", ".", "section", ":", "GATING", "NETWORK", "Softmax", "Gating", ":", "A", "simple", "choice", "of", "non", "-", "sparse", "gating", "function", "[", "reference", "]", "is", "to", "multiply", "the", "input", "by", "a", "trainable", "weight", "matrix", "W", "g", "and", "then", "apply", "the", "Sof", "tmax", "function", ".", "Noisy", "Top", "-", "K", "Gating", ":", "We", "add", "two", "components", "to", "the", "Softmax", "gating", "network", ":", "sparsity", "and", "noise", ".", "Before", "taking", "the", "softmax", "function", ",", "we", "add", "tunable", "Gaussian", "noise", ",", "then", "keep", "only", "the", "top", "k", "values", ",", "setting", "the", "rest", "to", "−∞", "(", "which", "causes", "the", "corresponding", "gate", "values", "to", "equal", "0", ")", ".", "The", "sparsity", "serves", "to", "save", "computation", ",", "as", "described", "above", ".", "While", "this", "form", "of", "sparsity", "creates", "some", "theoretically", "scary", "discontinuities", "in", "the", "output", "of", "gating", "function", ",", "we", "have", "not", "yet", "observed", "this", "to", "be", "a", "problem", "in", "practice", ".", "The", "noise", "term", "helps", "with", "load", "balancing", ",", "as", "will", "be", "discussed", "in", "Appendix", "A.", "The", "amount", "of", "noise", "per", "component", "is", "controlled", "by", "a", "second", "trainable", "weight", "matrix", "W", "noise", ".", "Training", "the", "Gating", "Network", "We", "train", "the", "gating", "network", "by", "simple", "back", "-", "propagation", ",", "along", "with", "the", "rest", "of", "the", "model", ".", "If", "we", "choose", "k", ">", "1", ",", "the", "gate", "values", "for", "the", "top", "k", "experts", "have", "nonzero", "derivatives", "with", "respect", "to", "the", "weights", "of", "the", "gating", "network", ".", "This", "type", "of", "occasionally", "-", "sensitive", "behavior", "is", "described", "in", "[", "reference", "]", ")", "with", "respect", "to", "noisy", "rectifiers", ".", "Gradients", "also", "backpropagate", "through", "the", "gating", "network", "to", "its", "inputs", ".", "Our", "method", "differs", "here", "from", "[", "reference", "]", "who", "use", "boolean", "gates", "and", "a", "REINFORCE", "-", "style", "approach", "to", "train", "the", "gating", "network", ".", "section", ":", "ADDRESSING", "PERFORMANCE", "CHALLENGES", "section", ":", "THE", "SHRINKING", "BATCH", "PROBLEM", "On", "modern", "CPUs", "and", "GPUs", ",", "large", "batch", "sizes", "are", "necessary", "for", "computational", "efficiency", ",", "so", "as", "to", "amortize", "the", "overhead", "of", "parameter", "loads", "and", "updates", ".", "If", "the", "gating", "network", "chooses", "k", "out", "of", "n", "experts", "for", "each", "example", ",", "then", "for", "a", "batch", "of", "b", "examples", ",", "each", "expert", "receives", "a", "much", "smaller", "batch", "of", "approximately", "kb", "n", "b", "examples", ".", "This", "causes", "a", "naive", "MoE", "implementation", "to", "become", "very", "inefficient", "as", "the", "number", "of", "experts", "increases", ".", "The", "solution", "to", "this", "shrinking", "batch", "problem", "is", "to", "make", "the", "original", "batch", "size", "as", "large", "as", "possible", ".", "However", ",", "batch", "size", "tends", "to", "be", "limited", "by", "the", "memory", "necessary", "to", "store", "activations", "between", "the", "forwards", "and", "backwards", "passes", ".", "We", "propose", "the", "following", "techniques", "for", "increasing", "the", "batch", "size", ":", "Mixing", "Data", "Parallelism", "and", "Model", "Parallelism", ":", "In", "a", "conventional", "distributed", "training", "setting", ",", "multiple", "copies", "of", "the", "model", "on", "different", "devices", "asynchronously", "process", "distinct", "batches", "of", "data", ",", "and", "parameters", "are", "synchronized", "through", "a", "set", "of", "parameter", "servers", ".", "In", "our", "technique", ",", "these", "different", "batches", "run", "synchronously", "so", "that", "they", "can", "be", "combined", "for", "the", "MoE", "layer", ".", "We", "distribute", "the", "standard", "layers", "of", "the", "model", "and", "the", "gating", "network", "according", "to", "conventional", "data", "-", "parallel", "schemes", ",", "but", "keep", "only", "one", "shared", "copy", "of", "each", "expert", ".", "Each", "expert", "in", "the", "MoE", "layer", "receives", "a", "combined", "batch", "consisting", "of", "the", "relevant", "examples", "from", "all", "of", "the", "data", "-", "parallel", "input", "batches", ".", "The", "same", "set", "of", "devices", "function", "as", "data", "-", "parallel", "replicas", "(", "for", "the", "standard", "layers", "and", "the", "gating", "networks", ")", "and", "as", "model", "-", "parallel", "shards", "(", "each", "hosting", "a", "subset", "of", "the", "experts", ")", ".", "If", "the", "model", "is", "distributed", "over", "d", "devices", ",", "and", "each", "device", "processes", "a", "batch", "of", "size", "b", ",", "each", "expert", "receives", "a", "batch", "of", "approximately", "kbd", "n", "examples", ".", "Thus", ",", "we", "achieve", "a", "factor", "of", "d", "improvement", "in", "expert", "batch", "size", ".", "In", "the", "case", "of", "a", "hierarchical", "MoE", "(", "Section", "B", ")", ",", "the", "primary", "gating", "network", "employs", "data", "parallelism", ",", "and", "the", "secondary", "MoEs", "employ", "model", "parallelism", ".", "Each", "secondary", "MoE", "resides", "on", "one", "device", ".", "This", "technique", "allows", "us", "to", "increase", "the", "number", "of", "experts", "(", "and", "hence", "the", "number", "of", "parameters", ")", "by", "proportionally", "increasing", "the", "number", "of", "devices", "in", "the", "training", "cluster", ".", "The", "total", "batch", "size", "increases", ",", "keeping", "the", "batch", "size", "per", "expert", "constant", ".", "The", "memory", "and", "bandwidth", "requirements", "per", "device", "also", "remain", "constant", ",", "as", "do", "the", "step", "times", ",", "as", "does", "the", "amount", "of", "time", "necessary", "to", "process", "a", "number", "of", "training", "examples", "equal", "to", "the", "number", "of", "parameters", "in", "the", "model", ".", "It", "is", "our", "goal", "to", "train", "a", "trillionparameter", "model", "on", "a", "trillion", "-", "word", "corpus", ".", "We", "have", "not", "scaled", "our", "systems", "this", "far", "as", "of", "the", "writing", "of", "this", "paper", ",", "but", "it", "should", "be", "possible", "by", "adding", "more", "hardware", ".", "Taking", "Advantage", "of", "Convolutionality", ":", "In", "our", "language", "models", ",", "we", "apply", "the", "same", "MoE", "to", "each", "time", "step", "of", "the", "previous", "layer", ".", "If", "we", "wait", "for", "the", "previous", "layer", "to", "finish", ",", "we", "can", "apply", "the", "MoE", "to", "all", "the", "time", "steps", "together", "as", "one", "big", "batch", ".", "Doing", "so", "increases", "the", "size", "of", "the", "input", "batch", "to", "the", "MoE", "layer", "by", "a", "factor", "of", "the", "number", "of", "unrolled", "time", "steps", ".", "Increasing", "Batch", "Size", "for", "a", "Recurrent", "MoE", ":", "We", "suspect", "that", "even", "more", "powerful", "models", "may", "involve", "applying", "a", "MoE", "recurrently", ".", "For", "example", ",", "the", "weight", "matrices", "of", "a", "LSTM", "or", "other", "RNN", "could", "be", "replaced", "by", "a", "MoE.", "Sadly", ",", "such", "models", "break", "the", "convolutional", "trick", "from", "the", "last", "paragraph", ",", "since", "the", "input", "to", "the", "MoE", "at", "one", "timestep", "depends", "on", "the", "output", "of", "the", "MoE", "at", "the", "previous", "timestep", ".", "[", "reference", "]", "describe", "a", "technique", "for", "drastically", "reducing", "the", "number", "of", "stored", "activations", "in", "an", "unrolled", "RNN", ",", "at", "the", "cost", "of", "recomputing", "forward", "activations", ".", "This", "would", "allow", "for", "a", "large", "increase", "in", "batch", "size", ".", "section", ":", "NETWORK", "BANDWIDTH", "Another", "major", "performance", "concern", "in", "distributed", "computing", "is", "network", "bandwidth", ".", "Since", "the", "experts", "are", "stationary", "(", "see", "above", ")", "and", "the", "number", "of", "gating", "parameters", "is", "small", ",", "most", "of", "the", "communication", "involves", "sending", "the", "inputs", "and", "outputs", "of", "the", "experts", "across", "the", "network", ".", "To", "maintain", "computational", "efficiency", ",", "the", "ratio", "of", "an", "expert", "'s", "computation", "to", "the", "size", "of", "its", "input", "and", "output", "must", "exceed", "the", "ratio", "of", "computational", "to", "network", "capacity", "of", "the", "computing", "device", ".", "For", "GPUs", ",", "this", "may", "be", "thousands", "to", "one", ".", "In", "our", "experiments", ",", "we", "use", "experts", "with", "one", "hidden", "layer", "containing", "thousands", "of", "RELU", "-", "activated", "units", ".", "Since", "the", "weight", "matrices", "in", "the", "expert", "have", "sizes", "input_size×hidden_size", "and", "hidden_size", "×", "output_size", ",", "the", "ratio", "of", "computation", "to", "input", "and", "output", "is", "equal", "to", "the", "size", "of", "the", "hidden", "layer", ".", "Conveniently", ",", "we", "can", "increase", "computational", "efficiency", "simply", "by", "using", "a", "larger", "hidden", "layer", ",", "or", "more", "hidden", "layers", ".", "section", ":", "BALANCING", "EXPERT", "UTILIZATION", "We", "have", "observed", "that", "the", "gating", "network", "tends", "to", "converge", "to", "a", "state", "where", "it", "always", "produces", "large", "weights", "for", "the", "same", "few", "experts", ".", "This", "imbalance", "is", "self", "-", "reinforcing", ",", "as", "the", "favored", "experts", "are", "trained", "more", "rapidly", "and", "thus", "are", "selected", "even", "more", "by", "the", "gating", "network", ".", "[", "reference", "]", "describe", "the", "same", "phenomenon", ",", "and", "use", "a", "hard", "constraint", "at", "the", "beginning", "of", "training", "to", "avoid", "this", "local", "minimum", ".", "[", "reference", "]", "include", "a", "soft", "constraint", "on", "the", "batch", "-", "wise", "average", "of", "each", "gate", ".", "[", "reference", "]", "We", "take", "a", "soft", "constraint", "approach", ".", "We", "define", "the", "importance", "of", "an", "expert", "relative", "to", "a", "batch", "of", "training", "examples", "to", "be", "the", "batchwise", "sum", "of", "the", "gate", "values", "for", "that", "expert", ".", "We", "define", "an", "additional", "loss", "L", "importance", ",", "which", "is", "added", "to", "the", "overall", "loss", "function", "for", "the", "model", ".", "This", "loss", "is", "equal", "to", "the", "square", "of", "the", "coefficient", "of", "variation", "of", "the", "set", "of", "importance", "values", ",", "multiplied", "by", "a", "hand", "-", "tuned", "scaling", "factor", "w", "importance", ".", "This", "additional", "loss", "encourages", "all", "experts", "to", "have", "equal", "importance", ".", "1", "[", "reference", "]", "also", "include", "two", "additional", "losses", ".", "One", "controls", "per", "-", "example", "sparsity", ",", "which", "we", "do", "not", "need", "since", "it", "is", "enforced", "by", "the", "fixed", "value", "of", "k.", "A", "third", "loss", "encourages", "diversity", "of", "gate", "values", ".", "In", "our", "experiments", ",", "we", "find", "that", "the", "gate", "values", "naturally", "diversify", "as", "the", "experts", "specialize", "(", "in", "a", "virtuous", "cycle", ")", ",", "and", "we", "do", "not", "need", "to", "enforce", "diversity", "of", "gate", "values", ".", "While", "this", "loss", "function", "can", "ensure", "equal", "importance", ",", "experts", "may", "still", "receive", "very", "different", "numbers", "of", "examples", ".", "For", "example", ",", "one", "expert", "may", "receive", "a", "few", "examples", "with", "large", "weights", ",", "and", "another", "may", "receive", "many", "examples", "with", "small", "weights", ".", "This", "can", "cause", "memory", "and", "performance", "problems", "on", "distributed", "hardware", ".", "To", "solve", "this", "problem", ",", "we", "introduce", "a", "second", "loss", "function", ",", "L", "load", ",", "which", "ensures", "balanced", "loads", ".", "Appendix", "A", "contains", "the", "definition", "of", "this", "function", ",", "along", "with", "experimental", "results", ".", "section", ":", "EXPERIMENTS", "section", ":", "1", "BILLION", "WORD", "LANGUAGE", "MODELING", "BENCHMARK", "Dataset", ":", "This", "dataset", ",", "introduced", "by", "[", "reference", "]", "consists", "of", "shuffled", "unique", "sentences", "from", "news", "articles", ",", "totaling", "approximately", "829", "million", "words", ",", "with", "a", "vocabulary", "of", "793", ",", "471", "words", ".", "Previous", "State", "-", "of", "-", "the", "-", "Art", ":", "The", "best", "previously", "published", "results", "[", "reference", "]", "use", "models", "consisting", "of", "one", "or", "more", "stacked", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "layers", "[", "reference", "][", "reference", "]", ".", "The", "number", "of", "parameters", "in", "the", "LSTM", "layers", "of", "these", "models", "vary", "from", "2", "million", "to", "151", "million", ".", "Quality", "increases", "greatly", "with", "parameter", "count", ",", "as", "do", "computational", "costs", ".", "Results", "for", "these", "models", "form", "the", "top", "line", "of", "Figure", "2", "-", "right", ".", "MoE", "Models", ":", "Our", "models", "consist", "of", "two", "stacked", "LSTM", "layers", "with", "a", "MoE", "layer", "between", "them", "(", "see", "Figure", "1", ")", ".", "We", "vary", "the", "sizes", "of", "the", "layers", "and", "the", "number", "of", "experts", ".", "For", "full", "details", "on", "model", "architecture", ",", "training", "regimen", ",", "additional", "baselines", "and", "results", ",", "see", "Appendix", "C.", "Low", "Computation", ",", "Varied", "Capacity", ":", "To", "investigate", "the", "effects", "of", "adding", "capacity", ",", "we", "trained", "a", "series", "of", "MoE", "models", "all", "with", "roughly", "equal", "computational", "costs", ":", "about", "8", "million", "multiply", "-", "andadds", "per", "training", "example", "per", "timestep", "in", "the", "forwards", "pass", ",", "excluding", "the", "softmax", "layer", ".", "We", "call", "this", "metric", "(", "ops", "/", "timestep", ")", ".", "We", "trained", "models", "with", "flat", "MoEs", "containing", "4", ",", "32", ",", "and", "256", "experts", ",", "and", "models", "with", "hierarchical", "MoEs", "containing", "256", ",", "1024", ",", "and", "4096", "experts", ".", "Each", "expert", "had", "about", "1", "million", "parameters", ".", "For", "all", "the", "MoE", "layers", ",", "4", "experts", "were", "active", "per", "input", ".", "The", "results", "of", "these", "models", "are", "shown", "in", "Figure", "2", "-", "left", ".", "The", "model", "with", "4", "always", "-", "active", "experts", "performed", "(", "unsurprisingly", ")", "similarly", "to", "the", "computationally", "-", "matched", "baseline", "models", ",", "while", "the", "largest", "of", "the", "models", "(", "4096", "experts", ")", "achieved", "an", "impressive", "24", "%", "lower", "perplexity", "on", "the", "test", "set", ".", "Varied", "Computation", ",", "High", "Capacity", ":", "In", "addition", "to", "the", "largest", "model", "from", "the", "previous", "section", ",", "we", "trained", "two", "more", "MoE", "models", "with", "similarly", "high", "capacity", "(", "4", "billion", "parameters", ")", ",", "but", "higher", "computation", "budgets", ".", "These", "models", "had", "larger", "LSTMs", ",", "and", "fewer", "but", "larger", "and", "experts", ".", "Details", "can", "be", "found", "in", "Appendix", "C.2", ".", "Results", "of", "these", "three", "models", "form", "the", "bottom", "line", "of", "Figure", "2", "-", "right", ".", "Table", "1", "compares", "the", "results", "of", "these", "models", "to", "the", "best", "previously", "-", "published", "result", "on", "this", "dataset", ".", "Even", "the", "fastest", "of", "these", "models", "beats", "the", "best", "published", "result", "(", "when", "controlling", "for", "the", "number", "of", "training", "epochs", ")", ",", "despite", "requiring", "only", "6", "%", "of", "the", "computation", ".", "Computational", "Efficiency", ":", "We", "trained", "our", "models", "using", "TensorFlow", "[", "reference", "]", "on", "clusters", "containing", "16", "-", "32", "Tesla", "K40", "GPUs", ".", "For", "each", "of", "our", "models", ",", "we", "determine", "computational", "efficiency", "in", "TFLOPS", "/", "GPU", "by", "dividing", "the", "number", "of", "floating", "point", "operations", "required", "to", "process", "one", "training", "batch", "by", "the", "observed", "step", "time", "and", "the", "number", "of", "GPUs", "in", "the", "cluster", ".", "The", "operation", "counts", "used", "here", "are", "higher", "than", "the", "ones", "we", "report", "in", "our", "ops", "/", "timestep", "numbers", "in", "that", "we", "include", "the", "backwards", "pass", ",", "we", "include", "the", "importance", "-", "sampling", "-", "based", "training", "of", "the", "softmax", "layer", ",", "and", "we", "count", "a", "multiply", "-", "and", "-", "add", "as", "two", "separate", "operations", ".", "For", "all", "of", "our", "MoE", "models", ",", "the", "floating", "point", "operations", "involved", "in", "the", "experts", "represent", "between", "37", "%", "and", "46", "%", "of", "the", "total", ".", "For", "our", "baseline", "models", "wtih", "no", "MoE", ",", "observed", "computational", "efficiency", "ranged", "from", "1.07", "-", "1.29", "TFLOPS", "/", "GPU", ".", "For", "our", "low", "-", "computation", "MoE", "models", ",", "computation", "efficiency", "ranged", "from", "0.74", "-", "0.90", "TFLOPS", "/", "GPU", ",", "except", "for", "the", "4", "-", "expert", "model", "which", "did", "not", "make", "full", "use", "of", "the", "available", "parallelism", ".", "Our", "highest", "-", "computation", "MoE", "model", "was", "more", "efficient", "at", "1.56", "TFLOPS", "/", "GPU", ",", "likely", "due", "to", "the", "larger", "matrices", ".", "These", "numbers", "represent", "a", "significant", "fraction", "of", "the", "theoretical", "maximum", "of", "4.29", "TFLOPS", "/", "GPU", "claimed", "by", "NVIDIA", ".", "Detailed", "results", "are", "in", "Appendix", "C", ",", "Table", "7", ".", "section", ":", "100", "BILLION", "WORD", "GOOGLE", "NEWS", "CORPUS", "Figure", "3", ":", "Language", "modeling", "on", "a", "100", "billion", "word", "corpus", ".", "Models", "have", "similar", "computational", "budgets", "(", "8", "million", "ops", "/", "timestep", ")", ".", "On", "the", "1", "-", "billion", "-", "word", "corpus", ",", "adding", "additional", "capacity", "seems", "to", "produce", "diminishing", "returns", "as", "the", "number", "of", "parameters", "in", "the", "MoE", "layer", "exceeds", "1", "billion", ",", "as", "can", "be", "seen", "in", "Figure", "2", "-", "left", ".", "We", "hypothesized", "that", "for", "a", "larger", "training", "set", ",", "even", "higher", "capacities", "would", "produce", "significant", "quality", "improvements", ".", "We", "constructed", "a", "similar", "training", "set", "consisting", "of", "shuffled", "unique", "sentences", "from", "Google", "'s", "internal", "news", "corpus", ",", "totalling", "roughly", "100", "billion", "words", ".", "Similarly", "to", "the", "previous", "section", ",", "we", "tested", "a", "series", "of", "models", "with", "similar", "computational", "costs", "of", "about", "8", "million", "ops", "/", "timestep", ".", "In", "addition", "to", "a", "baseline", "LSTM", "model", ",", "we", "trained", "models", "augmented", "with", "MoE", "layers", "containing", "[", "reference", "]", "experts", ".", "This", "corresponds", "to", "up", "to", "137", "billion", "parameters", "in", "the", "MoE", "layer", ".", "Details", "on", "architecture", ",", "training", ",", "and", "results", "are", "given", "in", "Appendix", "D.", "Results", ":", "Figure", "3", "shows", "test", "perplexity", "as", "a", "function", "of", "capacity", "after", "training", "on", "10", "billion", "words", "(", "top", "line", ")", "and", "100", "billion", "words", "(", "bottom", "line", ")", ".", "When", "training", "over", "the", "full", "100", "billion", "words", ",", "test", "perplexity", "improves", "significantly", "up", "to", "65536", "experts", "(", "68", "billion", "parameters", ")", ",", "dropping", "39", "%", "lower", "than", "the", "computationally", "matched", "baseline", ",", "but", "degrades", "at", "131072", "experts", ",", "possibly", "a", "result", "of", "too", "much", "sparsity", ".", "The", "widening", "gap", "between", "the", "two", "lines", "demonstrates", "(", "unsurprisingly", ")", "that", "increased", "model", "capacity", "helps", "more", "on", "larger", "training", "sets", ".", "Even", "at", "65536", "experts", "(", "99.994", "%", "layer", "sparsity", ")", ",", "computational", "efficiency", "for", "the", "model", "stays", "at", "a", "respectable", "0.72", "TFLOPS", "/", "GPU", ".", "section", ":", "MACHINE", "TRANSLATION", "(", "SINGLE", "LANGUAGE", "PAIR", ")", "Model", "Architecture", ":", "Our", "model", "was", "a", "modified", "version", "of", "the", "GNMT", "model", "described", "in", "[", "reference", "]", ".", "To", "reduce", "computation", ",", "we", "decreased", "the", "number", "of", "LSTM", "layers", "in", "the", "encoder", "and", "decoder", "from", "9", "and", "8", "to", "3", "and", "2", "respectively", ".", "We", "inserted", "MoE", "layers", "in", "both", "the", "encoder", "(", "between", "layers", "2", "and", "3", ")", "and", "the", "decoder", "(", "between", "layers", "1", "and", "2", ")", ".", "Each", "MoE", "layer", "contained", "up", "to", "2048", "experts", "each", "with", "about", "two", "million", "parameters", ",", "adding", "a", "total", "of", "about", "8", "billion", "parameters", "to", "the", "models", ".", "Further", "details", "on", "model", "architecture", ",", "testing", "procedure", "and", "results", "can", "be", "found", "in", "Appendix", "E.", "Datasets", ":", "We", "benchmarked", "our", "method", "on", "the", "WMT'14", "En→Fr", "and", "En→De", "corpora", ",", "whose", "training", "sets", "have", "36", "M", "sentence", "pairs", "and", "5", "M", "sentence", "pairs", ",", "respectively", ".", "The", "experimental", "protocols", "were", "also", "similar", "to", "those", "in", "[", "reference", "]", ":", "newstest2014", "was", "used", "as", "the", "test", "set", "to", "compare", "against", "previous", "work", "[", "reference", "][", "reference", "][", "reference", "]", ",", "while", "the", "combination", "of", "newstest2012", "and", "newstest2013", "was", "used", "as", "the", "development", "set", ".", "We", "also", "tested", "the", "same", "model", "on", "a", "Google", "'s", "Production", "English", "to", "French", "data", ".", "[", "reference", "]", "2.79", "39.22", "214", "M", "278", "M", "6", "days", "/", "96", "k80s", "GNMT", "+", "RL", "[", "reference", "]", "2.96", "39.92", "214", "M", "278", "M", "6", "days", "/", "96", "k80s", "PBMT", "[", "reference", "]", "37.0", "LSTM", "(", "6", "-", "layer", ")", "[", "reference", "]", "31.5", "LSTM", "(", "6", "-", "layer", "+", "PosUnk", ")", "[", "reference", "]", "33.1", "DeepAtt", "[", "reference", "]", "37.7", "DeepAtt", "+", "PosUnk", "[", "reference", "]", "39.2", "[", "reference", "]", "5.25", "24.91", "214", "M", "278", "M", "1", "day", "/", "96", "k80s", "GNMT", "+", "RL", "[", "reference", "]", "8.08", "24.66", "214", "M", "278", "M", "1", "day", "/", "96", "k80s", "PBMT", "[", "reference", "]", "20.7", "DeepAtt", "[", "reference", "]", "20.6", "Results", ":", "Tables", "2", ",", "3", ",", "and", "4", "show", "the", "results", "of", "our", "largest", "models", ",", "compared", "with", "published", "results", ".", "Our", "approach", "achieved", "BLEU", "scores", "of", "40.56", "and", "26.03", "on", "the", "WMT'14", "En→Fr", "and", "En→De", "benchmarks", ".", "As", "our", "models", "did", "not", "use", "RL", "refinement", ",", "these", "results", "constitute", "significant", "gains", "of", "1.34", "and", "1.12", "BLEU", "score", "on", "top", "of", "the", "strong", "baselines", "in", "[", "reference", "]", ".", "The", "perplexity", "scores", "are", "also", "better", ".", "2", "On", "the", "Google", "Production", "dataset", ",", "our", "model", "achieved", "1.01", "higher", "test", "BLEU", "score", "even", "after", "training", "for", "only", "one", "sixth", "of", "the", "time", ".", "section", ":", "MULTILINGUAL", "MACHINE", "TRANSLATION", "Dataset", ":", "(", "Johnson", "et", "al", ".", ",", "2016", ")", "train", "a", "single", "GNMT", "[", "reference", "]", ")", "model", "on", "a", "very", "large", "combined", "dataset", "of", "twelve", "language", "pairs", ".", "Results", "are", "somewhat", "worse", "than", "those", "for", "12", "separately", "trained", "single", "-", "pair", "GNMT", "models", ".", "This", "is", "not", "surprising", ",", "given", "that", "the", "twelve", "models", "have", "12", "times", "the", "capacity", "and", "twelve", "times", "the", "aggregate", "training", "of", "the", "one", "model", ".", "We", "repeat", "this", "experiment", "with", "a", "single", "MoE", "-", "augmented", "model", ".", "See", "Appendix", "E", "for", "details", "on", "model", "architecture", ".", "We", "train", "our", "model", "on", "the", "same", "dataset", "as", "[", "reference", "]", "and", "process", "the", "same", "number", "of", "training", "examples", "(", "about", "3", "billion", "sentence", "pairs", ")", ".", "Our", "training", "time", "was", "shorter", "due", "to", "the", "lower", "computational", "budget", "of", "our", "model", ".", "section", ":", "Results", ":", "Results", "for", "the", "single", "-", "pair", "GNMT", "models", ",", "the", "multilingual", "GNMT", "model", "and", "the", "multilingual", "MoE", "model", "are", "given", "in", "Table", "5", ".", "The", "MoE", "model", "achieves", "19", "%", "lower", "perplexity", "on", "the", "dev", "set", "than", "the", "multilingual", "GNMT", "model", ".", "On", "BLEU", "score", ",", "the", "MoE", "model", "significantly", "beats", "the", "multilingual", "GNMT", "model", "on", "11", "of", "the", "12", "language", "pairs", "(", "by", "as", "much", "as", "5.84", "points", ")", ",", "and", "even", "beats", "the", "monolingual", "GNMT", "models", "on", "8", "of", "12", "language", "pairs", ".", "The", "poor", "performance", "on", "English", "→", "Korean", "seems", "to", "be", "a", "result", "of", "severe", "overtraining", ",", "as", "for", "the", "rarer", "language", "pairs", "a", "small", "number", "of", "real", "examples", "were", "highly", "oversampled", "in", "the", "training", "corpus", ".", "section", ":", "CONCLUSION", "This", "work", "is", "the", "first", "to", "demonstrate", "major", "wins", "from", "conditional", "computation", "in", "deep", "networks", ".", "We", "carefully", "identified", "the", "design", "considerations", "and", "challenges", "of", "conditional", "computing", "and", "addressed", "them", "with", "a", "combination", "of", "algorithmic", "and", "engineering", "solutions", ".", "While", "we", "focused", "on", "text", ",", "conditional", "computation", "may", "help", "in", "other", "domains", "as", "well", ",", "provided", "sufficiently", "large", "training", "sets", ".", "We", "look", "forward", "to", "seeing", "many", "novel", "implementations", "and", "applications", "of", "conditional", "computation", "in", "the", "years", "to", "come", ".", "section", ":", "ACKNOWLEDGMENTS", "We", "would", "like", "to", "thank", "all", "of", "the", "members", "of", "the", "Google", "Brain", "and", "Google", "Translate", "teams", "who", "helped", "us", "with", "this", "project", ",", "in", "particular", "Zhifeng", "Chen", ",", "Yonghui", "Wu", ",", "and", "Melvin", "Johnson", ".", "Thanks", "also", "to", "our", "anonymous", "ICLR", "reviewers", "for", "the", "helpful", "suggestions", "on", "making", "this", "paper", "better", ".", "section", ":", "APPENDICES", "A", "LOAD", "-", "BALANCING", "LOSS", "As", "discussed", "in", "section", "4", ",", "for", "load", "-", "balancing", "purposes", ",", "we", "want", "to", "define", "an", "additional", "loss", "function", "to", "encourage", "experts", "to", "receive", "roughly", "equal", "numbers", "of", "training", "examples", ".", "Unfortunately", ",", "the", "number", "of", "examples", "received", "by", "an", "expert", "is", "a", "discrete", "quantity", ",", "so", "it", "can", "not", "be", "used", "in", "backpropagation", ".", "Instead", ",", "we", "define", "a", "smooth", "estimator", "Load", "(", "X", ")", "of", "the", "number", "of", "examples", "assigned", "to", "each", "expert", "for", "a", "batch", "X", "of", "inputs", ".", "The", "smoothness", "allows", "us", "to", "back", "-", "propagate", "gradients", "through", "the", "estimator", ".", "This", "is", "the", "purpose", "of", "the", "noise", "term", "in", "the", "gating", "function", ".", "We", "define", "P", "(", "x", ",", "i", ")", "as", "the", "probability", "that", "G", "(", "x", ")", "i", "is", "nonzero", ",", "given", "a", "new", "random", "choice", "of", "noise", "on", "element", "i", ",", "but", "keeping", "the", "already", "-", "sampled", "choices", "of", "noise", "on", "the", "other", "elements", ".", "To", "compute", "P", "(", "x", ",", "i", ")", ",", "we", "note", "that", "the", "G", "(", "x", ")", "i", "is", "nonzero", "if", "and", "only", "if", "H", "(", "x", ")", "i", "is", "greater", "than", "the", "k", "th", "-", "greatest", "element", "of", "H", "(", "x", ")", "excluding", "itself", ".", "The", "probability", "works", "out", "to", "be", ":", "Where", "kth_excluding", "(", "v", ",", "k", ",", "i", ")", "means", "the", "kth", "highest", "component", "of", "v", ",", "excluding", "component", "i.", "Simplifying", ",", "we", "get", ":", "Where", "Φ", "is", "the", "CDF", "of", "the", "standard", "normal", "distribution", ".", "We", "can", "now", "define", "the", "load", "loss", "to", "be", "the", "square", "of", "the", "coefficient", "of", "variation", "of", "the", "load", "vector", ",", "multiplied", "by", "a", "hand", "-", "tuned", "scaling", "factor", "w", "load", ".", "Initial", "Load", "Imbalance", ":", "To", "avoid", "out", "-", "of", "-", "memory", "errors", ",", "we", "need", "to", "initialize", "the", "network", "in", "a", "state", "of", "approximately", "equal", "expert", "load", "(", "since", "the", "soft", "constraints", "need", "some", "time", "to", "work", ")", ".", "To", "accomplish", "this", ",", "we", "initialize", "the", "matrices", "W", "g", "and", "W", "noise", "to", "all", "zeros", ",", "which", "yields", "no", "signal", "and", "some", "noise", ".", "section", ":", "Experiments", ":", "We", "trained", "a", "set", "of", "models", "with", "identical", "architecture", "(", "the", "MoE", "-", "256", "model", "described", "in", "Appendix", "C", ")", ",", "using", "different", "values", "of", "w", "importance", "and", "w", "load", ".", "We", "trained", "each", "model", "for", "10", "epochs", ",", "then", "measured", "perplexity", "on", "the", "test", "set", ".", "We", "also", "measured", "the", "coefficients", "of", "variation", "in", "Importance", "and", "Load", ",", "as", "well", "as", "ratio", "of", "the", "load", "on", "the", "most", "overloaded", "expert", "to", "the", "average", "load", ".", "This", "last", "value", "is", "significant", "for", "load", "balancing", "purposes", "on", "distributed", "hardware", ".", "All", "of", "these", "metrics", "were", "averaged", "over", "several", "training", "batches", ".", "Results", ":", "Results", "are", "reported", "in", "Table", "6", ".", "All", "the", "combinations", "containing", "at", "least", "one", "the", "two", "losses", "led", "to", "very", "similar", "model", "quality", ",", "where", "having", "no", "loss", "was", "much", "worse", ".", "Models", "with", "higher", "values", "of", "w", "load", "had", "lower", "loads", "on", "the", "most", "overloaded", "expert", ".", "section", ":", "B", "HIERACHICAL", "MIXTURE", "OF", "EXPERTS", "If", "the", "number", "of", "experts", "is", "very", "large", ",", "we", "can", "reduce", "the", "branching", "factor", "by", "using", "a", "two", "-", "level", "hierarchical", "MoE.", "In", "a", "hierarchical", "MoE", ",", "a", "primary", "gating", "network", "chooses", "a", "sparse", "weighted", "combination", "of", "\"", "experts", "\"", ",", "each", "of", "which", "is", "itself", "a", "secondary", "mixture", "-", "of", "-", "experts", "with", "its", "own", "gating", "network", ".", "3", "If", "the", "hierarchical", "MoE", "consists", "of", "a", "groups", "of", "b", "experts", "each", ",", "we", "denote", "the", "primary", "gating", "network", "by", "G", "primary", ",", "the", "secondary", "gating", "networks", "by", "(", "G", "1", ",", "G", "2", "..", "G", "a", ")", ",", "and", "the", "expert", "networks", "by", "(", "E", "0", ",", "0", ",", "E", "0", ",", "1", "..", "E", "a", ",", "b", ")", ".", "The", "output", "of", "the", "MoE", "is", "given", "by", ":", "Our", "metrics", "of", "expert", "utilization", "change", "to", "the", "following", ":", "Load", "primary", "and", "Load", "i", "deonte", "the", "Load", "functions", "for", "the", "primary", "gating", "network", "and", "i", "th", "secondary", "gating", "network", "respectively", ".", "X", "(", "i", ")", "denotes", "the", "subset", "of", "X", "for", "which", "G", "primary", "(", "x", ")", "i", ">", "0", ".", "It", "would", "seem", "simpler", "to", "let", "Load", "H", "(", "X", ")", "i", ",", "j", "=", "Load", "i", "(", "X", "i", ")", "j", ",", "but", "this", "would", "not", "have", "a", "gradient", "with", "respect", "to", "the", "primary", "gating", "network", ",", "so", "we", "use", "the", "formulation", "above", ".", "section", ":", "C", "1", "BILLION", "WORD", "LANGUAGE", "MODELING", "BENCHMARK", "-", "EXPERIMENTAL", "DETAILS", "C.1", "8", "-", "MILLION", "-", "OPERATIONS", "-", "PER", "-", "TIMESTEP", "MODELS", "Model", "Architecture", ":", "Our", "model", "consists", "of", "five", "layers", ":", "a", "word", "embedding", "layer", ",", "a", "recurrent", "Long", "Short", "-", "Term", "Memory", "(", "LSTM", ")", "layer", "[", "reference", "][", "reference", "]", ",", "a", "MoE", "layer", ",", "a", "second", "LSTM", "layer", ",", "and", "a", "softmax", "layer", ".", "The", "dimensionality", "of", "the", "embedding", "layer", ",", "the", "number", "of", "units", "in", "each", "LSTM", "layer", ",", "and", "the", "input", "and", "output", "dimensionality", "of", "the", "MoE", "layer", "are", "all", "equal", "to", "512", ".", "For", "every", "layer", "other", "than", "the", "softmax", ",", "we", "apply", "drouput", "[", "reference", "]", "to", "the", "layer", "output", ",", "dropping", "each", "activation", "with", "probability", "DropP", "rob", ",", "otherwise", "dividing", "by", "(", "1", "−", "DropP", "rob", ")", ".", "After", "dropout", ",", "the", "output", "of", "the", "previous", "layer", "is", "added", "to", "the", "layer", "output", ".", "This", "residual", "connection", "encourages", "gradient", "flow", "[", "reference", "]", ".", "For", "the", "hierarchical", "MoE", "layers", ",", "the", "first", "level", "branching", "factor", "was", "16", ",", "corresponding", "to", "the", "number", "of", "GPUs", "in", "our", "cluster", ".", "We", "use", "Noisy", "-", "Top", "-", "K", "Gating", "(", "see", "Section", "2.1", ")", "with", "k", "=", "4", "for", "the", "ordinary", "MoE", "layers", "and", "k", "=", "2", "at", "each", "level", "of", "the", "hierarchical", "MoE", "layers", ".", "Thus", ",", "each", "example", "is", "processed", "by", "exactly", "4", "experts", "for", "a", "total", "of", "4", "M", "ops", "/", "timestep", ".", "The", "two", "LSTM", "layers", "contribute", "2", "M", "ops", "/", "timestep", "each", "for", "the", "desired", "total", "of", "8M.", "section", ":", "Computationally", "-", "Matched", "Baselines", ":", "The", "MoE", "-", "4", "model", "does", "not", "employ", "sparsity", ",", "since", "all", "4", "experts", "are", "always", "used", ".", "In", "addition", ",", "we", "trained", "four", "more", "computationally", "-", "matched", "baseline", "models", "with", "no", "sparsity", ":", "•", "MoE", "-", "1", "-", "Wide", ":", "The", "MoE", "layer", "consists", "of", "a", "single", "\"", "expert", "\"", "containing", "one", "ReLU", "-", "activated", "hidden", "layer", "of", "size", "4096", ".", "•", "MoE", "-", "1", "-", "Deep", ":", "The", "MoE", "layer", "consists", "of", "a", "single", "\"", "expert", "\"", "containing", "four", "ReLU", "-", "activated", "hidden", "layers", ",", "each", "with", "size", "1024", ".", "•", "4xLSTM", "-", "512", ":", "We", "replace", "the", "MoE", "layer", "with", "two", "additional", "512", "-", "unit", "LSTM", "layers", ".", "•", "LSTM", "-", "2048", "-", "512", ":", "The", "model", "contains", "one", "2048", "-", "unit", "LSTM", "layer", "(", "and", "no", "MoE", ")", ".", "The", "output", "of", "the", "LSTM", "is", "projected", "down", "to", "512", "dimensions", "[", "reference", "]", ".", "The", "next", "timestep", "of", "the", "LSTM", "receives", "the", "projected", "output", ".", "This", "is", "identical", "to", "one", "of", "the", "models", "published", "in", "[", "reference", "]", ".", "We", "re", "-", "ran", "it", "to", "account", "for", "differences", "in", "training", "regimen", ",", "and", "obtained", "results", "very", "similar", "to", "the", "published", "ones", ".", "Training", ":", "The", "models", "were", "trained", "on", "a", "cluster", "of", "16", "K40", "GPUs", "using", "the", "synchronous", "method", "described", "in", "Section", "3", ".", "Each", "batch", "consisted", "of", "a", "set", "of", "sentences", "totaling", "roughly", "300", ",", "000", "words", ".", "In", "the", "interest", "of", "time", ",", "we", "limited", "training", "to", "10", "epochs", ",", "(", "27", ",", "000", "steps", ")", ".", "Training", "took", "12", "-", "16", "hours", "for", "all", "models", ",", "except", "for", "MoE", "-", "4", ",", "which", "took", "18", "hours", "(", "since", "all", "the", "expert", "computation", "was", "performed", "on", "only", "4", "of", "16", "GPUs", ")", ".", "We", "used", "the", "Adam", "optimizer", "[", "reference", "]", ".", "The", "base", "learning", "rate", "was", "increased", "linearly", "for", "the", "first", "1000", "training", "steps", ",", "and", "decreased", "after", "that", "so", "as", "to", "be", "proportional", "to", "the", "inverse", "square", "root", "of", "the", "step", "number", ".", "The", "Softmax", "output", "layer", "was", "trained", "efficiently", "using", "importance", "sampling", "similarly", "to", "the", "models", "in", "[", "reference", "]", ".", "For", "each", "model", ",", "we", "performed", "a", "hyper", "-", "parmeter", "search", "to", "find", "the", "best", "dropout", "probability", ",", "in", "increments", "of", "0.1", ".", "To", "ensure", "balanced", "expert", "utilization", "we", "set", "w", "importance", "=", "0.1", "and", "w", "load", "=", "0.1", ",", "as", "described", "in", "Section", "4", "and", "Appendix", "A.", "Results", ":", "We", "evaluate", "our", "model", "using", "perplexity", "on", "the", "holdout", "dataset", ",", "used", "by", "[", "reference", "][", "reference", "]", ".", "We", "follow", "the", "standard", "procedure", "and", "sum", "over", "all", "the", "words", "including", "the", "end", "of", "sentence", "symbol", ".", "Results", "are", "reported", "in", "Table", "7", ".", "For", "each", "model", ",", "we", "report", "the", "test", "perplexity", ",", "the", "computational", "budget", ",", "the", "parameter", "counts", ",", "the", "value", "of", "DropP", "rob", ",", "and", "the", "computational", "efficiency", ".", "section", ":", "C.2", "MORE", "EXPENSIVE", "MODELS", "We", "ran", "two", "additional", "models", "(", "MoE", "-", "34", "M", "and", "MoE", "-", "143", "M", ")", "to", "investigate", "the", "effects", "of", "adding", "more", "computation", "in", "the", "presence", "of", "a", "large", "MoE", "layer", ".", "These", "models", "have", "computation", "budgets", "of", "34", "M", "and", "143", "M", "ops", "/", "timestep", ".", "Similar", "to", "the", "models", "above", ",", "these", "models", "use", "a", "MoE", "layer", "between", "two", "LSTM", "layers", ".", "The", "dimensionality", "of", "the", "embedding", "layer", ",", "and", "the", "input", "and", "output", "dimensionality", "of", "the", "MoE", "layer", "are", "set", "to", "1024", "instead", "of", "512", ".", "For", "MoE", "-", "34", "M", ",", "the", "LSTM", "layers", "have", "1024", "units", ".", "For", "MoE", "-", "143", "M", ",", "the", "LSTM", "layers", "have", "4096", "units", "and", "an", "output", "projection", "of", "size", "1024", "[", "reference", "]", ".", "MoE", "-", "34", "M", "uses", "a", "hierarchical", "MoE", "layer", "with", "1024", "experts", ",", "each", "with", "a", "hidden", "layer", "of", "size", "2048", ".", "MoE", "-", "143", "M", "uses", "a", "hierarchical", "MoE", "layer", "with", "256", "experts", ",", "each", "with", "a", "hidden", "layer", "of", "size", "8192", ".", "Both", "models", "have", "4B", "parameters", "in", "the", "MoE", "layers", ".", "We", "searched", "for", "the", "best", "DropP", "rob", "for", "each", "model", ",", "and", "trained", "each", "model", "for", "10", "epochs", ".", "The", "two", "models", "achieved", "test", "perplexity", "of", "31.3", "and", "28.0", "respectively", ",", "showing", "that", "even", "in", "the", "presence", "of", "a", "large", "MoE", ",", "more", "computation", "is", "still", "useful", ".", "Results", "are", "reported", "at", "the", "bottom", "of", "Table", "7", ".", "The", "larger", "of", "the", "two", "models", "has", "a", "similar", "computational", "budget", "to", "the", "best", "published", "model", "from", "the", "literature", ",", "and", "training", "times", "are", "similar", ".", "Comparing", "after", "10", "epochs", ",", "our", "model", "has", "a", "lower", "test", "perplexity", "by", "18", "%", ".", "section", ":", "D", "100", "BILLION", "WORD", "GOOGLE", "NEWS", "CORPUS", "-", "EXPERIMENTAL", "DETAILS", "Model", "Architecture", ":", "The", "models", "are", "similar", "in", "structure", "to", "the", "8", "-", "million", "-", "operations", "-", "per", "-", "timestep", "models", "described", "in", "the", "previous", "section", ".", "We", "vary", "the", "number", "of", "experts", "between", "models", ",", "using", "an", "ordinary", "MoE", "layer", "with", "32", "experts", "and", "hierarchical", "MoE", "layers", "with", "256", ",", "1024", ",", "4096", ",", "16384", ",", "65536", "and", "131072", "experts", ".", "For", "the", "hierarchical", "MoE", "layers", ",", "the", "first", "level", "branching", "factors", "are", "[", "reference", "]", "Training", ":", "Models", "are", "trained", "on", "a", "cluster", "of", "32", "Tesla", "K40", "GPUs", ",", "except", "for", "the", "last", "two", "models", ",", "which", "are", "trained", "on", "clusters", "of", "64", "and", "128", "GPUs", "so", "as", "to", "have", "enough", "memory", "for", "all", "the", "parameters", ".", "For", "all", "models", ",", "training", "batch", "sizes", "are", "approximately", "2.5", "million", "words", ".", "Models", "are", "trained", "once", "-", "through", "over", "about", "100", "billion", "words", ".", "We", "implement", "several", "memory", "optimizations", "in", "order", "to", "fit", "up", "to", "1", "billion", "parameters", "per", "GPU", ".", "First", ",", "we", "do", "not", "store", "the", "activations", "of", "the", "hidden", "layers", "of", "the", "experts", ",", "but", "instead", "recompute", "them", "on", "the", "backwards", "pass", ".", "Secondly", ",", "we", "modify", "the", "optimizer", "on", "the", "expert", "parameters", "to", "require", "less", "auxiliary", "storage", ":", "The", "Adam", "optimizer", "[", "reference", "]", "keeps", "first", "and", "second", "moment", "estimates", "of", "the", "perparameter", "gradients", ".", "This", "triples", "the", "required", "memory", ".", "To", "avoid", "keeping", "a", "first", "-", "moment", "estimator", ",", "we", "set", "β", "1", "=", "0", ".", "To", "reduce", "the", "size", "of", "the", "second", "moment", "estimator", ",", "we", "replace", "it", "with", "a", "factored", "approximation", ".", "For", "a", "matrix", "of", "parameters", ",", "instead", "of", "maintaining", "a", "full", "matrix", "of", "second", "-", "moment", "estimators", ",", "we", "maintain", "vectors", "of", "row", "-", "wise", "and", "column", "-", "wise", "averages", "of", "that", "matrix", ".", "At", "each", "step", ",", "the", "matrix", "of", "estimators", "is", "taken", "to", "be", "the", "outer", "product", "of", "those", "two", "vectors", "divided", "by", "the", "mean", "of", "either", "one", ".", "This", "technique", "could", "similarly", "be", "applied", "to", "Adagrad", "[", "reference", "]", ".", "Results", ":", "We", "evaluate", "our", "model", "using", "perplexity", "on", "a", "holdout", "dataset", ".", "Results", "are", "reported", "in", "Table", "8", ".", "Perplexity", "after", "100", "billion", "training", "words", "is", "39", "%", "lower", "for", "the", "68", "-", "billion", "-", "parameter", "MoE", "model", "than", "for", "the", "baseline", "model", ".", "It", "is", "notable", "that", "the", "measured", "computational", "efficiency", "of", "the", "largest", "model", "(", "0.30", "TFLOPS", "/", "GPU", ")", "is", "very", "low", "compared", "to", "the", "other", "models", ".", "This", "is", "likely", "a", "result", "of", "the", "fact", "that", ",", "for", "purposes", "of", "comparison", "to", "the", "other", "models", ",", "we", "did", "not", "increase", "the", "training", "batch", "size", "proportionally", "to", "the", "number", "of", "GPUs", ".", "For", "comparison", ",", "we", "include", "results", "for", "a", "computationally", "matched", "baseline", "model", "consisting", "of", "4", "LSTMs", ",", "and", "for", "an", "unpruned", "5", "-", "gram", "model", "with", "Kneser", "-", "Ney", "smoothing", "[", "reference", "]", ".", "section", ":", "E", "MACHINE", "TRANSLATION", "-", "EXPERIMENTAL", "DETAILS", "Model", "Architecture", "for", "Single", "Language", "Pair", "MoE", "Models", ":", "Our", "model", "is", "a", "modified", "version", "of", "the", "GNMT", "model", "described", "in", "[", "reference", "]", ".", "To", "reduce", "computation", ",", "we", "decrease", "the", "number", "of", "LSTM", "layers", "in", "the", "encoder", "and", "decoder", "from", "9", "and", "8", "to", "3", "and", "2", "respectively", ".", "We", "insert", "MoE", "layers", "in", "both", "the", "encoder", "(", "between", "layers", "2", "and", "3", ")", "and", "the", "decoder", "(", "between", "layers", "1", "and", "2", ")", ".", "We", "use", "an", "attention", "mechanism", "between", "the", "encoder", "and", "decoder", ",", "with", "the", "first", "decoder", "LSTM", "receiving", "output", "from", "and", "providing", "input", "for", "the", "attention", "5", ".", "All", "of", "the", "layers", "in", "our", "model", "have", "input", "and", "output", "dimensionality", "of", "512", ".", "Our", "LSTM", "layers", "have", "2048", "hidden", "units", ",", "with", "a", "512", "-", "dimensional", "output", "projection", ".", "We", "add", "residual", "connections", "around", "all", "LSTM", "and", "MoE", "layers", "to", "encourage", "gradient", "flow", "[", "reference", "]", ".", "Similar", "to", "GNMT", ",", "to", "effectively", "deal", "with", "rare", "words", ",", "we", "used", "subword", "units", "(", "also", "known", "as", "\"", "wordpieces", "\"", ")", "(", "Schuster", "&", "Nakajima", ",", "2012", ")", "for", "inputs", "and", "outputs", "in", "our", "system", ".", "We", "use", "a", "shared", "source", "and", "target", "vocabulary", "of", "32", "K", "wordpieces", ".", "We", "also", "used", "the", "same", "beam", "search", "technique", "as", "proposed", "in", "[", "reference", "]", ".", "We", "train", "models", "with", "different", "numbers", "of", "experts", "in", "the", "MoE", "layers", ".", "In", "addition", "to", "a", "baseline", "model", "with", "no", "MoE", "layers", ",", "we", "train", "models", "with", "flat", "MoE", "layers", "containing", "32", "experts", ",", "and", "models", "with", "hierarchical", "MoE", "layers", "containing", "512", "and", "2048", "experts", ".", "The", "flat", "MoE", "layers", "use", "k", "=", "4", "and", "the", "hierarchical", "MoE", "models", "use", "k", "=", "2", "at", "each", "level", "of", "the", "gating", "network", ".", "Thus", ",", "each", "input", "is", "processed", "by", "exactly", "4", "experts", "in", "each", "MoE", "layer", ".", "Each", "expert", "in", "the", "MoE", "layer", "is", "a", "feed", "forward", "network", "with", "one", "hidden", "layer", "of", "size", "2048", "and", "ReLU", "activation", ".", "Thus", ",", "each", "expert", "contains", "[", "512", "*", "2048", "]", "+", "[", "2048", "*", "512", "]", "=", "2", "M", "parameters", ".", "The", "output", "of", "the", "MoE", "layer", "is", "passed", "through", "a", "sigmoid", "function", ".", "We", "use", "the", "strictly", "-", "balanced", "gating", "function", "described", "in", "Appendix", "F.", "section", ":", "Model", "Architecture", "for", "Multilingual", "MoE", "Model", ":", "We", "used", "the", "same", "model", "architecture", "as", "for", "the", "single", "-", "language", "-", "pair", "models", ",", "with", "the", "following", "exceptions", ":", "We", "used", "noisy", "-", "top", "-", "k", "gating", "as", "described", "in", "Section", "2.1", ",", "not", "the", "scheme", "from", "Appendix", "F.", "The", "MoE", "layers", "in", "the", "encoder", "and", "decoder", "are", "non", "-", "hierarchical", "MoEs", "with", "n", "=", "512", "experts", ",", "and", "k", "=", "2", ".", "Each", "expert", "has", "a", "larger", "hidden", "layer", "of", "size", "8192", ".", "This", "doubles", "the", "amount", "of", "computation", "in", "the", "MoE", "layers", ",", "raising", "the", "computational", "budget", "of", "the", "entire", "model", "from", "85", "M", "to", "102", "M", "ops", "/", "timestep", ".", "Training", ":", "We", "trained", "our", "networks", "using", "the", "Adam", "optimizer", "[", "reference", "]", ".", "The", "base", "learning", "rate", "was", "increased", "linearly", "for", "the", "first", "2000", "training", "steps", ",", "held", "constant", "for", "an", "additional", "8000", "steps", ",", "and", "decreased", "after", "that", "so", "as", "to", "be", "proportional", "to", "the", "inverse", "square", "root", "of", "the", "step", "number", ".", "For", "the", "single", "-", "language", "-", "pair", "models", ",", "similarly", "to", "[", "reference", "]", ",", "we", "applied", "dropout", "[", "reference", "]", "to", "the", "output", "of", "all", "embedding", ",", "LSTM", "and", "MoE", "layers", ",", "using", "DropP", "rob", "=", "0.4", ".", "Training", "was", "done", "synchronously", "on", "a", "cluster", "of", "up", "to", "64", "GPUs", "as", "described", "in", "section", "3", ".", "Each", "training", "batch", "consisted", "of", "a", "set", "of", "sentence", "pairs", "containing", "roughly", "16000", "words", "per", "GPU", ".", "To", "ensure", "balanced", "expert", "utilization", "we", "set", "w", "importance", "=", "0.01", "and", "w", "load", "=", "0.01", ",", "as", "described", "in", "Section", "4", "and", "Appendix", "A.", "section", ":", "Metrics", ":", "We", "evaluated", "our", "models", "using", "the", "perplexity", "and", "the", "standard", "BLEU", "score", "metric", ".", "We", "reported", "tokenized", "BLEU", "score", "as", "computed", "by", "the", "multi", "-", "bleu.pl", "script", ",", "downloaded", "from", "the", "public", "implementation", "of", "Moses", "(", "on", "Github", ")", ",", "which", "was", "also", "used", "in", "[", "reference", "]", ".", "Tables", "2", ",", "3", "and", "4", "in", "Section", "5.3", "show", "comparisons", "of", "our", "results", "to", "other", "published", "methods", ".", "Figure", "4", "shows", "test", "perplexity", "as", "a", "function", "of", "number", "of", "words", "in", "the", "(", "training", "data", "'s", ")", "source", "sentences", "processed", "for", "models", "with", "different", "numbers", "of", "experts", ".", "As", "can", "be", "seen", "from", "the", "Figure", ",", "as", "we", "increased", "the", "number", "of", "experts", "to", "approach", "2048", ",", "the", "test", "perplexity", "of", "our", "model", "continued", "to", "improve", ".", "Figure", "4", ":", "Perplexity", "on", "WMT'14", "En→", "Fr", "(", "left", ")", "and", "Google", "Production", "En→", "Fr", "(", "right", ")", "datasets", "as", "a", "function", "of", "number", "of", "words", "processed", ".", "The", "large", "differences", "between", "models", "at", "the", "beginning", "of", "training", "are", "due", "to", "different", "batch", "sizes", ".", "All", "models", "incur", "the", "same", "computational", "budget", "(", "85", "M", "ops", "/", "timestep", ")", "except", "the", "one", "with", "no", "experts", ".", "section", ":", "Results", ":", "We", "found", "that", "the", "experts", "indeed", "become", "highly", "specialized", "by", "syntax", "and", "/", "or", "semantics", ",", "as", "can", "be", "seen", "in", "Table", "9", ".", "For", "example", ",", "one", "expert", "is", "used", "when", "the", "indefinite", "article", "\"", "a", "\"", "introduces", "the", "direct", "object", "in", "a", "verb", "phrase", "indicating", "importance", "or", "leadership", ".", "section", ":", "F", "STRICTLY", "BALANCED", "GATING", "Due", "to", "some", "peculiarities", "in", "our", "infrastructure", "which", "have", "since", "been", "fixed", ",", "at", "the", "time", "we", "ran", "some", "of", "the", "machine", "translation", "experiments", ",", "our", "models", "ran", "faster", "if", "every", "expert", "received", "exactly", "the", "same", "batch", "size", ".", "To", "accommodate", "this", ",", "we", "used", "a", "different", "gating", "function", "which", "we", "describe", "below", ".", "Recall", "that", "we", "define", "the", "softmax", "gating", "function", "to", "be", ":", "Sparse", "Gating", "(", "alternate", "formulation", ")", ":", "To", "obtain", "a", "sparse", "gating", "vector", ",", "we", "multiply", "G", "σ", "(", "x", ")", "component", "-", "wise", "with", "a", "sparse", "mask", "M", "(", "G", "σ", "(", "x", ")", ")", "and", "normalize", "the", "output", ".", "The", "mask", "itself", "is", "a", "function", "of", "G", "σ", "(", "x", ")", "and", "specifies", "which", "experts", "are", "assigned", "to", "each", "input", "example", ":", "As", "our", "experiments", "suggest", "and", "also", "observed", "in", "[", "reference", "]", ",", "using", "a", "batchwise", "function", "during", "training", "(", "such", "as", "M", "batchwise", ")", "requires", "modifications", "to", "the", "inference", "when", "we", "may", "not", "have", "a", "large", "batch", "of", "examples", ".", "Our", "solution", "to", "this", "is", "to", "train", "a", "vector", "T", "of", "per", "-", "expert", "threshold", "values", "to", "approximate", "the", "effects", "of", "the", "batchwise", "mask", ".", "We", "use", "the", "following", "mask", "at", "inference", "time", ":", "To", "learn", "the", "threshold", "values", ",", "we", "apply", "an", "additional", "loss", "at", "training", "time", "which", "is", "minimized", "when", "the", "batchwise", "mask", "and", "the", "threshold", "mask", "are", "identical", ".", "G", "ATTENTION", "FUNCTION", "The", "attention", "mechanism", "described", "in", "GNMT", "[", "reference", "]", ")", "involves", "a", "learned", "\"", "Attention", "Function", "\"", "A", "(", "x", "i", ",", "y", "j", ")", "which", "takes", "a", "\"", "source", "vector", "\"", "x", "i", "and", "a", "\"", "target", "vector", "\"", "y", "j", ",", "and", "must", "be", "computed", "for", "every", "source", "time", "step", "i", "and", "target", "time", "step", "j.", "In", "GNMT", ",", "the", "attention", "function", "is", "implemented", "as", "a", "feed", "forward", "neural", "network", "with", "a", "hidden", "layer", "of", "size", "n.", "It", "can", "be", "expressed", "as", ":", "Where", "U", "and", "W", "are", "trainable", "weight", "matrices", "and", "V", "is", "a", "trainable", "weight", "vector", ".", "For", "performance", "reasons", ",", "in", "our", "models", ",", "we", "used", "a", "slightly", "different", "attention", "function", ":", "With", "our", "attention", "function", ",", "we", "can", "simultaneously", "compute", "the", "attention", "function", "on", "multiple", "source", "time", "steps", "and", "multiple", "target", "time", "steps", "using", "optimized", "matrix", "multiplications", ".", "We", "found", "little", "difference", "in", "quality", "between", "the", "two", "functions", ".", "section", ":"], "entity": [[0, 4, ["OUTRAGEOUSLY", "LARGE", "NEURAL", "NETWORKS"], "Method"], [6, 15, ["SPARSELY", "-", "GATED", "MIXTURE", "-", "OF", "-", "EXPERTS", "LAYER"], "Method"], [22, 24, ["neural", "network"], "Method"], [35, 37, ["Conditional", "computation"], "Task"], [63, 65, ["model", "capacity"], "Task"], [70, 71, ["computation"], "Task"], [99, 101, ["conditional", "computation"], "Task"], [108, 110, ["model", "capacity"], "Metric"], [115, 117, ["computational", "efficiency"], "Metric"], [119, 121, ["GPU", "clusters"], "Method"], [128, 134, ["Mixture", "-", "of", "-", "Experts", "layer"], "Method"], [135, 136, ["MoE"], "Method"], [144, 150, ["feed", "-", "forward", "sub", "-", "networks"], "Method"], [152, 155, ["trainable", "gating", "network"], "Method"], [171, 172, ["MoE"], "Method"], [176, 178, ["language", "modeling"], "Task"], [179, 181, ["machine", "translation"], "Task"], [183, 185, ["model", "capacity"], "Task"], [202, 204, ["model", "architectures"], "Method"], [207, 208, ["MoE"], "Method"], [216, 217, ["convolutionally"], "Method"], [247, 249, ["computational", "cost"], "Metric"], [258, 260, ["CONDITIONAL", "COMPUTATION"], "Task"], [267, 269, ["model", "size"], "Metric"], [276, 278, ["deep", "learning"], "Method"], [294, 296, ["neural", "networks"], "Method"], [300, 302, ["prediction", "accuracy"], "Metric"], [339, 342, ["deep", "learning", "models"], "Method"], [363, 365, ["training", "costs"], "Metric"], [369, 371, ["model", "size"], "Metric"], [384, 386, ["computing", "power"], "Metric"], [387, 389, ["distributed", "computation"], "Task"], [399, 401, ["conditional", "computation"], "Method"], [416, 418, ["computational", "costs"], "Metric"], [470, 472, ["reinforcement", "learning"], "Method"], [473, 476, ["back", "-", "propagation"], "Method"], [479, 483, ["trarining", "the", "gating", "decisions"], "Task"], [485, 487, ["Model", "capacity"], "Method"], [500, 502, ["conditional", "computation"], "Task"], [569, 571, ["conditional", "computation"], "Task"], [579, 581, ["model", "capacity"], "Metric"], [586, 588, ["computational", "efficiency"], "Metric"], [615, 624, ["SPARSELY", "-", "GATED", "MIXTURE", "-", "OF", "-", "EXPERTS", "LAYER"], "Method"], [627, 629, ["conditional", "computation"], "Task"], [636, 641, ["general", "purpose", "neural", "network", "component"], "Method"], [643, 652, ["Sparsely", "-", "Gated", "Mixture", "-", "of", "-", "Experts", "Layer"], "Method"], [657, 658, ["MoE"], "Method"], [663, 664, ["experts"], "Method"], [668, 673, ["feed", "-", "forward", "neural", "network"], "Method"], [676, 679, ["trainable", "gating", "network"], "Method"], [706, 709, ["back", "-", "propagation"], "Method"], [723, 725, ["language", "modeling"], "Task"], [762, 763, ["MoE"], "Method"], [807, 809, ["language", "modeling"], "Task"], [825, 827, ["computational", "cost"], "Metric"], [833, 836, ["MIXTURES", "OF", "EXPERTS"], "Task"], [851, 857, ["mixture", "-", "of", "-", "experts", "approach"], "Method"], [868, 870, ["expert", "architectures"], "Method"], [875, 876, ["SVMs"], "Method"], [880, 882, ["Gaussian", "Processes"], "Method"], [890, 892, ["Dirichlet", "Processes"], "Method"], [897, 899, ["deep", "networks"], "Method"], [938, 940, ["ensemble", "model"], "Method"], [944, 947, ["mixture", "of", "experts"], "Method"], [948, 950, ["machine", "translation"], "Task"], [952, 954, ["gating", "network"], "Method"], [961, 964, ["ensemble", "NMT", "model"], "Method"], [977, 980, ["mixture", "of", "experts"], "Method"], [994, 995, ["MoEs"], "Method"], [998, 1000, ["gating", "networks"], "Method"], [1004, 1006, ["deep", "model"], "Method"], [1046, 1047, ["MoEs"], "Method"], [1051, 1053, ["computational", "computation"], "Task"], [1061, 1062, ["MoEs"], "Method"], [1064, 1069, ["general", "purpose", "neural", "network", "component"], "Method"], [1076, 1078, ["stacked", "MoEs"], "Method"], [1087, 1089, ["convolutional", "application"], "Method"], [1091, 1092, ["MoE"], "Method"], [1107, 1109, ["sparse", "gating"], "Method"], [1120, 1122, ["model", "capacity"], "Metric"], [1129, 1135, ["MIXTURE", "-", "OF", "-", "EXPERTS", "LAYER"], "Method"], [1152, 1154, ["expert", "networks"], "Method"], [1168, 1170, ["gating", "network"], "Method"], [1196, 1198, ["neural", "networks"], "Method"], [1245, 1249, ["feed", "-", "forward", "networks"], "Method"], [1276, 1278, ["gating", "network"], "Method"], [1286, 1288, ["expert", "network"], "Method"], [1386, 1389, ["level", "hierarchical", "MoE."], "Method"], [1395, 1398, ["primary", "gating", "network"], "Method"], [1400, 1403, ["sparse", "weighted", "combination"], "Method"], [1423, 1425, ["gating", "network"], "Method"], [1432, 1434, ["ordinary", "MoEs"], "Task"], [1453, 1455, ["conditional", "computation"], "Task"], [1457, 1458, ["MoE"], "Method"], [1468, 1471, ["parameterized", "weight", "matrix"], "Method"], [1475, 1476, ["MoE"], "Method"], [1486, 1490, ["block", "-", "wise", "dropout"], "Method"], [1512, 1514, ["GATING", "NETWORK"], "Method"], [1514, 1516, ["Softmax", "Gating"], "Method"], [1521, 1526, ["non", "-", "sparse", "gating", "function"], "Method"], [1545, 1548, ["Sof", "tmax", "function"], "Method"], [1549, 1554, ["Noisy", "Top", "-", "K", "Gating"], "Method"], [1561, 1564, ["Softmax", "gating", "network"], "Method"], [1652, 1654, ["load", "balancing"], "Task"], [1673, 1678, ["trainable", "weight", "matrix", "W", "noise"], "Method"], [1681, 1683, ["Gating", "Network"], "Method"], [1686, 1688, ["gating", "network"], "Method"], [1690, 1693, ["back", "-", "propagation"], "Method"], [1727, 1729, ["gating", "network"], "Method"], [1747, 1749, ["noisy", "rectifiers"], "Method"], [1755, 1757, ["gating", "network"], "Method"], [1775, 1779, ["REINFORCE", "-", "style", "approach"], "Method"], [1782, 1784, ["gating", "network"], "Method"], [1788, 1790, ["PERFORMANCE", "CHALLENGES"], "Task"], [1793, 1796, ["SHRINKING", "BATCH", "PROBLEM"], "Task"], [1798, 1799, ["CPUs"], "Method"], [1800, 1801, ["GPUs"], "Method"], [1808, 1810, ["computational", "efficiency"], "Metric"], [1825, 1827, ["gating", "network"], "Method"], [1880, 1883, ["shrinking", "batch", "problem"], "Task"], [1925, 1927, ["batch", "size"], "Task"], [1928, 1931, ["Mixing", "Data", "Parallelism"], "Task"], [1932, 1934, ["Model", "Parallelism"], "Task"], [1938, 1941, ["distributed", "training", "setting"], "Task"], [1965, 1967, ["parameter", "servers"], "Method"], [1998, 2000, ["gating", "network"], "Method"], [2003, 2007, ["data", "-", "parallel", "schemes"], "Method"], [2050, 2054, ["data", "-", "parallel", "replicas"], "Method"], [2061, 2063, ["gating", "networks"], "Method"], [2066, 2070, ["model", "-", "parallel", "shards"], "Method"], [2120, 2123, ["expert", "batch", "size"], "Metric"], [2137, 2140, ["primary", "gating", "network"], "Method"], [2141, 2143, ["data", "parallelism"], "Method"], [2146, 2148, ["secondary", "MoEs"], "Method"], [2149, 2151, ["model", "parallelism"], "Method"], [2154, 2155, ["MoE"], "Method"], [2192, 2194, ["batch", "size"], "Metric"], [2205, 2209, ["memory", "and", "bandwidth", "requirements"], "Metric"], [2252, 2254, ["trillionparameter", "model"], "Method"], [2256, 2260, ["trillion", "-", "word", "corpus"], "Material"], [2290, 2291, ["Convolutionality"], "Method"], [2294, 2296, ["language", "models"], "Method"], [2301, 2302, ["MoE"], "Method"], [2325, 2326, ["MoE"], "Method"], [2361, 2364, ["Increasing", "Batch", "Size"], "Metric"], [2391, 2392, ["LSTM"], "Method"], [2394, 2395, ["RNN"], "Method"], [2400, 2401, ["MoE."], "Method"], [2407, 2409, ["convolutional", "trick"], "Method"], [2451, 2453, ["unrolled", "RNN"], "Task"], [2470, 2472, ["batch", "size"], "Metric"], [2482, 2484, ["distributed", "computing"], "Task"], [2525, 2527, ["computational", "efficiency"], "Metric"], [2532, 2535, ["expert", "'s", "computation"], "Method"], [2554, 2556, ["computing", "device"], "Method"], [2558, 2559, ["GPUs"], "Method"], [2573, 2574, ["experts"], "Method"], [2576, 2578, ["hidden", "layer"], "Method"], [2624, 2626, ["computational", "efficiency"], "Metric"], [2641, 2644, ["BALANCING", "EXPERT", "UTILIZATION"], "Task"], [2649, 2651, ["gating", "network"], "Method"], [2692, 2694, ["gating", "network"], "Method"], [2742, 2745, ["soft", "constraint", "approach"], "Method"], [2787, 2789, ["loss", "function"], "Metric"], [2802, 2805, ["coefficient", "of", "variation"], "Metric"], [2846, 2850, ["per", "-", "example", "sparsity"], "Task"], [2912, 2914, ["loss", "function"], "Method"], [2956, 2960, ["memory", "and", "performance", "problems"], "Task"], [3004, 3009, ["BILLION", "WORD", "LANGUAGE", "MODELING", "BENCHMARK"], "Task"], [3068, 3073, ["Long", "Short", "-", "Term", "Memory"], "Method"], [3074, 3075, ["LSTM"], "Method"], [3102, 3103, ["Quality"], "Metric"], [3106, 3108, ["parameter", "count"], "Metric"], [3111, 3113, ["computational", "costs"], "Metric"], [3207, 3209, ["computational", "costs"], "Metric"], [3228, 3230, ["softmax", "layer"], "Method"], [3281, 3283, ["MoE", "layers"], "Method"], [3321, 3324, ["matched", "baseline", "models"], "Method"], [3341, 3342, ["perplexity"], "Metric"], [3347, 3349, ["Varied", "Computation"], "Task"], [3389, 3390, ["LSTMs"], "Method"], [3471, 3473, ["Computational", "Efficiency"], "Metric"], [3479, 3480, ["TensorFlow"], "Method"], [3489, 3492, ["Tesla", "K40", "GPUs"], "Method"], [3501, 3503, ["computational", "efficiency"], "Metric"], [3504, 3507, ["TFLOPS", "/", "GPU"], "Metric"], [3564, 3570, ["importance", "-", "sampling", "-", "based", "training"], "Method"], [3572, 3574, ["softmax", "layer"], "Method"], [3621, 3622, ["MoE"], "Method"], [3624, 3626, ["computational", "efficiency"], "Metric"], [3643, 3645, ["computation", "efficiency"], "Metric"], [3652, 3653, ["GPU"], "Method"], [3657, 3661, ["4", "-", "expert", "model"], "Method"], [3734, 3736, ["Language", "modeling"], "Method"], [3739, 3742, ["billion", "word", "corpus"], "Material"], [3746, 3748, ["computational", "budgets"], "Metric"], [3758, 3764, ["1", "-", "billion", "-", "word", "corpus"], "Material"], [3811, 3813, ["quality", "improvements"], "Metric"], [3826, 3831, ["Google", "'s", "internal", "news", "corpus"], "Material"], [3852, 3854, ["computational", "costs"], "Metric"], [3875, 3877, ["MoE", "layers"], "Method"], [4016, 4018, ["layer", "sparsity"], "Metric"], [4020, 4022, ["computational", "efficiency"], "Metric"], [4036, 4038, ["MACHINE", "TRANSLATION"], "Task"], [4039, 4042, ["SINGLE", "LANGUAGE", "PAIR"], "Task"], [4043, 4045, ["Model", "Architecture"], "Method"], [4054, 4056, ["GNMT", "model"], "Method"], [4095, 4096, ["encoder"], "Method"], [4165, 4167, ["WMT'14", "En→Fr"], "Material"], [4168, 4170, ["En→De", "corpora"], "Material"], [4242, 4249, ["Google", "'s", "Production", "English", "to", "French", "data"], "Material"], [4264, 4265, ["GNMT"], "Method"], [4266, 4267, ["RL"], "Method"], [4286, 4287, ["LSTM"], "Method"], [4288, 4291, ["6", "-", "layer"], "Method"], [4386, 4388, ["BLEU", "scores"], "Metric"], [4394, 4396, ["WMT'14", "En→Fr"], "Material"], [4397, 4398, ["En→De"], "Material"], [4418, 4420, ["BLEU", "score"], "Metric"], [4432, 4434, ["perplexity", "scores"], "Metric"], [4441, 4444, ["Google", "Production", "dataset"], "Material"], [4466, 4469, ["MULTILINGUAL", "MACHINE", "TRANSLATION"], "Task"], [4482, 4483, ["GNMT"], "Method"], [4559, 4561, ["model", "architecture"], "Method"], [4591, 4593, ["training", "time"], "Metric"], [4599, 4601, ["computational", "budget"], "Metric"], [4640, 4641, ["perplexity"], "Metric"], [4652, 4654, ["BLEU", "score"], "Metric"], [4698, 4701, ["English", "→", "Korean"], "Material"], [4743, 4745, ["conditional", "computation"], "Method"], [4746, 4748, ["deep", "networks"], "Task"], [4758, 4760, ["conditional", "computing"], "Task"], [4767, 4771, ["algorithmic", "and", "engineering", "solutions"], "Method"], [4778, 4780, ["conditional", "computation"], "Task"], [4805, 4807, ["conditional", "computation"], "Task"], [4873, 4877, ["LOAD", "-", "BALANCING", "LOSS"], "Method"], [4884, 4888, ["load", "-", "balancing", "purposes"], "Task"], [4931, 4932, ["backpropagation"], "Method"], [4938, 4944, ["smooth", "estimator", "Load", "(", "X", ")"], "Method"], [4983, 4985, ["gating", "function"], "Method"], [5116, 5119, ["standard", "normal", "distribution"], "Method"], [5152, 5155, ["Initial", "Load", "Imbalance"], "Task"], [5227, 5229, ["identical", "architecture"], "Method"], [5261, 5262, ["perplexity"], "Method"], [5302, 5305, ["load", "balancing", "purposes"], "Task"], [5306, 5308, ["distributed", "hardware"], "Task"], [5343, 5345, ["model", "quality"], "Metric"], [5373, 5377, ["HIERACHICAL", "MIXTURE", "OF", "EXPERTS"], "Method"], [5390, 5392, ["branching", "factor"], "Metric"], [5397, 5400, ["level", "hierarchical", "MoE."], "Method"], [5406, 5409, ["primary", "gating", "network"], "Method"], [5411, 5414, ["sparse", "weighted", "combination"], "Method"], [5434, 5436, ["gating", "network"], "Method"], [5454, 5457, ["primary", "gating", "network"], "Method"], [5462, 5465, ["secondary", "gating", "networks"], "Method"], [5479, 5481, ["expert", "networks"], "Method"], [5503, 5504, ["MoE"], "Method"], [5511, 5513, ["expert", "utilization"], "Task"], [5529, 5532, ["primary", "gating", "network"], "Method"], [5535, 5538, ["secondary", "gating", "network"], "Method"], [5594, 5597, ["primary", "gating", "network"], "Method"], [5609, 5614, ["BILLION", "WORD", "LANGUAGE", "MODELING", "BENCHMARK"], "Task"], [5618, 5628, ["8", "-", "MILLION", "-", "OPERATIONS", "-", "PER", "-", "TIMESTEP", "MODELS"], "Method"], [5628, 5630, ["Model", "Architecture"], "Method"], [5671, 5673, ["softmax", "layer"], "Method"], [5716, 5717, ["drouput"], "Method"], [5763, 5765, ["gradient", "flow"], "Method"], [5795, 5801, ["Noisy", "-", "Top", "-", "K", "Gating"], "Method"], [5899, 5902, ["matched", "baseline", "models"], "Method"], [5925, 5930, ["ReLU", "-", "activated", "hidden", "layer"], "Method"], [5953, 5958, ["ReLU", "-", "activated", "hidden", "layers"], "Method"], [5965, 5968, ["4xLSTM", "-", "512"], "Method"], [6002, 6003, ["MoE"], "Method"], [6009, 6010, ["LSTM"], "Method"], [6025, 6026, ["LSTM"], "Method"], [6079, 6081, ["K40", "GPUs"], "Method"], [6083, 6085, ["synchronous", "method"], "Method"], [6125, 6126, ["Training"], "Task"], [6149, 6151, ["expert", "computation"], "Method"], [6164, 6166, ["Adam", "optimizer"], "Method"], [6171, 6174, ["base", "learning", "rate"], "Metric"], [6204, 6207, ["Softmax", "output", "layer"], "Method"], [6211, 6213, ["importance", "sampling"], "Method"], [6229, 6233, ["hyper", "-", "parmeter", "search"], "Method"], [6277, 6278, ["perplexity"], "Method"], [6280, 6282, ["holdout", "dataset"], "Material"], [6323, 6325, ["test", "perplexity"], "Metric"], [6327, 6329, ["computational", "budget"], "Metric"], [6337, 6339, ["DropP", "rob"], "Metric"], [6342, 6344, ["computational", "efficiency"], "Metric"], [6387, 6389, ["computation", "budgets"], "Metric"], [6502, 6503, ["M"], "Method"], [6515, 6517, ["hidden", "layer"], "Method"], [6555, 6556, ["perplexity"], "Metric"], [6598, 6600, ["computational", "budget"], "Metric"], [6610, 6612, ["training", "times"], "Metric"], [6625, 6627, ["test", "perplexity"], "Metric"], [6634, 6637, ["100", "BILLION", "WORD"], "Material"], [6637, 6640, ["GOOGLE", "NEWS", "CORPUS"], "Material"], [6643, 6645, ["Model", "Architecture"], "Method"], [6654, 6664, ["8", "-", "million", "-", "operations", "-", "per", "-", "timestep", "models"], "Method"], [6790, 6792, ["memory", "optimizations"], "Method"], [6834, 6835, ["optimizer"], "Method"], [6846, 6848, ["Adam", "optimizer"], "Method"], [6855, 6857, ["moment", "estimates"], "Method"], [6872, 6876, ["first", "-", "moment", "estimator"], "Method"], [6890, 6893, ["second", "moment", "estimator"], "Method"], [6899, 6901, ["factored", "approximation"], "Method"], [6912, 6919, ["full", "matrix", "of", "second", "-", "moment", "estimators"], "Method"], [6943, 6944, ["estimators"], "Method"], [6970, 6971, ["Adagrad"], "Method"], [6982, 6983, ["perplexity"], "Method"], [6985, 6987, ["holdout", "dataset"], "Material"], [6995, 6996, ["Perplexity"], "Metric"], [7026, 7028, ["computational", "efficiency"], "Metric"], [7034, 7035, ["TFLOPS"], "Method"], [7036, 7037, ["GPU"], "Method"], [7071, 7074, ["training", "batch", "size"], "Metric"], [7089, 7093, ["computationally", "matched", "baseline", "model"], "Method"], [7096, 7097, ["LSTMs"], "Method"], [7101, 7106, ["unpruned", "5", "-", "gram", "model"], "Method"], [7107, 7111, ["Kneser", "-", "Ney", "smoothing"], "Method"], [7118, 7120, ["MACHINE", "TRANSLATION"], "Task"], [7123, 7125, ["Model", "Architecture"], "Method"], [7140, 7142, ["GNMT", "model"], "Method"], [7181, 7182, ["encoder"], "Method"], [7203, 7205, ["attention", "mechanism"], "Method"], [7207, 7210, ["encoder", "and", "decoder"], "Method"], [7278, 7279, ["GNMT"], "Method"], [7332, 7335, ["beam", "search", "technique"], "Method"], [7359, 7361, ["baseline", "model"], "Method"], [7411, 7413, ["gating", "network"], "Method"], [7437, 7440, ["feed", "forward", "network"], "Method"], [7442, 7444, ["hidden", "layer"], "Method"], [7448, 7450, ["ReLU", "activation"], "Method"], [7482, 7484, ["sigmoid", "function"], "Method"], [7488, 7493, ["strictly", "-", "balanced", "gating", "function"], "Method"], [7499, 7501, ["Model", "Architecture"], "Method"], [7510, 7512, ["model", "architecture"], "Method"], [7515, 7521, ["single", "-", "language", "-", "pair", "models"], "Method"], [7529, 7535, ["noisy", "-", "top", "-", "k", "gating"], "Method"], [7548, 7550, ["MoE", "layers"], "Method"], [7552, 7555, ["encoder", "and", "decoder"], "Method"], [7558, 7560, ["hierarchical", "MoEs"], "Method"], [7595, 7597, ["computational", "budget"], "Metric"], [7619, 7621, ["Adam", "optimizer"], "Method"], [7627, 7629, ["learning", "rate"], "Metric"], [7668, 7674, ["single", "-", "language", "-", "pair", "models"], "Task"], [7683, 7684, ["dropout"], "Method"], [7692, 7693, ["embedding"], "Method"], [7694, 7695, ["LSTM"], "Method"], [7696, 7698, ["MoE", "layers"], "Method"], [7705, 7706, ["Training"], "Task"], [7767, 7768, ["Metrics"], "Metric"], [7775, 7776, ["perplexity"], "Method"], [7792, 7796, ["multi", "-", "bleu.pl", "script"], "Method"], [7803, 7804, ["Moses"], "Method"], [7899, 7900, ["Perplexity"], "Metric"], [8021, 8024, ["STRICTLY", "BALANCED", "GATING"], "Method"], [8045, 8047, ["machine", "translation"], "Task"], [8071, 8073, ["gating", "function"], "Method"], [8089, 8091, ["Sparse", "Gating"], "Method"], [8092, 8094, ["alternate", "formulation"], "Method"], [8115, 8118, ["sparse", "mask", "M"], "Method"], [8167, 8169, ["batchwise", "function"], "Method"], [8170, 8171, ["training"], "Method"], [8174, 8176, ["M", "batchwise"], "Method"], [8181, 8182, ["inference"], "Task"], [8201, 8209, ["vector", "T", "of", "per", "-", "expert", "threshold", "values"], "Method"], [8256, 8258, ["ATTENTION", "FUNCTION"], "Method"], [8259, 8261, ["attention", "mechanism"], "Method"], [8263, 8264, ["GNMT"], "Method"], [8272, 8274, ["Attention", "Function"], "Method"], [8317, 8318, ["GNMT"], "Method"], [8320, 8322, ["attention", "function"], "Method"], [8326, 8330, ["feed", "forward", "neural", "network"], "Method"], [8377, 8379, ["attention", "function"], "Method"], [8399, 8401, ["matrix", "multiplications"], "Method"], [219, 220, ["LSTM"], "Method"], [224, 226, ["language", "modeling"], "Task"], [227, 229, ["machine", "translation"], "Task"], [602, 604, ["language", "modeling"], "Task"], [653, 654, ["MoE"], "Method"], [726, 728, ["machine", "translation"], "Task"], [746, 747, ["MoE"], "Method"], [750, 751, ["LSTM"], "Method"], [810, 812, ["machine", "translation"], "Task"], [1142, 1143, ["MoE"], "Method"], [1189, 1190, ["MoE"], "Method"], [1298, 1299, ["MoE"], "Method"], [1392, 1393, ["MoE"], "Method"], [1863, 1864, ["MoE"], "Method"], [1985, 1986, ["MoE"], "Method"], [2022, 2023, ["MoE"], "Method"], [2130, 2131, ["MoE"], "Method"], [2348, 2349, ["MoE"], "Method"], [2367, 2368, ["MoE"], "Method"], [2380, 2381, ["MoE"], "Method"], [2419, 2420, ["MoE"], "Method"], [2429, 2430, ["MoE"], "Method"], [3089, 3090, ["LSTM"], "Method"], [3128, 3129, ["MoE"], "Method"], [3137, 3138, ["LSTM"], "Method"], [3141, 3142, ["MoE"], "Method"], [3168, 3170, ["model", "architecture"], "Method"], [3201, 3202, ["MoE"], "Method"], [3368, 3369, ["MoE"], "Method"], [3593, 3594, ["MoE"], "Method"], [3640, 3641, ["MoE"], "Method"], [3676, 3677, ["MoE"], "Method"], [3725, 3728, ["100", "BILLION", "WORD"], "Material"], [3728, 3731, ["GOOGLE", "NEWS", "CORPUS"], "Material"], [3780, 3781, ["MoE"], "Method"], [3867, 3868, ["LSTM"], "Method"], [3893, 3894, ["MoE"], "Method"], [4071, 4072, ["LSTM"], "Method"], [4090, 4091, ["MoE"], "Method"], [4115, 4116, ["MoE"], "Method"], [4144, 4146, ["model", "architecture"], "Method"], [4296, 4297, ["LSTM"], "Method"], [4334, 4335, ["GNMT"], "Method"], [4336, 4337, ["RL"], "Method"], [4406, 4407, ["RL"], "Method"], [4451, 4453, ["BLEU", "score"], "Metric"], [4512, 4513, ["GNMT"], "Method"], [4548, 4549, ["MoE"], "Method"], [4615, 4616, ["GNMT"], "Method"], [4620, 4621, ["GNMT"], "Method"], [4625, 4626, ["MoE"], "Method"], [4634, 4635, ["MoE"], "Method"], [4648, 4649, ["GNMT"], "Method"], [4656, 4657, ["MoE"], "Method"], [4662, 4663, ["GNMT"], "Method"], [4685, 4686, ["GNMT"], "Method"], [5231, 5232, ["MoE"], "Method"], [5403, 5404, ["MoE"], "Method"], [5441, 5442, ["MoE"], "Method"], [5640, 5641, ["embedding"], "Method"], [5645, 5650, ["Long", "Short", "-", "Term", "Memory"], "Method"], [5651, 5652, ["LSTM"], "Method"], [5661, 5662, ["MoE"], "Method"], [5666, 5667, ["LSTM"], "Method"], [5678, 5679, ["embedding"], "Method"], [5687, 5688, ["LSTM"], "Method"], [5698, 5699, ["MoE"], "Method"], [5772, 5773, ["MoE"], "Method"], [5813, 5814, ["MoE"], "Method"], [5825, 5826, ["MoE"], "Method"], [5850, 5851, ["LSTM"], "Method"], [5873, 5874, ["MoE"], "Method"], [5907, 5908, ["MoE"], "Method"], [5914, 5915, ["MoE"], "Method"], [5935, 5936, ["MoE"], "Method"], [5942, 5943, ["MoE"], "Method"], [5972, 5973, ["MoE"], "Method"], [5980, 5981, ["LSTM"], "Method"], [5984, 5985, ["LSTM"], "Method"], [5997, 5998, ["LSTM"], "Method"], [6137, 6138, ["MoE"], "Method"], [6357, 6358, ["MoE"], "Method"], [6362, 6363, ["MoE"], "Method"], [6381, 6382, ["MoE"], "Method"], [6409, 6410, ["MoE"], "Method"], [6413, 6414, ["LSTM"], "Method"], [6420, 6421, ["embedding"], "Method"], [6431, 6432, ["MoE"], "Method"], [6442, 6443, ["MoE"], "Method"], [6448, 6449, ["LSTM"], "Method"], [6455, 6456, ["MoE"], "Method"], [6461, 6462, ["LSTM"], "Method"], [6477, 6478, ["MoE"], "Method"], [6484, 6485, ["MoE"], "Method"], [6499, 6500, ["MoE"], "Method"], [6506, 6507, ["MoE"], "Method"], [6528, 6529, ["MoE"], "Method"], [6571, 6572, ["MoE"], "Method"], [6682, 6683, ["MoE"], "Method"], [6689, 6690, ["MoE"], "Method"], [6708, 6709, ["MoE"], "Method"], [6730, 6733, ["Tesla", "K40", "GPUs"], "Method"], [7012, 7013, ["MoE"], "Method"], [7017, 7019, ["baseline", "model"], "Method"], [7129, 7130, ["MoE"], "Method"], [7157, 7158, ["LSTM"], "Method"], [7176, 7177, ["MoE"], "Method"], [7215, 7216, ["LSTM"], "Method"], [7243, 7244, ["LSTM"], "Method"], [7264, 7265, ["LSTM"], "Method"], [7266, 7267, ["MoE"], "Method"], [7352, 7353, ["MoE"], "Method"], [7363, 7364, ["MoE"], "Method"], [7371, 7372, ["MoE"], "Method"], [7381, 7382, ["MoE"], "Method"], [7391, 7392, ["MoE"], "Method"], [7400, 7401, ["MoE"], "Method"], [7426, 7427, ["MoE"], "Method"], [7433, 7434, ["MoE"], "Method"], [7476, 7477, ["MoE"], "Method"], [7503, 7504, ["MoE"], "Method"], [7590, 7591, ["MoE"], "Method"], [7779, 7781, ["BLEU", "score"], "Metric"], [7786, 7788, ["BLEU", "score"], "Metric"]]} {"doc_id": "44078d0daed8b13114cffb15b368acc467f96351", "text": ["document", ":", "Triplet", "Probabilistic", "Embedding", "for", "Face", "Verification", "and", "Clustering", "Despite", "significant", "progress", "made", "over", "the", "past", "twenty", "five", "years", ",", "unconstrained", "face", "verification", "remains", "a", "challenging", "problem", ".", "This", "paper", "proposes", "an", "approach", "that", "couples", "a", "deep", "CNN", "-", "based", "approach", "with", "a", "low", "-", "dimensional", "discriminative", "embedding", "step", ",", "learned", "using", "triplet", "probability", "constraints", "to", "address", "the", "unconstrained", "face", "verification", "problem", ".", "Aside", "from", "yielding", "performance", "improvements", ",", "this", "embedding", "provides", "significant", "advantages", "in", "terms", "of", "memory", "and", "for", "post", "-", "processing", "operations", "like", "subject", "specific", "clustering", ".", "Experiments", "on", "the", "challenging", "IJB", "-", "A", "dataset", "show", "that", "the", "proposed", "algorithm", "performs", "close", "to", "the", "state", "of", "the", "art", "methods", "in", "verification", "and", "identification", "metrics", ",", "while", "requiring", "much", "less", "training", "data", "and", "training", "/", "test", "time", ".", "The", "superior", "performance", "of", "the", "proposed", "method", "on", "the", "CFP", "dataset", "shows", "that", "the", "representation", "learned", "by", "our", "deep", "CNN", "is", "robust", "to", "large", "pose", "variation", ".", "Furthermore", ",", "we", "demonstrate", "the", "robustness", "of", "deep", "features", "to", "challenges", "including", "age", ",", "pose", ",", "blur", "and", "clutter", "by", "performing", "simple", "clustering", "experiments", "on", "both", "IJB", "-", "A", "and", "LFW", "datasets", ".", "section", ":", "Introduction", "Recently", ",", "with", "the", "advent", "of", "curated", "face", "datasets", "like", "Labeled", "faces", "in", "the", "Wild", "(", "LFW", ")", "and", "advances", "in", "learning", "algorithms", "like", "Deep", "neural", "nets", ",", "there", "is", "more", "hope", "that", "the", "unconstrained", "face", "verification", "problem", "can", "be", "solved", ".", "A", "face", "verification", "algorithm", "compares", "two", "given", "templates", "that", "are", "typically", "not", "seen", "during", "training", ".", "Research", "in", "face", "verification", "has", "progressed", "well", "over", "the", "past", "few", "years", ",", "resulting", "in", "the", "saturation", "of", "performance", "on", "the", "LFW", "dataset", ",", "yet", "the", "problem", "of", "unconstrained", "face", "verification", "remains", "a", "challenge", ".", "This", "is", "evident", "by", "the", "performance", "of", "traditional", "algorithms", "on", "the", "publicly", "available", "IJB", "-", "A", "dataset", "(", ",", ")", "that", "was", "released", "recently", ".", "Moreover", ",", "despite", "the", "superb", "performance", "of", "CNN", "-", "based", "approaches", "compared", "to", "traditional", "methods", ",", "a", "drawback", "of", "such", "methods", "is", "the", "long", "training", "time", "needed", ".", "In", "this", "work", ",", "we", "present", "a", "Deep", "CNN", "(", "DCNN", ")", "architecture", "that", "ensures", "faster", "training", ",", "and", "investigate", "how", "much", "the", "performance", "can", "be", "improved", "if", "we", "are", "provided", "domain", "specific", "data", ".", "Specifically", ",", "our", "contributions", "are", "as", "follows", ":", "We", "propose", "a", "deep", "network", "architecture", "and", "a", "training", "scheme", "that", "ensures", "faster", "training", "time", ".", "We", "formulate", "a", "triplet", "probability", "embedding", "learning", "method", "to", "improve", "the", "performance", "of", "deep", "features", "for", "face", "verification", "and", "subject", "clustering", ".", "During", "training", ",", "we", "use", "a", "publicly", "available", "face", "dataset", "to", "train", "our", "deep", "architecture", ".", "Each", "image", "is", "pre", "-", "processed", "and", "aligned", "to", "a", "canonical", "view", "before", "passing", "it", "to", "the", "deep", "network", "whose", "features", "are", "used", "to", "represent", "the", "image", ".", "In", "the", "case", "of", "IJB", "-", "A", "dataset", ",", "the", "data", "is", "divided", "into", "10", "splits", ",", "each", "split", "containing", "a", "training", "set", "and", "a", "test", "set", ".", "Hence", ",", "to", "further", "improve", "performance", ",", "we", "learn", "the", "proposed", "triplet", "probability", "embedding", "using", "the", "training", "set", "provided", "with", "each", "split", "over", "the", "features", "extracted", "from", "our", "DCNN", "model", ".", "During", "the", "deployment", "phase", ",", "given", "a", "face", "template", ",", "we", "extract", "the", "deep", "features", "using", "the", "raw", "CNN", "model", "after", "implementing", "automatic", "pre", "-", "processing", "steps", "such", "as", "face", "detection", "and", "fiducial", "extraction", ".", "The", "deep", "features", "are", "projected", "onto", "a", "low", "-", "dimensional", "space", "using", "the", "embedding", "matrix", "learned", "during", "training", "(", "note", "that", "the", "projection", "involves", "only", "matrix", "multiplication", ")", ".", "We", "use", "the", "128", "-", "dimensional", "feature", "as", "the", "final", "representation", "of", "the", "given", "face", "template", ".", "This", "paper", "is", "organized", "as", "follows", ":", "Section", "[", "reference", "]", "places", "our", "work", "among", "the", "recently", "proposed", "approaches", "for", "face", "verification", ".", "Section", "[", "reference", "]", "details", "the", "network", "architecture", "and", "the", "training", "scheme", ".", "The", "triplet", "probabilistic", "embedding", "learning", "method", "is", "described", "in", "Section", "[", "reference", "]", "followed", "by", "results", "on", "IJB", "-", "A", "and", "CFP", "datasets", "and", "a", "brief", "discussion", "in", "Section", "[", "reference", "]", ".", "In", "Section", "[", "reference", "]", ",", "we", "demonstrate", "the", "ability", "of", "the", "proposed", "method", "to", "cluster", "a", "media", "collection", "from", "LFW", "and", "IJB", "-", "A", "datasets", ".", "section", ":", "Related", "Work", "In", "the", "past", "few", "years", ",", "there", "have", "been", "numerous", "works", "in", "using", "deep", "features", "for", "tasks", "related", "to", "face", "verification", ".", "The", "DeepFace", "approach", "uses", "a", "carefully", "crafted", "3D", "alignment", "procedure", "to", "preprocess", "face", "images", "and", "feeds", "them", "to", "a", "deep", "network", "that", "is", "trained", "using", "a", "large", "training", "set", ".", "More", "recently", ",", "Facenet", "uses", "a", "large", "private", "dataset", "to", "train", "several", "deep", "network", "models", "using", "a", "triplet", "distance", "loss", "function", ".", "The", "training", "time", "for", "this", "network", "is", "of", "the", "order", "of", "few", "weeks", ".", "Since", "the", "release", "of", "the", "IJB", "-", "A", "dataset", ",", "there", "have", "been", "several", "works", "that", "have", "published", "verification", "results", "for", "this", "dataset", ".", "Previous", "approaches", "presented", "in", "and", "train", "deep", "networks", "using", "the", "CASIA", "-", "WebFace", "dataset", "and", "the", "VGG", "-", "Face", "dataset", "respectively", ",", "requiring", "substantial", "training", "time", ".", "This", "paper", "proposes", "a", "network", "architecture", "and", "a", "training", "scheme", "that", "needs", "shorter", "training", "time", "and", "a", "small", "query", "time", ".", "The", "idea", "of", "learning", "a", "compact", "and", "discriminative", "representation", "has", "been", "around", "for", "decades", ".", "Weinberger", "et", "al", ".", "used", "a", "Semi", "Definite", "Programming", "(", "SDP", ")-", "based", "formulation", "to", "learn", "a", "metric", "satisfying", "pairwise", "and", "triplet", "distance", "constraints", "in", "a", "large", "margin", "framework", ".", "More", "recently", ",", "this", "idea", "has", "been", "successfully", "applied", "to", "face", "verification", "by", "integrating", "the", "loss", "function", "within", "the", "deep", "network", "architecture", "(", ",", ")", ".", "Joint", "Bayesian", "metric", "learning", "is", "also", "another", "popular", "metric", "used", "for", "face", "verification", "(", ",", ")", ".", "These", "methods", "either", "require", "a", "large", "dataset", "for", "convergence", "or", "learn", "a", "metric", "directly", "and", "therefore", "are", "not", "amenable", "to", "subsequent", "operations", "like", "discriminative", "clustering", "or", "hashing", ".", "Classic", "methods", "like", "t", "-", "SNE", ",", "t", "-", "STE", "and", "Crowd", "Kernel", "Learning", "(", "CKL", ")", "perform", "extremely", "well", "when", "used", "to", "visualize", "or", "cluster", "a", "given", "data", "collection", ".", "They", "either", "operate", "on", "the", "data", "matrix", "directly", "or", "the", "distance", "matrix", "generated", "from", "data", "by", "generating", "a", "large", "set", "of", "pairwise", "or", "triplet", "constraints", ".", "While", "these", "methods", "perform", "very", "well", "on", "a", "given", "set", "of", "data", "points", ",", "they", "do", "not", "generalize", "to", "out", "-", "of", "-", "sample", "data", ".", "In", "the", "current", "work", ",", "we", "aim", "to", "generalize", "such", "formulations", ",", "to", "a", "more", "traditional", "classification", "setting", ",", "where", "domain", "specific", "training", "and", "testing", "data", "is", "provided", ".", "We", "formulate", "an", "optimization", "problem", "based", "on", "triplet", "probabilities", "that", "performs", "dimensionality", "reduction", "aside", "from", "improving", "the", "discriminative", "ability", "of", "the", "test", "data", ".", "The", "embedding", "scheme", "described", "in", "this", "work", "is", "a", "more", "general", "framework", "that", "can", "be", "applied", "to", "any", "setting", "where", "labeled", "training", "data", "is", "available", ".", "section", ":", "Network", "Architecture", "This", "section", "details", "the", "architecture", "and", "training", "algorithm", "for", "the", "deep", "network", "used", "in", "our", "work", ".", "Our", "architecture", "consists", "of", "7", "convolutional", "layers", "with", "varying", "kernel", "sizes", ".", "The", "initial", "layers", "have", "a", "larger", "size", "rapidly", "subsampling", "the", "image", "and", "reducing", "the", "parameters", "while", "subsequent", "layers", "consist", "of", "small", "filter", "sizes", ",", "which", "has", "proved", "to", "be", "very", "useful", "in", "face", "recognition", "tasks", "(", ",", ")", ".", "Furthermore", ",", "we", "use", "the", "Parametric", "Rectifier", "Linear", "units", "(", "PReLUs", ")", "instead", "of", "ReLUs", ",", "since", "they", "allow", "a", "negative", "value", "for", "the", "output", "based", "on", "a", "learned", "threshold", "and", "have", "been", "shown", "to", "improve", "the", "convergence", "rate", ".", "tableDeep", "Network", "architecture", "details", "The", "top", "three", "convolutional", "layers", "(", "conv1", "-", "conv3", ")", "are", "initialized", "with", "the", "weights", "from", "the", "AlexNet", "model", "trained", "on", "the", "ImageNet", "challenge", "dataset", ".", "Several", "recent", "works", "(", ",", ")", "have", "empirically", "shown", "that", "this", "transfer", "of", "knowledge", "across", "different", "networks", ",", "albeit", "for", "a", "different", "objective", ",", "improves", "performance", "and", "more", "significantly", "reduces", "the", "need", "to", "train", "over", "a", "large", "number", "of", "iterations", ".", "The", "compared", "methods", "either", "learn", "their", "deep", "models", "from", "scratch", "(", ",", ")", "or", "finetune", "only", "the", "last", "layer", "of", "fully", "pre", "-", "trained", "models", ".", "The", "former", "results", "in", "large", "training", "time", "and", "the", "latter", "does", "not", "generalize", "well", "to", "the", "task", "at", "hand", "(", "face", "verification", ")", "and", "hence", "resulting", "in", "sub", "optimal", "performance", ".", "In", "the", "current", "work", ",", "even", "though", "we", "use", "a", "pre", "-", "trained", "model", "(", "AlexNet", ")", "to", "initialize", "the", "proposed", "deep", "network", ",", "we", "do", "so", "only", "for", "the", "first", "three", "convolutional", "layers", ",", "since", "they", "retain", "more", "generic", "information", "(", ")", ".", "Subsequent", "layers", "learn", "representations", "which", "are", "more", "specific", "to", "the", "task", "at", "hand", ".", "Thus", ",", "to", "learn", "more", "task", "specific", "information", ",", "we", "add", "4", "convolutional", "layers", "each", "consisting", "of", "512", "kernels", "of", "size", ".", "The", "layers", "conv4", "-", "conv7", "do", "not", "downsample", "the", "input", "thereby", "learning", "more", "complex", "higher", "dimensional", "representations", ".", "This", "hybrid", "architecture", "proves", "to", "be", "extremely", "effective", "as", "our", "raw", "CNN", "representation", "outperforms", "some", "very", "deep", "CNN", "models", "on", "the", "IJB", "-", "A", "dataset", "(", "Table", "2", "in", "Results", ")", ".", "In", "addition", ",", "we", "achieve", "that", "performance", "by", "training", "the", "proposed", "deep", "network", "using", "the", "relatively", "smaller", "CASIA", "-", "WebFace", "dataset", ".", "The", "architecture", "of", "our", "network", "is", "shown", "in", "Table", "[", "reference", "]", ".", "Layers", "conv4", "-", "conv7", "and", "the", "fully", "connected", "layers", "fc6", "-", "fc8", "are", "initialized", "from", "scratch", "using", "random", "Gaussian", "distributions", ".", "PReLU", "activation", "functions", "are", "added", "between", "each", "layer", ".", "Since", "the", "network", "is", "used", "as", "a", "feature", "extractor", ",", "the", "last", "layer", "fc8", "is", "removed", "during", "deployment", ",", "thus", "reducing", "the", "number", "of", "parameters", "to", "29M.", "The", "inputs", "to", "the", "network", "are", "227x227x3", "RGB", "images", ".", "When", "the", "network", "is", "deployed", ",", "the", "features", "are", "extracted", "from", "the", "fc7", "layer", "resulting", "in", "a", "dimensionality", "of", "512", ".", "The", "network", "is", "trained", "using", "the", "Softmax", "loss", "function", "for", "multiclass", "classification", "using", "the", "Caffe", "deep", "learning", "platform", ".", "section", ":", "Learning", "a", "Discriminative", "Embedding", "In", "this", "section", ",", "we", "describe", "our", "algorithm", "for", "learning", "a", "low", "-", "dimensional", "embedding", "such", "that", "the", "resulting", "projections", "are", "more", "discriminative", ".", "Aside", "from", "an", "improved", "performance", ",", "this", "embedding", "provides", "significant", "advantages", "in", "terms", "of", "memory", "and", "enables", "post", "-", "processing", "operations", "like", "visualization", "and", "clustering", ".", "Consider", "a", "triplet", ",", "where", "(", "anchor", ")", "and", "(", "positive", ")", "are", "from", "the", "same", "class", ",", "but", "(", "negative", ")", "belongs", "to", "a", "different", "class", ".", "Consider", "a", "function", "that", "is", "parameterized", "by", "the", "matrix", ",", "that", "measures", "the", "similarity", "between", "two", "vectors", ".", "Ideally", ",", "for", "all", "triplets", "that", "exist", "in", "the", "training", "set", ",", "we", "would", "like", "the", "following", "constraint", "to", "be", "satisfied", ":", "Thus", ",", "the", "probability", "of", "a", "given", "triplet", "satisfying", "(", "[", "reference", "]", ")", "can", "be", "written", "as", ":", "The", "specific", "form", "of", "the", "similarity", "function", "is", "given", "as", ":", ".", "In", "our", "case", ",", "and", "are", "deep", "features", "normalized", "to", "unit", "length", ".", "To", "learn", "the", "embedding", "from", "a", "given", "set", "of", "triplets", ",", "we", "solve", "the", "following", "optimization", ":", "(", "[", "reference", "]", ")", "can", "be", "interpreted", "as", "maximizing", "the", "likelihood", "(", "[", "reference", "]", ")", "or", "minimizing", "the", "negative", "log", "-", "likelihood", "(", "NLL", ")", "over", "the", "triplet", "set", ".", "In", "practice", ",", "the", "above", "problem", "is", "solved", "in", "a", "Large", "-", "Margin", "framework", "using", "Stochastic", "Gradient", "Descent", "(", "SGD", ")", "and", "the", "triplets", "are", "sampled", "online", ".", "The", "gradient", "update", "for", "is", "given", "as", ":", "where", "is", "the", "estimate", "at", "iteration", ",", "is", "the", "updated", "estimate", ",", "is", "the", "triplet", "sampled", "at", "the", "current", "iteration", "and", "is", "the", "learning", "rate", ".", "By", "choosing", "the", "dimension", "of", "as", "with", ",", "we", "achieve", "dimensionality", "reduction", "in", "addition", "to", "improved", "performance", ".", "For", "our", "work", ",", "we", "fix", "based", "on", "cross", "validation", "and", "is", "the", "dimensionality", "of", "our", "deep", "features", ".", "is", "initialized", "with", "the", "first", "principal", "components", "of", "the", "training", "data", ".", "At", "each", "iteration", ",", "a", "random", "anchor", "and", "a", "random", "positive", "data", "point", "are", "chosen", ".", "To", "choose", "the", "negative", ",", "we", "perform", "hard", "negative", "mining", ",", "ie", ".", "we", "choose", "the", "data", "point", "that", "has", "the", "least", "likelihood", "(", "[", "reference", "]", ")", "among", "the", "randomly", "chosen", "2000", "negative", "instances", "at", "each", "iteration", ".", "Since", "we", "compute", "the", "embedding", "matrix", "by", "optimizing", "over", "triplet", "probabilities", ",", "we", "call", "this", "method", "Triplet", "Probability", "Embedding", "(", "TPE", ")", ".", "The", "technique", "closest", "to", "the", "one", "presented", "in", "this", "section", ",", "which", "is", "used", "in", "recent", "works", "(", ",", ")", "computes", "the", "embedding", "based", "on", "satisfying", "a", "hinge", "loss", "constraint", ":", "acts", "a", "margin", "parameter", "for", "the", "loss", "function", ".", "To", "be", "consistent", "with", "the", "terminology", "used", "in", "this", "paper", ",", "we", "call", "it", "Triplet", "Distance", "Embedding", "(", "TDE", ")", ".", "To", "appreciate", "the", "difference", "between", "the", "two", "approaches", ",", "Figure", "[", "reference", "]", "shows", "the", "case", "where", "the", "gradient", "update", "for", "the", "TDE", "method", "(", "[", "reference", "]", ")", "occurs", ".", "If", "the", "value", "of", "is", "not", "appropriately", "chosen", ",", "a", "triplet", "is", "considered", "good", "even", "if", "the", "positive", "and", "negative", "are", "very", "close", "to", "one", "another", ".", "But", "under", "the", "proposed", "formulation", ",", "both", "cases", "referred", "to", "in", "Figure", "[", "reference", "]", "will", "update", "the", "gradient", "but", "their", "contribution", "to", "the", "gradient", "will", "be", "modulated", "by", "the", "probability", "with", "which", "they", "violate", "the", "constraint", "in", "(", "[", "reference", "]", ")", ".", "This", "modulation", "factor", "is", "specified", "by", "the", "term", "in", "the", "gradient", "update", "for", "TPE", "in", "(", "[", "reference", "]", ")", "implying", "that", "if", "the", "likelihood", "of", "a", "sampled", "triplet", "satisfying", "(", "[", "reference", "]", ")", "is", "high", ",", "then", "the", "gradient", "update", "is", "given", "a", "lower", "weight", "and", "vice", "-", "versa", ".", "Thus", ",", "in", "our", "method", ",", "the", "margin", "parameter", "(", ")", "is", "automatically", "set", "based", "on", "the", "likelihood", ".", "To", "compare", "the", "relative", "performances", "of", "the", "raw", "features", "before", "projection", ",", "with", "TDE", "and", "with", "TPE", "(", "proposed", "method", ")", ",", "we", "plot", "the", "traditional", "ROC", "curve", "(", "TAR", "(", "vs", ")", "FAR", ")", "for", "split", "1", "of", "the", "IJB", "-", "A", "verify", "protocol", "for", "the", "three", "methods", "in", "Figure", "[", "reference", "]", ".", "The", "Equal", "Error", "Rate", "(", "EER", ")", "metric", "is", "specified", "for", "each", "method", ".", "The", "performance", "improvement", "due", "to", "TPE", "is", "significant", ",", "especially", "at", "regions", "of", "FAR", ".", "We", "observed", "a", "similar", "behaviour", "for", "all", "the", "ten", "splits", "of", "the", "IJB", "-", "A", "dataset", ".", "section", ":", "Experimental", "setup", "and", "Results", "In", "this", "section", "we", "evaluate", "the", "proposed", "method", "on", "two", "challenging", "datasets", ":", "IARPA", "Janus", "Benchmark", "-", "A", "(", "IJB", "-", "A", ")", ":", "This", "dataset", "contains", "500", "subjects", "with", "a", "total", "of", "25", ",", "813", "images", "(", "5", ",", "399", "still", "images", "and", "20", ",", "414", "video", "frames", "sampled", "at", "a", "rate", "of", "1", "in", "60", ")", ".", "The", "faces", "in", "the", "IJB", "-", "A", "dataset", "contain", "extreme", "poses", "and", "illuminations", ",", "more", "challenging", "than", "LFW", ".", "Some", "sample", "images", "from", "the", "IJB", "-", "A", "dataset", "are", "shown", "in", "Figure", "[", "reference", "]", ".", "An", "additional", "challenge", "of", "the", "IJB", "-", "A", "verification", "protocol", "is", "that", "the", "template", "comparisons", "include", "image", "to", "image", ",", "image", "to", "set", "and", "set", "to", "set", "comparisons", ".", "In", "this", "work", ",", "for", "a", "given", "test", "template", "of", "the", "IJB", "-", "A", "data", "we", "perform", "two", "kinds", "of", "pooling", "to", "produce", "its", "final", "representation", ":", "Average", "pooling", "(", "CNN", ")", ":", "The", "deep", "features", "of", "the", "images", "and", "/", "or", "frames", "present", "in", "the", "template", "are", "combined", "by", "taking", "a", "componentwise", "average", "to", "produce", "one", "feature", "vector", ".", "Thus", "each", "feature", "equally", "contributes", "to", "the", "final", "representation", ".", "Media", "pooling", "(", "CNN", ")", ":", "The", "deep", "features", "are", "combined", "keeping", "in", "mind", "the", "media", "source", "they", "come", "from", ".", "The", "metadata", "provided", "with", "IJB", "-", "A", "gives", "us", "the", "media", "i", "d", "for", "each", "item", "of", "the", "template", ".", "Thus", "to", "get", "the", "final", "feature", "vector", ",", "we", "first", "take", "an", "intra", "-", "media", "average", "and", "then", "combine", "these", "by", "taking", "the", "inter", "-", "media", "average", ".", "Thus", "each", "feature", "’s", "contribution", "to", "the", "final", "representation", "is", "weighted", "based", "on", "its", "source", ".", "Celebrities", "in", "Frontal", "-", "Profile", "(", "CFP", ")", "[", "]", ":", "This", "dataset", "contains", "7000", "images", "of", "500", "subjects", ".", "The", "dataset", "is", "used", "for", "evaluating", "how", "face", "verification", "approaches", "handle", "pose", "variation", ".", "Hence", ",", "it", "consists", "of", "5000", "images", "in", "frontal", "view", "and", "2000", "images", "in", "extreme", "profile", ".", "The", "data", "is", "organized", "into", "10", "splits", ",", "each", "containing", "equal", "number", "of", "frontal", "-", "frontal", "and", "frontal", "-", "profile", "comparisons", ".", "Sample", "comparison", "pairs", "of", "the", "CFP", "dataset", "are", "shown", "in", "Figure", "[", "reference", "]", ".", ".5", ".25", ".25", "subsection", ":", "Pre", "-", "processing", "In", "the", "training", "phase", ",", "given", "an", "input", "image", ",", "we", "use", "the", "HyperFace", "method", "for", "face", "detection", "and", "fiducial", "point", "extraction", ".", "The", "HyperFace", "detector", "automatically", "extracts", "many", "faces", "from", "a", "given", "image", ".", "For", "the", "IJB", "-", "A", "dataset", ",", "since", "most", "images", "contain", "more", "than", "one", "face", ",", "we", "use", "the", "bounding", "boxes", "provided", "along", "with", "the", "dataset", "to", "select", "the", "person", "of", "interest", "from", "the", "list", "of", "automatic", "detections", ".", "We", "select", "the", "detection", "that", "has", "the", "maximum", "area", "overlap", "with", "the", "manually", "provided", "bounding", "box", ".", "In", "the", "IJB", "-", "A", "dataset", ",", "there", "are", "few", "images", "for", "which", "the", "HyperFace", "detector", "can", "not", "find", "the", "relevant", "face", ".", "For", "the", "missed", "cases", ",", "we", "crop", "the", "face", "using", "the", "bounding", "box", "information", "provided", "with", "the", "dataset", "and", "pass", "it", "to", "HyperFace", "to", "extract", "the", "fiducials", ".", "We", "use", "six", "fiducial", "points", "(", "eyes", "and", "mouth", "corners", ")", "to", "align", "the", "detected", "image", "to", "a", "canonical", "view", "using", "the", "similarity", "transform", ".", "For", "the", "CFP", "dataset", ",", "since", "the", "six", "keypoints", "can", "not", "be", "computed", "for", "profile", "faces", "we", "only", "use", "three", "keypoints", "on", "one", "side", "of", "the", "face", "for", "aligning", "them", ".", "tableIdentification", "and", "Verification", "results", "on", "the", "IJB", "-", "A", "dataset", ".", "For", "identification", ",", "the", "scores", "reported", "are", "TPIR", "values", "at", "the", "indicated", "points", ".", "The", "results", "are", "averages", "over", "10", "splits", "and", "the", "standard", "deviation", "is", "given", "in", "the", "brackets", "for", "methods", "which", "have", "reported", "them", ".", "implies", "that", "the", "result", "is", "not", "reported", "for", "that", "method", ".", "The", "best", "results", "are", "given", "in", "bold", ".", "tableResults", "on", "the", "CFP", "dataset", ".", "The", "numbers", "are", "averaged", "over", "ten", "test", "splits", "and", "the", "numbers", "in", "brackets", "indicate", "standard", "deviations", "of", "those", "runs", ".", "The", "best", "results", "are", "given", "in", "bold", ".", "subsection", ":", "Parameters", "and", "training", "times", "The", "training", "of", "the", "proposed", "deep", "architecture", "is", "done", "using", "SGD", "with", "momentum", ",", "which", "is", "set", "to", "0.9", "and", "the", "learning", "rate", "is", "set", "to", "1e", "-", "3", "and", "decreased", "uniformly", "by", "a", "factor", "of", "10", "every", "50", "K", "iterations", ".", "The", "weight", "decay", "is", "set", "to", "5e", "-", "4", "for", "all", "layers", ".", "The", "training", "batch", "size", "is", "set", "to", "256", ".", "The", "training", "time", "for", "our", "deep", "network", "is", "24", "hours", "on", "a", "single", "NVIDIA", "TitanX", "GPU", ".", "For", "the", "IJB", "-", "A", "dataset", ",", "we", "use", "the", "training", "data", "provided", "with", "each", "split", "to", "obtain", "the", "triplet", "embedding", "which", "takes", "3", "mins", "per", "split", ".", "This", "is", "the", "only", "additional", "splitwise", "processing", "that", "is", "done", "by", "the", "proposed", "approach", ".", "During", "deployment", ",", "the", "average", "enrollment", "time", "per", "image", "after", "pre", "-", "processing", ",", "including", "alignment", "and", "feature", "extraction", "is", "8ms", ".", "subsection", ":", "Evaluation", "Pipeline", "Given", "an", "image", ",", "we", "pre", "-", "process", "it", "as", "described", "in", "Section", "5.1", ".", "The", "deep", "features", "are", "computed", "as", "an", "average", "of", "the", "image", "and", "its", "flip", ".", "Given", "two", "deep", "features", "to", "compare", ",", "we", "compute", "their", "cosine", "similarity", "score", ".", "More", "specifically", ",", "for", "the", "IJB", "-", "A", "dataset", ",", "given", "a", "template", "containing", "multiple", "faces", ",", "we", "flatten", "the", "template", "features", "by", "average", "pooling", "or", "media", "pooling", "to", "obtain", "a", "vector", "representation", ".", "For", "each", "split", ",", "we", "learn", "the", "TPE", "projection", "using", "the", "provided", "training", "data", ".", "Given", "two", "templates", "for", "comparison", ",", "we", "compute", "the", "cosine", "similarity", "score", "using", "the", "projected", "128", "-", "dimensional", "representations", ".", "matrix", ".", "subsection", ":", "Evaluation", "Metrics", "We", "report", "two", "types", "of", "results", "for", "the", "IJB", "-", "A", "dataset", ":", "Verification", "and", "Identification", ".", "For", "the", "verification", "protocol", ",", "we", "report", "the", "False", "Non", "-", "Match", "Rate", "(", "FNMR", ")", "values", "at", "several", "False", "Match", "Rates", "(", "FMR", ")", ".", "For", "the", "identification", "results", ",", "we", "report", "open", "set", "and", "closed", "set", "metrics", ".", "For", "the", "open", "set", "metrics", ",", "the", "True", "Positive", "Identification", "Rate", "quantifies", "the", "fraction", "of", "subjects", "that", "are", "classified", "correctly", "among", "the", "ones", "that", "exist", "in", "probe", "but", "not", "in", "gallery", ".", "For", "the", "closed", "set", "metrics", ",", "we", "report", "the", "CMC", "numbers", "at", "different", "values", "of", "False", "Positive", "Identification", "Rates", "(", "FPIRs", ")", "and", "Ranks", ".", "More", "details", "on", "the", "evaluation", "metrics", "for", "the", "IJB", "-", "A", "protocol", "can", "be", "found", "in", ".", "For", "the", "CFP", "dataset", ",", "following", "the", "protocol", "set", "in", ",", "we", "report", "the", "Area", "under", "the", "curve", "(", "AUC", ")", "and", "Equal", "Error", "Rate", "(", "EER", ")", "values", "as", "averages", "across", "splits", ",", "in", "addition", "to", "the", "classification", "accuracy", ".", "To", "obtain", "the", "accuracy", "for", "each", "split", ",", "we", "threshold", "our", "CNN", "similarity", "scores", "where", "the", "threshold", "is", "set", "to", "the", "value", "that", "provides", "the", "highest", "classification", "accuracy", "over", "the", "training", "data", "for", "each", "split", ".", "subsection", ":", "Discussion", "subsubsection", ":", "Performance", "on", "IJB", "-", "A", "Table", "[", "reference", "]", "presents", "the", "results", "for", "the", "proposed", "methods", "compared", "to", "existing", "results", "for", "the", "IJB", "-", "A", "Verification", "and", "Identification", "protocol", ".", "The", "compared", "methods", "are", "described", "below", ":", "Government", "-", "of", "-", "the", "-", "Shelf", "(", "GOTS", ")", "is", "the", "baseline", "performance", "provided", "along", "with", "the", "IJB", "-", "A", "dataset", ".", "Parkhi", "et", "al", ".", "train", "a", "very", "deep", "network", "(", "22", "layers", ")", "over", "the", "VGG", "-", "Face", "dataset", "which", "contains", "2.6", "M", "images", "from", "2622", "subjects", ".", "The", "Neural", "Aggregation", "network", "(", "NAN", ")", "is", "trained", "over", "large", "amount", "of", "videos", "from", "the", "CELEB", "-", "1000", "dataset", "starting", "from", "the", "GoogleNet", "architecture", ".", "Masi", "et", "al", ".", "use", "a", "deep", "CNN", "based", "approach", "that", "includes", "a", "combination", "of", "in", "-", "plane", "aligned", "images", ",", "3D", "rendered", "images", "to", "augment", "their", "performance", ".", "The", "3D", "rendered", "images", "are", "also", "generated", "during", "test", "time", "per", "template", "comparison", ".", "It", "should", "be", "noted", "that", "many", "test", "images", "of", "the", "IJB", "-", "A", "dataset", "contain", "extreme", "poses", ",", "harsh", "illumination", "conditions", "and", "significant", "blur", ".", "Crosswhite", "et", "al", ".", "use", "template", "adaptation", "to", "tune", "the", "performance", "of", "their", "raw", "features", "specifically", "to", "the", "IJB", "-", "A", "dataset", ".", "Compared", "to", "these", "methods", ",", "the", "proposed", "method", "trains", "a", "single", "CNN", "model", "on", "the", "CASIA", "-", "WebFace", "dataset", "which", "consists", "of", "about", "500", "K", "images", "and", "requires", "much", "shorter", "training", "time", "and", "has", "a", "very", "fast", "query", "time", "(", "0.08s", "after", "face", "detection", "per", "image", "pair", ")", ".", "As", "shown", "in", "Table", "[", "reference", "]", ",", "our", "raw", "CNN", "features", "after", "media", "pooling", "perform", "better", "than", "most", "compared", "methods", "across", "both", "the", "verification", "and", "identification", "protocols", "of", "the", "IJB", "-", "A", "dataset", ",", "with", "the", "exception", "of", "the", "template", "adaptation", "method", "by", "Crosswhite", "et", "al", ".", "which", "is", "discussed", "below", ".", "The", "TPE", "method", "provides", "significant", "improvement", "for", "both", "identification", "and", "verification", "tasks", "as", "shown", "in", "Table", "[", "reference", "]", ".", "The", "method", "by", "Crosswhite", "et", "al", ".", "uses", "the", "VGG", "-", "Face", "network", "descriptors", "(", "4096", "-", "d", ")", "as", "the", "raw", "features", ".", "They", "use", "the", "concept", "of", "template", "adaptation", "to", "improve", "their", "performance", "as", "follows", ":", "when", "pooling", "multiple", "faces", "of", "a", "given", "template", ",", "they", "train", "a", "linear", "SVM", "with", "the", "features", "of", "this", "template", "as", "positive", "and", "a", "fixed", "set", "of", "negatives", "extracted", "from", "the", "training", "data", "of", "the", "IJB", "-", "A", "splits", ".", "Let", "’s", "denote", "the", "pooled", "template", "feature", "and", "classifier", "pair", "as", ".", "Then", ",", "at", "query", "time", "when", "comparing", "two", "templates", "and", ",", "the", "similarity", "score", "is", "computed", "as", ":", ".", "Even", "when", "using", "a", "carefully", "engineered", "fast", "linear", "classifier", "training", "algorithm", ",", "this", "procedure", "increases", "the", "run", "time", "of", "the", "pooling", "procedure", ".", "The", "query", "time", "per", "template", "comparison", "is", "also", "higher", "due", "to", "the", "high", "dimensionality", "of", "the", "input", "features", ".", "In", "contrast", ",", "the", "proposed", "approach", "requires", "a", "matrix", "multiplication", "and", "a", "vector", "dot", "product", "per", "comparison", ".", "By", "using", "a", "simple", "neural", "network", "architecture", ",", "a", "relatively", "smaller", "training", "dataset", "and", "a", "fast", "embedding", "method", "we", "have", "realized", "a", "faster", "and", "more", "efficient", "end", "-", "to", "-", "end", "system", ".", "To", "improve", "our", "performance", "further", ",", "we", "are", "currently", "incorporating", "the", "use", "of", "video", "data", "into", "our", "approach", ".", "subsubsection", ":", "Performance", "on", "CFP", "On", "the", "CFP", "dataset", ",", "we", "achieve", "a", "new", "state", "-", "of", "-", "art", "on", "both", "Frontal", "-", "Frontal", "and", "Frontal", "-", "Profile", "comparisons", ",", "the", "latter", "by", "a", "large", "margin", ".", "More", "specifically", ",", "for", "the", "Frontal", "-", "Profile", "case", ",", "we", "manage", "to", "reduce", "the", "error", "rate", "by", "40.8", "%", ".", "It", "should", "be", "noted", "that", "for", "a", "fair", "comparison", "we", "have", "used", "our", "raw", "CNN", "features", "without", "performing", "TPE", ".", "This", "shows", "that", "the", "raw", "CNN", "features", "we", "learn", "are", "effective", "even", "at", "extreme", "pose", "variations", ".", "section", ":", "Clustering", "Faces", ".5", ".5", ".5", ".5", "This", "section", "illustrates", "how", "the", "proposed", "TPE", "method", "can", "be", "used", "to", "cluster", "a", "given", "data", "collection", ".", "We", "perform", "two", "clustering", "experiments", ":", "We", "perform", "clustering", "on", "the", "entire", "LFW", "dataset", "that", "consists", "of", "13233", "images", "of", "5749", "subjects", ".", "It", "should", "be", "noted", "that", "about", "4169", "subjects", "have", "only", "one", "image", ".", "We", "use", "the", "IJB", "-", "A", "dataset", "and", "cluster", "the", "templates", "corresponding", "to", "the", "query", "set", "for", "each", "split", "in", "the", "IJB", "-", "A", "verify", "protocol", ".", "For", "evaluating", "the", "clustering", "results", ",", "we", "use", "the", "metrics", "defined", "in", ".", "These", "are", "summarized", "below", ":", "Pairwise", "Precision", "(", "P⁢pair", ")", ":", "The", "fraction", "of", "pairs", "of", "samples", "within", "a", "cluster", "among", "all", "possible", "pairs", "which", "are", "of", "the", "same", "class", ",", "over", "the", "total", "number", "of", "same", "cluster", "pairs", ".", "Pairwise", "Recall", "(", "R⁢pair", ")", ":", "The", "fraction", "of", "pairs", "of", "samples", "within", "a", "class", "among", "all", "possible", "pairs", "which", "are", "placed", "in", "the", "same", "cluster", ",", "over", "the", "total", "number", "of", "same", "-", "class", "pairs", ".", "Using", "these", "metrics", ",", "the", "F", "-", "score", "is", "computed", "as", ":", "The", "simplest", "way", "we", "found", "to", "demonstrate", "the", "effectiveness", "of", "our", "deep", "features", "and", "the", "proposed", "TPE", "method", ",", "is", "to", "use", "the", "standard", "MATLAB", "implementation", "of", "the", "agglomerative", "clustering", "algorithm", "with", "the", "average", "linkage", "metric", ".", "We", "use", "the", "cosine", "similarity", "as", "our", "basic", "clustering", "metric", ".", "The", "simple", "clustering", "algorithm", "that", "we", "have", "used", "here", "has", "computational", "complexity", "of", ".", "In", "its", "current", "form", ",", "this", "does", "not", "scale", "to", "large", "datasets", "with", "millions", "of", "images", ".", "We", "are", "currently", "working", "on", "a", "more", "efficient", "and", "scalable", "(", "yet", "approximate", ")", "version", "of", "this", "algorithm", ".", "paragraph", ":", "Clustering", "LFW:", "-", "The", "images", "in", "the", "LFW", "dataset", "are", "pre", "-", "processed", "as", "described", "in", "Section", "5.1", ".", "For", "each", "image", "and", "its", "flip", ",", "the", "deep", "features", "are", "extracted", "using", "the", "proposed", "architecture", ",", "averaged", "and", "normalized", "to", "unit", "norm", ".", "We", "run", "the", "clustering", "algorithm", "over", "the", "entire", "data", "in", "a", "single", "shot", ".", "The", "clustering", "algorithm", "takes", "as", "input", "a", "cut", "-", "off", "parameter", "which", "acts", "as", "a", "distance", "threshold", "(", "below", "which", "any", "two", "clusters", "will", "not", "be", "merged", ")", ".", "In", "our", "experiments", ",", "we", "vary", "this", "cut", "-", "off", "parameter", "over", "a", "small", "range", "and", "evaluate", "the", "resulting", "clustering", "using", "the", "-", "score", ".", "We", "pick", "the", "result", "that", "yields", "the", "best", "-", "score", ".", "Table", "[", "reference", "]", "shows", "the", "result", "of", "our", "approach", "and", "compares", "it", "to", "a", "recently", "released", "clustering", "approach", "based", "on", "approximate", "Rank", "-", "order", "clustering", ".", "It", "should", "be", "noted", "that", ",", "in", "the", "case", "of", ",", "the", "clustering", "result", "is", "chosen", "by", "varying", "the", "number", "of", "clusters", "and", "picking", "the", "one", "with", "the", "best", "-", "score", ".", "In", "our", "approach", ",", "we", "vary", "the", "cut", "-", "off", "threshold", "which", "is", "the", "property", "of", "deep", "features", "and", "hence", "is", "a", "more", "intuitive", "parameter", "to", "tune", ".", "We", "see", "from", "Table", "[", "reference", "]", "that", "aside", "from", "better", "performance", ",", "our", "total", "cluster", "estimate", "is", "closer", "to", "the", "ground", "truth", "value", "of", "5749", "than", ".", "table", "-", "score", "for", "comparison", "of", "the", "two", "clustering", "schemes", "on", "the", "LFW", "dataset", ".", "The", "ground", "truth", "cluster", "number", "is", "5749", ".", "tableClustering", "metrics", "over", "the", "IJB", "-", "A", "1:1", "protocol", ".", "The", "standard", "deviation", "is", "indicated", "in", "brackets", ".", "The", "ground", "truth", "subjects", "per", "each", "split", "is", "167", ".", "paragraph", ":", "Clustering", "IJB", "-", "A:", "-", "The", "IJB", "-", "A", "dataset", "is", "processed", "as", "described", "in", "Section", "5", ".", "In", "this", "section", ",", "we", "aim", "to", "cluster", "the", "query", "templates", "provided", "with", "each", "split", "for", "the", "verify", "protocol", ".", "We", "report", "the", "results", "of", "two", "experiments", ":", "with", "the", "raw", "CNN", "features", "(", "CNN", "in", "Table", "2", ")", "and", "with", "the", "projected", "CNN", "features", ",", "where", "the", "projection", "matrix", "is", "learned", "through", "the", "proposed", "TPE", "method", "(", "CNN", "+", "TPE", "in", "Table", "2", ")", ".", "The", "cut", "-", "off", "threshold", "required", "for", "our", "clustering", "algorithm", "is", "learned", "automatically", "based", "on", "the", "training", "data", ",", "i.e.", "we", "choose", "the", "threshold", "that", "gives", "the", "maximum", "-", "score", "over", "the", "training", "data", ".", "The", "scores", "reported", "in", "Table", "[", "reference", "]", "are", "average", "values", "over", "ten", "splits", ".", "As", "expected", ",", "the", "TPE", "method", "improves", "the", "clustering", "performance", "of", "raw", "features", ".", "The", "subject", "estimate", "is", "the", "number", "of", "clusters", "produced", "as", "a", "direct", "result", "of", "our", "clustering", "algorithm", ".", "The", "pruned", "estimate", "is", "obtained", "by", "ignoring", "clusters", "that", "have", "fewer", "than", "3", "images", ".", "For", "a", "more", "complete", "evaluation", "of", "our", "performance", "over", "varying", "threshold", "values", ",", "we", "plot", "the", "Precision", "-", "Recall", "(", "PR", ")", "curve", "for", "the", "IJB", "-", "A", "clustering", "experiment", "in", "Figure", "[", "reference", "]", ".", "As", "can", "be", "observed", ",", "the", "PR", "curve", "for", "clustering", "the", "IJB", "-", "A", "data", "using", "embedded", "features", "exhibits", "a", "better", "performance", "at", "all", "operating", "points", ".", "This", "is", "a", "more", "transparent", "evaluation", "than", "reporting", "only", "the", "-", "score", "since", "the", "latter", "effectively", "fixes", "the", "operating", "point", "but", "the", "PR", "curve", "reveals", "the", "performance", "at", "all", "operating", "points", ".", "section", ":", "Conclusion", "and", "Future", "Work", "In", "this", "paper", ",", "we", "proposed", "a", "deep", "CNN", "-", "based", "approach", "coupled", "with", "a", "low", "-", "dimensional", "discriminative", "embedding", "learned", "using", "triplet", "probability", "constraints", "in", "a", "large", "margin", "fashion", ".", "The", "proposed", "pipeline", "enables", "a", "faster", "training", "time", "and", "improves", "face", "verification", "performance", "especially", "at", "low", "FMRs", ".", "We", "demonstrated", "the", "effectiveness", "of", "the", "proposed", "method", "on", "two", "challenging", "datasets", ":", "IJB", "-", "A", "and", "CFP", "and", "achieved", "performance", "close", "to", "the", "state", "of", "the", "art", "while", "using", "a", "deep", "model", "which", "is", "more", "compact", "and", "trained", "using", "a", "moderately", "sized", "dataset", ".", "We", "demonstrated", "the", "robustness", "of", "our", "features", "using", "a", "simple", "clustering", "algorithm", "on", "the", "LFW", "and", "IJB", "-", "A", "datasets", ".", "For", "future", "work", ",", "we", "plan", "to", "use", "videos", "directly", "during", "training", "and", "also", "embed", "our", "TPE", "approach", "into", "training", "the", "deep", "network", ".", "We", "intend", "to", "scale", "our", "clustering", "algorithm", "to", "handle", "large", "scale", "scenarios", "such", "as", "large", "impostor", "sets", "of", "the", "order", "of", "millions", ".", "section", ":", "Acknowledgement", "This", "research", "is", "based", "upon", "work", "supported", "by", "the", "Office", "of", "the", "Director", "of", "National", "Intelligence", "(", "ODNI", ")", ",", "Intelligence", "Advanced", "Research", "Projects", "Activity", "(", "IARPA", "),", "via", "IARPA", "R", "&", "D", "Contract", "No", ".", "2014", "-", "14071600012", ".", "The", "views", "and", "conclusions", "contained", "herein", "are", "those", "of", "the", "authors", "and", "should", "not", "be", "interpreted", "as", "necessarily", "representing", "the", "official", "policies", "or", "endorsements", ",", "either", "expressed", "or", "implied", ",", "of", "the", "ODNI", ",", "IARPA", ",", "or", "the", "U.S.", "Government", ".", "The", "U.S.", "Government", "is", "authorized", "to", "reproduce", "and", "distribute", "reprints", "for", "Governmental", "purposes", "notwithstanding", "any", "copyright", "annotation", "thereon", ".", "bibliography", ":", "References"], "entity": [[2, 5, ["Triplet", "Probabilistic", "Embedding"], "Method"], [6, 8, ["Face", "Verification"], "Task"], [9, 10, ["Clustering"], "Task"], [21, 24, ["unconstrained", "face", "verification"], "Task"], [44, 50, ["low", "-", "dimensional", "discriminative", "embedding", "step"], "Method"], [59, 63, ["unconstrained", "face", "verification", "problem"], "Task"], [78, 79, ["memory"], "Metric"], [81, 85, ["post", "-", "processing", "operations"], "Task"], [86, 89, ["subject", "specific", "clustering"], "Method"], [94, 98, ["IJB", "-", "A", "dataset"], "Material"], [113, 117, ["verification", "and", "identification", "metrics"], "Metric"], [125, 129, ["training", "/", "test", "time"], "Metric"], [144, 145, ["representation"], "Method"], [179, 180, ["clustering"], "Method"], [183, 186, ["IJB", "-", "A"], "Material"], [187, 189, ["LFW", "datasets"], "Material"], [203, 208, ["Labeled", "faces", "in", "the", "Wild"], "Material"], [209, 210, ["LFW"], "Material"], [214, 216, ["learning", "algorithms"], "Method"], [217, 220, ["Deep", "neural", "nets"], "Method"], [227, 231, ["unconstrained", "face", "verification", "problem"], "Task"], [236, 239, ["face", "verification", "algorithm"], "Method"], [253, 255, ["face", "verification"], "Task"], [279, 282, ["unconstrained", "face", "verification"], "Task"], [299, 303, ["IJB", "-", "A", "dataset"], "Material"], [335, 337, ["training", "time"], "Metric"], [346, 348, ["Deep", "CNN"], "Method"], [349, 350, ["DCNN"], "Method"], [355, 356, ["training"], "Task"], [385, 388, ["deep", "network", "architecture"], "Method"], [390, 392, ["training", "scheme"], "Method"], [395, 397, ["training", "time"], "Metric"], [401, 406, ["triplet", "probability", "embedding", "learning", "method"], "Method"], [414, 416, ["face", "verification"], "Task"], [417, 419, ["subject", "clustering"], "Task"], [421, 422, ["training"], "Task"], [433, 435, ["deep", "architecture"], "Method"], [453, 455, ["deep", "network"], "Method"], [468, 472, ["IJB", "-", "A", "dataset"], "Material"], [503, 506, ["triplet", "probability", "embedding"], "Method"], [545, 550, ["automatic", "pre", "-", "processing", "steps"], "Method"], [552, 554, ["face", "detection"], "Task"], [555, 557, ["fiducial", "extraction"], "Task"], [624, 626, ["face", "verification"], "Task"], [633, 635, ["network", "architecture"], "Method"], [637, 639, ["training", "scheme"], "Method"], [641, 646, ["triplet", "probabilistic", "embedding", "learning", "method"], "Method"], [657, 660, ["IJB", "-", "A"], "Material"], [693, 694, ["LFW"], "Material"], [695, 699, ["IJB", "-", "A", "datasets"], "Material"], [723, 725, ["face", "verification"], "Task"], [727, 729, ["DeepFace", "approach"], "Method"], [733, 736, ["3D", "alignment", "procedure"], "Method"], [745, 747, ["deep", "network"], "Method"], [759, 760, ["Facenet"], "Material"], [768, 771, ["deep", "network", "models"], "Method"], [779, 781, ["training", "time"], "Metric"], [797, 801, ["IJB", "-", "A", "dataset"], "Material"], [810, 811, ["verification"], "Task"], [822, 824, ["deep", "networks"], "Method"], [826, 830, ["CASIA", "-", "WebFace", "dataset"], "Material"], [832, 836, ["VGG", "-", "Face", "dataset"], "Material"], [840, 842, ["training", "time"], "Metric"], [847, 849, ["network", "architecture"], "Method"], [851, 853, ["training", "scheme"], "Method"], [856, 858, ["training", "time"], "Metric"], [861, 863, ["query", "time"], "Metric"], [869, 873, ["compact", "and", "discriminative", "representation"], "Method"], [885, 893, ["Semi", "Definite", "Programming", "(", "SDP", ")-", "based", "formulation"], "Method"], [905, 908, ["large", "margin", "framework"], "Method"], [919, 921, ["face", "verification"], "Task"], [928, 931, ["deep", "network", "architecture"], "Method"], [935, 939, ["Joint", "Bayesian", "metric", "learning"], "Method"], [946, 948, ["face", "verification"], "Task"], [975, 977, ["discriminative", "clustering"], "Method"], [978, 979, ["hashing"], "Method"], [983, 986, ["t", "-", "SNE"], "Method"], [987, 990, ["t", "-", "STE"], "Method"], [991, 994, ["Crowd", "Kernel", "Learning"], "Method"], [995, 996, ["CKL"], "Method"], [1079, 1081, ["classification", "setting"], "Task"], [1095, 1097, ["optimization", "problem"], "Task"], [1103, 1105, ["dimensionality", "reduction"], "Task"], [1109, 1111, ["discriminative", "ability"], "Metric"], [1117, 1119, ["embedding", "scheme"], "Method"], [1144, 1146, ["Network", "Architecture"], "Method"], [1152, 1154, ["training", "algorithm"], "Method"], [1156, 1158, ["deep", "network"], "Method"], [1168, 1170, ["convolutional", "layers"], "Method"], [1207, 1210, ["face", "recognition", "tasks"], "Task"], [1219, 1223, ["Parametric", "Rectifier", "Linear", "units"], "Method"], [1224, 1225, ["PReLUs"], "Method"], [1228, 1229, ["ReLUs"], "Method"], [1251, 1253, ["convergence", "rate"], "Metric"], [1254, 1257, ["tableDeep", "Network", "architecture"], "Method"], [1261, 1263, ["convolutional", "layers"], "Method"], [1264, 1267, ["conv1", "-", "conv3"], "Method"], [1275, 1277, ["AlexNet", "model"], "Method"], [1280, 1283, ["ImageNet", "challenge", "dataset"], "Material"], [1331, 1333, ["deep", "models"], "Method"], [1342, 1350, ["last", "layer", "of", "fully", "pre", "-", "trained", "models"], "Method"], [1356, 1358, ["training", "time"], "Metric"], [1371, 1373, ["face", "verification"], "Task"], [1392, 1396, ["pre", "-", "trained", "model"], "Method"], [1397, 1398, ["AlexNet"], "Method"], [1403, 1405, ["deep", "network"], "Method"], [1414, 1416, ["convolutional", "layers"], "Method"], [1452, 1454, ["convolutional", "layers"], "Method"], [1464, 1467, ["conv4", "-", "conv7"], "Method"], [1501, 1505, ["IJB", "-", "A", "dataset"], "Material"], [1523, 1525, ["deep", "network"], "Method"], [1529, 1533, ["CASIA", "-", "WebFace", "dataset"], "Material"], [1548, 1551, ["conv4", "-", "conv7"], "Method"], [1553, 1559, ["fully", "connected", "layers", "fc6", "-", "fc8"], "Method"], [1564, 1567, ["random", "Gaussian", "distributions"], "Method"], [1584, 1586, ["feature", "extractor"], "Method"], [1588, 1591, ["last", "layer", "fc8"], "Method"], [1626, 1628, ["fc7", "layer"], "Method"], [1631, 1632, ["dimensionality"], "Metric"], [1641, 1644, ["Softmax", "loss", "function"], "Method"], [1645, 1647, ["multiclass", "classification"], "Task"], [1649, 1653, ["Caffe", "deep", "learning", "platform"], "Method"], [1658, 1660, ["Discriminative", "Embedding"], "Method"], [1671, 1675, ["low", "-", "dimensional", "embedding"], "Task"], [1698, 1699, ["memory"], "Metric"], [1701, 1705, ["post", "-", "processing", "operations"], "Task"], [1706, 1707, ["visualization"], "Task"], [1708, 1709, ["clustering"], "Task"], [1837, 1838, ["optimization"], "Task"], [1864, 1865, ["NLL"], "Method"], [1881, 1885, ["Large", "-", "Margin", "framework"], "Method"], [1886, 1889, ["Stochastic", "Gradient", "Descent"], "Method"], [1890, 1891, ["SGD"], "Method"], [1900, 1902, ["gradient", "update"], "Method"], [1930, 1932, ["learning", "rate"], "Metric"], [1943, 1945, ["dimensionality", "reduction"], "Task"], [1959, 1961, ["cross", "validation"], "Metric"], [1964, 1965, ["dimensionality"], "Metric"], [1975, 1977, ["principal", "components"], "Method"], [2005, 2008, ["hard", "negative", "mining"], "Task"], [2053, 2056, ["Triplet", "Probability", "Embedding"], "Method"], [2057, 2058, ["TPE"], "Method"], [2082, 2083, ["embedding"], "Task"], [2114, 2117, ["Triplet", "Distance", "Embedding"], "Method"], [2118, 2119, ["TDE"], "Method"], [2139, 2141, ["gradient", "update"], "Method"], [2143, 2145, ["TDE", "method"], "Method"], [2233, 2235, ["gradient", "update"], "Method"], [2236, 2237, ["TPE"], "Method"], [2263, 2265, ["gradient", "update"], "Method"], [2304, 2305, ["projection"], "Method"], [2307, 2308, ["TDE"], "Method"], [2310, 2311, ["TPE"], "Method"], [2320, 2322, ["ROC", "curve"], "Metric"], [2323, 2324, ["TAR"], "Metric"], [2327, 2328, ["FAR"], "Metric"], [2350, 2357, ["Equal", "Error", "Rate", "(", "EER", ")", "metric"], "Metric"], [2368, 2369, ["TPE"], "Method"], [2390, 2394, ["IJB", "-", "A", "dataset"], "Material"], [2414, 2419, ["IARPA", "Janus", "Benchmark", "-", "A"], "Material"], [2420, 2423, ["IJB", "-", "A"], "Material"], [2442, 2444, ["still", "images"], "Material"], [2464, 2468, ["IJB", "-", "A", "dataset"], "Material"], [2477, 2478, ["LFW"], "Material"], [2484, 2488, ["IJB", "-", "A", "dataset"], "Material"], [2545, 2546, ["pooling"], "Method"], [2552, 2554, ["Average", "pooling"], "Method"], [2555, 2556, ["CNN"], "Method"], [2577, 2579, ["componentwise", "average"], "Method"], [2595, 2597, ["Media", "pooling"], "Method"], [2598, 2599, ["CNN"], "Method"], [2620, 2623, ["IJB", "-", "A"], "Material"], [2648, 2652, ["intra", "-", "media", "average"], "Method"], [2659, 2663, ["inter", "-", "media", "average"], "Method"], [2707, 2710, ["face", "verification", "approaches"], "Method"], [2711, 2713, ["pose", "variation"], "Task"], [2773, 2776, ["Pre", "-", "processing"], "Task"], [2778, 2780, ["training", "phase"], "Task"], [2789, 2791, ["HyperFace", "method"], "Method"], [2792, 2794, ["face", "detection"], "Task"], [2795, 2798, ["fiducial", "point", "extraction"], "Task"], [2800, 2802, ["HyperFace", "detector"], "Method"], [2813, 2817, ["IJB", "-", "A", "dataset"], "Material"], [2847, 2849, ["automatic", "detections"], "Task"], [2869, 2873, ["IJB", "-", "A", "dataset"], "Material"], [2881, 2883, ["HyperFace", "detector"], "Method"], [2912, 2913, ["HyperFace"], "Method"], [2940, 2942, ["similarity", "transform"], "Method"], [2976, 2977, ["Verification"], "Task"], [2980, 2984, ["IJB", "-", "A", "dataset"], "Material"], [2986, 2987, ["identification"], "Task"], [3086, 3088, ["deep", "architecture"], "Method"], [3091, 3092, ["SGD"], "Method"], [3093, 3094, ["momentum"], "Method"], [3102, 3104, ["learning", "rate"], "Metric"], [3146, 3148, ["training", "time"], "Metric"], [3150, 3152, ["deep", "network"], "Method"], [3164, 3168, ["IJB", "-", "A", "dataset"], "Material"], [3181, 3183, ["triplet", "embedding"], "Method"], [3195, 3197, ["splitwise", "processing"], "Task"], [3206, 3207, ["deployment"], "Task"], [3210, 3212, ["enrollment", "time"], "Metric"], [3215, 3218, ["pre", "-", "processing"], "Task"], [3220, 3221, ["alignment"], "Task"], [3222, 3224, ["feature", "extraction"], "Task"], [3229, 3231, ["Evaluation", "Pipeline"], "Metric"], [3271, 3274, ["cosine", "similarity", "score"], "Metric"], [3280, 3284, ["IJB", "-", "A", "dataset"], "Material"], [3298, 3300, ["average", "pooling"], "Method"], [3301, 3303, ["media", "pooling"], "Method"], [3306, 3308, ["vector", "representation"], "Method"], [3316, 3318, ["TPE", "projection"], "Method"], [3333, 3336, ["cosine", "similarity", "score"], "Metric"], [3338, 3343, ["projected", "128", "-", "dimensional", "representations"], "Method"], [3348, 3350, ["Evaluation", "Metrics"], "Metric"], [3358, 3362, ["IJB", "-", "A", "dataset"], "Material"], [3363, 3364, ["Verification"], "Task"], [3365, 3366, ["Identification"], "Task"], [3369, 3371, ["verification", "protocol"], "Task"], [3375, 3384, ["False", "Non", "-", "Match", "Rate", "(", "FNMR", ")", "values"], "Metric"], [3386, 3389, ["False", "Match", "Rates"], "Metric"], [3390, 3391, ["FMR"], "Method"], [3395, 3396, ["identification"], "Task"], [3400, 3406, ["open", "set", "and", "closed", "set", "metrics"], "Metric"], [3409, 3412, ["open", "set", "metrics"], "Metric"], [3414, 3418, ["True", "Positive", "Identification", "Rate"], "Metric"], [3441, 3444, ["closed", "set", "metrics"], "Metric"], [3448, 3450, ["CMC", "numbers"], "Metric"], [3454, 3458, ["False", "Positive", "Identification", "Rates"], "Metric"], [3459, 3460, ["FPIRs"], "Metric"], [3462, 3463, ["Ranks"], "Metric"], [3468, 3470, ["evaluation", "metrics"], "Metric"], [3500, 3501, ["AUC"], "Metric"], [3503, 3506, ["Equal", "Error", "Rate"], "Metric"], [3507, 3508, ["EER"], "Metric"], [3519, 3521, ["classification", "accuracy"], "Metric"], [3525, 3526, ["accuracy"], "Metric"], [3548, 3550, ["classification", "accuracy"], "Metric"], [3565, 3568, ["IJB", "-", "A"], "Material"], [3600, 3608, ["Government", "-", "of", "-", "the", "-", "Shelf", "("], "Method"], [3608, 3609, ["GOTS"], "Method"], [3618, 3622, ["IJB", "-", "A", "dataset"], "Material"], [3630, 3632, ["deep", "network"], "Method"], [3638, 3642, ["VGG", "-", "Face", "dataset"], "Material"], [3652, 3655, ["Neural", "Aggregation", "network"], "Method"], [3656, 3657, ["NAN"], "Method"], [3667, 3671, ["CELEB", "-", "1000", "dataset"], "Material"], [3674, 3676, ["GoogleNet", "architecture"], "Method"], [3698, 3701, ["3D", "rendered", "images"], "Material"], [3730, 3734, ["IJB", "-", "A", "dataset"], "Material"], [3750, 3752, ["template", "adaptation"], "Method"], [3763, 3767, ["IJB", "-", "A", "dataset"], "Material"], [3783, 3787, ["CASIA", "-", "WebFace", "dataset"], "Material"], [3798, 3800, ["training", "time"], "Metric"], [3805, 3807, ["query", "time"], "Metric"], [3810, 3812, ["face", "detection"], "Task"], [3830, 3832, ["media", "pooling"], "Method"], [3841, 3845, ["verification", "and", "identification", "protocols"], "Task"], [3847, 3851, ["IJB", "-", "A", "dataset"], "Material"], [3857, 3860, ["template", "adaptation", "method"], "Method"], [3871, 3873, ["TPE", "method"], "Method"], [3878, 3882, ["identification", "and", "verification", "tasks"], "Task"], [3899, 3904, ["VGG", "-", "Face", "network", "descriptors"], "Method"], [3919, 3921, ["template", "adaptation"], "Method"], [3940, 3942, ["linear", "SVM"], "Method"], [3963, 3966, ["IJB", "-", "A"], "Material"], [3992, 3994, ["similarity", "score"], "Metric"], [4005, 4010, ["fast", "linear", "classifier", "training", "algorithm"], "Method"], [4015, 4017, ["run", "time"], "Metric"], [4019, 4021, ["pooling", "procedure"], "Method"], [4023, 4025, ["query", "time"], "Metric"], [4026, 4028, ["template", "comparison"], "Task"], [4049, 4051, ["matrix", "multiplication"], "Method"], [4063, 4066, ["neural", "network", "architecture"], "Method"], [4075, 4077, ["embedding", "method"], "Method"], [4089, 4091, ["end", "system"], "Task"], [4111, 4112, ["subsubsection"], "Method"], [4115, 4116, ["CFP"], "Material"], [4132, 4140, ["Frontal", "-", "Frontal", "and", "Frontal", "-", "Profile", "comparisons"], "Task"], [4153, 4157, ["Frontal", "-", "Profile", "case"], "Task"], [4163, 4165, ["error", "rate"], "Metric"], [4187, 4188, ["TPE"], "Method"], [4220, 4222, ["TPE", "method"], "Method"], [4229, 4231, ["data", "collection"], "Task"], [4235, 4236, ["clustering"], "Task"], [4240, 4241, ["clustering"], "Method"], [4271, 4275, ["IJB", "-", "A", "dataset"], "Material"], [4298, 4299, ["clustering"], "Task"], [4313, 4315, ["Pairwise", "Precision"], "Metric"], [4348, 4350, ["Pairwise", "Recall"], "Metric"], [4390, 4393, ["F", "-", "score"], "Metric"], [4408, 4410, ["deep", "features"], "Method"], [4413, 4415, ["TPE", "method"], "Method"], [4421, 4423, ["MATLAB", "implementation"], "Method"], [4425, 4428, ["agglomerative", "clustering", "algorithm"], "Method"], [4430, 4433, ["average", "linkage", "metric"], "Method"], [4442, 4444, ["clustering", "metric"], "Metric"], [4447, 4449, ["clustering", "algorithm"], "Method"], [4455, 4457, ["computational", "complexity"], "Metric"], [4543, 4545, ["clustering", "algorithm"], "Method"], [4555, 4557, ["clustering", "algorithm"], "Method"], [4602, 4603, ["clustering"], "Method"], [4636, 4638, ["clustering", "approach"], "Method"], [4640, 4645, ["approximate", "Rank", "-", "order", "clustering"], "Method"], [4658, 4659, ["clustering"], "Task"], [4721, 4723, ["cluster", "estimate"], "Method"], [4742, 4744, ["clustering", "schemes"], "Method"], [4750, 4754, ["ground", "truth", "cluster", "number"], "Metric"], [4757, 4759, ["tableClustering", "metrics"], "Metric"], [4788, 4792, ["IJB", "-", "A:", "-"], "Material"], [4793, 4797, ["IJB", "-", "A", "dataset"], "Material"], [4822, 4824, ["verify", "protocol"], "Method"], [4839, 4840, ["CNN"], "Method"], [4860, 4862, ["TPE", "method"], "Method"], [4863, 4864, ["CNN"], "Method"], [4865, 4866, ["TPE"], "Method"], [4879, 4881, ["clustering", "algorithm"], "Method"], [4925, 4927, ["TPE", "method"], "Method"], [4929, 4930, ["clustering"], "Task"], [4950, 4952, ["clustering", "algorithm"], "Method"], [4954, 4956, ["pruned", "estimate"], "Method"], [4984, 4987, ["Precision", "-", "Recall"], "Metric"], [4988, 4989, ["PR"], "Metric"], [4993, 4997, ["IJB", "-", "A", "clustering"], "Method"], [5013, 5014, ["clustering"], "Task"], [5084, 5089, ["low", "-", "dimensional", "discriminative", "embedding"], "Method"], [5106, 5108, ["training", "time"], "Metric"], [5110, 5112, ["face", "verification"], "Task"], [5116, 5117, ["FMRs"], "Metric"], [5131, 5134, ["IJB", "-", "A"], "Material"], [5135, 5136, ["CFP"], "Material"], [5149, 5151, ["deep", "model"], "Method"], [5173, 5175, ["clustering", "algorithm"], "Method"], [5200, 5202, ["TPE", "approach"], "Method"], [5205, 5207, ["deep", "network"], "Method"], [5213, 5215, ["clustering", "algorithm"], "Method"], [5217, 5220, ["large", "scale", "scenarios"], "Task"], [38, 39, ["CNN"], "Method"], [139, 140, ["CFP"], "Material"], [149, 150, ["CNN"], "Method"], [272, 273, ["LFW"], "Material"], [318, 319, ["CNN"], "Method"], [520, 521, ["DCNN"], "Method"], [541, 542, ["CNN"], "Method"], [661, 662, ["CFP"], "Material"], [1491, 1492, ["CNN"], "Method"], [1497, 1498, ["CNN"], "Method"], [2334, 2337, ["IJB", "-", "A"], "Material"], [2376, 2377, ["FAR"], "Metric"], [2501, 2504, ["IJB", "-", "A"], "Material"], [2536, 2539, ["IJB", "-", "A"], "Material"], [2686, 2687, ["CFP"], "Material"], [2758, 2759, ["CFP"], "Material"], [2945, 2946, ["CFP"], "Material"], [3044, 3045, ["CFP"], "Material"], [3472, 3475, ["IJB", "-", "A"], "Material"], [3483, 3484, ["CFP"], "Material"], [3533, 3534, ["CNN"], "Method"], [3585, 3588, ["IJB", "-", "A"], "Material"], [3684, 3685, ["CNN"], "Method"], [3779, 3780, ["CNN"], "Method"], [3827, 3828, ["CNN"], "Method"], [4118, 4119, ["CFP"], "Material"], [4183, 4184, ["CNN"], "Method"], [4194, 4195, ["CNN"], "Method"], [4208, 4209, ["Clustering"], "Task"], [4244, 4245, ["LFW"], "Material"], [4289, 4292, ["IJB", "-", "A"], "Material"], [4497, 4498, ["Clustering"], "Task"], [4504, 4505, ["LFW"], "Material"], [4746, 4747, ["LFW"], "Material"], [4761, 4764, ["IJB", "-", "A"], "Material"], [4787, 4788, ["Clustering"], "Task"], [4836, 4837, ["CNN"], "Method"], [4848, 4849, ["CNN"], "Method"], [5010, 5011, ["PR"], "Metric"], [5015, 5018, ["IJB", "-", "A"], "Material"], [5053, 5054, ["PR"], "Metric"], [5077, 5078, ["CNN"], "Method"], [5177, 5178, ["LFW"], "Material"], [5179, 5182, ["IJB", "-", "A"], "Material"]]} {"doc_id": "45429c281e30f9e87ebcd1ae42e0656d2ead24d1", "text": ["document", ":", "High", "-", "Resolution", "Image", "Synthesis", "and", "Semantic", "Manipulation", "with", "Conditional", "GANs", "We", "present", "a", "new", "method", "for", "synthesizing", "high", "-", "resolution", "photo", "-", "realistic", "images", "from", "semantic", "label", "maps", "using", "conditional", "generative", "adversarial", "networks", "(", "conditional", "GANs", ")", ".", "Conditional", "GANs", "have", "enabled", "a", "variety", "of", "applications", ",", "but", "the", "results", "are", "often", "limited", "to", "low", "-", "resolution", "and", "still", "far", "from", "realistic", ".", "In", "this", "work", ",", "we", "generate", "visually", "appealing", "results", "with", "a", "novel", "adversarial", "loss", ",", "as", "well", "as", "new", "multi", "-", "scale", "generator", "and", "discriminator", "architectures", ".", "Furthermore", ",", "we", "extend", "our", "framework", "to", "interactive", "visual", "manipulation", "with", "two", "additional", "features", ".", "First", ",", "we", "incorporate", "object", "instance", "segmentation", "information", ",", "which", "enables", "object", "manipulations", "such", "as", "removing", "/", "adding", "objects", "and", "changing", "the", "object", "category", ".", "Second", ",", "we", "propose", "a", "method", "to", "generate", "diverse", "results", "given", "the", "same", "input", ",", "allowing", "users", "to", "edit", "the", "object", "appearance", "interactively", ".", "Human", "opinion", "studies", "demonstrate", "that", "our", "method", "significantly", "outperforms", "existing", "methods", ",", "advancing", "both", "the", "quality", "and", "the", "resolution", "of", "deep", "image", "synthesis", "and", "editing", ".", "section", ":", "Introduction", "Photo", "-", "realistic", "image", "rendering", "using", "standard", "graphics", "techniques", "is", "involved", ",", "since", "geometry", ",", "materials", ",", "and", "light", "transport", "must", "be", "simulated", "explicitly", ".", "Although", "existing", "graphics", "algorithms", "excel", "at", "the", "task", ",", "building", "and", "editing", "virtual", "environments", "is", "expensive", "and", "time", "-", "consuming", ".", "That", "is", "because", "we", "have", "to", "model", "every", "aspect", "of", "the", "world", "explicitly", ".", "If", "we", "were", "able", "to", "render", "photo", "-", "realistic", "images", "using", "a", "model", "learned", "from", "data", ",", "we", "could", "turn", "the", "process", "of", "graphics", "rendering", "into", "a", "model", "learning", "and", "inference", "problem", ".", "Then", ",", "we", "could", "simplify", "the", "process", "of", "creating", "new", "virtual", "worlds", "by", "training", "models", "on", "new", "datasets", ".", "We", "could", "even", "make", "it", "easier", "to", "customize", "environments", "by", "allowing", "users", "to", "simply", "specify", "overall", "semantic", "structure", "rather", "than", "modeling", "geometry", ",", "materials", ",", "or", "lighting", ".", "In", "this", "paper", ",", "we", "discuss", "a", "new", "approach", "that", "produces", "high", "-", "resolution", "images", "from", "semantic", "label", "maps", ".", "This", "method", "has", "a", "wide", "range", "of", "applications", ".", "For", "example", ",", "we", "can", "use", "it", "to", "create", "synthetic", "training", "data", "for", "training", "visual", "recognition", "algorithms", ",", "since", "it", "is", "much", "easier", "to", "create", "semantic", "labels", "for", "desired", "scenarios", "than", "to", "generate", "training", "images", ".", "Using", "semantic", "segmentation", "methods", ",", "we", "can", "transform", "images", "into", "a", "semantic", "label", "domain", ",", "edit", "the", "objects", "in", "the", "label", "domain", ",", "and", "then", "transform", "them", "back", "to", "the", "image", "domain", ".", "This", "method", "also", "gives", "us", "new", "tools", "for", "higher", "-", "level", "image", "editing", ",", "e.g.", ",", "adding", "objects", "to", "images", "or", "changing", "the", "appearance", "of", "existing", "objects", ".", "To", "synthesize", "images", "from", "semantic", "labels", ",", "one", "can", "use", "the", "pix2pix", "method", ",", "an", "image", "-", "to", "-", "image", "translation", "framework", "which", "leverages", "generative", "adversarial", "networks", "(", "GANs", ")", "in", "a", "conditional", "setting", ".", "Recently", ",", "Chen", "and", "Koltun", "suggest", "that", "adversarial", "training", "might", "be", "unstable", "and", "prone", "to", "failure", "for", "high", "-", "resolution", "image", "generation", "tasks", ".", "Instead", ",", "they", "adopt", "a", "modified", "perceptual", "loss", "to", "synthesize", "images", ",", "which", "are", "high", "-", "resolution", "but", "often", "lack", "fine", "details", "and", "realistic", "textures", ".", "Here", "we", "address", "two", "main", "issues", "of", "the", "above", "state", "-", "of", "-", "the", "-", "art", "methods", ":", "(", "1", ")", "the", "difficulty", "of", "generating", "high", "-", "resolution", "images", "with", "GANs", "and", "(", "2", ")", "the", "lack", "of", "details", "and", "realistic", "textures", "in", "the", "previous", "high", "-", "resolution", "results", ".", "We", "show", "that", "through", "a", "new", ",", "robust", "adversarial", "learning", "objective", "together", "with", "new", "multi", "-", "scale", "generator", "and", "discriminator", "architectures", ",", "we", "can", "synthesize", "photo", "-", "realistic", "images", "at", "resolution", ",", "which", "are", "more", "visually", "appealing", "than", "those", "computed", "by", "previous", "methods", ".", "We", "first", "obtain", "our", "results", "with", "adversarial", "training", "only", ",", "without", "relying", "on", "any", "hand", "-", "crafted", "losses", "or", "pre", "-", "trained", "networks", "(", "e.g.", "VGGNet", ")", "for", "perceptual", "losses", "(", "Figs", ".", "[", "reference", "]", "c", ",", "[", "reference", "]", "b", ")", ".", "Then", "we", "show", "that", "adding", "perceptual", "losses", "from", "pre", "-", "trained", "networks", "can", "slightly", "improve", "the", "results", "in", "some", "circumstances", "(", "Figs", ".", "[", "reference", "]", "d", ",", "[", "reference", "]", "c", ")", ",", "if", "a", "pre", "-", "trained", "network", "is", "available", ".", "Both", "results", "outperform", "previous", "works", "substantially", "in", "terms", "of", "image", "quality", ".", "Furthermore", ",", "to", "support", "interactive", "semantic", "manipulation", ",", "we", "extend", "our", "method", "in", "two", "directions", ".", "First", ",", "we", "use", "instance", "-", "level", "object", "segmentation", "information", ",", "which", "can", "separate", "different", "object", "instances", "within", "the", "same", "category", ".", "This", "enables", "flexible", "object", "manipulations", ",", "such", "as", "adding", "/", "removing", "objects", "and", "changing", "object", "types", ".", "Second", ",", "we", "propose", "a", "method", "to", "generate", "diverse", "results", "given", "the", "same", "input", "label", "map", ",", "allowing", "the", "user", "to", "edit", "the", "appearance", "of", "the", "same", "object", "interactively", ".", "We", "compare", "against", "state", "-", "of", "-", "the", "-", "art", "visual", "synthesis", "systems", ",", "and", "show", "that", "our", "method", "outperforms", "these", "approaches", "regarding", "both", "quantitative", "evaluations", "and", "human", "perception", "studies", ".", "We", "also", "perform", "an", "ablation", "study", "regarding", "the", "training", "objectives", "and", "the", "importance", "of", "instance", "-", "level", "segmentation", "information", ".", "In", "addition", "to", "semantic", "manipulation", ",", "we", "test", "our", "method", "on", "edge2photo", "applications", "(", "Figs", ".", "[", "reference", "]", ",", "[", "reference", "]", ")", ",", "which", "shows", "the", "generalizability", "of", "our", "approach", ".", "Code", "and", "data", "are", "available", "at", "our", ".", "section", ":", "Related", "Work", "paragraph", ":", "Generative", "adversarial", "networks", "Generative", "adversarial", "networks", "(", "GANs", ")", "aim", "to", "model", "the", "natural", "image", "distribution", "by", "forcing", "the", "generated", "samples", "to", "be", "indistinguishable", "from", "natural", "images", ".", "GANs", "enable", "a", "wide", "variety", "of", "applications", "such", "as", "image", "generation", ",", "representation", "learning", ",", "image", "manipulation", ",", "object", "detection", ",", "and", "video", "applications", ".", "Various", "coarse", "-", "to", "-", "fine", "schemes", "have", "been", "proposed", "to", "synthesize", "larger", "images", "(", "e.g.", ")", "in", "an", "unconditional", "setting", ".", "Inspired", "by", "their", "successes", ",", "we", "propose", "a", "new", "coarse", "-", "to", "-", "fine", "generator", "and", "multi", "-", "scale", "discriminator", "architectures", "suitable", "for", "conditional", "image", "generation", "at", "a", "much", "higher", "resolution", ".", "paragraph", ":", "Image", "-", "to", "-", "image", "translation", "Many", "researchers", "have", "leveraged", "adversarial", "learning", "for", "image", "-", "to", "-", "image", "translation", ",", "whose", "goal", "is", "to", "translate", "an", "input", "image", "from", "one", "domain", "to", "another", "domain", "given", "input", "-", "output", "image", "pairs", "as", "training", "data", ".", "Compared", "to", "loss", ",", "which", "often", "leads", "to", "blurry", "images", ",", "the", "adversarial", "loss", "has", "become", "a", "popular", "choice", "for", "many", "image", "-", "to", "-", "image", "tasks", ".", "The", "reason", "is", "that", "the", "discriminator", "can", "learn", "a", "trainable", "loss", "function", "and", "automatically", "adapt", "to", "the", "differences", "between", "the", "generated", "and", "real", "images", "in", "the", "target", "domain", ".", "For", "example", ",", "the", "recent", "pix2pix", "framework", "used", "image", "-", "conditional", "GANs", "for", "different", "applications", ",", "such", "as", "transforming", "Google", "maps", "to", "satellite", "views", "and", "generating", "cats", "from", "user", "sketches", ".", "Various", "methods", "have", "also", "been", "proposed", "to", "learn", "an", "image", "-", "to", "-", "image", "translation", "in", "the", "absence", "of", "training", "pairs", ".", "Recently", ",", "Chen", "and", "Koltun", "suggest", "that", "it", "might", "be", "hard", "for", "conditional", "GANs", "to", "generate", "high", "-", "resolution", "images", "due", "to", "the", "training", "instability", "and", "optimization", "issues", ".", "To", "avoid", "this", "difficulty", ",", "they", "use", "a", "direct", "regression", "objective", "based", "on", "a", "perceptual", "loss", "and", "produce", "the", "first", "model", "that", "can", "synthesize", "images", ".", "The", "generated", "results", "are", "high", "-", "resolution", "but", "often", "lack", "fine", "details", "and", "realistic", "textures", ".", "Motivated", "by", "their", "success", ",", "we", "show", "that", "using", "our", "new", "objective", "function", "as", "well", "as", "novel", "multi", "-", "scale", "generators", "and", "discriminators", ",", "we", "not", "only", "largely", "stabilize", "the", "training", "of", "conditional", "GANs", "on", "high", "-", "resolution", "images", ",", "but", "also", "achieve", "significantly", "better", "results", "compared", "to", "Chen", "and", "Koltun", ".", "Side", "-", "by", "-", "side", "comparisons", "clearly", "show", "our", "advantage", "(", "Figs", ".", "[", "reference", "]", ",", "[", "reference", "]", ",", "[", "reference", "]", ",", "[", "reference", "]", ")", ".", "paragraph", ":", "Deep", "visual", "manipulation", "Recently", ",", "deep", "neural", "networks", "have", "obtained", "promising", "results", "in", "various", "image", "processing", "tasks", ",", "such", "as", "style", "transfer", ",", "inpainting", ",", "colorization", ",", "and", "restoration", ".", "However", ",", "most", "of", "these", "works", "lack", "an", "interface", "for", "users", "to", "adjust", "the", "current", "result", "or", "explore", "the", "output", "space", ".", "To", "address", "this", "issue", ",", "Zhu", "developed", "an", "optimization", "method", "for", "editing", "the", "object", "appearance", "based", "on", "the", "priors", "learned", "by", "GANs", ".", "Recent", "works", "also", "provide", "user", "interfaces", "for", "creating", "novel", "imagery", "from", "low", "-", "level", "cues", "such", "as", "color", "and", "sketch", ".", "All", "of", "the", "prior", "works", "report", "results", "on", "low", "-", "resolution", "images", ".", "Our", "system", "shares", "the", "same", "spirit", "as", "this", "past", "work", ",", "but", "we", "focus", "on", "object", "-", "level", "semantic", "editing", ",", "allowing", "users", "to", "interact", "with", "the", "entire", "scene", "and", "manipulate", "individual", "objects", "in", "the", "image", ".", "As", "a", "result", ",", "users", "can", "quickly", "create", "a", "new", "scene", "with", "minimal", "effort", ".", "Our", "interface", "is", "inspired", "by", "prior", "data", "-", "driven", "graphics", "systems", ".", "But", "our", "system", "allows", "more", "flexible", "manipulations", "and", "produces", "high", "-", "res", "results", "in", "real", "-", "time", ".", "section", ":", "Instance", "-", "Level", "Image", "Synthesis", "We", "propose", "a", "conditional", "adversarial", "framework", "for", "generating", "high", "-", "resolution", "photo", "-", "realistic", "images", "from", "semantic", "label", "maps", ".", "We", "first", "review", "our", "baseline", "model", "pix2pix", "(", "Sec", ".", "[", "reference", "]", ")", ".", "We", "then", "describe", "how", "we", "increase", "the", "photo", "-", "realism", "and", "resolution", "of", "the", "results", "with", "our", "improved", "objective", "function", "and", "network", "design", "(", "Sec", ".", "[", "reference", "]", ")", ".", "Next", ",", "we", "use", "additional", "instance", "-", "level", "object", "semantic", "information", "to", "further", "improve", "the", "image", "quality", "(", "Sec", ".", "[", "reference", "]", ")", ".", "Finally", ",", "we", "introduce", "an", "instance", "-", "level", "feature", "embedding", "scheme", "to", "better", "handle", "the", "multi", "-", "modal", "nature", "of", "image", "synthesis", ",", "which", "enables", "interactive", "object", "editing", "(", "Sec", ".", "[", "reference", "]", ")", ".", "subsection", ":", "The", "pix2pix", "Baseline", "The", "pix2pix", "method", "is", "a", "conditional", "GAN", "framework", "for", "image", "-", "to", "-", "image", "translation", ".", "It", "consists", "of", "a", "generator", "and", "a", "discriminator", ".", "For", "our", "task", ",", "the", "objective", "of", "the", "generator", "is", "to", "translate", "semantic", "label", "maps", "to", "realistic", "-", "looking", "images", ",", "while", "the", "discriminator", "aims", "to", "distinguish", "real", "images", "from", "the", "translated", "ones", ".", "The", "framework", "operates", "in", "a", "supervised", "setting", ".", "In", "other", "words", ",", "the", "training", "dataset", "is", "given", "as", "a", "set", "of", "pairs", "of", "corresponding", "images", ",", "where", "is", "a", "semantic", "label", "map", "and", "is", "a", "corresponding", "natural", "photo", ".", "Conditional", "GANs", "aim", "to", "model", "the", "conditional", "distribution", "of", "real", "images", "given", "the", "input", "semantic", "label", "maps", "via", "the", "following", "minimax", "game", ":", "where", "the", "objective", "function", "is", "given", "by", "The", "pix2pix", "method", "adopts", "U", "-", "Net", "as", "the", "generator", "and", "a", "patch", "-", "based", "fully", "convolutional", "network", "as", "the", "discriminator", ".", "The", "input", "to", "the", "discriminator", "is", "a", "channel", "-", "wise", "concatenation", "of", "the", "semantic", "label", "map", "and", "the", "corresponding", "image", ".", "However", ",", "the", "resolution", "of", "the", "generated", "images", "on", "Cityscapes", "is", "up", "to", ".", "We", "tested", "directly", "applying", "the", "pix2pix", "framework", "to", "generate", "high", "-", "resolution", "images", "but", "found", "the", "training", "unstable", "and", "the", "quality", "of", "generated", "images", "unsatisfactory", ".", "Therefore", ",", "we", "describe", "how", "we", "improve", "the", "pix2pix", "framework", "in", "the", "next", "subsection", ".", "subsection", ":", "Improving", "Photorealism", "and", "Resolution", "We", "improve", "the", "pix2pix", "framework", "by", "using", "a", "coarse", "-", "to", "-", "fine", "generator", ",", "a", "multi", "-", "scale", "discriminator", "architecture", ",", "and", "a", "robust", "adversarial", "learning", "objective", "function", ".", "Coarse", "-", "to", "-", "fine", "generator", "We", "decompose", "the", "generator", "into", "two", "sub", "-", "networks", ":", "and", ".", "We", "term", "as", "the", "global", "generator", "network", "and", "as", "the", "local", "enhancer", "network", ".", "The", "generator", "is", "then", "given", "by", "the", "tuple", "as", "visualized", "in", "Fig", ".", "[", "reference", "]", ".", "The", "global", "generator", "network", "operates", "at", "a", "resolution", "of", ",", "and", "the", "local", "enhancer", "network", "outputs", "an", "image", "with", "a", "resolution", "that", "is", "the", "output", "size", "of", "the", "previous", "one", "(", "along", "each", "image", "dimension", ")", ".", "For", "synthesizing", "images", "at", "an", "even", "higher", "resolution", ",", "additional", "local", "enhancer", "networks", "could", "be", "utilized", ".", "For", "example", ",", "the", "output", "image", "resolution", "of", "the", "generator", "is", ",", "and", "the", "output", "image", "resolution", "of", "is", ".", "Our", "global", "generator", "is", "built", "on", "the", "architecture", "proposed", "by", "Johnson", ",", "which", "has", "been", "proven", "successful", "for", "neural", "style", "transfer", "on", "images", "up", "to", ".", "It", "consists", "of", "components", ":", "a", "convolutional", "front", "-", "end", ",", "a", "set", "of", "residual", "blocks", ",", "and", "a", "transposed", "convolutional", "back", "-", "end", ".", "A", "semantic", "label", "map", "of", "resolution", "is", "passed", "through", "the", "3", "components", "sequentially", "to", "output", "an", "image", "of", "resolution", ".", "The", "local", "enhancer", "network", "also", "consists", "of", "3", "components", ":", "a", "convolutional", "front", "-", "end", ",", "a", "set", "of", "residual", "blocks", ",", "and", "a", "transposed", "convolutional", "back", "-", "end", ".", "The", "resolution", "of", "the", "input", "label", "map", "to", "is", ".", "Different", "from", "the", "global", "generator", "network", ",", "the", "input", "to", "the", "residual", "block", "is", "the", "element", "-", "wise", "sum", "of", "two", "feature", "maps", ":", "the", "output", "feature", "map", "of", ",", "and", "the", "last", "feature", "map", "of", "the", "back", "-", "end", "of", "the", "global", "generator", "network", ".", "This", "helps", "integrating", "the", "global", "information", "from", "to", ".", "During", "training", ",", "we", "first", "train", "the", "global", "generator", "and", "then", "train", "the", "local", "enhancer", "in", "the", "order", "of", "their", "resolutions", ".", "We", "then", "jointly", "fine", "-", "tune", "all", "the", "networks", "together", ".", "We", "use", "this", "generator", "design", "to", "effectively", "aggregate", "global", "and", "local", "information", "for", "the", "image", "synthesis", "task", ".", "We", "note", "that", "such", "a", "multi", "-", "resolution", "pipeline", "is", "a", "well", "-", "established", "practice", "in", "computer", "vision", "and", "two", "-", "scale", "is", "often", "enough", ".", "Similar", "ideas", "but", "different", "architectures", "could", "be", "found", "in", "recent", "unconditional", "GANs", "and", "conditional", "image", "generation", ".", "Multi", "-", "scale", "discriminators", "High", "-", "resolution", "image", "synthesis", "poses", "a", "significant", "challenge", "to", "the", "GAN", "discriminator", "design", ".", "To", "differentiate", "high", "-", "resolution", "real", "and", "synthesized", "images", ",", "the", "discriminator", "needs", "to", "have", "a", "large", "receptive", "field", ".", "This", "would", "require", "either", "a", "deeper", "network", "or", "larger", "convolutional", "kernels", ",", "both", "of", "which", "would", "increase", "the", "network", "capacity", "and", "potentially", "cause", "overfitting", ".", "Also", ",", "both", "choices", "demand", "a", "larger", "memory", "footprint", "for", "training", ",", "which", "is", "already", "a", "scarce", "resource", "for", "high", "-", "resolution", "image", "generation", ".", "To", "address", "the", "issue", ",", "we", "propose", "using", "multi", "-", "scale", "discriminators", ".", "We", "use", "discriminators", "that", "have", "an", "identical", "network", "structure", "but", "operate", "at", "different", "image", "scales", ".", "We", "will", "refer", "to", "the", "discriminators", "as", ",", "and", ".", "Specifically", ",", "we", "downsample", "the", "real", "and", "synthesized", "high", "-", "resolution", "images", "by", "a", "factor", "of", "and", "to", "create", "an", "image", "pyramid", "of", "3", "scales", ".", "The", "discriminators", ",", "and", "are", "then", "trained", "to", "differentiate", "real", "and", "synthesized", "images", "at", "the", "different", "scales", ",", "respectively", ".", "Although", "the", "discriminators", "have", "an", "identical", "architecture", ",", "the", "one", "that", "operates", "at", "the", "coarsest", "scale", "has", "the", "largest", "receptive", "field", ".", "It", "has", "a", "more", "global", "view", "of", "the", "image", "and", "can", "guide", "the", "generator", "to", "generate", "globally", "consistent", "images", ".", "On", "the", "other", "hand", ",", "the", "discriminator", "at", "the", "finest", "scale", "encourages", "the", "generator", "to", "produce", "finer", "details", ".", "This", "also", "makes", "training", "the", "coarse", "-", "to", "-", "fine", "generator", "easier", ",", "since", "extending", "a", "low", "-", "resolution", "model", "to", "a", "higher", "resolution", "only", "requires", "adding", "a", "discriminator", "at", "the", "finest", "level", ",", "rather", "than", "retraining", "from", "scratch", ".", "Without", "the", "multi", "-", "scale", "discriminators", ",", "we", "observe", "that", "many", "repeated", "patterns", "often", "appear", "in", "the", "generated", "images", ".", "With", "the", "discriminators", ",", "the", "learning", "problem", "in", "Eq", ".", "(", "[", "reference", "]", ")", "then", "becomes", "a", "multi", "-", "task", "learning", "problem", "of", "Using", "multiple", "GAN", "discriminators", "at", "the", "same", "image", "scale", "has", "been", "proposed", "in", "unconditional", "GANs", ".", "Iizuka", "et", "al", ".", "add", "a", "global", "image", "classifier", "to", "conditional", "GANs", "to", "synthesize", "globally", "coherent", "content", "for", "inpainting", ".", "Here", "we", "extend", "the", "design", "to", "multiple", "discriminators", "at", "different", "image", "scales", "for", "modeling", "high", "-", "resolution", "images", ".", "Improved", "adversarial", "loss", "We", "improve", "the", "GAN", "loss", "in", "Eq", ".", "(", "[", "reference", "]", ")", "by", "incorporating", "a", "feature", "matching", "loss", "based", "on", "the", "discriminator", ".", "This", "loss", "stabilizes", "the", "training", "as", "the", "generator", "has", "to", "produce", "natural", "statistics", "at", "multiple", "scales", ".", "Specifically", ",", "we", "extract", "features", "from", "multiple", "layers", "of", "the", "discriminator", "and", "learn", "to", "match", "these", "intermediate", "representations", "from", "the", "real", "and", "the", "synthesized", "image", ".", "For", "ease", "of", "presentation", ",", "we", "denote", "the", "th", "-", "layer", "feature", "extractor", "of", "discriminator", "as", "(", "from", "input", "to", "the", "th", "layer", "of", ")", ".", "The", "feature", "matching", "loss", "is", "then", "calculated", "as", ":", "where", "is", "the", "total", "number", "of", "layers", "and", "denotes", "the", "number", "of", "elements", "in", "each", "layer", ".", "Our", "GAN", "discriminator", "feature", "matching", "loss", "is", "related", "to", "the", "perceptual", "loss", ",", "which", "has", "been", "shown", "to", "be", "useful", "for", "image", "super", "-", "resolution", "and", "style", "transfer", ".", "In", "our", "experiments", ",", "we", "discuss", "how", "the", "discriminator", "feature", "matching", "loss", "and", "the", "perceptual", "loss", "can", "be", "jointly", "used", "for", "further", "improving", "the", "performance", ".", "We", "note", "that", "a", "similar", "loss", "is", "used", "in", "VAE", "-", "GANs", ".", "Our", "full", "objective", "combines", "both", "GAN", "loss", "and", "feature", "matching", "loss", "as", ":", "where", "controls", "the", "importance", "of", "the", "two", "terms", ".", "Note", "that", "for", "the", "feature", "matching", "loss", ",", "only", "serves", "as", "a", "feature", "extractor", "and", "does", "not", "maximize", "the", "loss", ".", "subsection", ":", "Using", "Instance", "Maps", "Existing", "image", "synthesis", "methods", "only", "utilize", "semantic", "label", "maps", ",", "an", "image", "where", "each", "pixel", "value", "represents", "the", "object", "class", "of", "the", "pixel", ".", "This", "map", "does", "not", "differentiate", "objects", "of", "the", "same", "category", ".", "On", "the", "other", "hand", ",", "an", "instance", "-", "level", "semantic", "label", "map", "contains", "a", "unique", "object", "ID", "for", "each", "individual", "object", ".", "To", "incorporate", "the", "instance", "map", ",", "one", "can", "directly", "pass", "it", "into", "the", "network", ",", "or", "encode", "it", "into", "a", "one", "-", "hot", "vector", ".", "However", ",", "both", "approaches", "are", "difficult", "to", "implement", "in", "practice", ",", "since", "different", "images", "may", "contain", "different", "numbers", "of", "objects", "of", "the", "same", "category", ".", "Alternatively", ",", "one", "can", "pre", "-", "allocate", "a", "fixed", "number", "of", "channels", "(", "e.g.", ",", ")", "for", "each", "class", ",", "but", "this", "method", "fails", "when", "the", "number", "is", "set", "too", "small", ",", "and", "wastes", "memory", "when", "the", "number", "is", "too", "large", ".", "Instead", ",", "we", "argue", "that", "the", "most", "critical", "information", "the", "instance", "map", "provides", ",", "which", "is", "not", "available", "in", "the", "semantic", "label", "map", ",", "is", "the", "object", "boundary", ".", "For", "example", ",", "when", "objects", "of", "the", "same", "class", "are", "next", "to", "one", "another", ",", "looking", "at", "the", "semantic", "label", "map", "alone", "can", "not", "tell", "them", "apart", ".", "This", "is", "especially", "true", "for", "the", "street", "scene", "since", "many", "parked", "cars", "or", "walking", "pedestrians", "are", "often", "next", "to", "one", "another", ",", "as", "shown", "in", "Fig", ".", "[", "reference", "]", "a.", "However", ",", "with", "the", "instance", "map", ",", "separating", "these", "objects", "becomes", "an", "easier", "task", ".", "Therefore", ",", "to", "extract", "this", "information", ",", "we", "first", "compute", "the", "instance", "boundary", "map", "(", "Fig", ".", "[", "reference", "]", "b", ")", ".", "In", "our", "implementation", ",", "a", "pixel", "in", "the", "instance", "boundary", "map", "is", "if", "its", "object", "ID", "is", "different", "from", "any", "of", "its", "-", "neighbors", ",", "and", "otherwise", ".", "The", "instance", "boundary", "map", "is", "then", "concatenated", "with", "the", "one", "-", "hot", "vector", "representation", "of", "the", "semantic", "label", "map", ",", "and", "fed", "into", "the", "generator", "network", ".", "Similarly", ",", "the", "input", "to", "the", "discriminator", "is", "the", "channel", "-", "wise", "concatenation", "of", "instance", "boundary", "map", ",", "semantic", "label", "map", ",", "and", "the", "real", "/", "synthesized", "image", ".", "Figure", "[", "reference", "]", "b", "shows", "an", "example", "demonstrating", "the", "improvement", "by", "using", "object", "boundaries", ".", "Our", "user", "study", "in", "Sec", ".", "[", "reference", "]", "also", "shows", "the", "model", "trained", "with", "instance", "boundary", "maps", "renders", "more", "photo", "-", "realistic", "object", "boundaries", ".", "subsection", ":", "Learning", "an", "Instance", "-", "level", "Feature", "Embedding", "Image", "synthesis", "from", "semantic", "label", "maps", "is", "a", "one", "-", "to", "-", "many", "mapping", "problem", ".", "An", "ideal", "image", "synthesis", "algorithm", "should", "be", "able", "to", "generate", "diverse", ",", "realistic", "images", "using", "the", "same", "semantic", "label", "map", ".", "Recently", ",", "several", "works", "learn", "to", "produce", "a", "fixed", "number", "of", "discrete", "outputs", "given", "the", "same", "input", "or", "synthesize", "diverse", "modes", "controlled", "by", "a", "latent", "code", "that", "encodes", "the", "entire", "image", ".", "Although", "these", "approaches", "tackle", "the", "multi", "-", "modal", "image", "synthesis", "problem", ",", "they", "are", "unsuitable", "for", "our", "image", "manipulation", "task", "mainly", "for", "two", "reasons", ".", "First", ",", "the", "user", "has", "no", "intuitive", "control", "over", "which", "kinds", "of", "images", "the", "model", "would", "produce", ".", "Second", ",", "these", "methods", "focus", "on", "global", "color", "and", "texture", "changes", "and", "allow", "no", "object", "-", "level", "control", "on", "the", "generated", "contents", ".", "To", "generate", "diverse", "images", "and", "allow", "instance", "-", "level", "control", ",", "we", "propose", "adding", "additional", "low", "-", "dimensional", "feature", "channels", "as", "the", "input", "to", "the", "generator", "network", ".", "We", "show", "that", ",", "by", "manipulating", "these", "features", ",", "we", "can", "have", "flexible", "control", "over", "the", "image", "synthesis", "process", ".", "Furthermore", ",", "note", "that", "since", "the", "feature", "channels", "are", "continuous", "quantities", ",", "our", "model", "is", ",", "in", "principle", ",", "capable", "of", "generating", "infinitely", "many", "images", ".", "To", "generate", "the", "low", "-", "dimensional", "features", ",", "we", "train", "an", "encoder", "network", "to", "find", "a", "low", "-", "dimensional", "feature", "vector", "that", "corresponds", "to", "the", "ground", "truth", "target", "for", "each", "instance", "in", "the", "image", ".", "Our", "feature", "encoder", "architecture", "is", "a", "standard", "encoder", "-", "decoder", "network", ".", "To", "ensure", "the", "features", "are", "consistent", "within", "each", "instance", ",", "we", "add", "an", "instance", "-", "wise", "average", "pooling", "layer", "to", "the", "output", "of", "the", "encoder", "to", "compute", "the", "average", "feature", "for", "the", "object", "instance", ".", "The", "average", "feature", "is", "then", "broadcast", "to", "all", "the", "pixel", "locations", "of", "the", "instance", ".", "Figure", "[", "reference", "]", "visualizes", "an", "example", "of", "the", "encoded", "features", ".", "We", "replace", "with", "in", "Eq", ".", "(", "[", "reference", "]", ")", "and", "train", "the", "encoder", "jointly", "with", "the", "generators", "and", "discriminators", ".", "After", "the", "encoder", "is", "trained", ",", "we", "run", "it", "on", "all", "instances", "in", "the", "training", "images", "and", "record", "the", "obtained", "features", ".", "Then", "we", "perform", "a", "-", "means", "clustering", "on", "these", "features", "for", "each", "semantic", "category", ".", "Each", "cluster", "thus", "encodes", "the", "features", "for", "a", "specific", "style", ",", "for", "example", ",", "the", "asphalt", "or", "cobblestone", "texture", "for", "a", "road", ".", "At", "inference", "time", ",", "we", "randomly", "pick", "one", "of", "the", "cluster", "centers", "and", "use", "it", "as", "the", "encoded", "features", ".", "These", "features", "are", "concatenated", "with", "the", "label", "map", "and", "used", "as", "the", "input", "to", "our", "generator", ".", "We", "tried", "to", "enforce", "the", "Kullback", "-", "Leibler", "loss", "on", "the", "feature", "space", "for", "better", "test", "-", "time", "sampling", "as", "used", "in", "the", "recent", "work", "but", "found", "it", "quite", "involved", "for", "users", "to", "adjust", "the", "latent", "vectors", "for", "each", "object", "directly", ".", "Instead", ",", "for", "each", "object", "instance", ",", "we", "present", "modes", "for", "users", "to", "choose", "from", ".", "section", ":", "Results", "We", "first", "provide", "a", "quantitative", "comparison", "against", "leading", "methods", "in", "Sec", ".", "[", "reference", "]", ".", "We", "then", "report", "a", "subjective", "human", "perceptual", "study", "in", "Sec", ".", "[", "reference", "]", ".", "Finally", ",", "we", "show", "a", "few", "examples", "of", "interactive", "object", "editing", "results", "in", "Sec", ".", "[", "reference", "]", ".", "Implementation", "details", "We", "use", "LSGANs", "for", "stable", "training", ".", "In", "all", "experiments", ",", "we", "set", "the", "weight", "(", "Eq", ".", "(", "[", "reference", "]", ")", ")", "and", "for", "K", "-", "means", ".", "We", "use", "-", "dimensional", "vectors", "to", "encode", "features", "for", "each", "object", "instance", ".", "We", "experimented", "with", "adding", "a", "perceptual", "loss", "to", "our", "objective", "(", "Eq", ".", "(", "[", "reference", "]", ")", ")", ",", "where", "and", "denotes", "the", "-", "th", "layer", "with", "elements", "of", "the", "VGG", "network", ".", "We", "observe", "that", "this", "loss", "slightly", "improves", "the", "results", ".", "We", "name", "these", "two", "variants", "as", "ours", "and", "ours", "(", "w", "/", "o", "VGG", "loss", ")", ".", "Please", "find", "more", "training", "and", "architecture", "details", "in", "the", "appendix", ".", "Datasets", "We", "conduct", "extensive", "comparisons", "and", "ablation", "studies", "on", "Cityscapes", "dataset", "and", "NYU", "Indoor", "RGBD", "dataset", ".", "We", "report", "additional", "qualitative", "results", "on", "ADE20", "K", "dataset", "and", "Helen", "Face", "dataset", ".", "Baselines", "We", "compare", "our", "method", "with", "two", "state", "-", "of", "-", "the", "-", "art", "algorithms", ":", "pix2pix", "and", "CRN", ".", "We", "train", "pix2pix", "models", "on", "high", "-", "res", "images", "with", "the", "default", "setting", ".", "We", "produce", "the", "high", "-", "res", "CRN", "images", "via", "the", "authors", "’", "publicly", "available", "model", ".", "subsection", ":", "Quantitative", "Comparisons", "We", "adopt", "the", "same", "evaluation", "protocol", "from", "previous", "image", "-", "to", "-", "image", "translation", "works", ".", "To", "quantify", "the", "quality", "of", "our", "results", ",", "we", "perform", "semantic", "segmentation", "on", "the", "synthesized", "images", "and", "compare", "how", "well", "the", "predicted", "segments", "match", "the", "input", ".", "The", "intuition", "is", "that", "if", "we", "can", "produce", "realistic", "images", "that", "correspond", "to", "the", "input", "label", "map", ",", "an", "off", "-", "the", "-", "shelf", "semantic", "segmentation", "model", "(", "e.g.", ",", "PSPNet", "that", "we", "use", ")", "should", "be", "able", "to", "predict", "the", "ground", "truth", "label", ".", "Table", "[", "reference", "]", "reports", "the", "calculated", "segmentation", "accuracy", ".", "As", "can", "be", "seen", ",", "for", "both", "pixel", "-", "wise", "accuracy", "and", "mean", "intersection", "-", "over", "-", "union", "(", "IoU", ")", ",", "our", "method", "outperforms", "the", "other", "methods", "by", "a", "large", "margin", ".", "Moreover", ",", "our", "result", "is", "very", "close", "to", "the", "result", "of", "the", "original", "images", ",", "the", "theoretical", "“", "upper", "bound", "”", "of", "the", "realism", "we", "can", "achieve", ".", "This", "justifies", "the", "superiority", "of", "our", "algorithm", ".", "subsection", ":", "Human", "Perceptual", "Study", "We", "further", "evaluate", "our", "algorithm", "via", "a", "human", "subjective", "study", ".", "We", "perform", "pairwise", "A", "/", "B", "tests", "deployed", "on", "the", "Amazon", "Mechanical", "Turk", "(", "MTurk", ")", "platform", "on", "the", "Cityscapes", "dataset", ".", "We", "follow", "the", "same", "experimental", "procedure", "as", "described", "in", "Chen", "and", "Koltun", ".", "More", "specifically", ",", "two", "different", "kinds", "of", "experiments", "are", "conducted", ":", "unlimited", "time", "and", "limited", "time", ",", "as", "explained", "below", ".", "Unlimited", "time", "For", "this", "task", ",", "workers", "are", "given", "two", "images", "at", "once", ",", "each", "of", "which", "is", "synthesized", "by", "a", "different", "method", "for", "the", "same", "label", "map", ".", "We", "give", "them", "unlimited", "time", "to", "select", "which", "image", "looks", "more", "natural", ".", "The", "left", "-", "right", "order", "and", "the", "image", "order", "are", "randomized", "to", "ensure", "fair", "comparisons", ".", "All", "Cityscapes", "test", "images", "are", "compared", "times", ",", "resulting", "in", "human", "judgments", "for", "each", "method", ".", "In", "this", "experiment", ",", "we", "use", "the", "model", "trained", "on", "labels", "only", "(", "without", "instance", "maps", ")", "to", "ensure", "a", "fair", "comparison", ".", "Table", "[", "reference", "]", "shows", "that", "both", "variants", "of", "our", "method", "outperform", "the", "other", "methods", "significantly", ".", "Limited", "time", "Next", ",", "for", "the", "limited", "time", "experiment", ",", "we", "compare", "our", "result", "with", "CRN", "and", "the", "original", "image", "(", "ground", "truth", ")", ".", "In", "each", "comparison", ",", "we", "show", "results", "of", "two", "methods", "for", "a", "short", "period", "of", "time", ".", "We", "randomly", "select", "a", "duration", "between", "seconds", "and", "seconds", ",", "as", "adopted", "by", "prior", "work", ".", "This", "evaluates", "how", "quickly", "the", "difference", "between", "the", "images", "can", "be", "perceived", ".", "Fig", ".", "[", "reference", "]", "shows", "the", "comparison", "results", "at", "different", "time", "intervals", ".", "As", "the", "given", "time", "becomes", "longer", "and", "longer", ",", "the", "differences", "between", "these", "three", "types", "of", "images", "become", "more", "apparent", "and", "easier", "to", "observe", ".", "Figures", "[", "reference", "]", "and", "[", "reference", "]", "show", "some", "example", "results", ".", "Analysis", "of", "the", "loss", "function", "We", "also", "study", "the", "importance", "of", "each", "term", "in", "our", "objective", "function", "using", "the", "unlimited", "time", "experiment", ".", "Specifically", ",", "our", "final", "loss", "contains", "three", "components", ":", "GAN", "loss", ",", "discriminator", "-", "based", "feature", "matching", "loss", ",", "and", "VGG", "perceptual", "loss", ".", "We", "compare", "our", "final", "implementation", "to", "the", "results", "using", "(", ")", "only", "GAN", "loss", ",", "and", "(", ")", "GAN", "feature", "matching", "loss", "(", "i.e.", ",", "without", "VGG", "loss", ")", ".", "The", "obtained", "preference", "rates", "are", "and", ",", "respectively", ".", "As", "can", "be", "seen", ",", "adding", "the", "feature", "matching", "loss", "substantially", "improves", "the", "performance", ",", "while", "adding", "perceptual", "loss", "further", "enhances", "the", "results", ".", "However", ",", "note", "that", "using", "the", "perceptual", "loss", "is", "not", "critical", ",", "and", "we", "are", "still", "able", "to", "generate", "visually", "appealing", "results", "even", "without", "it", "(", "e.g.", ",", "Figs", ".", "[", "reference", "]", "c", ",", "[", "reference", "]", "b", ")", ".", "Using", "instance", "maps", "We", "compare", "the", "results", "using", "instance", "maps", "to", "results", "without", "using", "them", ".", "We", "highlight", "the", "car", "regions", "in", "the", "images", "and", "ask", "the", "participants", "to", "choose", "which", "region", "looks", "more", "realistic", ".", "We", "obtain", "a", "preference", "rate", "of", ",", "which", "indicates", "that", "using", "instance", "maps", "improves", "the", "realism", "of", "our", "results", ",", "especially", "around", "the", "object", "boundaries", ".", "Analysis", "of", "the", "generator", "We", "compare", "results", "of", "different", "generators", "with", "all", "the", "other", "components", "fixed", ".", "In", "particular", ",", "we", "compare", "our", "generator", "with", "two", "state", "-", "of", "-", "the", "-", "art", "generator", "architectures", ":", "U", "-", "Net", "and", "CRN", ".", "We", "evaluate", "the", "performance", "regarding", "both", "semantic", "segmentation", "scores", "and", "human", "perceptual", "study", "results", ".", "Table", "[", "reference", "]", "and", "Table", "[", "reference", "]", "show", "that", "our", "coarse", "-", "to", "-", "fine", "generator", "outperforms", "other", "networks", "by", "a", "large", "margin", ".", "Analysis", "of", "the", "discriminator", "Next", ",", "we", "also", "compare", "results", "using", "our", "multi", "-", "scale", "discriminators", "and", "results", "using", "only", "one", "discriminator", "while", "we", "keep", "the", "generator", "and", "the", "loss", "function", "fixed", ".", "The", "segmentation", "scores", "on", "Cityscapes", "(", "Table", "[", "reference", "]", ")", "demonstrate", "that", "using", "multi", "-", "scale", "discriminators", "helps", "produce", "higher", "quality", "results", "as", "well", "as", "stabilize", "the", "adversarial", "training", ".", "We", "also", "perform", "pairwise", "A", "/", "B", "tests", "on", "the", "Amazon", "Mechanical", "Turk", "platform", ".", "of", "the", "participants", "prefer", "our", "results", "with", "multi", "-", "scale", "discriminators", "over", "the", "results", "trained", "with", "a", "single", "-", "scale", "discriminator", "(", "Chance", "is", ")", ".", "Additional", "datasets", "To", "further", "evaluate", "our", "method", ",", "we", "perform", "unlimited", "time", "comparisons", "on", "the", "NYU", "dataset", ".", "We", "obtain", "and", "against", "pix2pix", "and", "CRN", ",", "respectively", ".", "Fig", ".", "[", "reference", "]", "show", "some", "example", "images", ".", "Finally", ",", "we", "show", "results", "on", "the", "ADE20", "K", "dataset", "(", "Fig", ".", "[", "reference", "]", ")", ".", "subsection", ":", "Interactive", "Object", "Editing", "Our", "feature", "encoder", "allows", "us", "to", "perform", "interactive", "instance", "editing", "on", "the", "resulting", "images", ".", "For", "example", ",", "we", "can", "change", "the", "object", "labels", "in", "the", "image", "to", "quickly", "create", "novel", "scenes", ",", "such", "as", "replacing", "trees", "with", "buildings", "(", "Fig", ".", "[", "reference", "]", "b", ")", ".", "We", "can", "also", "change", "the", "colors", "of", "individual", "cars", "or", "the", "textures", "of", "the", "road", "(", "Fig", ".", "[", "reference", "]", "c", ")", ".", "Please", "check", "out", "our", "interactive", "demos", "on", "our", "website", ".", "Besides", ",", "we", "implement", "our", "interactive", "object", "editing", "feature", "on", "the", "Helen", "Face", "dataset", "where", "labels", "for", "different", "facial", "parts", "are", "available", "(", "Fig", ".", "[", "reference", "]", ")", ".", "This", "makes", "it", "easy", "to", "edit", "human", "portraits", ",", "e.g.", ",", "changing", "the", "face", "color", "to", "mimic", "different", "make", "-", "up", "effects", "or", "adding", "beard", "to", "a", "face", ".", "section", ":", "Discussion", "and", "Conclusion", "The", "results", "in", "this", "paper", "suggest", "that", "conditional", "GANs", "can", "synthesize", "high", "-", "resolution", "photo", "-", "realistic", "imagery", "without", "any", "hand", "-", "crafted", "losses", "or", "pre", "-", "trained", "networks", ".", "We", "have", "observed", "that", "incorporating", "a", "perceptual", "loss", "can", "slightly", "improve", "the", "results", ".", "Our", "method", "allows", "many", "applications", "and", "will", "be", "potentially", "useful", "for", "domains", "where", "high", "-", "resolution", "results", "are", "in", "demand", "but", "pre", "-", "trained", "networks", "are", "not", "available", "(", "e.g.", ",", "medical", "imaging", "and", "biology", ")", ".", "This", "paper", "also", "shows", "that", "an", "image", "-", "to", "-", "image", "synthesis", "pipeline", "can", "be", "extended", "to", "produce", "diverse", "outputs", ",", "and", "enable", "interactive", "image", "manipulation", "given", "appropriate", "training", "input", "-", "output", "pairs", "(", "e.g.", ",", "instance", "maps", "in", "our", "case", ")", ".", "Without", "ever", "been", "told", "what", "a", "“", "texture", "”", "is", ",", "our", "model", "learns", "to", "stylize", "different", "objects", ",", "which", "may", "be", "generalized", "to", "other", "datasets", "as", "well", "(", "i.e.", ",", "using", "textures", "in", "one", "dataset", "to", "synthesize", "images", "in", "another", "dataset", ")", ".", "We", "believe", "these", "extensions", "can", "be", "potentially", "applied", "to", "other", "image", "synthesis", "problems", ".", "Acknowledgements", "We", "thank", "Taesung", "Park", ",", "Phillip", "Isola", ",", "Tinghui", "Zhou", ",", "Richard", "Zhang", ",", "Rafael", "Valle", "and", "Alexei", "A.", "Efros", "for", "helpful", "comments", ".", "We", "also", "thank", "Chen", "and", "Koltun", "and", "Isola", "et", "al", ".", "for", "sharing", "their", "code", ".", "JYZ", "is", "supported", "by", "a", "Facebook", "graduate", "fellowship", ".", "bibliography", ":", "References", "appendix", ":", "Training", "Details", "All", "the", "networks", "were", "trained", "from", "scratch", ",", "using", "the", "Adam", "solver", "and", "a", "learning", "rate", "of", ".", "We", "keep", "the", "same", "learning", "rate", "for", "the", "first", "epochs", "and", "linearly", "decay", "the", "rate", "to", "zero", "over", "the", "next", "epochs", ".", "Weights", "were", "initialized", "from", "a", "Gaussian", "distribution", "with", "mean", "and", "standard", "deviation", ".", "We", "train", "all", "our", "models", "on", "an", "NVIDIA", "Quadro", "M6000", "GPU", "with", "GB", "GPU", "memory", ".", "The", "inference", "time", "is", "between", "milliseconds", "per", "input", "image", "on", "an", "NVIDIA", "1080Ti", "GPU", "with", "GB", "GPU", "memory", ".", "This", "real", "-", "time", "performance", "allows", "us", "to", "develop", "interactive", "image", "editing", "applications", ".", "Below", "we", "discuss", "the", "details", "of", "the", "datasets", "we", "used", ".", "Cityscapes", "dataset", "[", "]", ":", "training", "images", "from", "the", "Cityscapes", "training", "set", "with", "image", "size", ".", "We", "use", "the", "Cityscapes", "validation", "set", "for", "testing", ",", "which", "consists", "of", "500", "images", ".", "NYU", "Indoor", "RGBD", "dataset", "[", "]", ":", "training", "images", "and", "test", "images", ",", "all", "at", "resolution", "of", ".", "ADE20", "K", "dataset", "[", "]", ":", "training", "images", "and", "test", "images", "with", "varying", "image", "sizes", ".", "We", "scale", "the", "width", "of", "all", "images", "to", "before", "training", "and", "inference", ".", "Helen", "Face", "dataset", "[", "]", ":", "training", "images", "and", "test", "images", "with", "varying", "image", "sizes", ".", "We", "resize", "all", "images", "to", "before", "training", "and", "inference", ".", "appendix", ":", "Generator", "Architectures", "Our", "generator", "consists", "of", "a", "global", "generator", "network", "and", "a", "local", "enhancer", "network", ".", "we", "follow", "the", "naming", "convention", "used", "in", "Johnson", "el", "al", ".", "and", "CycleGAN", ".", "Let", "c7s1", "-", "k", "denote", "a", "Convolution", "-", "InstanceNorm", "-", "ReLU", "layer", "with", "filters", "and", "stride", ".", "dk", "denotes", "a", "Convolution", "-", "InstanceNorm", "-", "ReLU", "layer", "with", "filters", ",", "and", "stride", ".", "We", "use", "reflection", "padding", "to", "reduce", "boundary", "artifacts", ".", "Rk", "denotes", "a", "residual", "block", "that", "contains", "two", "convolutional", "layers", "with", "the", "same", "number", "of", "filters", "on", "both", "layers", ".", "uk", "denotes", "a", "fractional", "-", "strided", "-", "Convolution", "-", "InstanceNorm", "-", "ReLU", "layer", "with", "filters", ",", "and", "stride", ".", "Recall", "that", "we", "have", "two", "generators", ":", "the", "global", "generator", "and", "the", "local", "enhancer", ".", "Our", "global", "network", ":", "c7s1", "-", "64", ",", "d128", ",", "d256", ",", "d512", ",", "d1024", ",", "R1024", ",", "R1024", ",", "R1024", ",", "R1024", ",", "R1024", ",", "R1024", ",", "R1024", ",", "R1024", ",", "R1024", ",", "u512", ",", "u256", ",", "u128", ",", "u64", ",", "c7s1", "-", "3", "Our", "local", "enhancer", ":", "c7s1", "-", "32", ",", "d64We", "add", "the", "last", "feature", "map", "(", "u64", ")", "in", "our", "global", "network", "to", "the", "output", "of", "this", "layer.", ",", "R64", ",", "R64", ",", "R64", ",", "u32", ",", "c7s1", "-", "3", "appendix", ":", "Discriminator", "Architectures", "For", "discriminator", "networks", ",", "we", "use", "PatchGAN", ".", "Let", "Ck", "denote", "a", "Convolution", "-", "InstanceNorm", "-", "LeakyReLU", "layer", "with", "k", "filters", "and", "stride", ".", "After", "the", "last", "layer", ",", "we", "apply", "a", "convolution", "to", "produce", "a", "dimensional", "output", ".", "We", "do", "not", "use", "InstanceNorm", "for", "the", "first", "C64", "layer", ".", "We", "use", "leaky", "ReLUs", "with", "slope", ".", "All", "our", "three", "discriminators", "have", "the", "identical", "architecture", "as", "follows", ":", "C64", "-", "C128", "-", "C256", "-", "C512", "appendix", ":", "Change", "log", "paragraph", ":", "v1", "initial", "preprint", "release", "paragraph", ":", "v2", "CVPR", "camera", "ready", ",", "adding", "more", "results", "for", "edge", "-", "to", "-", "photo", "examples", "."], "entity": [[2, 7, ["High", "-", "Resolution", "Image", "Synthesis"], "Task"], [8, 10, ["Semantic", "Manipulation"], "Task"], [11, 13, ["Conditional", "GANs"], "Method"], [19, 27, ["synthesizing", "high", "-", "resolution", "photo", "-", "realistic", "images"], "Task"], [28, 31, ["semantic", "label", "maps"], "Task"], [32, 36, ["conditional", "generative", "adversarial", "networks"], "Method"], [37, 39, ["conditional", "GANs"], "Method"], [41, 43, ["Conditional", "GANs"], "Method"], [78, 80, ["adversarial", "loss"], "Method"], [85, 92, ["multi", "-", "scale", "generator", "and", "discriminator", "architectures"], "Method"], [100, 103, ["interactive", "visual", "manipulation"], "Task"], [172, 173, ["quality"], "Metric"], [175, 176, ["resolution"], "Metric"], [177, 182, ["deep", "image", "synthesis", "and", "editing"], "Task"], [186, 191, ["Photo", "-", "realistic", "image", "rendering"], "Task"], [193, 195, ["graphics", "techniques"], "Method"], [213, 215, ["graphics", "algorithms"], "Method"], [220, 225, ["building", "and", "editing", "virtual", "environments"], "Task"], [269, 271, ["graphics", "rendering"], "Task"], [273, 278, ["model", "learning", "and", "inference", "problem"], "Task"], [287, 291, ["creating", "new", "virtual", "worlds"], "Task"], [337, 341, ["high", "-", "resolution", "images"], "Task"], [342, 345, ["semantic", "label", "maps"], "Task"], [369, 372, ["visual", "recognition", "algorithms"], "Task"], [392, 395, ["semantic", "segmentation", "methods"], "Method"], [432, 437, ["higher", "-", "level", "image", "editing"], "Task"], [463, 465, ["pix2pix", "method"], "Method"], [476, 479, ["generative", "adversarial", "networks"], "Method"], [480, 481, ["GANs"], "Method"], [484, 486, ["conditional", "setting"], "Method"], [494, 496, ["adversarial", "training"], "Method"], [504, 510, ["high", "-", "resolution", "image", "generation", "tasks"], "Task"], [517, 519, ["perceptual", "loss"], "Method"], [561, 566, ["generating", "high", "-", "resolution", "images"], "Task"], [567, 568, ["GANs"], "Method"], [595, 598, ["adversarial", "learning", "objective"], "Method"], [601, 608, ["multi", "-", "scale", "generator", "and", "discriminator", "architectures"], "Method"], [637, 639, ["adversarial", "training"], "Method"], [650, 654, ["pre", "-", "trained", "networks"], "Method"], [656, 657, ["VGGNet"], "Method"], [685, 687, ["trained", "networks"], "Method"], [711, 715, ["pre", "-", "trained", "network"], "Method"], [727, 729, ["image", "quality"], "Metric"], [734, 737, ["interactive", "semantic", "manipulation"], "Task"], [771, 773, ["object", "manipulations"], "Task"], [825, 828, ["visual", "synthesis", "systems"], "Method"], [839, 841, ["quantitative", "evaluations"], "Metric"], [842, 845, ["human", "perception", "studies"], "Task"], [854, 856, ["training", "objectives"], "Metric"], [869, 871, ["semantic", "manipulation"], "Task"], [877, 879, ["edge2photo", "applications"], "Task"], [913, 916, ["Generative", "adversarial", "networks"], "Method"], [916, 919, ["Generative", "adversarial", "networks"], "Method"], [920, 921, ["GANs"], "Method"], [941, 942, ["GANs"], "Method"], [950, 952, ["image", "generation"], "Task"], [953, 955, ["representation", "learning"], "Task"], [956, 958, ["image", "manipulation"], "Task"], [959, 961, ["object", "detection"], "Task"], [963, 965, ["video", "applications"], "Task"], [967, 973, ["coarse", "-", "to", "-", "fine", "schemes"], "Method"], [985, 987, ["unconditional", "setting"], "Task"], [997, 1003, ["coarse", "-", "to", "-", "fine", "generator"], "Method"], [1004, 1009, ["multi", "-", "scale", "discriminator", "architectures"], "Method"], [1011, 1014, ["conditional", "image", "generation"], "Task"], [1022, 1028, ["Image", "-", "to", "-", "image", "translation"], "Task"], [1032, 1034, ["adversarial", "learning"], "Method"], [1035, 1041, ["image", "-", "to", "-", "image", "translation"], "Task"], [1068, 1069, ["loss"], "Method"], [1078, 1080, ["adversarial", "loss"], "Method"], [1087, 1093, ["image", "-", "to", "-", "image", "tasks"], "Task"], [1099, 1100, ["discriminator"], "Method"], [1103, 1106, ["trainable", "loss", "function"], "Method"], [1131, 1135, ["image", "-", "conditional", "GANs"], "Method"], [1142, 1144, ["Google", "maps"], "Method"], [1148, 1150, ["generating", "cats"], "Task"], [1163, 1169, ["image", "-", "to", "-", "image", "translation"], "Task"], [1188, 1190, ["conditional", "GANs"], "Method"], [1192, 1196, ["high", "-", "resolution", "images"], "Task"], [1199, 1201, ["training", "instability"], "Task"], [1202, 1204, ["optimization", "issues"], "Task"], [1213, 1216, ["direct", "regression", "objective"], "Method"], [1264, 1268, ["multi", "-", "scale", "generators"], "Method"], [1269, 1270, ["discriminators"], "Method"], [1279, 1281, ["conditional", "GANs"], "Method"], [1331, 1334, ["Deep", "visual", "manipulation"], "Task"], [1336, 1339, ["deep", "neural", "networks"], "Method"], [1345, 1348, ["image", "processing", "tasks"], "Task"], [1351, 1353, ["style", "transfer"], "Task"], [1354, 1355, ["inpainting"], "Task"], [1356, 1357, ["colorization"], "Task"], [1359, 1360, ["restoration"], "Task"], [1391, 1393, ["optimization", "method"], "Method"], [1394, 1398, ["editing", "the", "object", "appearance"], "Task"], [1404, 1405, ["GANs"], "Method"], [1414, 1416, ["novel", "imagery"], "Task"], [1455, 1460, ["object", "-", "level", "semantic", "editing"], "Task"], [1498, 1503, ["data", "-", "driven", "graphics", "systems"], "Method"], [1524, 1529, ["Instance", "-", "Level", "Image", "Synthesis"], "Task"], [1532, 1535, ["conditional", "adversarial", "framework"], "Method"], [1537, 1544, ["high", "-", "resolution", "photo", "-", "realistic", "images"], "Task"], [1545, 1548, ["semantic", "label", "maps"], "Task"], [1555, 1556, ["pix2pix"], "Method"], [1571, 1574, ["photo", "-", "realism"], "Metric"], [1575, 1576, ["resolution"], "Metric"], [1582, 1584, ["objective", "function"], "Metric"], [1585, 1587, ["network", "design"], "Method"], [1610, 1612, ["image", "quality"], "Metric"], [1625, 1631, ["instance", "-", "level", "feature", "embedding", "scheme"], "Method"], [1635, 1642, ["multi", "-", "modal", "nature", "of", "image", "synthesis"], "Task"], [1645, 1648, ["interactive", "object", "editing"], "Task"], [1662, 1664, ["pix2pix", "method"], "Method"], [1666, 1669, ["conditional", "GAN", "framework"], "Method"], [1670, 1676, ["image", "-", "to", "-", "image", "translation"], "Task"], [1681, 1682, ["generator"], "Method"], [1684, 1685, ["discriminator"], "Method"], [1698, 1701, ["semantic", "label", "maps"], "Task"], [1709, 1710, ["discriminator"], "Method"], [1725, 1727, ["supervised", "setting"], "Task"], [1759, 1761, ["Conditional", "GANs"], "Method"], [1779, 1781, ["minimax", "game"], "Method"], [1784, 1786, ["objective", "function"], "Metric"], [1790, 1792, ["pix2pix", "method"], "Method"], [1793, 1796, ["U", "-", "Net"], "Method"], [1798, 1799, ["generator"], "Method"], [1801, 1807, ["patch", "-", "based", "fully", "convolutional", "network"], "Method"], [1809, 1810, ["discriminator"], "Method"], [1815, 1816, ["discriminator"], "Method"], [1818, 1822, ["channel", "-", "wise", "concatenation"], "Method"], [1824, 1827, ["semantic", "label", "map"], "Method"], [1841, 1842, ["Cityscapes"], "Material"], [1890, 1893, ["Photorealism", "and", "Resolution"], "Metric"], [1901, 1907, ["coarse", "-", "to", "-", "fine", "generator"], "Method"], [1909, 1914, ["multi", "-", "scale", "discriminator", "architecture"], "Method"], [1917, 1922, ["robust", "adversarial", "learning", "objective", "function"], "Method"], [1923, 1929, ["Coarse", "-", "to", "-", "fine", "generator"], "Task"], [1932, 1933, ["generator"], "Method"], [1935, 1938, ["sub", "-", "networks"], "Method"], [1945, 1948, ["global", "generator", "network"], "Method"], [1951, 1954, ["local", "enhancer", "network"], "Method"], [1956, 1957, ["generator"], "Method"], [1973, 1976, ["global", "generator", "network"], "Method"], [1984, 1987, ["local", "enhancer", "network"], "Method"], [2010, 2012, ["synthesizing", "images"], "Task"], [2019, 2022, ["local", "enhancer", "networks"], "Method"], [2035, 2036, ["generator"], "Method"], [2041, 2043, ["image", "resolution"], "Metric"], [2047, 2049, ["global", "generator"], "Method"], [2056, 2057, ["Johnson"], "Method"], [2064, 2067, ["neural", "style", "transfer"], "Task"], [2078, 2082, ["convolutional", "front", "-", "end"], "Method"], [2086, 2088, ["residual", "blocks"], "Method"], [2091, 2096, ["transposed", "convolutional", "back", "-", "end"], "Method"], [2113, 2116, ["image", "of", "resolution"], "Task"], [2118, 2121, ["local", "enhancer", "network"], "Method"], [2128, 2132, ["convolutional", "front", "-", "end"], "Method"], [2141, 2146, ["transposed", "convolutional", "back", "-", "end"], "Method"], [2160, 2163, ["global", "generator", "network"], "Method"], [2199, 2202, ["global", "generator", "network"], "Method"], [2219, 2221, ["global", "generator"], "Method"], [2225, 2227, ["local", "enhancer"], "Method"], [2248, 2250, ["generator", "design"], "Method"], [2259, 2262, ["image", "synthesis", "task"], "Task"], [2268, 2272, ["multi", "-", "resolution", "pipeline"], "Method"], [2279, 2281, ["computer", "vision"], "Task"], [2282, 2285, ["two", "-", "scale"], "Task"], [2299, 2301, ["unconditional", "GANs"], "Method"], [2302, 2305, ["conditional", "image", "generation"], "Task"], [2306, 2315, ["Multi", "-", "scale", "discriminators", "High", "-", "resolution", "image", "synthesis"], "Task"], [2321, 2324, ["GAN", "discriminator", "design"], "Method"], [2336, 2337, ["discriminator"], "Method"], [2350, 2352, ["deeper", "network"], "Method"], [2354, 2356, ["convolutional", "kernels"], "Method"], [2380, 2381, ["training"], "Task"], [2389, 2394, ["high", "-", "resolution", "image", "generation"], "Task"], [2403, 2407, ["multi", "-", "scale", "discriminators"], "Method"], [2410, 2411, ["discriminators"], "Method"], [2482, 2483, ["discriminators"], "Method"], [2515, 2516, ["generator"], "Method"], [2535, 2536, ["generator"], "Method"], [2546, 2552, ["coarse", "-", "to", "-", "fine", "generator"], "Task"], [2557, 2561, ["low", "-", "resolution", "model"], "Method"], [2569, 2570, ["discriminator"], "Method"], [2583, 2587, ["multi", "-", "scale", "discriminators"], "Method"], [2603, 2604, ["discriminators"], "Method"], [2606, 2608, ["learning", "problem"], "Task"], [2619, 2624, ["multi", "-", "task", "learning", "problem"], "Task"], [2627, 2629, ["GAN", "discriminators"], "Method"], [2638, 2640, ["unconditional", "GANs"], "Method"], [2647, 2650, ["global", "image", "classifier"], "Method"], [2651, 2653, ["conditional", "GANs"], "Method"], [2659, 2660, ["inpainting"], "Task"], [2668, 2669, ["discriminators"], "Method"], [2674, 2679, ["modeling", "high", "-", "resolution", "images"], "Task"], [2681, 2683, ["adversarial", "loss"], "Metric"], [2686, 2688, ["GAN", "loss"], "Metric"], [2699, 2702, ["feature", "matching", "loss"], "Method"], [2705, 2706, ["discriminator"], "Method"], [2711, 2712, ["training"], "Task"], [2714, 2715, ["generator"], "Method"], [2734, 2735, ["discriminator"], "Method"], [2740, 2742, ["intermediate", "representations"], "Method"], [2758, 2763, ["th", "-", "layer", "feature", "extractor"], "Method"], [2764, 2765, ["discriminator"], "Method"], [2777, 2780, ["feature", "matching", "loss"], "Metric"], [2803, 2808, ["GAN", "discriminator", "feature", "matching", "loss"], "Method"], [2823, 2830, ["image", "super", "-", "resolution", "and", "style", "transfer"], "Task"], [2839, 2843, ["discriminator", "feature", "matching", "loss"], "Method"], [2866, 2869, ["VAE", "-", "GANs"], "Method"], [2875, 2877, ["GAN", "loss"], "Metric"], [2878, 2881, ["feature", "matching", "loss"], "Task"], [2896, 2899, ["feature", "matching", "loss"], "Task"], [2904, 2906, ["feature", "extractor"], "Method"], [2911, 2912, ["loss"], "Metric"], [2916, 2918, ["Instance", "Maps"], "Method"], [2919, 2922, ["image", "synthesis", "methods"], "Method"], [2924, 2927, ["semantic", "label", "maps"], "Method"], [3087, 3090, ["semantic", "label", "map"], "Task"], [3130, 3132, ["street", "scene"], "Task"], [3159, 3161, ["instance", "map"], "Method"], [3222, 3225, ["instance", "boundary", "map"], "Method"], [3230, 3240, ["one", "-", "hot", "vector", "representation", "of", "the", "semantic", "label", "map"], "Method"], [3245, 3247, ["generator", "network"], "Method"], [3254, 3255, ["discriminator"], "Method"], [3257, 3265, ["channel", "-", "wise", "concatenation", "of", "instance", "boundary", "map"], "Method"], [3266, 3269, ["semantic", "label", "map"], "Method"], [3308, 3311, ["instance", "boundary", "maps"], "Method"], [3323, 3328, ["Instance", "-", "level", "Feature", "Embedding"], "Method"], [3328, 3330, ["Image", "synthesis"], "Task"], [3331, 3334, ["semantic", "label", "maps"], "Task"], [3336, 3343, ["one", "-", "to", "-", "many", "mapping", "problem"], "Task"], [3346, 3349, ["image", "synthesis", "algorithm"], "Method"], [3389, 3391, ["latent", "code"], "Method"], [3402, 3408, ["multi", "-", "modal", "image", "synthesis", "problem"], "Task"], [3414, 3417, ["image", "manipulation", "task"], "Task"], [3469, 3473, ["instance", "-", "level", "control"], "Task"], [3488, 3490, ["generator", "network"], "Method"], [3507, 3510, ["image", "synthesis", "process"], "Task"], [3548, 3550, ["encoder", "network"], "Method"], [3573, 3576, ["feature", "encoder", "architecture"], "Method"], [3579, 3583, ["encoder", "-", "decoder", "network"], "Method"], [3597, 3603, ["instance", "-", "wise", "average", "pooling", "layer"], "Method"], [3660, 3661, ["encoder"], "Method"], [3664, 3667, ["generators", "and", "discriminators"], "Method"], [3670, 3671, ["encoder"], "Method"], [3694, 3697, ["-", "means", "clustering"], "Method"], [3780, 3784, ["test", "-", "time", "sampling"], "Task"], [3865, 3868, ["interactive", "object", "editing"], "Task"], [3880, 3881, ["LSGANs"], "Method"], [3882, 3884, ["stable", "training"], "Task"], [3904, 3907, ["K", "-", "means"], "Method"], [3926, 3928, ["perceptual", "loss"], "Metric"], [3952, 3954, ["VGG", "network"], "Method"], [3975, 3980, ["w", "/", "o", "VGG", "loss"], "Method"], [4002, 4004, ["Cityscapes", "dataset"], "Material"], [4005, 4009, ["NYU", "Indoor", "RGBD", "dataset"], "Material"], [4016, 4019, ["ADE20", "K", "dataset"], "Material"], [4020, 4023, ["Helen", "Face", "dataset"], "Material"], [4040, 4041, ["pix2pix"], "Method"], [4042, 4043, ["CRN"], "Method"], [4046, 4048, ["pix2pix", "models"], "Method"], [4086, 4092, ["image", "-", "to", "-", "image", "translation"], "Task"], [4104, 4106, ["semantic", "segmentation"], "Task"], [4145, 4148, ["semantic", "segmentation", "model"], "Method"], [4151, 4152, ["PSPNet"], "Method"], [4173, 4175, ["segmentation", "accuracy"], "Metric"], [4183, 4187, ["pixel", "-", "wise", "accuracy"], "Metric"], [4189, 4194, ["intersection", "-", "over", "-", "union"], "Metric"], [4195, 4196, ["IoU"], "Metric"], [4232, 4233, ["realism"], "Metric"], [4247, 4250, ["Human", "Perceptual", "Study"], "Task"], [4257, 4260, ["human", "subjective", "study"], "Metric"], [4271, 4274, ["Amazon", "Mechanical", "Turk"], "Method"], [4275, 4276, ["MTurk"], "Method"], [4280, 4282, ["Cityscapes", "dataset"], "Material"], [4437, 4440, ["limited", "time", "experiment"], "Task"], [4446, 4447, ["CRN"], "Method"], [4569, 4571, ["objective", "function"], "Metric"], [4586, 4588, ["GAN", "loss"], "Method"], [4589, 4595, ["discriminator", "-", "based", "feature", "matching", "loss"], "Method"], [4597, 4600, ["VGG", "perceptual", "loss"], "Method"], [4633, 4635, ["preference", "rates"], "Metric"], [4647, 4650, ["feature", "matching", "loss"], "Method"], [4706, 4708, ["instance", "maps"], "Method"], [4713, 4715, ["instance", "maps"], "Method"], [4744, 4746, ["preference", "rate"], "Metric"], [4752, 4754, ["instance", "maps"], "Method"], [4800, 4802, ["generator", "architectures"], "Method"], [4803, 4806, ["U", "-", "Net"], "Method"], [4807, 4808, ["CRN"], "Method"], [4815, 4818, ["semantic", "segmentation", "scores"], "Metric"], [4836, 4842, ["coarse", "-", "to", "-", "fine", "generator"], "Method"], [4862, 4866, ["multi", "-", "scale", "discriminators"], "Method"], [4876, 4877, ["generator"], "Method"], [4884, 4886, ["segmentation", "scores"], "Metric"], [4887, 4888, ["Cityscapes"], "Material"], [4897, 4901, ["multi", "-", "scale", "discriminators"], "Method"], [4911, 4913, ["adversarial", "training"], "Method"], [4917, 4922, ["pairwise", "A", "/", "B", "tests"], "Metric"], [4924, 4928, ["Amazon", "Mechanical", "Turk", "platform"], "Method"], [4936, 4940, ["multi", "-", "scale", "discriminators"], "Method"], [4948, 4950, ["scale", "discriminator"], "Method"], [4970, 4972, ["NYU", "dataset"], "Material"], [4979, 4980, ["CRN"], "Method"], [5000, 5003, ["ADE20", "K", "dataset"], "Material"], [5013, 5016, ["Interactive", "Object", "Editing"], "Task"], [5017, 5019, ["feature", "encoder"], "Method"], [5023, 5026, ["interactive", "instance", "editing"], "Task"], [5103, 5107, ["interactive", "object", "editing", "feature"], "Method"], [5109, 5112, ["Helen", "Face", "dataset"], "Material"], [5169, 5171, ["conditional", "GANs"], "Method"], [5187, 5191, ["pre", "-", "trained", "networks"], "Method"], [5229, 5231, ["trained", "networks"], "Method"], [5249, 5256, ["image", "-", "to", "-", "image", "synthesis", "pipeline"], "Method"], [5266, 5269, ["interactive", "image", "manipulation"], "Task"], [5340, 5343, ["image", "synthesis", "problems"], "Task"], [5385, 5386, ["JYZ"], "Method"], [5411, 5413, ["Adam", "solver"], "Method"], [5415, 5417, ["learning", "rate"], "Metric"], [5423, 5425, ["learning", "rate"], "Metric"], [5446, 5448, ["Gaussian", "distribution"], "Method"], [5462, 5465, ["Quadro", "M6000", "GPU"], "Method"], [5471, 5473, ["inference", "time"], "Metric"], [5481, 5484, ["NVIDIA", "1080Ti", "GPU"], "Method"], [5485, 5488, ["GB", "GPU", "memory"], "Method"], [5498, 5502, ["interactive", "image", "editing", "applications"], "Task"], [5514, 5516, ["Cityscapes", "dataset"], "Material"], [5545, 5549, ["NYU", "Indoor", "RGBD", "dataset"], "Material"], [5563, 5564, ["ADE20"], "Material"], [5590, 5591, ["inference"], "Task"], [5592, 5595, ["Helen", "Face", "dataset"], "Material"], [5616, 5617, ["inference"], "Task"], [5620, 5622, ["Generator", "Architectures"], "Method"], [5627, 5630, ["global", "generator", "network"], "Method"], [5632, 5635, ["local", "enhancer", "network"], "Method"], [5651, 5654, ["c7s1", "-", "k"], "Method"], [5656, 5662, ["Convolution", "-", "InstanceNorm", "-", "ReLU", "layer"], "Method"], [5663, 5664, ["filters"], "Method"], [5667, 5668, ["dk"], "Method"], [5670, 5676, ["Convolution", "-", "InstanceNorm", "-", "ReLU", "layer"], "Method"], [5677, 5678, ["filters"], "Method"], [5691, 5692, ["Rk"], "Method"], [5699, 5701, ["convolutional", "layers"], "Method"], [5711, 5712, ["uk"], "Method"], [5714, 5724, ["fractional", "-", "strided", "-", "Convolution", "-", "InstanceNorm", "-", "ReLU", "layer"], "Method"], [5725, 5726, ["filters"], "Method"], [5738, 5740, ["global", "generator"], "Method"], [5742, 5744, ["local", "enhancer"], "Method"], [5746, 5748, ["global", "network"], "Method"], [5791, 5793, ["local", "enhancer"], "Method"], [5809, 5811, ["global", "network"], "Method"], [5831, 5833, ["Discriminator", "Architectures"], "Method"], [5834, 5836, ["discriminator", "networks"], "Method"], [5839, 5840, ["PatchGAN"], "Method"], [5842, 5843, ["Ck"], "Method"], [5845, 5851, ["Convolution", "-", "InstanceNorm", "-", "LeakyReLU", "layer"], "Method"], [5853, 5854, ["filters"], "Method"], [5865, 5866, ["convolution"], "Method"], [5876, 5877, ["InstanceNorm"], "Method"], [5880, 5882, ["C64", "layer"], "Method"], [5921, 5923, ["CVPR", "camera"], "Method"], [5929, 5935, ["edge", "-", "to", "-", "photo", "examples"], "Task"], [467, 473, ["image", "-", "to", "-", "image", "translation"], "Task"], [1128, 1129, ["pix2pix"], "Method"], [1659, 1660, ["pix2pix"], "Method"], [1851, 1852, ["pix2pix"], "Method"], [1880, 1881, ["pix2pix"], "Method"], [1896, 1897, ["pix2pix"], "Method"], [4376, 4377, ["Cityscapes"], "Material"], [4977, 4978, ["pix2pix"], "Method"], [5466, 5469, ["GB", "GPU", "memory"], "Method"], [5523, 5524, ["Cityscapes"], "Material"], [5533, 5534, ["Cityscapes"], "Material"]]} {"doc_id": "455da02e5048dffb51fb6ab5eb8aeca5926c9d9a", "text": ["document", ":", "Spatial", "Pyramid", "Pooling", "in", "Deep", "Convolutional", "Networks", "for", "Visual", "Recognition", "Existing", "deep", "convolutional", "neural", "networks", "(", "CNNs", ")", "require", "a", "fixed", "-", "size", "(", ",", "224", "224", ")", "input", "image", ".", "This", "requirement", "is", "“", "artificial", "”", "and", "may", "reduce", "the", "recognition", "accuracy", "for", "the", "images", "or", "sub", "-", "images", "of", "an", "arbitrary", "size", "/", "scale", ".", "In", "this", "work", ",", "we", "equip", "the", "networks", "with", "another", "pooling", "strategy", ",", "“", "spatial", "pyramid", "pooling", "”", ",", "to", "eliminate", "the", "above", "requirement", ".", "The", "new", "network", "structure", ",", "called", "SPP", "-", "net", ",", "can", "generate", "a", "fixed", "-", "length", "representation", "regardless", "of", "image", "size", "/", "scale", ".", "Pyramid", "pooling", "is", "also", "robust", "to", "object", "deformations", ".", "With", "these", "advantages", ",", "SPP", "-", "net", "should", "in", "general", "improve", "all", "CNN", "-", "based", "image", "classification", "methods", ".", "On", "the", "ImageNet", "2012", "dataset", ",", "we", "demonstrate", "that", "SPP", "-", "net", "boosts", "the", "accuracy", "of", "a", "variety", "of", "CNN", "architectures", "despite", "their", "different", "designs", ".", "On", "the", "Pascal", "VOC", "2007", "and", "Caltech101", "datasets", ",", "SPP", "-", "net", "achieves", "state", "-", "of", "-", "the", "-", "art", "classification", "results", "using", "a", "single", "full", "-", "image", "representation", "and", "no", "fine", "-", "tuning", ".", "The", "power", "of", "SPP", "-", "net", "is", "also", "significant", "in", "object", "detection", ".", "Using", "SPP", "-", "net", ",", "we", "compute", "the", "feature", "maps", "from", "the", "entire", "image", "only", "once", ",", "and", "then", "pool", "features", "in", "arbitrary", "regions", "(", "sub", "-", "images", ")", "to", "generate", "fixed", "-", "length", "representations", "for", "training", "the", "detectors", ".", "This", "method", "avoids", "repeatedly", "computing", "the", "convolutional", "features", ".", "In", "processing", "test", "images", ",", "our", "method", "is", "24", "-", "102", "faster", "than", "the", "R", "-", "CNN", "method", ",", "while", "achieving", "better", "or", "comparable", "accuracy", "on", "Pascal", "VOC", "2007", ".", "In", "ImageNet", "Large", "Scale", "Visual", "Recognition", "Challenge", "(", "ILSVRC", ")", "2014", ",", "our", "methods", "rank", "#", "2", "in", "object", "detection", "and", "#", "3", "in", "image", "classification", "among", "all", "38", "teams", ".", "This", "manuscript", "also", "introduces", "the", "improvement", "made", "for", "this", "competition", ".", "onvolutional", "Neural", "Networks", ",", "Spatial", "Pyramid", "Pooling", ",", "Image", "Classification", ",", "Object", "Detection", "section", ":", "Introduction", "We", "are", "witnessing", "a", "rapid", ",", "revolutionary", "change", "in", "our", "vision", "community", ",", "mainly", "caused", "by", "deep", "convolutional", "neural", "networks", "(", "CNNs", ")", "and", "the", "availability", "of", "large", "scale", "training", "data", ".", "Deep", "-", "networks", "-", "based", "approaches", "have", "recently", "been", "substantially", "improving", "upon", "the", "state", "of", "the", "art", "in", "image", "classification", ",", "object", "detection", ",", "many", "other", "recognition", "tasks", ",", "and", "even", "non", "-", "recognition", "tasks", ".", "However", ",", "there", "is", "a", "technical", "issue", "in", "the", "training", "and", "testing", "of", "the", "CNNs", ":", "the", "prevalent", "CNNs", "require", "a", "fixed", "input", "image", "size", "(", ",", "224", "224", ")", ",", "which", "limits", "both", "the", "aspect", "ratio", "and", "the", "scale", "of", "the", "input", "image", ".", "When", "applied", "to", "images", "of", "arbitrary", "sizes", ",", "current", "methods", "mostly", "fit", "the", "input", "image", "to", "the", "fixed", "size", ",", "either", "via", "cropping", "or", "via", "warping", ",", "as", "shown", "in", "Figure", "[", "reference", "]", "(", "top", ")", ".", "But", "the", "cropped", "region", "may", "not", "contain", "the", "entire", "object", ",", "while", "the", "warped", "content", "may", "result", "in", "unwanted", "geometric", "distortion", ".", "Recognition", "accuracy", "can", "be", "compromised", "due", "to", "the", "content", "loss", "or", "distortion", ".", "Besides", ",", "a", "pre", "-", "defined", "scale", "may", "not", "be", "suitable", "when", "object", "scales", "vary", ".", "Fixing", "input", "sizes", "overlooks", "the", "issues", "involving", "scales", ".", "So", "why", "do", "CNNs", "require", "a", "fixed", "input", "size", "?", "A", "CNN", "mainly", "consists", "of", "two", "parts", ":", "convolutional", "layers", ",", "and", "fully", "-", "connected", "layers", "that", "follow", ".", "The", "convolutional", "layers", "operate", "in", "a", "sliding", "-", "window", "manner", "and", "output", "feature", "maps", "which", "represent", "the", "spatial", "arrangement", "of", "the", "activations", "(", "Figure", "[", "reference", "]", ")", ".", "In", "fact", ",", "convolutional", "layers", "do", "not", "require", "a", "fixed", "image", "size", "and", "can", "generate", "feature", "maps", "of", "any", "sizes", ".", "On", "the", "other", "hand", ",", "the", "fully", "-", "connected", "layers", "need", "to", "have", "fixed", "-", "size", "/", "length", "input", "by", "their", "definition", ".", "Hence", ",", "the", "fixed", "-", "size", "constraint", "comes", "only", "from", "the", "fully", "-", "connected", "layers", ",", "which", "exist", "at", "a", "deeper", "stage", "of", "the", "network", ".", "In", "this", "paper", ",", "we", "introduce", "a", "spatial", "pyramid", "pooling", "(", "SPP", ")", "layer", "to", "remove", "the", "fixed", "-", "size", "constraint", "of", "the", "network", ".", "Specifically", ",", "we", "add", "an", "SPP", "layer", "on", "top", "of", "the", "last", "convolutional", "layer", ".", "The", "SPP", "layer", "pools", "the", "features", "and", "generates", "fixed", "-", "length", "outputs", ",", "which", "are", "then", "fed", "into", "the", "fully", "-", "connected", "layers", "(", "or", "other", "classifiers", ")", ".", "In", "other", "words", ",", "we", "perform", "some", "information", "“", "aggregation", "”", "at", "a", "deeper", "stage", "of", "the", "network", "hierarchy", "(", "between", "convolutional", "layers", "and", "fully", "-", "connected", "layers", ")", "to", "avoid", "the", "need", "for", "cropping", "or", "warping", "at", "the", "beginning", ".", "Figure", "[", "reference", "]", "(", "bottom", ")", "shows", "the", "change", "of", "the", "network", "architecture", "by", "introducing", "the", "SPP", "layer", ".", "We", "call", "the", "new", "network", "structure", "SPP", "-", "net", ".", "Spatial", "pyramid", "pooling", "(", "popularly", "known", "as", "spatial", "pyramid", "matching", "or", "SPM", ")", ",", "as", "an", "extension", "of", "the", "Bag", "-", "of", "-", "Words", "(", "BoW", ")", "model", ",", "is", "one", "of", "the", "most", "successful", "methods", "in", "computer", "vision", ".", "It", "partitions", "the", "image", "into", "divisions", "from", "finer", "to", "coarser", "levels", ",", "and", "aggregates", "local", "features", "in", "them", ".", "SPP", "has", "long", "been", "a", "key", "component", "in", "the", "leading", "and", "competition", "-", "winning", "systems", "for", "classification", "(", ",", ")", "and", "detection", "(", ",", ")", "before", "the", "recent", "prevalence", "of", "CNNs", ".", "Nevertheless", ",", "SPP", "has", "not", "been", "considered", "in", "the", "context", "of", "CNNs", ".", "We", "note", "that", "SPP", "has", "several", "remarkable", "properties", "for", "deep", "CNNs", ":", "1", ")", "SPP", "is", "able", "to", "generate", "a", "fixed", "-", "length", "output", "regardless", "of", "the", "input", "size", ",", "while", "the", "sliding", "window", "pooling", "used", "in", "the", "previous", "deep", "networks", "can", "not", ";", "2", ")", "SPP", "uses", "multi", "-", "level", "spatial", "bins", ",", "while", "the", "sliding", "window", "pooling", "uses", "only", "a", "single", "window", "size", ".", "Multi", "-", "level", "pooling", "has", "been", "shown", "to", "be", "robust", "to", "object", "deformations", ";", "3", ")", "SPP", "can", "pool", "features", "extracted", "at", "variable", "scales", "thanks", "to", "the", "flexibility", "of", "input", "scales", ".", "Through", "experiments", "we", "show", "that", "all", "these", "factors", "elevate", "the", "recognition", "accuracy", "of", "deep", "networks", ".", "SPP", "-", "net", "not", "only", "makes", "it", "possible", "to", "generate", "representations", "from", "arbitrarily", "sized", "images", "/", "windows", "for", "testing", ",", "but", "also", "allows", "us", "to", "feed", "images", "with", "varying", "sizes", "or", "scales", "during", "training", ".", "Training", "with", "variable", "-", "size", "images", "increases", "scale", "-", "invariance", "and", "reduces", "over", "-", "fitting", ".", "We", "develop", "a", "simple", "multi", "-", "size", "training", "method", ".", "For", "a", "single", "network", "to", "accept", "variable", "input", "sizes", ",", "we", "approximate", "it", "by", "multiple", "networks", "that", "share", "all", "parameters", ",", "while", "each", "of", "these", "networks", "is", "trained", "using", "a", "fixed", "input", "size", ".", "In", "each", "epoch", "we", "train", "the", "network", "with", "a", "given", "input", "size", ",", "and", "switch", "to", "another", "input", "size", "for", "the", "next", "epoch", ".", "Experiments", "show", "that", "this", "multi", "-", "size", "training", "converges", "just", "as", "the", "traditional", "single", "-", "size", "training", ",", "and", "leads", "to", "better", "testing", "accuracy", ".", "The", "advantages", "of", "SPP", "are", "orthogonal", "to", "the", "specific", "CNN", "designs", ".", "In", "a", "series", "of", "controlled", "experiments", "on", "the", "ImageNet", "2012", "dataset", ",", "we", "demonstrate", "that", "SPP", "improves", "four", "different", "CNN", "architectures", "in", "existing", "publications", "(", "or", "their", "modifications", ")", ",", "over", "the", "no", "-", "SPP", "counterparts", ".", "These", "architectures", "have", "various", "filter", "numbers", "/", "sizes", ",", "strides", ",", "depths", ",", "or", "other", "designs", ".", "It", "is", "thus", "reasonable", "for", "us", "to", "conjecture", "that", "SPP", "should", "improve", "more", "sophisticated", "(", "deeper", "and", "larger", ")", "convolutional", "architectures", ".", "SPP", "-", "net", "also", "shows", "state", "-", "of", "-", "the", "-", "art", "classification", "results", "on", "Caltech101", "and", "Pascal", "VOC", "2007", "using", "only", "a", "single", "full", "-", "image", "representation", "and", "no", "fine", "-", "tuning", ".", "SPP", "-", "net", "also", "shows", "great", "strength", "in", "object", "detection", ".", "In", "the", "leading", "object", "detection", "method", "R", "-", "CNN", ",", "the", "features", "from", "candidate", "windows", "are", "extracted", "via", "deep", "convolutional", "networks", ".", "This", "method", "shows", "remarkable", "detection", "accuracy", "on", "both", "the", "VOC", "and", "ImageNet", "datasets", ".", "But", "the", "feature", "computation", "in", "R", "-", "CNN", "is", "time", "-", "consuming", ",", "because", "it", "repeatedly", "applies", "the", "deep", "convolutional", "networks", "to", "the", "raw", "pixels", "of", "thousands", "of", "warped", "regions", "per", "image", ".", "In", "this", "paper", ",", "we", "show", "that", "we", "can", "run", "the", "convolutional", "layers", "only", "once", "on", "the", "entire", "image", "(", "regardless", "of", "the", "number", "of", "windows", ")", ",", "and", "then", "extract", "features", "by", "SPP", "-", "net", "on", "the", "feature", "maps", ".", "This", "method", "yields", "a", "speedup", "of", "over", "one", "hundred", "times", "over", "R", "-", "CNN", ".", "Note", "that", "training", "/", "running", "a", "detector", "on", "the", "feature", "maps", "(", "rather", "than", "image", "regions", ")", "is", "actually", "a", "more", "popular", "idea", ".", "But", "SPP", "-", "net", "inherits", "the", "power", "of", "the", "deep", "CNN", "feature", "maps", "and", "also", "the", "flexibility", "of", "SPP", "on", "arbitrary", "window", "sizes", ",", "which", "leads", "to", "outstanding", "accuracy", "and", "efficiency", ".", "In", "our", "experiment", ",", "the", "SPP", "-", "net", "-", "based", "system", "(", "built", "upon", "the", "R", "-", "CNN", "pipeline", ")", "computes", "features", "24", "-", "102", "faster", "than", "R", "-", "CNN", ",", "while", "has", "better", "or", "comparable", "accuracy", ".", "With", "the", "recent", "fast", "proposal", "method", "of", "EdgeBoxes", ",", "our", "system", "takes", "0.5", "seconds", "processing", "an", "image", "(", "including", "all", "steps", ")", ".", "This", "makes", "our", "method", "practical", "for", "real", "-", "world", "applications", ".", "A", "preliminary", "version", "of", "this", "manuscript", "has", "been", "published", "in", "ECCV", "2014", ".", "Based", "on", "this", "work", ",", "we", "attended", "the", "competition", "of", "ILSVRC", "2014", ",", "and", "ranked", "#", "2", "in", "object", "detection", "and", "#", "3", "in", "image", "classification", "(", "both", "are", "provided", "-", "data", "-", "only", "tracks", ")", "among", "all", "38", "teams", ".", "There", "are", "a", "few", "modifications", "made", "for", "ILSVRC", "2014", ".", "We", "show", "that", "the", "SPP", "-", "nets", "can", "boost", "various", "networks", "that", "are", "deeper", "and", "larger", "(", "Sec", ".", "[", "reference", "]", "-", "[", "reference", "]", ")", "over", "the", "no", "-", "SPP", "counterparts", ".", "Further", ",", "driven", "by", "our", "detection", "framework", ",", "we", "find", "that", "multi", "-", "view", "testing", "on", "feature", "maps", "with", "flexibly", "located", "/", "sized", "windows", "(", "Sec", ".", "[", "reference", "]", ")", "can", "increase", "the", "classification", "accuracy", ".", "This", "manuscript", "also", "provides", "the", "details", "of", "these", "modifications", ".", "We", "have", "released", "the", "code", "to", "facilitate", "future", "research", "(", "http:", "//", "research.microsoft.com", "/", "en", "-", "us", "/", "um", "/", "people", "/", "kahe", "/", ")", ".", "section", ":", "Deep", "Networks", "with", "Spatial", "Pyramid", "Pooling", "subsection", ":", "Convolutional", "Layers", "and", "Feature", "Maps", "Consider", "the", "popular", "seven", "-", "layer", "architectures", ".", "The", "first", "five", "layers", "are", "convolutional", ",", "some", "of", "which", "are", "followed", "by", "pooling", "layers", ".", "These", "pooling", "layers", "can", "also", "be", "considered", "as", "“", "convolutional", "”", ",", "in", "the", "sense", "that", "they", "are", "using", "sliding", "windows", ".", "The", "last", "two", "layers", "are", "fully", "connected", ",", "with", "an", "N", "-", "way", "softmax", "as", "the", "output", ",", "where", "N", "is", "the", "number", "of", "categories", ".", "The", "deep", "network", "described", "above", "needs", "a", "fixed", "image", "size", ".", "However", ",", "we", "notice", "that", "the", "requirement", "of", "fixed", "sizes", "is", "only", "due", "to", "the", "fully", "-", "connected", "layers", "that", "demand", "fixed", "-", "length", "vectors", "as", "inputs", ".", "On", "the", "other", "hand", ",", "the", "convolutional", "layers", "accept", "inputs", "of", "arbitrary", "sizes", ".", "The", "convolutional", "layers", "use", "sliding", "filters", ",", "and", "their", "outputs", "have", "roughly", "the", "same", "aspect", "ratio", "as", "the", "inputs", ".", "These", "outputs", "are", "known", "as", "feature", "maps", "-", "they", "involve", "not", "only", "the", "strength", "of", "the", "responses", ",", "but", "also", "their", "spatial", "positions", ".", "In", "Figure", "[", "reference", "]", ",", "we", "visualize", "some", "feature", "maps", ".", "They", "are", "generated", "by", "some", "filters", "of", "the", "conv", "layer", ".", "Figure", "[", "reference", "]", "(", "c", ")", "shows", "the", "strongest", "activated", "images", "of", "these", "filters", "in", "the", "ImageNet", "dataset", ".", "We", "see", "a", "filter", "can", "be", "activated", "by", "some", "semantic", "content", ".", "For", "example", ",", "the", "55", "-", "th", "filter", "(", "Figure", "[", "reference", "]", ",", "bottom", "left", ")", "is", "most", "activated", "by", "a", "circle", "shape", ";", "the", "66", "-", "th", "filter", "(", "Figure", "[", "reference", "]", ",", "top", "right", ")", "is", "most", "activated", "by", "a", "-", "shape", ";", "and", "the", "118", "-", "th", "filter", "(", "Figure", "[", "reference", "]", ",", "bottom", "right", ")", "is", "most", "activated", "by", "a", "-", "shape", ".", "These", "shapes", "in", "the", "input", "images", "(", "Figure", "[", "reference", "]", "(", "a", ")", ")", "activate", "the", "feature", "maps", "at", "the", "corresponding", "positions", "(", "the", "arrows", "in", "Figure", "[", "reference", "]", ")", ".", "It", "is", "worth", "noticing", "that", "we", "generate", "the", "feature", "maps", "in", "Figure", "[", "reference", "]", "without", "fixing", "the", "input", "size", ".", "These", "feature", "maps", "generated", "by", "deep", "convolutional", "layers", "are", "analogous", "to", "the", "feature", "maps", "in", "traditional", "methods", ".", "In", "those", "methods", ",", "SIFT", "vectors", "or", "image", "patches", "are", "densely", "extracted", "and", "then", "encoded", ",", ",", "by", "vector", "quantization", ",", "sparse", "coding", ",", "or", "Fisher", "kernels", ".", "These", "encoded", "features", "consist", "of", "the", "feature", "maps", ",", "and", "are", "then", "pooled", "by", "Bag", "-", "of", "-", "Words", "(", "BoW", ")", "or", "spatial", "pyramids", ".", "Analogously", ",", "the", "deep", "convolutional", "features", "can", "be", "pooled", "in", "a", "similar", "way", ".", "subsection", ":", "The", "Spatial", "Pyramid", "Pooling", "Layer", "The", "convolutional", "layers", "accept", "arbitrary", "input", "sizes", ",", "but", "they", "produce", "outputs", "of", "variable", "sizes", ".", "The", "classifiers", "(", "SVM", "/", "softmax", ")", "or", "fully", "-", "connected", "layers", "require", "fixed", "-", "length", "vectors", ".", "Such", "vectors", "can", "be", "generated", "by", "the", "Bag", "-", "of", "-", "Words", "(", "BoW", ")", "approach", "that", "pools", "the", "features", "together", ".", "Spatial", "pyramid", "pooling", "improves", "BoW", "in", "that", "it", "can", "maintain", "spatial", "information", "by", "pooling", "in", "local", "spatial", "bins", ".", "These", "spatial", "bins", "have", "sizes", "proportional", "to", "the", "image", "size", ",", "so", "the", "number", "of", "bins", "is", "fixed", "regardless", "of", "the", "image", "size", ".", "This", "is", "in", "contrast", "to", "the", "sliding", "window", "pooling", "of", "the", "previous", "deep", "networks", ",", "where", "the", "number", "of", "sliding", "windows", "depends", "on", "the", "input", "size", ".", "To", "adopt", "the", "deep", "network", "for", "images", "of", "arbitrary", "sizes", ",", "we", "replace", "the", "last", "pooling", "layer", "(", ",", "pool", ",", "after", "the", "last", "convolutional", "layer", ")", "with", "a", "spatial", "pyramid", "pooling", "layer", ".", "Figure", "[", "reference", "]", "illustrates", "our", "method", ".", "In", "each", "spatial", "bin", ",", "we", "pool", "the", "responses", "of", "each", "filter", "(", "throughout", "this", "paper", "we", "use", "max", "pooling", ")", ".", "The", "outputs", "of", "the", "spatial", "pyramid", "pooling", "are", "-", "dimensional", "vectors", "with", "the", "number", "of", "bins", "denoted", "as", "(", "is", "the", "number", "of", "filters", "in", "the", "last", "convolutional", "layer", ")", ".", "The", "fixed", "-", "dimensional", "vectors", "are", "the", "input", "to", "the", "fully", "-", "connected", "layer", ".", "With", "spatial", "pyramid", "pooling", ",", "the", "input", "image", "can", "be", "of", "any", "sizes", ".", "This", "not", "only", "allows", "arbitrary", "aspect", "ratios", ",", "but", "also", "allows", "arbitrary", "scales", ".", "We", "can", "resize", "the", "input", "image", "to", "any", "scale", "(", ",", "=", "180", ",", "224", ",", "…", ")", "and", "apply", "the", "same", "deep", "network", ".", "When", "the", "input", "image", "is", "at", "different", "scales", ",", "the", "network", "(", "with", "the", "same", "filter", "sizes", ")", "will", "extract", "features", "at", "different", "scales", ".", "The", "scales", "play", "important", "roles", "in", "traditional", "methods", ",", ",", "the", "SIFT", "vectors", "are", "often", "extracted", "at", "multiple", "scales", "(", "determined", "by", "the", "sizes", "of", "the", "patches", "and", "Gaussian", "filters", ")", ".", "We", "will", "show", "that", "the", "scales", "are", "also", "important", "for", "the", "accuracy", "of", "deep", "networks", ".", "Interestingly", ",", "the", "coarsest", "pyramid", "level", "has", "a", "single", "bin", "that", "covers", "the", "entire", "image", ".", "This", "is", "in", "fact", "a", "“", "global", "pooling", "”", "operation", ",", "which", "is", "also", "investigated", "in", "several", "concurrent", "works", ".", "In", "a", "global", "average", "pooling", "is", "used", "to", "reduce", "the", "model", "size", "and", "also", "reduce", "overfitting", ";", "in", ",", "a", "global", "average", "pooling", "is", "used", "on", "the", "testing", "stage", "after", "all", "fc", "layers", "to", "improve", "accuracy", ";", "in", ",", "a", "global", "max", "pooling", "is", "used", "for", "weakly", "supervised", "object", "recognition", ".", "The", "global", "pooling", "operation", "corresponds", "to", "the", "traditional", "Bag", "-", "of", "-", "Words", "method", ".", "subsection", ":", "Training", "the", "Network", "Theoretically", ",", "the", "above", "network", "structure", "can", "be", "trained", "with", "standard", "back", "-", "propagation", ",", "regardless", "of", "the", "input", "image", "size", ".", "But", "in", "practice", "the", "GPU", "implementations", "(", "such", "as", "cuda", "-", "convnet", "and", "Caffe", ")", "are", "preferably", "run", "on", "fixed", "input", "images", ".", "Next", "we", "describe", "our", "training", "solution", "that", "takes", "advantage", "of", "these", "GPU", "implementations", "while", "still", "preserving", "the", "spatial", "pyramid", "pooling", "behaviors", ".", "subsubsection", ":", "Single", "-", "size", "training", "As", "in", "previous", "works", ",", "we", "first", "consider", "a", "network", "taking", "a", "fixed", "-", "size", "input", "(", "224", "224", ")", "cropped", "from", "images", ".", "The", "cropping", "is", "for", "the", "purpose", "of", "data", "augmentation", ".", "For", "an", "image", "with", "a", "given", "size", ",", "we", "can", "pre", "-", "compute", "the", "bin", "sizes", "needed", "for", "spatial", "pyramid", "pooling", ".", "Consider", "the", "feature", "maps", "after", "conv", "that", "have", "a", "size", "of", "(", ",", "13", "13", ")", ".", "With", "a", "pyramid", "level", "of", "bins", ",", "we", "implement", "this", "pooling", "level", "as", "a", "sliding", "window", "pooling", ",", "where", "the", "window", "size", "and", "stride", "with", "and", "denoting", "ceiling", "and", "floor", "operations", ".", "With", "an", "-", "level", "pyramid", ",", "we", "implement", "such", "layers", ".", "The", "next", "fully", "-", "connected", "layer", "(", "fc", ")", "will", "concatenate", "the", "outputs", ".", "Figure", "[", "reference", "]", "shows", "an", "example", "configuration", "of", "3", "-", "level", "pyramid", "pooling", "(", "3", "3", ",", "2", "2", ",", "1", "1", ")", "in", "the", "cuda", "-", "convnet", "style", ".", "The", "main", "purpose", "of", "our", "single", "-", "size", "training", "is", "to", "enable", "the", "multi", "-", "level", "pooling", "behavior", ".", "Experiments", "show", "that", "this", "is", "one", "reason", "for", "the", "gain", "of", "accuracy", ".", "subsubsection", ":", "Multi", "-", "size", "training", "Our", "network", "with", "SPP", "is", "expected", "to", "be", "applied", "on", "images", "of", "any", "sizes", ".", "To", "address", "the", "issue", "of", "varying", "image", "sizes", "in", "training", ",", "we", "consider", "a", "set", "of", "pre", "-", "defined", "sizes", ".", "We", "consider", "two", "sizes", ":", "180", "180", "in", "addition", "to", "224", "224", ".", "Rather", "than", "crop", "a", "smaller", "180", "180", "region", ",", "we", "resize", "the", "aforementioned", "224", "224", "region", "to", "180", "180", ".", "So", "the", "regions", "at", "both", "scales", "differ", "only", "in", "resolution", "but", "not", "in", "content", "/", "layout", ".", "For", "the", "network", "to", "accept", "180", "180", "inputs", ",", "we", "implement", "another", "fixed", "-", "size", "-", "input", "(", "180", "180", ")", "network", ".", "The", "feature", "map", "size", "after", "conv", "is", "10", "in", "this", "case", ".", "Then", "we", "still", "use", "and", "to", "implement", "each", "pyramid", "pooling", "level", ".", "The", "output", "of", "the", "spatial", "pyramid", "pooling", "layer", "of", "this", "180", "-", "network", "has", "the", "same", "fixed", "length", "as", "the", "224", "-", "network", ".", "As", "such", ",", "this", "180", "-", "network", "has", "exactly", "the", "same", "parameters", "as", "the", "224", "-", "network", "in", "each", "layer", ".", "In", "other", "words", ",", "during", "training", "we", "implement", "the", "varying", "-", "input", "-", "size", "SPP", "-", "net", "by", "two", "fixed", "-", "size", "networks", "that", "share", "parameters", ".", "To", "reduce", "the", "overhead", "to", "switch", "from", "one", "network", "(", ",", "224", ")", "to", "the", "other", "(", ",", "180", ")", ",", "we", "train", "each", "full", "epoch", "on", "one", "network", ",", "and", "then", "switch", "to", "the", "other", "one", "(", "keeping", "all", "weights", ")", "for", "the", "next", "full", "epoch", ".", "This", "is", "iterated", ".", "In", "experiments", ",", "we", "find", "the", "convergence", "rate", "of", "this", "multi", "-", "size", "training", "to", "be", "similar", "to", "the", "above", "single", "-", "size", "training", ".", "The", "main", "purpose", "of", "our", "multi", "-", "size", "training", "is", "to", "simulate", "the", "varying", "input", "sizes", "while", "still", "leveraging", "the", "existing", "well", "-", "optimized", "fixed", "-", "size", "implementations", ".", "Besides", "the", "above", "two", "-", "scale", "implementation", ",", "we", "have", "also", "tested", "a", "variant", "using", "as", "input", "where", "is", "randomly", "and", "uniformly", "sampled", "from", "at", "each", "epoch", ".", "We", "report", "the", "results", "of", "both", "variants", "in", "the", "experiment", "section", ".", "Note", "that", "the", "above", "single", "/", "multi", "-", "size", "solutions", "are", "for", "training", "only", ".", "At", "the", "testing", "stage", ",", "it", "is", "straightforward", "to", "apply", "SPP", "-", "net", "on", "images", "of", "any", "sizes", ".", "section", ":", "SPP", "-", "net", "for", "Image", "Classification", "subsection", ":", "Experiments", "on", "ImageNet", "2012", "Classification", "We", "train", "the", "networks", "on", "the", "1000", "-", "category", "training", "set", "of", "ImageNet", "2012", ".", "Our", "training", "algorithm", "follows", "the", "practices", "of", "previous", "work", ".", "The", "images", "are", "resized", "so", "that", "the", "smaller", "dimension", "is", "256", ",", "and", "a", "224", "224", "crop", "is", "picked", "from", "the", "center", "or", "the", "four", "corners", "from", "the", "entire", "image", ".", "The", "data", "are", "augmented", "by", "horizontal", "flipping", "and", "color", "altering", ".", "Dropout", "is", "used", "on", "the", "two", "fully", "-", "connected", "layers", ".", "The", "learning", "rate", "starts", "from", "0.01", ",", "and", "is", "divided", "by", "10", "(", "twice", ")", "when", "the", "error", "plateaus", ".", "Our", "implementation", "is", "based", "on", "the", "publicly", "available", "code", "of", "cuda", "-", "convnet", "and", "Caffe", ".", "All", "networks", "in", "this", "paper", "can", "be", "trained", "on", "a", "single", "GeForce", "GTX", "Titan", "GPU", "(", "6", "GB", "memory", ")", "within", "two", "to", "four", "weeks", ".", "subsubsection", ":", "Baseline", "Network", "Architectures", "The", "advantages", "of", "SPP", "are", "independent", "of", "the", "convolutional", "network", "architectures", "used", ".", "We", "investigate", "four", "different", "network", "architectures", "in", "existing", "publications", "(", "or", "their", "modifications", ")", ",", "and", "we", "show", "SPP", "improves", "the", "accuracy", "of", "all", "these", "architectures", ".", "These", "baseline", "architectures", "are", "in", "Table", "[", "reference", "]", "and", "briefly", "introduced", "below", ":", "ZF", "-", "5", ":", "this", "architecture", "is", "based", "on", "Zeiler", "and", "Fergus", "’s", "(", "ZF", ")", "“", "fast", "”", "(", "smaller", ")", "model", ".", "The", "number", "indicates", "five", "convolutional", "layers", ".", "Convnet*", "-", "5", ":", "this", "is", "a", "modification", "on", "Krizhevsky", "’s", "network", ".", "We", "put", "the", "two", "pooling", "layers", "after", "conv", "and", "conv", "(", "instead", "of", "after", "conv", "and", "conv", ")", ".", "As", "a", "result", ",", "the", "feature", "maps", "after", "each", "layer", "have", "the", "same", "size", "as", "ZF", "-", "5", ".", "Overfeat", "-", "5", "/", "7", ":", "this", "architecture", "is", "based", "on", "the", "Overfeat", "paper", ",", "with", "some", "modifications", "as", "in", ".", "In", "contrast", "to", "ZF", "-", "5", "/", "Convnet*", "-", "5", ",", "this", "architecture", "produces", "a", "larger", "feature", "map", "(", "instead", "of", ")", "before", "the", "last", "pooling", "layer", ".", "A", "larger", "filter", "number", "(", "512", ")", "is", "used", "in", "conv", "and", "the", "following", "convolutional", "layers", ".", "We", "also", "investigate", "a", "deeper", "architecture", "with", "7", "convolutional", "layers", ",", "where", "conv", "to", "conv", "have", "the", "same", "structures", ".", "In", "the", "baseline", "models", ",", "the", "pooling", "layer", "after", "the", "last", "convolutional", "layer", "generates", "feature", "maps", ",", "with", "two", "4096", "-", "d", "fc", "layers", "and", "a", "1000", "-", "way", "softmax", "layer", "following", ".", "Our", "replications", "of", "these", "baseline", "networks", "are", "in", "Table", "[", "reference", "]", "(", "a", ")", ".", "We", "train", "70", "epochs", "for", "ZF", "-", "5", "and", "90", "epochs", "for", "the", "others", ".", "Our", "replication", "of", "ZF", "-", "5", "is", "better", "than", "the", "one", "reported", "in", ".", "This", "gain", "is", "because", "the", "corner", "crops", "are", "from", "the", "entire", "image", ",", "as", "is", "also", "reported", "in", ".", "subsubsection", ":", "Multi", "-", "level", "Pooling", "Improves", "Accuracy", "In", "Table", "[", "reference", "]", "(", "b", ")", "we", "show", "the", "results", "using", "single", "-", "size", "training", ".", "The", "training", "and", "testing", "sizes", "are", "both", "224", "224", ".", "In", "these", "networks", ",", "the", "convolutional", "layers", "have", "the", "same", "structures", "as", "the", "corresponding", "baseline", "models", ",", "whereas", "the", "pooling", "layer", "after", "the", "final", "convolutional", "layer", "is", "replaced", "with", "the", "SPP", "layer", ".", "For", "the", "results", "in", "Table", "[", "reference", "]", ",", "we", "use", "a", "4", "-", "level", "pyramid", ".", "The", "pyramid", "is", "{", "6", "6", ",", "3", "3", ",", "2", "2", ",", "1", "1", "}", "(", "totally", "50", "bins", ")", ".", "For", "fair", "comparison", ",", "we", "still", "use", "the", "standard", "10", "-", "view", "prediction", "with", "each", "view", "a", "224", "224", "crop", ".", "Our", "results", "in", "Table", "[", "reference", "]", "(", "b", ")", "show", "considerable", "improvement", "over", "the", "no", "-", "SPP", "baselines", "in", "Table", "[", "reference", "]", "(", "a", ")", ".", "Interestingly", ",", "the", "largest", "gain", "of", "top", "-", "1", "error", "(", "1.65", "%", ")", "is", "given", "by", "the", "most", "accurate", "architecture", ".", "Since", "we", "are", "still", "using", "the", "same", "10", "cropped", "views", "as", "in", "(", "a", ")", ",", "these", "gains", "are", "solely", "because", "of", "multi", "-", "level", "pooling", ".", "It", "is", "worth", "noticing", "that", "the", "gain", "of", "multi", "-", "level", "pooling", "is", "not", "simply", "due", "to", "more", "parameters", ";", "rather", ",", "it", "is", "because", "the", "multi", "-", "level", "pooling", "is", "robust", "to", "the", "variance", "in", "object", "deformations", "and", "spatial", "layout", ".", "To", "show", "this", ",", "we", "train", "another", "ZF", "-", "5", "network", "with", "a", "different", "4", "-", "level", "pyramid", ":", "{", "4", "4", ",", "3", "3", ",", "2", "2", ",", "1", "1", "}", "(", "totally", "30", "bins", ")", ".", "This", "network", "has", "fewer", "parameters", "than", "its", "no", "-", "SPP", "counterpart", ",", "because", "its", "fc", "layer", "has", "30", "256", "-", "d", "inputs", "instead", "of", "36", "256", "-", "d", ".", "The", "top", "-", "1", "/", "top", "-", "5", "errors", "of", "this", "network", "are", "35.06", "/", "14.04", ".", "This", "result", "is", "similar", "to", "the", "50", "-", "bin", "pyramid", "above", "(", "34.98", "/", "14.14", ")", ",", "but", "considerably", "better", "than", "the", "no", "-", "SPP", "counterpart", "(", "35.99", "/", "14.76", ")", ".", "subsubsection", ":", "Multi", "-", "size", "Training", "Improves", "Accuracy", "Table", "[", "reference", "]", "(", "c", ")", "shows", "our", "results", "using", "multi", "-", "size", "training", ".", "The", "training", "sizes", "are", "224", "and", "180", ",", "while", "the", "testing", "size", "is", "still", "224", ".", "We", "still", "use", "the", "standard", "10", "-", "view", "prediction", ".", "The", "top", "-", "1", "/", "top", "-", "5", "errors", "of", "all", "architectures", "further", "drop", ".", "The", "top", "-", "1", "error", "of", "SPP", "-", "net", "(", "Overfeat", "-", "7", ")", "drops", "to", "29.68", "%", ",", "which", "is", "2.33", "%", "better", "than", "its", "no", "-", "SPP", "counterpart", "and", "0.68", "%", "better", "than", "its", "single", "-", "size", "trained", "counterpart", ".", "Besides", "using", "the", "two", "discrete", "sizes", "of", "180", "and", "224", ",", "we", "have", "also", "evaluated", "using", "a", "random", "size", "uniformly", "sampled", "from", ".", "The", "top", "-", "1", "/", "5", "error", "of", "SPP", "-", "net", "(", "Overfeat", "-", "7", ")", "is", "30.06%", "/", "10.96", "%", ".", "The", "top", "-", "1", "error", "is", "slightly", "worse", "than", "the", "two", "-", "size", "version", ",", "possibly", "because", "the", "size", "of", "224", "(", "which", "is", "used", "for", "testing", ")", "is", "visited", "less", ".", "But", "the", "results", "are", "still", "better", "the", "single", "-", "size", "version", ".", "There", "are", "previous", "CNN", "solutions", "that", "deal", "with", "various", "scales", "/", "sizes", ",", "but", "they", "are", "mostly", "based", "on", "testing", ".", "In", "Overfeat", "and", "Howard", "’s", "method", ",", "the", "single", "network", "is", "applied", "at", "multiple", "scales", "in", "the", "testing", "stage", ",", "and", "the", "scores", "are", "averaged", ".", "Howard", "further", "trains", "two", "different", "networks", "on", "low", "/", "high", "-", "resolution", "image", "regions", "and", "averages", "the", "scores", ".", "To", "our", "knowledge", ",", "our", "method", "is", "the", "first", "one", "that", "trains", "a", "single", "network", "with", "input", "images", "of", "multiple", "sizes", ".", "subsubsection", ":", "Full", "-", "image", "Representations", "Improve", "Accuracy", "Next", "we", "investigate", "the", "accuracy", "of", "the", "full", "-", "image", "views", ".", "We", "resize", "the", "image", "so", "that", "=", "256", "while", "maintaining", "its", "aspect", "ratio", ".", "The", "SPP", "-", "net", "is", "applied", "on", "this", "full", "image", "to", "compute", "the", "scores", "of", "the", "full", "view", ".", "For", "fair", "comparison", ",", "we", "also", "evaluate", "the", "accuracy", "of", "the", "single", "view", "in", "the", "center", "224", "224", "crop", "(", "which", "is", "used", "in", "the", "above", "evaluations", ")", ".", "The", "comparisons", "of", "single", "-", "view", "testing", "accuracy", "are", "in", "Table", "[", "reference", "]", ".", "Here", "we", "evaluate", "ZF", "-", "5", "/", "Overfeat", "-", "7", ".", "The", "top", "-", "1", "error", "rates", "are", "all", "reduced", "by", "the", "full", "-", "view", "representation", ".", "This", "shows", "the", "importance", "of", "maintaining", "the", "complete", "content", ".", "Even", "though", "our", "network", "is", "trained", "using", "square", "images", "only", ",", "it", "generalizes", "well", "to", "other", "aspect", "ratios", ".", "Comparing", "Table", "[", "reference", "]", "and", "Table", "[", "reference", "]", ",", "we", "find", "that", "the", "combination", "of", "multiple", "views", "is", "substantially", "better", "than", "the", "single", "full", "-", "image", "view", ".", "However", ",", "the", "full", "-", "image", "representations", "are", "still", "of", "good", "merits", ".", "First", ",", "we", "empirically", "find", "that", "(", "discussed", "in", "the", "next", "subsection", ")", "even", "for", "the", "combination", "of", "dozens", "of", "views", ",", "the", "additional", "two", "full", "-", "image", "views", "(", "with", "flipping", ")", "can", "still", "boost", "the", "accuracy", "by", "about", "0.2", "%", ".", "Second", ",", "the", "full", "-", "image", "view", "is", "methodologically", "consistent", "with", "the", "traditional", "methods", "where", "the", "encoded", "SIFT", "vectors", "of", "the", "entire", "image", "are", "pooled", "together", ".", "Third", ",", "in", "other", "applications", "such", "as", "image", "retrieval", ",", "an", "image", "representation", ",", "rather", "than", "a", "classification", "score", ",", "is", "required", "for", "similarity", "ranking", ".", "A", "full", "-", "image", "representation", "can", "be", "preferred", ".", "subsubsection", ":", "Multi", "-", "view", "Testing", "on", "Feature", "Maps", "Inspired", "by", "our", "detection", "algorithm", "(", "described", "in", "the", "next", "section", ")", ",", "we", "further", "propose", "a", "multi", "-", "view", "testing", "method", "on", "the", "feature", "maps", ".", "Thanks", "to", "the", "flexibility", "of", "SPP", ",", "we", "can", "easily", "extract", "the", "features", "from", "windows", "(", "views", ")", "of", "arbitrary", "sizes", "from", "the", "convolutional", "feature", "maps", ".", "On", "the", "testing", "stage", ",", "we", "resize", "an", "image", "so", "where", "represents", "a", "predefined", "scale", "(", "like", "256", ")", ".", "Then", "we", "compute", "the", "convolutional", "feature", "maps", "from", "the", "entire", "image", ".", "For", "the", "usage", "of", "flipped", "views", ",", "we", "also", "compute", "the", "feature", "maps", "of", "the", "flipped", "image", ".", "Given", "any", "view", "(", "window", ")", "in", "the", "image", ",", "we", "map", "this", "window", "to", "the", "feature", "maps", "(", "the", "way", "of", "mapping", "is", "in", "Appendix", ")", ",", "and", "then", "use", "SPP", "to", "pool", "the", "features", "from", "this", "window", "(", "see", "Figure", "[", "reference", "]", ")", ".", "The", "pooled", "features", "are", "then", "fed", "into", "the", "fc", "layers", "to", "compute", "the", "softmax", "score", "of", "this", "window", ".", "These", "scores", "are", "averaged", "for", "the", "final", "prediction", ".", "For", "the", "standard", "10", "-", "view", ",", "we", "use", "and", "the", "views", "are", "224", "224", "windows", "on", "the", "corners", "or", "center", ".", "Experiments", "show", "that", "the", "top", "-", "5", "error", "of", "the", "10", "-", "view", "prediction", "on", "feature", "maps", "is", "within", "0.1", "%", "around", "the", "original", "10", "-", "view", "prediction", "on", "image", "crops", ".", "We", "further", "apply", "this", "method", "to", "extract", "multiple", "views", "from", "multiple", "scales", ".", "We", "resize", "the", "image", "to", "six", "scales", "and", "compute", "the", "feature", "maps", "on", "the", "entire", "image", "for", "each", "scale", ".", "We", "use", "as", "the", "view", "size", "for", "any", "scale", ",", "so", "these", "views", "have", "different", "relative", "sizes", "on", "the", "original", "image", "for", "different", "scales", ".", "We", "use", "18", "views", "for", "each", "scale", ":", "one", "at", "the", "center", ",", "four", "at", "the", "corners", ",", "and", "four", "on", "the", "middle", "of", "each", "side", ",", "with", "/", "without", "flipping", "(", "when", "=", "224", "there", "are", "6", "different", "views", ")", ".", "The", "combination", "of", "these", "96", "views", "reduces", "the", "top", "-", "5", "error", "from", "10.95", "%", "to", "9.36", "%", ".", "Combining", "the", "two", "full", "-", "image", "views", "(", "with", "flipping", ")", "further", "reduces", "the", "top", "-", "5", "error", "to", "9.14", "%", ".", "In", "the", "Overfeat", "paper", ",", "the", "views", "are", "also", "extracted", "from", "the", "convolutional", "feature", "maps", "instead", "of", "image", "crops", ".", "However", ",", "their", "views", "can", "not", "have", "arbitrary", "sizes", ";", "rather", ",", "the", "windows", "are", "those", "where", "the", "pooled", "features", "match", "the", "desired", "dimensionality", ".", "We", "empirically", "find", "that", "these", "restricted", "windows", "are", "less", "beneficial", "than", "our", "flexibly", "located", "/", "sized", "windows", ".", "subsubsection", ":", "Summary", "and", "Results", "for", "ILSVRC", "2014", "In", "Table", "[", "reference", "]", "we", "compare", "with", "previous", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "Krizhevsky", "’s", "is", "the", "winning", "method", "in", "ILSVRC", "2012", ";", "Overfeat", ",", "Howard", "’s", ",", "and", "Zeiler", "and", "Fergus", "’s", "are", "the", "leading", "methods", "in", "ILSVRC", "2013", ".", "We", "only", "consider", "single", "-", "network", "performance", "for", "manageable", "comparisons", ".", "Our", "best", "single", "network", "achieves", "9.14", "%", "top", "-", "5", "error", "on", "the", "validation", "set", ".", "This", "is", "exactly", "the", "single", "-", "model", "entry", "we", "submitted", "to", "ILSVRC", "2014", ".", "The", "top", "-", "5", "error", "is", "9.08", "%", "on", "the", "testing", "set", "(", "ILSVRC", "2014", "has", "the", "same", "training", "/", "validation", "/", "testing", "data", "as", "ILSVRC", "2012", ")", ".", "After", "combining", "eleven", "models", ",", "our", "team", "’s", "result", "(", "8.06", "%", ")", "is", "ranked", "#", "3", "among", "all", "38", "teams", "attending", "ILSVRC", "2014", "(", "Table", "[", "reference", "]", ")", ".", "Since", "the", "advantages", "of", "SPP", "-", "net", "should", "be", "in", "general", "independent", "of", "architectures", ",", "we", "expect", "that", "it", "will", "further", "improve", "the", "deeper", "and", "larger", "convolutional", "architectures", ".", "subsection", ":", "Experiments", "on", "VOC", "2007", "Classification", "Our", "method", "can", "generate", "a", "full", "-", "view", "image", "representation", ".", "With", "the", "above", "networks", "pre", "-", "trained", "on", "ImageNet", ",", "we", "extract", "these", "representations", "from", "the", "images", "in", "the", "target", "datasets", "and", "re", "-", "train", "SVM", "classifiers", ".", "In", "the", "SVM", "training", ",", "we", "intentionally", "do", "not", "use", "any", "data", "augmentation", "(", "flip", "/", "multi", "-", "view", ")", ".", "We", "l", "-", "normalize", "the", "features", "for", "SVM", "training", ".", "The", "classification", "task", "in", "Pascal", "VOC", "2007", "involves", "9", ",", "963", "images", "in", "20", "categories", ".", "5", ",", "011", "images", "are", "for", "training", ",", "and", "the", "rest", "are", "for", "testing", ".", "The", "performance", "is", "evaluated", "by", "mean", "Average", "Precision", "(", "mAP", ")", ".", "Table", "[", "reference", "]", "summarizes", "the", "results", ".", "We", "start", "from", "a", "baseline", "in", "Table", "[", "reference", "]", "(", "a", ")", ".", "The", "model", "is", "ZF", "-", "5", "without", "SPP", ".", "To", "apply", "this", "model", ",", "we", "resize", "the", "image", "so", "that", "its", "smaller", "dimension", "is", "224", ",", "and", "crop", "the", "center", "224", "224", "region", ".", "The", "SVM", "is", "trained", "via", "the", "features", "of", "a", "layer", ".", "On", "this", "dataset", ",", "the", "deeper", "the", "layer", "is", ",", "the", "better", "the", "result", "is", ".", "In", "Table", "[", "reference", "]", "(", "b", ")", ",", "we", "replace", "the", "no", "-", "SPP", "net", "with", "our", "SPP", "-", "net", ".", "As", "a", "first", "-", "step", "comparison", ",", "we", "still", "apply", "the", "SPP", "-", "net", "on", "the", "center", "224", "224", "crop", ".", "The", "results", "of", "the", "fc", "layers", "improve", ".", "This", "gain", "is", "mainly", "due", "to", "multi", "-", "level", "pooling", ".", "Table", "[", "reference", "]", "(", "c", ")", "shows", "our", "results", "on", "full", "images", ",", "where", "the", "images", "are", "resized", "so", "that", "the", "shorter", "side", "is", "224", ".", "We", "find", "that", "the", "results", "are", "considerably", "improved", "(", "78.39", "%", "76.45", "%", ")", ".", "This", "is", "due", "to", "the", "full", "-", "image", "representation", "that", "maintains", "the", "complete", "content", ".", "Because", "the", "usage", "of", "our", "network", "does", "not", "depend", "on", "scale", ",", "we", "resize", "the", "images", "so", "that", "the", "smaller", "dimension", "is", "and", "use", "the", "same", "network", "to", "extract", "features", ".", "We", "find", "that", "gives", "the", "best", "results", "(", "Table", "[", "reference", "]", "(", "d", ")", ")", "based", "on", "the", "validation", "set", ".", "This", "is", "mainly", "because", "the", "objects", "occupy", "smaller", "regions", "in", "VOC", "2007", "but", "larger", "regions", "in", "ImageNet", ",", "so", "the", "relative", "object", "scales", "are", "different", "between", "the", "two", "sets", ".", "These", "results", "indicate", "scale", "matters", "in", "the", "classification", "tasks", ",", "and", "SPP", "-", "net", "can", "partially", "address", "this", "“", "scale", "mismatch", "”", "issue", ".", "In", "Table", "[", "reference", "]", "(", "e", ")", "the", "network", "architecture", "is", "replaced", "with", "our", "best", "model", "(", "Overfeat", "-", "7", ",", "multi", "-", "size", "trained", ")", ",", "and", "the", "mAP", "increases", "to", "82.44", "%", ".", "Table", "[", "reference", "]", "summarizes", "our", "results", "and", "the", "comparisons", "with", "the", "state", "-", "of", "-", "the", "-", "art", "methods", ".", "Among", "these", "methods", ",", "VQ", ",", "LCC", ",", "and", "FK", "are", "all", "based", "on", "spatial", "pyramids", "matching", ",", "and", "are", "based", "on", "deep", "networks", ".", "In", "these", "results", ",", "Oquab", "’s", "(", "77.7", "%", ")", "and", "Chatfield", "’s", "(", "82.42", "%", ")", "are", "obtained", "by", "network", "fine", "-", "tuning", "and", "multi", "-", "view", "testing", ".", "Our", "result", "is", "comparable", "with", "the", "state", "of", "the", "art", ",", "using", "only", "a", "single", "full", "-", "image", "representation", "and", "without", "fine", "-", "tuning", ".", "subsection", ":", "Experiments", "on", "Caltech101", "The", "Caltech101", "dataset", "contains", "9", ",", "144", "images", "in", "102", "categories", "(", "one", "background", ")", ".", "We", "randomly", "sample", "30", "images", "per", "category", "for", "training", "and", "up", "to", "50", "images", "per", "category", "for", "testing", ".", "We", "repeat", "10", "random", "splits", "and", "average", "the", "accuracy", ".", "Table", "[", "reference", "]", "summarizes", "our", "results", ".", "There", "are", "some", "common", "observations", "in", "the", "Pascal", "VOC", "2007", "and", "Caltech101", "results", ":", "SPP", "-", "net", "is", "better", "than", "the", "no", "-", "SPP", "net", "(", "Table", "[", "reference", "]", "(", "b", ")", "(", "a", ")", ")", ",", "and", "the", "full", "-", "view", "representation", "is", "better", "than", "the", "crop", "(", "(", "c", ")", "(", "b", ")", ")", ".", "But", "the", "results", "in", "Caltech101", "have", "some", "differences", "with", "Pascal", "VOC", ".", "The", "fully", "-", "connected", "layers", "are", "less", "accurate", ",", "and", "the", "SPP", "layers", "are", "better", ".", "This", "is", "possibly", "because", "the", "object", "categories", "in", "Caltech101", "are", "less", "related", "to", "those", "in", "ImageNet", ",", "and", "the", "deeper", "layers", "are", "more", "category", "-", "specialized", ".", "Further", ",", "we", "find", "that", "the", "scale", "224", "has", "the", "best", "performance", "among", "the", "scales", "we", "tested", "on", "this", "dataset", ".", "This", "is", "mainly", "because", "the", "objects", "in", "Caltech101", "also", "occupy", "large", "regions", "of", "the", "images", ",", "as", "is", "the", "case", "of", "ImageNet", ".", "Besides", "cropping", ",", "we", "also", "evaluate", "warping", "the", "image", "to", "fit", "the", "224", "224", "size", ".", "This", "solution", "maintains", "the", "complete", "content", ",", "but", "introduces", "distortion", ".", "On", "the", "SPP", "(", "ZF", "-", "5", ")", "model", ",", "the", "accuracy", "is", "89.91", "%", "using", "the", "SPP", "layer", "as", "features", "-", "lower", "than", "91.44", "%", "which", "uses", "the", "same", "model", "on", "the", "undistorted", "full", "image", ".", "Table", "[", "reference", "]", "summarizes", "our", "results", "compared", "with", "the", "state", "-", "of", "-", "the", "-", "art", "methods", "on", "Caltech101", ".", "Our", "result", "(", "93.42", "%", ")", "exceeds", "the", "previous", "record", "(", "88.54", "%", ")", "by", "a", "substantial", "margin", "(", "4.88", "%", ")", ".", "section", ":", "SPP", "-", "net", "for", "Object", "Detection", "Deep", "networks", "have", "been", "used", "for", "object", "detection", ".", "We", "briefly", "review", "the", "recent", "state", "-", "of", "-", "the", "-", "art", "R", "-", "CNN", "method", ".", "R", "-", "CNN", "first", "extracts", "about", "2", ",", "000", "candidate", "windows", "from", "each", "image", "via", "selective", "search", ".", "Then", "the", "image", "region", "in", "each", "window", "is", "warped", "to", "a", "fixed", "size", "(", "227", "227", ")", ".", "A", "pre", "-", "trained", "deep", "network", "is", "used", "to", "extract", "the", "feature", "of", "each", "window", ".", "A", "binary", "SVM", "classifier", "is", "then", "trained", "on", "these", "features", "for", "detection", ".", "R", "-", "CNN", "generates", "results", "of", "compelling", "quality", "and", "substantially", "outperforms", "previous", "methods", ".", "However", ",", "because", "R", "-", "CNN", "repeatedly", "applies", "the", "deep", "convolutional", "network", "to", "about", "2", ",", "000", "windows", "per", "image", ",", "it", "is", "time", "-", "consuming", ".", "Feature", "extraction", "is", "the", "major", "timing", "bottleneck", "in", "testing", ".", "Our", "SPP", "-", "net", "can", "also", "be", "used", "for", "object", "detection", ".", "We", "extract", "the", "feature", "maps", "from", "the", "entire", "image", "only", "once", "(", "possibly", "at", "multiple", "scales", ")", ".", "Then", "we", "apply", "the", "spatial", "pyramid", "pooling", "on", "each", "candidate", "window", "of", "the", "feature", "maps", "to", "pool", "a", "fixed", "-", "length", "representation", "of", "this", "window", "(", "see", "Figure", "[", "reference", "]", ")", ".", "Because", "the", "time", "-", "consuming", "convolutions", "are", "only", "applied", "once", ",", "our", "method", "can", "run", "orders", "of", "magnitude", "faster", ".", "Our", "method", "extracts", "window", "-", "wise", "features", "from", "regions", "of", "the", "feature", "maps", ",", "while", "R", "-", "CNN", "extracts", "directly", "from", "image", "regions", ".", "In", "previous", "works", ",", "the", "Deformable", "Part", "Model", "(", "DPM", ")", "extracts", "features", "from", "windows", "in", "HOG", "feature", "maps", ",", "and", "the", "Selective", "Search", "(", "SS", ")", "method", "extracts", "from", "windows", "in", "encoded", "SIFT", "feature", "maps", ".", "The", "Overfeat", "detection", "method", "also", "extracts", "from", "windows", "of", "deep", "convolutional", "feature", "maps", ",", "but", "needs", "to", "pre", "-", "define", "the", "window", "size", ".", "On", "the", "contrary", ",", "our", "method", "enables", "feature", "extraction", "in", "arbitrary", "windows", "from", "the", "deep", "convolutional", "feature", "maps", ".", "subsection", ":", "Detection", "Algorithm", "We", "use", "the", "“", "fast", "”", "mode", "of", "selective", "search", "to", "generate", "about", "2", ",", "000", "candidate", "windows", "per", "image", ".", "Then", "we", "resize", "the", "image", "such", "that", ",", "and", "extract", "the", "feature", "maps", "from", "the", "entire", "image", ".", "We", "use", "the", "SPP", "-", "net", "model", "of", "ZF", "-", "5", "(", "single", "-", "size", "trained", ")", "for", "the", "time", "being", ".", "In", "each", "candidate", "window", ",", "we", "use", "a", "4", "-", "level", "spatial", "pyramid", "(", "1", "1", ",", "2", "2", ",", "3", "3", ",", "6", "6", ",", "totally", "50", "bins", ")", "to", "pool", "the", "features", ".", "This", "generates", "a", "12", ",", "800", "-", "d", "(", "256", "50", ")", "representation", "for", "each", "window", ".", "These", "representations", "are", "provided", "to", "the", "fully", "-", "connected", "layers", "of", "the", "network", ".", "Then", "we", "train", "a", "binary", "linear", "SVM", "classifier", "for", "each", "category", "on", "these", "features", ".", "Our", "implementation", "of", "the", "SVM", "training", "follows", ".", "We", "use", "the", "ground", "-", "truth", "windows", "to", "generate", "the", "positive", "samples", ".", "The", "negative", "samples", "are", "those", "overlapping", "a", "positive", "window", "by", "at", "most", "30", "%", "(", "measured", "by", "the", "intersection", "-", "over", "-", "union", "(", "IoU", ")", "ratio", ")", ".", "Any", "negative", "sample", "is", "removed", "if", "it", "overlaps", "another", "negative", "sample", "by", "more", "than", "70", "%", ".", "We", "apply", "the", "standard", "hard", "negative", "mining", "to", "train", "the", "SVM", ".", "This", "step", "is", "iterated", "once", ".", "It", "takes", "less", "than", "1", "hour", "to", "train", "SVMs", "for", "all", "20", "categories", ".", "In", "testing", ",", "the", "classifier", "is", "used", "to", "score", "the", "candidate", "windows", ".", "Then", "we", "use", "non", "-", "maximum", "suppression", "(", "threshold", "of", "30", "%", ")", "on", "the", "scored", "windows", ".", "Our", "method", "can", "be", "improved", "by", "multi", "-", "scale", "feature", "extraction", ".", "We", "resize", "the", "image", "such", "that", ",", "and", "compute", "the", "feature", "maps", "of", "conv", "for", "each", "scale", ".", "One", "strategy", "of", "combining", "the", "features", "from", "these", "scales", "is", "to", "pool", "them", "channel", "-", "by", "-", "channel", ".", "But", "we", "empirically", "find", "that", "another", "strategy", "provides", "better", "results", ".", "For", "each", "candidate", "window", ",", "we", "choose", "a", "single", "scale", "such", "that", "the", "scaled", "candidate", "window", "has", "a", "number", "of", "pixels", "closest", "to", "224", "224", ".", "Then", "we", "only", "use", "the", "feature", "maps", "extracted", "from", "this", "scale", "to", "compute", "the", "feature", "of", "this", "window", ".", "If", "the", "pre", "-", "defined", "scales", "are", "dense", "enough", "and", "the", "window", "is", "approximately", "square", ",", "our", "method", "is", "roughly", "equivalent", "to", "resizing", "the", "window", "to", "224", "224", "and", "then", "extracting", "features", "from", "it", ".", "Nevertheless", ",", "our", "method", "only", "requires", "computing", "the", "feature", "maps", "once", "(", "at", "each", "scale", ")", "from", "the", "entire", "image", ",", "regardless", "of", "the", "number", "of", "candidate", "windows", ".", "We", "also", "fine", "-", "tune", "our", "pre", "-", "trained", "network", ",", "following", ".", "Since", "our", "features", "are", "pooled", "from", "the", "conv", "feature", "maps", "from", "windows", "of", "any", "sizes", ",", "for", "simplicity", "we", "only", "fine", "-", "tune", "the", "fully", "-", "connected", "layers", ".", "In", "this", "case", ",", "the", "data", "layer", "accepts", "the", "fixed", "-", "length", "pooled", "features", "after", "conv", ",", "and", "the", "fc", "layers", "and", "a", "new", "21", "-", "way", "(", "one", "extra", "negative", "category", ")", "fc", "layer", "follow", ".", "The", "fc", "weights", "are", "initialized", "with", "a", "Gaussian", "distribution", "of", "=", "0.01", ".", "We", "fix", "all", "the", "learning", "rates", "to", "1e", "-", "4", "and", "then", "adjust", "to", "1e", "-", "5", "for", "all", "three", "layers", ".", "During", "fine", "-", "tuning", ",", "the", "positive", "samples", "are", "those", "overlapping", "with", "a", "ground", "-", "truth", "window", "by", ",", "and", "the", "negative", "samples", "by", ".", "In", "each", "mini", "-", "batch", ",", "25", "%", "of", "the", "samples", "are", "positive", ".", "We", "train", "250k", "mini", "-", "batches", "using", "the", "learning", "rate", "1e", "-", "4", ",", "and", "then", "50k", "mini", "-", "batches", "using", "1e", "-", "5", ".", "Because", "we", "only", "fine", "-", "tune", "the", "fc", "layers", ",", "the", "training", "is", "very", "fast", "and", "takes", "about", "2", "hours", "on", "the", "GPU", "(", "excluding", "pre", "-", "caching", "feature", "maps", "which", "takes", "about", "1", "hour", ")", ".", "Also", "following", ",", "we", "use", "bounding", "box", "regression", "to", "post", "-", "process", "the", "prediction", "windows", ".", "The", "features", "used", "for", "regression", "are", "the", "pooled", "features", "from", "conv", "(", "as", "a", "counterpart", "of", "the", "pool", "features", "used", "in", ")", ".", "The", "windows", "used", "for", "the", "regression", "training", "are", "those", "overlapping", "with", "a", "ground", "-", "truth", "window", "by", "at", "least", "50", "%", ".", "subsection", ":", "Detection", "Results", "We", "evaluate", "our", "method", "on", "the", "detection", "task", "of", "the", "Pascal", "VOC", "2007", "dataset", ".", "Table", "[", "reference", "]", "shows", "our", "results", "on", "various", "layers", ",", "by", "using", "1", "-", "scale", "(", "=", "688", ")", "or", "5", "-", "scale", ".", "Here", "the", "R", "-", "CNN", "results", "are", "as", "reported", "in", "using", "the", "AlexNet", "with", "5", "conv", "layers", ".", "Using", "the", "pool", "layers", "(", "in", "our", "case", "the", "pooled", "features", ")", ",", "our", "result", "(", "44.9", "%", ")", "is", "comparable", "with", "R", "-", "CNN", "’s", "result", "(", "44.2", "%", ")", ".", "But", "using", "the", "non", "-", "fine", "-", "tuned", "fc", "layers", ",", "our", "results", "are", "inferior", ".", "An", "explanation", "is", "that", "our", "fc", "layers", "are", "pre", "-", "trained", "using", "image", "regions", ",", "while", "in", "the", "detection", "case", "they", "are", "used", "on", "the", "feature", "map", "regions", ".", "The", "feature", "map", "regions", "can", "have", "strong", "activations", "near", "the", "window", "boundaries", ",", "while", "the", "image", "regions", "may", "not", ".", "This", "difference", "of", "usages", "can", "be", "addressed", "by", "fine", "-", "tuning", ".", "Using", "the", "fine", "-", "tuned", "fc", "layers", "(", "ftfc", ")", ",", "our", "results", "are", "comparable", "with", "or", "slightly", "better", "than", "the", "fine", "-", "tuned", "results", "of", "R", "-", "CNN", ".", "After", "bounding", "box", "regression", ",", "our", "5", "-", "scale", "result", "(", "59.2", "%", ")", "is", "0.7", "%", "better", "than", "R", "-", "CNN", "(", "58.5", "%", ")", ",", "and", "our", "1", "-", "scale", "result", "(", "58.0", "%", ")", "is", "0.5", "%", "worse", ".", "In", "Table", "[", "reference", "]", "we", "further", "compare", "with", "R", "-", "CNN", "using", "the", "same", "pre", "-", "trained", "model", "of", "SPPnet", "(", "ZF", "-", "5", ")", ".", "In", "this", "case", ",", "our", "method", "and", "R", "-", "CNN", "have", "comparable", "averaged", "scores", ".", "The", "R", "-", "CNN", "result", "is", "boosted", "by", "this", "pre", "-", "trained", "model", ".", "This", "is", "because", "of", "the", "better", "architecture", "of", "ZF", "-", "5", "than", "AlexNet", ",", "and", "also", "because", "of", "the", "multi", "-", "level", "pooling", "of", "SPPnet", "(", "if", "using", "the", "no", "-", "SPP", "ZF", "-", "5", ",", "the", "R", "-", "CNN", "result", "drops", ")", ".", "Table", "[", "reference", "]", "shows", "the", "results", "for", "each", "category", ".", "Table", "[", "reference", "]", "also", "includes", "additional", "methods", ".", "Selective", "Search", "(", "SS", ")", "applies", "spatial", "pyramid", "matching", "on", "SIFT", "feature", "maps", ".", "DPM", "and", "Regionlet", "are", "based", "on", "HOG", "features", ".", "The", "Regionlet", "method", "improves", "to", "46.1", "%", "by", "combining", "various", "features", "including", "conv", ".", "DetectorNet", "trains", "a", "deep", "network", "that", "outputs", "pixel", "-", "wise", "object", "masks", ".", "This", "method", "only", "needs", "to", "apply", "the", "deep", "network", "once", "to", "the", "entire", "image", ",", "as", "is", "the", "case", "for", "our", "method", ".", "But", "this", "method", "has", "lower", "mAP", "(", "30.5", "%", ")", ".", "subsection", ":", "Complexity", "and", "Running", "Time", "Despite", "having", "comparable", "accuracy", ",", "our", "method", "is", "much", "faster", "than", "R", "-", "CNN", ".", "The", "complexity", "of", "the", "convolutional", "feature", "computation", "in", "R", "-", "CNN", "is", "with", "the", "window", "number", "(", "2000", ")", ".", "This", "complexity", "of", "our", "method", "is", "at", "a", "scale", ",", "where", "is", "the", "aspect", "ratio", ".", "Assume", "is", "about", "4", "/", "3", ".", "In", "the", "single", "-", "scale", "version", "when", ",", "this", "complexity", "is", "about", "1", "/", "160", "of", "R", "-", "CNN", "’s", ";", "in", "the", "5", "-", "scale", "version", ",", "this", "complexity", "is", "about", "1", "/", "24", "of", "R", "-", "CNN", "’s", ".", "In", "Table", "[", "reference", "]", ",", "we", "provide", "a", "fair", "comparison", "on", "the", "running", "time", "of", "the", "feature", "computation", "using", "the", "same", "SPP", "(", "ZF", "-", "5", ")", "model", ".", "The", "implementation", "of", "R", "-", "CNN", "is", "from", "the", "code", "published", "by", "the", "authors", "implemented", "in", "Caffe", ".", "We", "also", "implement", "our", "feature", "computation", "in", "Caffe", ".", "In", "Table", "[", "reference", "]", "we", "evaluate", "the", "average", "time", "of", "100", "random", "VOC", "images", "using", "GPU", ".", "R", "-", "CNN", "takes", "14.37s", "per", "image", "for", "convolutions", ",", "while", "our", "1", "-", "scale", "version", "takes", "only", "0.053s", "per", "image", ".", "So", "ours", "is", "270", "faster", "than", "R", "-", "CNN", ".", "Our", "5", "-", "scale", "version", "takes", "0.293s", "per", "image", "for", "convolutions", ",", "so", "is", "49", "faster", "than", "R", "-", "CNN", ".", "Our", "convolutional", "feature", "computation", "is", "so", "fast", "that", "the", "computational", "time", "of", "fc", "layers", "takes", "a", "considerable", "portion", ".", "Table", "[", "reference", "]", "shows", "that", "the", "GPU", "time", "of", "computing", "the", "4", ",", "096", "-", "d", "fc", "features", "is", "0.089s", "per", "image", ".", "Considering", "both", "convolutional", "and", "fully", "-", "connected", "features", ",", "our", "1", "-", "scale", "version", "is", "102×", "faster", "than", "R", "-", "CNN", "and", "is", "1.2", "%", "inferior", ";", "our", "5", "-", "scale", "version", "is", "38×", "faster", "and", "has", "comparable", "results", ".", "We", "also", "compares", "the", "running", "time", "in", "Table", "[", "reference", "]", "where", "R", "-", "CNN", "uses", "AlexNet", "as", "is", "in", "the", "original", "paper", ".", "Our", "method", "is", "24", "to", "64", "faster", ".", "Note", "that", "the", "AlexNet", "has", "the", "same", "number", "of", "filters", "as", "our", "ZF", "-", "5", "on", "each", "conv", "layer", ".", "The", "AlexNet", "is", "faster", "because", "it", "uses", "splitting", "on", "some", "layers", ",", "which", "was", "designed", "for", "two", "GPUs", "in", ".", "We", "further", "achieve", "an", "efficient", "full", "system", "with", "the", "help", "of", "the", "recent", "window", "proposal", "method", ".", "The", "Selective", "Search", "(", "SS", ")", "proposal", "takes", "about", "1", "-", "2", "seconds", "per", "image", "on", "a", "CPU", ".", "The", "method", "of", "EdgeBoxes", "only", "takes", "0.2s", ".", "Note", "that", "it", "is", "sufficient", "to", "use", "a", "fast", "proposal", "method", "during", "testing", "only", ".", "Using", "the", "same", "model", "trained", "as", "above", "(", "using", "SS", ")", ",", "we", "test", "proposals", "generated", "by", "EdgeBoxes", "only", ".", "The", "mAP", "is", "52.8", "without", "bounding", "box", "regression", ".", "This", "is", "reasonable", "considering", "that", "EdgeBoxes", "are", "not", "used", "for", "training", ".", "Then", "we", "use", "both", "SS", "and", "EdgeBox", "as", "proposals", "in", "the", "training", "stage", ",", "and", "adopt", "only", "EdgeBoxes", "in", "the", "testing", "stage", ".", "The", "mAP", "is", "56.3", "without", "bounding", "box", "regression", ",", "which", "is", "better", "than", "55.2", "(", "Table", "[", "reference", "]", ")", "due", "to", "additional", "training", "samples", ".", "In", "this", "case", ",", "the", "overall", "testing", "time", "is", "0.5s", "per", "image", "including", "all", "steps", "(", "proposal", "and", "recognition", ")", ".", "This", "makes", "our", "method", "practical", "for", "real", "-", "world", "applications", ".", "subsection", ":", "Model", "Combination", "for", "Detection", "Model", "combination", "is", "an", "important", "strategy", "for", "boosting", "CNN", "-", "based", "classification", "accuracy", ".", "We", "propose", "a", "simple", "combination", "method", "for", "detection", ".", "We", "pre", "-", "train", "another", "network", "in", "ImageNet", ",", "using", "the", "same", "structure", "but", "different", "random", "initializations", ".", "Then", "we", "repeat", "the", "above", "detection", "algorithm", ".", "Table", "[", "reference", "]", "(", "SPP", "-", "net", "(", "2", ")", ")", "shows", "the", "results", "of", "this", "network", ".", "Its", "mAP", "is", "comparable", "with", "the", "first", "network", "(", "59.1", "%", "59.2", "%", ")", ",", "and", "outperforms", "the", "first", "network", "in", "11", "categories", ".", "Given", "the", "two", "models", ",", "we", "first", "use", "either", "model", "to", "score", "all", "candidate", "windows", "on", "the", "test", "image", ".", "Then", "we", "perform", "non", "-", "maximum", "suppression", "on", "the", "union", "of", "the", "two", "sets", "of", "candidate", "windows", "(", "with", "their", "scores", ")", ".", "A", "more", "confident", "window", "given", "by", "one", "method", "can", "suppress", "those", "less", "confident", "given", "by", "the", "other", "method", ".", "After", "combination", ",", "the", "mAP", "is", "boosted", "to", "60.9", "%", "(", "Table", "[", "reference", "]", ")", ".", "In", "17", "out", "of", "all", "20", "categories", "the", "combination", "performs", "better", "than", "either", "individual", "model", ".", "This", "indicates", "that", "the", "two", "models", "are", "complementary", ".", "We", "further", "find", "that", "the", "complementarity", "is", "mainly", "because", "of", "the", "convolutional", "layers", ".", "We", "have", "tried", "to", "combine", "two", "randomly", "initialized", "fine", "-", "tuned", "results", "of", "the", "same", "convolutional", "model", ",", "and", "found", "no", "gain", ".", "subsection", ":", "ILSVRC", "2014", "Detection", "The", "ILSVRC", "2014", "detection", "task", "involves", "200", "categories", ".", "There", "are", "450k", "/", "20k", "/", "40k", "images", "in", "the", "training", "/", "validation", "/", "testing", "sets", ".", "We", "focus", "on", "the", "task", "of", "the", "provided", "-", "data", "-", "only", "track", "(", "the", "1000", "-", "category", "CLS", "training", "data", "is", "not", "allowed", "to", "use", ")", ".", "There", "are", "three", "major", "differences", "between", "the", "detection", "(", "DET", ")", "and", "classification", "(", "CLS", ")", "training", "datasets", ",", "which", "greatly", "impacts", "the", "pre", "-", "training", "quality", ".", "First", ",", "the", "DET", "training", "data", "is", "merely", "1", "/", "3", "of", "the", "CLS", "training", "data", ".", "This", "seems", "to", "be", "a", "fundamental", "challenge", "of", "the", "provided", "-", "data", "-", "only", "DET", "task", ".", "Second", ",", "the", "category", "number", "of", "DET", "is", "1", "/", "5", "of", "CLS", ".", "To", "overcome", "this", "problem", ",", "we", "harness", "the", "provided", "subcategory", "labels", "for", "pre", "-", "training", ".", "There", "are", "totally", "499", "non", "-", "overlapping", "subcategories", "(", ",", "the", "leaf", "nodes", "in", "the", "provided", "category", "hierarchy", ")", ".", "So", "we", "pre", "-", "train", "a", "499", "-", "category", "network", "on", "the", "DET", "training", "set", ".", "Third", ",", "the", "distributions", "of", "object", "scales", "are", "different", "between", "DET", "/", "CLS", "training", "sets", ".", "The", "dominant", "object", "scale", "in", "CLS", "is", "about", "0.8", "of", "the", "image", "length", ",", "but", "in", "DET", "is", "about", "0.5", ".", "To", "address", "the", "scale", "difference", ",", "we", "resize", "each", "training", "image", "to", "(", "instead", "of", ")", ",", "and", "randomly", "crop", "views", "for", "training", ".", "A", "crop", "is", "only", "used", "when", "it", "overlaps", "with", "a", "ground", "truth", "object", "by", "at", "least", "50", "%", ".", "We", "verify", "the", "effect", "of", "pre", "-", "training", "on", "Pascal", "VOC", "2007", ".", "For", "a", "CLS", "-", "pre", "-", "training", "baseline", ",", "we", "consider", "the", "pool", "features", "(", "mAP", "43.0", "%", "in", "Table", "[", "reference", "]", ")", ".", "Replaced", "with", "a", "200", "-", "category", "network", "pre", "-", "trained", "on", "DET", ",", "the", "mAP", "significantly", "drops", "to", "32.7", "%", ".", "A", "499", "-", "category", "pre", "-", "trained", "network", "improves", "the", "result", "to", "35.9", "%", ".", "Interestingly", ",", "even", "if", "the", "amount", "of", "training", "data", "do", "not", "increase", ",", "training", "a", "network", "of", "more", "categories", "boosts", "the", "feature", "quality", ".", "Finally", ",", "training", "with", "instead", "of", "further", "improves", "the", "mAP", "to", "37.8", "%", ".", "Even", "so", ",", "we", "see", "that", "there", "is", "still", "a", "considerable", "gap", "to", "the", "CLS", "-", "pre", "-", "training", "result", ".", "This", "indicates", "the", "importance", "of", "big", "data", "to", "deep", "learning", ".", "For", "ILSVRC", "2014", ",", "we", "train", "a", "499", "-", "category", "Overfeat", "-", "7", "SPP", "-", "net", ".", "The", "remaining", "steps", "are", "similar", "to", "the", "VOC", "2007", "case", ".", "Following", ",", "we", "use", "the", "validation", "set", "to", "generate", "the", "positive", "/", "negative", "samples", ",", "with", "windows", "proposed", "by", "the", "selective", "search", "fast", "mode", ".", "The", "training", "set", "only", "contributes", "positive", "samples", "using", "the", "ground", "truth", "windows", ".", "We", "fine", "-", "tune", "the", "fc", "layers", "and", "then", "train", "the", "SVMs", "using", "the", "samples", "in", "both", "validation", "and", "training", "sets", ".", "The", "bounding", "box", "regression", "is", "trained", "on", "the", "validation", "set", ".", "Our", "single", "model", "leads", "to", "31.84", "%", "mAP", "in", "the", "ILSVRC", "2014", "testing", "set", ".", "We", "combine", "six", "similar", "models", "using", "the", "strategy", "introduced", "in", "this", "paper", ".", "The", "mAP", "is", "35.11", "%", "in", "the", "testing", "set", ".", "This", "result", "ranks", "#", "2", "in", "the", "provided", "-", "data", "-", "only", "track", "of", "ILSVRC", "2014", "(", "Table", "[", "reference", "]", ")", ".", "The", "winning", "result", "is", "37.21", "%", "from", "NUS", ",", "which", "uses", "contextual", "information", ".", "Our", "system", "still", "shows", "great", "advantages", "on", "speed", "for", "this", "dataset", ".", "It", "takes", "our", "single", "model", "0.6", "seconds", "(", "0.5", "for", "conv", ",", "0.1", "for", "fc", ",", "excluding", "proposals", ")", "per", "testing", "image", "on", "a", "GPU", "extracting", "convolutional", "features", "from", "all", "5", "scales", ".", "Using", "the", "same", "model", ",", "it", "takes", "32", "seconds", "per", "image", "in", "the", "way", "of", "RCNN", ".", "For", "the", "40k", "testing", "images", ",", "our", "method", "requires", "8", "GPU", "hours", "to", "compute", "convolutional", "features", ",", "while", "RCNN", "would", "require", "15", "GPU", "days", ".", "section", ":", "Conclusion", "SPP", "is", "a", "flexible", "solution", "for", "handling", "different", "scales", ",", "sizes", ",", "and", "aspect", "ratios", ".", "These", "issues", "are", "important", "in", "visual", "recognition", ",", "but", "received", "little", "consideration", "in", "the", "context", "of", "deep", "networks", ".", "We", "have", "suggested", "a", "solution", "to", "train", "a", "deep", "network", "with", "a", "spatial", "pyramid", "pooling", "layer", ".", "The", "resulting", "SPP", "-", "net", "shows", "outstanding", "accuracy", "in", "classification", "/", "detection", "tasks", "and", "greatly", "accelerates", "DNN", "-", "based", "detection", ".", "Our", "studies", "also", "show", "that", "many", "time", "-", "proven", "techniques", "/", "insights", "in", "computer", "vision", "can", "still", "play", "important", "roles", "in", "deep", "-", "networks", "-", "based", "recognition", ".", "section", ":", "In", "the", "appendix", ",", "we", "describe", "some", "implementation", "details", ":", "Mean", "Subtraction", ".", "The", "224", "224", "cropped", "training", "/", "testing", "images", "are", "often", "pre", "-", "processed", "by", "subtracting", "the", "per", "-", "pixel", "mean", ".", "When", "input", "images", "are", "in", "any", "sizes", ",", "the", "fixed", "-", "size", "mean", "image", "is", "not", "directly", "applicable", ".", "In", "the", "ImageNet", "dataset", ",", "we", "warp", "the", "224", "224", "mean", "image", "to", "the", "desired", "size", "and", "then", "subtract", "it", ".", "In", "Pascal", "VOC", "2007", "and", "Caltech101", ",", "we", "use", "the", "constant", "mean", "(", "128", ")", "in", "all", "the", "experiments", ".", "Implementation", "of", "Pooling", "Bins", ".", "We", "use", "the", "following", "implementation", "to", "handle", "all", "bins", "when", "applying", "the", "network", ".", "Denote", "the", "width", "and", "height", "of", "the", "conv", "feature", "maps", "(", "can", "be", "the", "full", "image", "or", "a", "window", ")", "as", "and", ".", "For", "a", "pyramid", "level", "with", "bins", ",", "the", "-", "th", "bin", "is", "in", "the", "range", "of", ".", "Intuitively", ",", "if", "rounding", "is", "needed", ",", "we", "take", "the", "floor", "operation", "on", "the", "left", "/", "top", "boundary", "and", "ceiling", "on", "the", "right", "/", "bottom", "boundary", ".", "Mapping", "a", "Window", "to", "Feature", "Maps", ".", "In", "the", "detection", "algorithm", "(", "and", "multi", "-", "view", "testing", "on", "feature", "maps", ")", ",", "a", "window", "is", "given", "in", "the", "image", "domain", ",", "and", "we", "use", "it", "to", "crop", "the", "convolutional", "feature", "maps", "(", ",", "conv", ")", "which", "have", "been", "sub", "-", "sampled", "several", "times", ".", "So", "we", "need", "to", "align", "the", "window", "on", "the", "feature", "maps", ".", "In", "our", "implementation", ",", "we", "project", "the", "corner", "point", "of", "a", "window", "onto", "a", "pixel", "in", "the", "feature", "maps", ",", "such", "that", "this", "corner", "point", "in", "the", "image", "domain", "is", "closest", "to", "the", "center", "of", "the", "receptive", "field", "of", "that", "feature", "map", "pixel", ".", "The", "mapping", "is", "complicated", "by", "the", "padding", "of", "all", "convolutional", "and", "pooling", "layers", ".", "To", "simplify", "the", "implementation", ",", "during", "deployment", "we", "pad", "pixels", "for", "a", "layer", "with", "a", "filter", "size", "of", ".", "As", "such", ",", "for", "a", "response", "centered", "at", ",", "its", "effective", "receptive", "field", "in", "the", "image", "domain", "is", "centered", "at", "where", "is", "the", "product", "of", "all", "previous", "strides", ".", "In", "our", "models", ",", "for", "ZF", "-", "5", "on", "conv", ",", "and", "for", "Overfeat", "-", "5", "/", "7", "on", "conv", ".", "Given", "a", "window", "in", "the", "image", "domain", ",", "we", "project", "the", "left", "(", "top", ")", "boundary", "by", ":", "and", "the", "right", "(", "bottom", ")", "boundary", ".", "If", "the", "padding", "is", "not", ",", "we", "need", "to", "add", "a", "proper", "offset", "to", ".", "bibliography", ":", "References", "section", ":", "Changelog", "arXiv", "v1", ".", "Initial", "technical", "report", "for", "ECCV", "2014", "paper", ".", "arXiv", "v2", ".", "Submitted", "version", "for", "TPAMI", ".", "Includes", "extra", "experiments", "of", "SPP", "on", "various", "architectures", ".", "Includes", "details", "for", "ILSVRC", "2014", ".", "arXiv", "v3", ".", "Accepted", "version", "for", "TPAMI", ".", "Includes", "comparisons", "with", "R", "-", "CNN", "using", "the", "same", "architecture", ".", "Includes", "detection", "experiments", "using", "EdgeBoxes", ".", "arXiv", "v4", ".", "Revised", "“", "Mapping", "a", "Window", "to", "Feature", "Maps", "”", "in", "Appendix", "for", "easier", "implementation", "."], "entity": [[2, 5, ["Spatial", "Pyramid", "Pooling"], "Method"], [6, 9, ["Deep", "Convolutional", "Networks"], "Method"], [10, 12, ["Visual", "Recognition"], "Task"], [13, 17, ["deep", "convolutional", "neural", "networks"], "Method"], [18, 19, ["CNNs"], "Method"], [43, 45, ["recognition", "accuracy"], "Metric"], [69, 71, ["pooling", "strategy"], "Method"], [73, 76, ["spatial", "pyramid", "pooling"], "Method"], [86, 88, ["network", "structure"], "Method"], [90, 93, ["SPP", "-", "net"], "Method"], [97, 101, ["fixed", "-", "length", "representation"], "Method"], [108, 110, ["Pyramid", "pooling"], "Method"], [121, 124, ["SPP", "-", "net"], "Method"], [129, 135, ["CNN", "-", "based", "image", "classification", "methods"], "Method"], [138, 141, ["ImageNet", "2012", "dataset"], "Material"], [145, 148, ["SPP", "-", "net"], "Method"], [150, 151, ["accuracy"], "Metric"], [155, 157, ["CNN", "architectures"], "Method"], [164, 167, ["Pascal", "VOC", "2007"], "Material"], [168, 170, ["Caltech101", "datasets"], "Material"], [171, 174, ["SPP", "-", "net"], "Method"], [182, 183, ["classification"], "Task"], [187, 191, ["full", "-", "image", "representation"], "Method"], [193, 196, ["fine", "-", "tuning"], "Method"], [200, 203, ["SPP", "-", "net"], "Method"], [207, 209, ["object", "detection"], "Task"], [211, 214, ["SPP", "-", "net"], "Method"], [241, 245, ["fixed", "-", "length", "representations"], "Method"], [273, 277, ["R", "-", "CNN", "method"], "Method"], [283, 284, ["accuracy"], "Metric"], [285, 288, ["Pascal", "VOC", "2007"], "Material"], [290, 296, ["ImageNet", "Large", "Scale", "Visual", "Recognition", "Challenge"], "Task"], [297, 300, ["ILSVRC", ")", "2014"], "Task"], [307, 309, ["object", "detection"], "Task"], [313, 315, ["image", "classification"], "Task"], [331, 334, ["onvolutional", "Neural", "Networks"], "Method"], [335, 338, ["Spatial", "Pyramid", "Pooling"], "Method"], [339, 341, ["Image", "Classification"], "Task"], [342, 344, ["Object", "Detection"], "Task"], [357, 359, ["vision", "community"], "Task"], [363, 367, ["deep", "convolutional", "neural", "networks"], "Method"], [368, 369, ["CNNs"], "Method"], [379, 385, ["Deep", "-", "networks", "-", "based", "approaches"], "Method"], [397, 399, ["image", "classification"], "Task"], [400, 402, ["object", "detection"], "Task"], [405, 407, ["recognition", "tasks"], "Task"], [410, 414, ["non", "-", "recognition", "tasks"], "Task"], [424, 427, ["training", "and", "testing"], "Task"], [429, 430, ["CNNs"], "Method"], [432, 434, ["prevalent", "CNNs"], "Method"], [482, 483, ["cropping"], "Method"], [520, 522, ["Recognition", "accuracy"], "Metric"], [561, 562, ["CNNs"], "Method"], [569, 570, ["CNN"], "Method"], [576, 578, ["convolutional", "layers"], "Method"], [580, 584, ["fully", "-", "connected", "layers"], "Method"], [588, 590, ["convolutional", "layers"], "Method"], [593, 597, ["sliding", "-", "window", "manner"], "Method"], [619, 621, ["convolutional", "layers"], "Method"], [693, 696, ["spatial", "pyramid", "pooling"], "Method"], [697, 698, ["SPP"], "Method"], [723, 725, ["convolutional", "layer"], "Method"], [745, 749, ["fully", "-", "connected", "layers"], "Method"], [752, 753, ["classifiers"], "Method"], [762, 765, ["information", "“", "aggregation"], "Method"], [808, 810, ["network", "architecture"], "Method"], [826, 829, ["Spatial", "pyramid", "pooling"], "Method"], [833, 836, ["spatial", "pyramid", "matching"], "Method"], [837, 838, ["SPM"], "Method"], [845, 854, ["Bag", "-", "of", "-", "Words", "(", "BoW", ")", "model"], "Method"], [863, 865, ["computer", "vision"], "Task"], [885, 886, ["SPP"], "Method"], [894, 900, ["leading", "and", "competition", "-", "winning", "systems"], "Task"], [901, 907, ["classification", "(", ",", ")", "and", "detection"], "Task"], [915, 916, ["CNNs"], "Method"], [919, 920, ["SPP"], "Method"], [928, 929, ["CNNs"], "Method"], [933, 934, ["SPP"], "Method"], [939, 941, ["deep", "CNNs"], "Method"], [944, 945, ["SPP"], "Method"], [962, 965, ["sliding", "window", "pooling"], "Method"], [969, 971, ["deep", "networks"], "Method"], [976, 977, ["SPP"], "Method"], [986, 989, ["sliding", "window", "pooling"], "Method"], [996, 1000, ["Multi", "-", "level", "pooling"], "Method"], [1012, 1013, ["SPP"], "Method"], [1038, 1040, ["recognition", "accuracy"], "Metric"], [1041, 1043, ["deep", "networks"], "Method"], [1044, 1047, ["SPP", "-", "net"], "Method"], [1091, 1094, ["over", "-", "fitting"], "Task"], [1099, 1104, ["multi", "-", "size", "training", "method"], "Method"], [1167, 1171, ["multi", "-", "size", "training"], "Method"], [1176, 1180, ["single", "-", "size", "training"], "Method"], [1185, 1187, ["testing", "accuracy"], "Metric"], [1191, 1192, ["SPP"], "Method"], [1197, 1199, ["CNN", "designs"], "Method"], [1208, 1211, ["ImageNet", "2012", "dataset"], "Material"], [1215, 1216, ["SPP"], "Method"], [1219, 1221, ["CNN", "architectures"], "Method"], [1263, 1264, ["SPP"], "Method"], [1273, 1275, ["convolutional", "architectures"], "Method"], [1276, 1279, ["SPP", "-", "net"], "Method"], [1288, 1289, ["classification"], "Metric"], [1291, 1292, ["Caltech101"], "Material"], [1293, 1296, ["Pascal", "VOC", "2007"], "Material"], [1300, 1304, ["full", "-", "image", "representation"], "Method"], [1306, 1309, ["fine", "-", "tuning"], "Method"], [1310, 1313, ["SPP", "-", "net"], "Method"], [1318, 1320, ["object", "detection"], "Task"], [1339, 1342, ["deep", "convolutional", "networks"], "Method"], [1347, 1349, ["detection", "accuracy"], "Metric"], [1352, 1353, ["VOC"], "Material"], [1354, 1356, ["ImageNet", "datasets"], "Material"], [1359, 1361, ["feature", "computation"], "Task"], [1362, 1365, ["R", "-", "CNN"], "Method"], [1375, 1378, ["deep", "convolutional", "networks"], "Method"], [1401, 1403, ["convolutional", "layers"], "Method"], [1423, 1426, ["SPP", "-", "net"], "Method"], [1442, 1445, ["R", "-", "CNN"], "Method"], [1452, 1453, ["detector"], "Method"], [1471, 1474, ["SPP", "-", "net"], "Method"], [1479, 1483, ["deep", "CNN", "feature", "maps"], "Method"], [1488, 1489, ["SPP"], "Method"], [1498, 1499, ["accuracy"], "Metric"], [1517, 1521, ["R", "-", "CNN", "pipeline"], "Method"], [1529, 1532, ["R", "-", "CNN"], "Method"], [1538, 1539, ["accuracy"], "Metric"], [1544, 1546, ["proposal", "method"], "Method"], [1547, 1548, ["EdgeBoxes"], "Method"], [1569, 1573, ["real", "-", "world", "applications"], "Task"], [1597, 1599, ["ILSVRC", "2014"], "Task"], [1605, 1607, ["object", "detection"], "Task"], [1611, 1613, ["image", "classification"], "Task"], [1635, 1637, ["ILSVRC", "2014"], "Task"], [1677, 1679, ["detection", "framework"], "Method"], [1683, 1687, ["multi", "-", "view", "testing"], "Task"], [1706, 1708, ["classification", "accuracy"], "Metric"], [1747, 1749, ["Deep", "Networks"], "Method"], [1750, 1753, ["Spatial", "Pyramid", "Pooling"], "Method"], [1755, 1757, ["Convolutional", "Layers"], "Method"], [1758, 1760, ["Feature", "Maps"], "Method"], [1763, 1767, ["seven", "-", "layer", "architectures"], "Method"], [1773, 1774, ["convolutional"], "Method"], [1781, 1783, ["pooling", "layers"], "Method"], [1785, 1787, ["pooling", "layers"], "Method"], [1816, 1820, ["N", "-", "way", "softmax"], "Method"], [1833, 1835, ["deep", "network"], "Method"], [1858, 1862, ["fully", "-", "connected", "layers"], "Method"], [1877, 1879, ["convolutional", "layers"], "Method"], [1886, 1888, ["convolutional", "layers"], "Method"], [1889, 1891, ["sliding", "filters"], "Method"], [1946, 1951, ["filters", "of", "the", "conv", "layer"], "Method"], [1969, 1971, ["ImageNet", "dataset"], "Material"], [1975, 1976, ["filter"], "Method"], [2033, 2037, ["118", "-", "th", "filter"], "Method"], [2113, 2116, ["deep", "convolutional", "layers"], "Method"], [2144, 2146, ["vector", "quantization"], "Method"], [2147, 2149, ["sparse", "coding"], "Method"], [2151, 2153, ["Fisher", "kernels"], "Method"], [2168, 2176, ["Bag", "-", "of", "-", "Words", "(", "BoW", ")"], "Method"], [2177, 2179, ["spatial", "pyramids"], "Method"], [2202, 2204, ["convolutional", "layers"], "Method"], [2218, 2219, ["classifiers"], "Method"], [2220, 2223, ["SVM", "/", "softmax"], "Method"], [2225, 2229, ["fully", "-", "connected", "layers"], "Method"], [2242, 2251, ["Bag", "-", "of", "-", "Words", "(", "BoW", ")", "approach"], "Method"], [2257, 2260, ["Spatial", "pyramid", "pooling"], "Method"], [2261, 2262, ["BoW"], "Method"], [2306, 2309, ["sliding", "window", "pooling"], "Method"], [2312, 2314, ["deep", "networks"], "Method"], [2330, 2332, ["deep", "network"], "Method"], [2342, 2344, ["pooling", "layer"], "Method"], [2351, 2353, ["convolutional", "layer"], "Method"], [2387, 2389, ["max", "pooling"], "Method"], [2395, 2398, ["spatial", "pyramid", "pooling"], "Method"], [2418, 2420, ["convolutional", "layer"], "Method"], [2432, 2436, ["fully", "-", "connected", "layer"], "Method"], [2438, 2441, ["spatial", "pyramid", "pooling"], "Method"], [2487, 2489, ["deep", "network"], "Method"], [2543, 2545, ["Gaussian", "filters"], "Method"], [2558, 2559, ["accuracy"], "Metric"], [2560, 2562, ["deep", "networks"], "Task"], [2601, 2604, ["global", "average", "pooling"], "Method"], [2619, 2622, ["global", "average", "pooling"], "Method"], [2630, 2632, ["fc", "layers"], "Method"], [2634, 2635, ["accuracy"], "Metric"], [2639, 2642, ["global", "max", "pooling"], "Method"], [2645, 2649, ["weakly", "supervised", "object", "recognition"], "Task"], [2651, 2654, ["global", "pooling", "operation"], "Method"], [2658, 2664, ["Bag", "-", "of", "-", "Words", "method"], "Method"], [2674, 2676, ["network", "structure"], "Method"], [2681, 2684, ["back", "-", "propagation"], "Method"], [2696, 2698, ["GPU", "implementations"], "Method"], [2701, 2706, ["cuda", "-", "convnet", "and", "Caffe"], "Method"], [2719, 2721, ["training", "solution"], "Method"], [2726, 2728, ["GPU", "implementations"], "Method"], [2737, 2738, ["subsubsection"], "Method"], [2739, 2743, ["Single", "-", "size", "training"], "Method"], [2768, 2769, ["cropping"], "Task"], [2774, 2776, ["data", "augmentation"], "Task"], [2795, 2798, ["spatial", "pyramid", "pooling"], "Method"], [2804, 2805, ["conv"], "Method"], [2826, 2828, ["pooling", "level"], "Method"], [2830, 2833, ["sliding", "window", "pooling"], "Method"], [2850, 2853, ["-", "level", "pyramid"], "Method"], [2861, 2865, ["fully", "-", "connected", "layer"], "Method"], [2866, 2867, ["fc"], "Method"], [2899, 2903, ["cuda", "-", "convnet", "style"], "Method"], [2909, 2913, ["single", "-", "size", "training"], "Method"], [2934, 2935, ["accuracy"], "Metric"], [2938, 2942, ["Multi", "-", "size", "training"], "Task"], [2945, 2946, ["SPP"], "Method"], [2966, 2967, ["training"], "Task"], [3052, 3055, ["feature", "map", "size"], "Metric"], [3056, 3057, ["conv"], "Method"], [3125, 3126, ["training"], "Task"], [3139, 3143, ["fixed", "-", "size", "networks"], "Method"], [3205, 3207, ["convergence", "rate"], "Metric"], [3209, 3213, ["multi", "-", "size", "training"], "Method"], [3219, 3223, ["single", "-", "size", "training"], "Method"], [3229, 3233, ["multi", "-", "size", "training"], "Method"], [3248, 3252, ["fixed", "-", "size", "implementations"], "Method"], [3258, 3260, ["scale", "implementation"], "Method"], [3299, 3303, ["multi", "-", "size", "solutions"], "Method"], [3305, 3306, ["training"], "Task"], [3318, 3321, ["SPP", "-", "net"], "Method"], [3329, 3332, ["SPP", "-", "net"], "Method"], [3333, 3335, ["Image", "Classification"], "Task"], [3339, 3342, ["ImageNet", "2012", "Classification"], "Task"], [3354, 3356, ["ImageNet", "2012"], "Material"], [3358, 3360, ["training", "algorithm"], "Method"], [3409, 3410, ["Dropout"], "Method"], [3415, 3419, ["fully", "-", "connected", "layers"], "Method"], [3421, 3423, ["learning", "rate"], "Metric"], [3450, 3453, ["cuda", "-", "convnet"], "Method"], [3454, 3455, ["Caffe"], "Method"], [3467, 3471, ["GeForce", "GTX", "Titan", "GPU"], "Method"], [3484, 3487, ["Baseline", "Network", "Architectures"], "Method"], [3490, 3491, ["SPP"], "Method"], [3495, 3498, ["convolutional", "network", "architectures"], "Method"], [3504, 3506, ["network", "architectures"], "Method"], [3518, 3519, ["SPP"], "Method"], [3521, 3522, ["accuracy"], "Metric"], [3541, 3544, ["ZF", "-", "5"], "Method"], [3569, 3571, ["convolutional", "layers"], "Method"], [3572, 3575, ["Convnet*", "-", "5"], "Method"], [3581, 3584, ["Krizhevsky", "’s", "network"], "Method"], [3589, 3591, ["pooling", "layers"], "Method"], [3592, 3593, ["conv"], "Method"], [3594, 3595, ["conv"], "Method"], [3599, 3600, ["conv"], "Method"], [3601, 3602, ["conv"], "Method"], [3647, 3654, ["ZF", "-", "5", "/", "Convnet*", "-", "5"], "Method"], [3669, 3671, ["pooling", "layer"], "Method"], [3686, 3688, ["convolutional", "layers"], "Method"], [3693, 3695, ["deeper", "architecture"], "Method"], [3697, 3699, ["convolutional", "layers"], "Method"], [3715, 3717, ["pooling", "layer"], "Method"], [3720, 3722, ["convolutional", "layer"], "Method"], [3728, 3733, ["4096", "-", "d", "fc", "layers"], "Method"], [3735, 3740, ["1000", "-", "way", "softmax", "layer"], "Method"], [3776, 3779, ["ZF", "-", "5"], "Method"], [3808, 3812, ["Multi", "-", "level", "Pooling"], "Method"], [3813, 3814, ["Accuracy"], "Metric"], [3827, 3831, ["single", "-", "size", "training"], "Method"], [3847, 3849, ["convolutional", "layers"], "Method"], [3856, 3858, ["baseline", "models"], "Method"], [3861, 3863, ["pooling", "layer"], "Method"], [3866, 3868, ["convolutional", "layer"], "Method"], [3923, 3927, ["10", "-", "view", "prediction"], "Method"], [3969, 3973, ["top", "-", "1", "error"], "Metric"], [4007, 4011, ["multi", "-", "level", "pooling"], "Method"], [4020, 4024, ["multi", "-", "level", "pooling"], "Method"], [4038, 4042, ["multi", "-", "level", "pooling"], "Method"], [4061, 4065, ["ZF", "-", "5", "network"], "Method"], [4106, 4108, ["fc", "layer"], "Method"], [4122, 4130, ["top", "-", "1", "/", "top", "-", "5", "errors"], "Metric"], [4172, 4176, ["Multi", "-", "size", "Training"], "Method"], [4177, 4178, ["Accuracy"], "Metric"], [4189, 4193, ["multi", "-", "size", "training"], "Method"], [4215, 4219, ["10", "-", "view", "prediction"], "Method"], [4221, 4229, ["top", "-", "1", "/", "top", "-", "5", "errors"], "Metric"], [4236, 4240, ["top", "-", "1", "error"], "Metric"], [4241, 4244, ["SPP", "-", "net"], "Method"], [4271, 4276, ["single", "-", "size", "trained", "counterpart"], "Method"], [4301, 4307, ["top", "-", "1", "/", "5", "error"], "Metric"], [4308, 4311, ["SPP", "-", "net"], "Method"], [4323, 4327, ["top", "-", "1", "error"], "Metric"], [4369, 4371, ["CNN", "solutions"], "Method"], [4390, 4393, ["Howard", "’s", "method"], "Method"], [4395, 4397, ["single", "network"], "Method"], [4456, 4460, ["Full", "-", "image", "Representations"], "Method"], [4461, 4462, ["Accuracy"], "Task"], [4466, 4467, ["accuracy"], "Metric"], [4489, 4492, ["SPP", "-", "net"], "Method"], [4515, 4516, ["accuracy"], "Metric"], [4539, 4544, ["single", "-", "view", "testing", "accuracy"], "Metric"], [4563, 4568, ["top", "-", "1", "error", "rates"], "Metric"], [4573, 4577, ["full", "-", "view", "representation"], "Method"], [4632, 4636, ["full", "-", "image", "view"], "Method"], [4640, 4644, ["full", "-", "image", "representations"], "Method"], [4687, 4688, ["accuracy"], "Metric"], [4696, 4700, ["full", "-", "image", "view"], "Task"], [4727, 4729, ["image", "retrieval"], "Task"], [4731, 4733, ["image", "representation"], "Task"], [4737, 4739, ["classification", "score"], "Metric"], [4743, 4745, ["similarity", "ranking"], "Task"], [4747, 4751, ["full", "-", "image", "representation"], "Method"], [4757, 4761, ["Multi", "-", "view", "Testing"], "Task"], [4762, 4764, ["Feature", "Maps"], "Task"], [4767, 4769, ["detection", "algorithm"], "Method"], [4781, 4786, ["multi", "-", "view", "testing", "method"], "Method"], [4796, 4797, ["SPP"], "Method"], [4814, 4817, ["convolutional", "feature", "maps"], "Method"], [4842, 4845, ["convolutional", "feature", "maps"], "Method"], [4899, 4900, ["SPP"], "Method"], [4923, 4925, ["fc", "layers"], "Method"], [4928, 4930, ["softmax", "score"], "Metric"], [4941, 4942, ["prediction"], "Task"], [4946, 4949, ["10", "-", "view"], "Method"], [4969, 4973, ["top", "-", "5", "error"], "Metric"], [4975, 4979, ["10", "-", "view", "prediction"], "Task"], [4980, 4982, ["feature", "maps"], "Task"], [4989, 4993, ["10", "-", "view", "prediction"], "Method"], [4994, 4996, ["image", "crops"], "Task"], [5105, 5109, ["top", "-", "5", "error"], "Metric"], [5130, 5134, ["top", "-", "5", "error"], "Metric"], [5150, 5153, ["convolutional", "feature", "maps"], "Method"], [5227, 5229, ["Krizhevsky", "’s"], "Method"], [5234, 5235, ["ILSVRC"], "Task"], [5252, 5254, ["ILSVRC", "2013"], "Task"], [5273, 5277, ["top", "-", "5", "error"], "Metric"], [5297, 5301, ["top", "-", "5", "error"], "Metric"], [5309, 5311, ["ILSVRC", "2014"], "Task"], [5321, 5323, ["ILSVRC", "2012"], "Material"], [5360, 5363, ["SPP", "-", "net"], "Method"], [5382, 5384, ["convolutional", "architectures"], "Method"], [5397, 5402, ["full", "-", "view", "image", "representation"], "Task"], [5411, 5412, ["ImageNet"], "Material"], [5428, 5430, ["SVM", "classifiers"], "Method"], [5433, 5435, ["SVM", "training"], "Task"], [5442, 5444, ["data", "augmentation"], "Method"], [5459, 5461, ["SVM", "training"], "Task"], [5463, 5465, ["classification", "task"], "Task"], [5466, 5469, ["Pascal", "VOC", "2007"], "Material"], [5498, 5501, ["mean", "Average", "Precision"], "Metric"], [5502, 5503, ["mAP"], "Metric"], [5534, 5535, ["SPP"], "Method"], [5562, 5563, ["SVM"], "Method"], [5606, 5609, ["SPP", "-", "net"], "Method"], [5621, 5624, ["SPP", "-", "net"], "Method"], [5635, 5637, ["fc", "layers"], "Method"], [5645, 5649, ["multi", "-", "level", "pooling"], "Method"], [5697, 5701, ["full", "-", "image", "representation"], "Method"], [5776, 5777, ["ImageNet"], "Material"], [5797, 5799, ["classification", "tasks"], "Task"], [5801, 5804, ["SPP", "-", "net"], "Method"], [5823, 5825, ["network", "architecture"], "Method"], [5844, 5845, ["mAP"], "Metric"], [5875, 5876, ["VQ"], "Method"], [5877, 5878, ["LCC"], "Method"], [5880, 5881, ["FK"], "Method"], [5885, 5888, ["spatial", "pyramids", "matching"], "Method"], [5893, 5895, ["deep", "networks"], "Method"], [5907, 5909, ["Chatfield", "’s"], "Metric"], [5916, 5920, ["network", "fine", "-", "tuning"], "Method"], [5921, 5925, ["multi", "-", "view", "testing"], "Method"], [5941, 5945, ["full", "-", "image", "representation"], "Method"], [5947, 5950, ["fine", "-", "tuning"], "Method"], [5955, 5956, ["Caltech101"], "Material"], [5999, 6000, ["accuracy"], "Metric"], [6016, 6019, ["Pascal", "VOC", "2007"], "Material"], [6020, 6021, ["Caltech101"], "Material"], [6023, 6026, ["SPP", "-", "net"], "Method"], [6049, 6053, ["full", "-", "view", "representation"], "Method"], [6057, 6058, ["crop"], "Method"], [6071, 6072, ["Caltech101"], "Material"], [6080, 6084, ["fully", "-", "connected", "layers"], "Method"], [6103, 6104, ["Caltech101"], "Material"], [6110, 6111, ["ImageNet"], "Material"], [6114, 6116, ["deeper", "layers"], "Method"], [6128, 6130, ["scale", "224"], "Method"], [6150, 6151, ["Caltech101"], "Material"], [6164, 6165, ["ImageNet"], "Material"], [6167, 6168, ["cropping"], "Task"], [6204, 6205, ["accuracy"], "Metric"], [6249, 6250, ["Caltech101"], "Material"], [6276, 6279, ["SPP", "-", "net"], "Method"], [6280, 6282, ["Object", "Detection"], "Task"], [6282, 6284, ["Deep", "networks"], "Method"], [6288, 6290, ["object", "detection"], "Task"], [6303, 6307, ["R", "-", "CNN", "method"], "Method"], [6308, 6311, ["R", "-", "CNN"], "Method"], [6323, 6325, ["selective", "search"], "Method"], [6348, 6350, ["deep", "network"], "Method"], [6361, 6364, ["binary", "SVM", "classifier"], "Method"], [6371, 6372, ["detection"], "Task"], [6373, 6376, ["R", "-", "CNN"], "Method"], [6390, 6393, ["R", "-", "CNN"], "Method"], [6396, 6399, ["deep", "convolutional", "network"], "Method"], [6414, 6416, ["Feature", "extraction"], "Task"], [6422, 6423, ["testing"], "Task"], [6425, 6428, ["SPP", "-", "net"], "Method"], [6433, 6435, ["object", "detection"], "Task"], [6458, 6461, ["spatial", "pyramid", "pooling"], "Method"], [6492, 6493, ["convolutions"], "Method"], [6522, 6525, ["R", "-", "CNN"], "Method"], [6536, 6539, ["Deformable", "Part", "Model"], "Method"], [6540, 6541, ["DPM"], "Method"], [6553, 6559, ["Selective", "Search", "(", "SS", ")", "method"], "Method"], [6569, 6572, ["Overfeat", "detection", "method"], "Method"], [6575, 6581, ["windows", "of", "deep", "convolutional", "feature", "maps"], "Method"], [6599, 6601, ["feature", "extraction"], "Task"], [6606, 6610, ["deep", "convolutional", "feature", "maps"], "Method"], [6613, 6615, ["Detection", "Algorithm"], "Method"], [6623, 6625, ["selective", "search"], "Method"], [6662, 6670, ["ZF", "-", "5", "(", "single", "-", "size", "trained"], "Method"], [6734, 6741, ["fully", "-", "connected", "layers", "of", "the", "network"], "Method"], [6746, 6750, ["binary", "linear", "SVM", "classifier"], "Method"], [6761, 6763, ["SVM", "training"], "Method"], [6796, 6805, ["intersection", "-", "over", "-", "union", "(", "IoU", ")", "ratio"], "Metric"], [6828, 6831, ["hard", "negative", "mining"], "Method"], [6834, 6835, ["SVM"], "Method"], [6850, 6851, ["SVMs"], "Method"], [6860, 6861, ["classifier"], "Method"], [6872, 6876, ["non", "-", "maximum", "suppression"], "Method"], [6893, 6898, ["multi", "-", "scale", "feature", "extraction"], "Method"], [6912, 6913, ["conv"], "Method"], [7062, 7066, ["pre", "-", "trained", "network"], "Method"], [7076, 7079, ["conv", "feature", "maps"], "Method"], [7093, 7097, ["fully", "-", "connected", "layers"], "Method"], [7103, 7105, ["data", "layer"], "Method"], [7113, 7114, ["conv"], "Method"], [7117, 7119, ["fc", "layers"], "Method"], [7131, 7133, ["fc", "layer"], "Method"], [7136, 7138, ["fc", "weights"], "Method"], [7142, 7144, ["Gaussian", "distribution"], "Method"], [7152, 7154, ["learning", "rates"], "Metric"], [7171, 7174, ["fine", "-", "tuning"], "Task"], [7217, 7222, ["learning", "rate", "1e", "-", "4"], "Metric"], [7241, 7243, ["fc", "layers"], "Method"], [7245, 7246, ["training"], "Task"], [7276, 7279, ["bounding", "box", "regression"], "Method"], [7291, 7292, ["regression"], "Task"], [7297, 7298, ["conv"], "Method"], [7315, 7317, ["regression", "training"], "Task"], [7334, 7335, ["Detection"], "Task"], [7342, 7344, ["detection", "task"], "Task"], [7372, 7375, ["5", "-", "scale"], "Method"], [7378, 7381, ["R", "-", "CNN"], "Method"], [7388, 7389, ["AlexNet"], "Method"], [7391, 7393, ["conv", "layers"], "Method"], [7416, 7419, ["R", "-", "CNN"], "Method"], [7431, 7436, ["fine", "-", "tuned", "fc", "layers"], "Method"], [7447, 7449, ["fc", "layers"], "Method"], [7460, 7462, ["detection", "case"], "Task"], [7499, 7502, ["fine", "-", "tuning"], "Method"], [7505, 7510, ["fine", "-", "tuned", "fc", "layers"], "Method"], [7511, 7512, ["ftfc"], "Method"], [7529, 7532, ["R", "-", "CNN"], "Method"], [7534, 7537, ["bounding", "box", "regression"], "Method"], [7539, 7543, ["5", "-", "scale", "result"], "Metric"], [7552, 7555, ["R", "-", "CNN"], "Method"], [7562, 7566, ["1", "-", "scale", "result"], "Metric"], [7584, 7587, ["R", "-", "CNN"], "Method"], [7590, 7594, ["pre", "-", "trained", "model"], "Method"], [7595, 7596, ["SPPnet"], "Method"], [7597, 7600, ["ZF", "-", "5"], "Method"], [7609, 7612, ["R", "-", "CNN"], "Method"], [7614, 7616, ["averaged", "scores"], "Metric"], [7618, 7621, ["R", "-", "CNN"], "Task"], [7639, 7642, ["ZF", "-", "5"], "Method"], [7643, 7644, ["AlexNet"], "Method"], [7650, 7654, ["multi", "-", "level", "pooling"], "Method"], [7655, 7656, ["SPPnet"], "Method"], [7668, 7671, ["R", "-", "CNN"], "Method"], [7695, 7697, ["Selective", "Search"], "Method"], [7698, 7699, ["SS"], "Method"], [7701, 7704, ["spatial", "pyramid", "matching"], "Method"], [7709, 7710, ["DPM"], "Method"], [7711, 7712, ["Regionlet"], "Method"], [7719, 7721, ["Regionlet", "method"], "Method"], [7732, 7733, ["DetectorNet"], "Method"], [7735, 7737, ["deep", "network"], "Method"], [7752, 7754, ["deep", "network"], "Method"], [7773, 7774, ["mAP"], "Metric"], [7781, 7782, ["Complexity"], "Metric"], [7783, 7785, ["Running", "Time"], "Metric"], [7788, 7789, ["accuracy"], "Metric"], [7796, 7799, ["R", "-", "CNN"], "Method"], [7801, 7802, ["complexity"], "Metric"], [7804, 7807, ["convolutional", "feature", "computation"], "Method"], [7808, 7811, ["R", "-", "CNN"], "Method"], [7859, 7863, ["R", "-", "CNN", "’s"], "Method"], [7879, 7883, ["R", "-", "CNN", "’s"], "Method"], [7897, 7899, ["running", "time"], "Metric"], [7901, 7903, ["feature", "computation"], "Method"], [7917, 7920, ["R", "-", "CNN"], "Method"], [7930, 7931, ["Caffe"], "Method"], [7936, 7938, ["feature", "computation"], "Method"], [7939, 7940, ["Caffe"], "Method"], [7949, 7951, ["average", "time"], "Metric"], [7957, 7958, ["GPU"], "Method"], [7959, 7962, ["R", "-", "CNN"], "Method"], [7967, 7968, ["convolutions"], "Task"], [7971, 7975, ["1", "-", "scale", "version"], "Method"], [7987, 7990, ["R", "-", "CNN"], "Method"], [7992, 7996, ["5", "-", "scale", "version"], "Method"], [8001, 8002, ["convolutions"], "Method"], [8008, 8011, ["R", "-", "CNN"], "Method"], [8013, 8016, ["convolutional", "feature", "computation"], "Method"], [8021, 8023, ["computational", "time"], "Metric"], [8024, 8026, ["fc", "layers"], "Method"], [8038, 8040, ["GPU", "time"], "Metric"], [8057, 8063, ["convolutional", "and", "fully", "-", "connected", "features"], "Method"], [8065, 8069, ["1", "-", "scale", "version"], "Method"], [8073, 8076, ["R", "-", "CNN"], "Method"], [8083, 8087, ["5", "-", "scale", "version"], "Method"], [8099, 8101, ["running", "time"], "Metric"], [8107, 8110, ["R", "-", "CNN"], "Method"], [8111, 8112, ["AlexNet"], "Method"], [8130, 8131, ["AlexNet"], "Method"], [8139, 8142, ["ZF", "-", "5"], "Method"], [8144, 8146, ["conv", "layer"], "Method"], [8148, 8149, ["AlexNet"], "Method"], [8172, 8174, ["full", "system"], "Method"], [8179, 8183, ["recent", "window", "proposal", "method"], "Method"], [8185, 8191, ["Selective", "Search", "(", "SS", ")", "proposal"], "Method"], [8206, 8207, ["EdgeBoxes"], "Method"], [8220, 8222, ["proposal", "method"], "Method"], [8223, 8224, ["testing"], "Task"], [8235, 8236, ["SS"], "Method"], [8247, 8248, ["mAP"], "Metric"], [8251, 8254, ["bounding", "box", "regression"], "Method"], [8265, 8266, ["training"], "Task"], [8271, 8272, ["SS"], "Method"], [8291, 8292, ["mAP"], "Metric"], [8295, 8298, ["bounding", "box", "regression"], "Method"], [8322, 8324, ["testing", "time"], "Metric"], [8334, 8335, ["recognition"], "Task"], [8343, 8347, ["real", "-", "world", "applications"], "Task"], [8350, 8352, ["Model", "Combination"], "Method"], [8353, 8354, ["Detection"], "Task"], [8354, 8356, ["Model", "combination"], "Method"], [8361, 8367, ["boosting", "CNN", "-", "based", "classification", "accuracy"], "Task"], [8372, 8374, ["combination", "method"], "Method"], [8375, 8376, ["detection"], "Task"], [8384, 8385, ["ImageNet"], "Material"], [8400, 8402, ["detection", "algorithm"], "Method"], [8408, 8411, ["SPP", "-", "net"], "Method"], [8423, 8424, ["mAP"], "Metric"], [8469, 8473, ["non", "-", "maximum", "suppression"], "Method"], [8509, 8510, ["combination"], "Task"], [8512, 8513, ["mAP"], "Metric"], [8561, 8563, ["convolutional", "layers"], "Method"], [8579, 8581, ["convolutional", "model"], "Method"], [8633, 8639, ["1000", "-", "category", "CLS", "training", "data"], "Material"], [8653, 8656, ["detection", "(", "DET"], "Task"], [8658, 8661, ["classification", "(", "CLS"], "Task"], [8669, 8673, ["pre", "-", "training", "quality"], "Metric"], [8677, 8680, ["DET", "training", "data"], "Material"], [8687, 8690, ["CLS", "training", "data"], "Material"], [8700, 8707, ["provided", "-", "data", "-", "only", "DET", "task"], "Task"], [8734, 8737, ["pre", "-", "training"], "Task"], [8764, 8768, ["499", "-", "category", "network"], "Method"], [8770, 8773, ["DET", "training", "set"], "Material"], [8784, 8789, ["DET", "/", "CLS", "training", "sets"], "Material"], [8795, 8796, ["CLS"], "Method"], [8859, 8862, ["pre", "-", "training"], "Task"], [8863, 8866, ["Pascal", "VOC", "2007"], "Material"], [8869, 8875, ["CLS", "-", "pre", "-", "training", "baseline"], "Task"], [8882, 8883, ["mAP"], "Metric"], [8895, 8899, ["200", "-", "category", "network"], "Method"], [8903, 8904, ["DET"], "Method"], [8906, 8907, ["mAP"], "Metric"], [8917, 8921, ["pre", "-", "trained", "network"], "Method"], [8949, 8951, ["feature", "quality"], "Metric"], [8961, 8962, ["mAP"], "Metric"], [8980, 8981, ["CLS"], "Method"], [8995, 8997, ["deep", "learning"], "Task"], [8999, 9001, ["ILSVRC", "2014"], "Task"], [9008, 9011, ["Overfeat", "-", "7"], "Method"], [9011, 9014, ["SPP", "-", "net"], "Method"], [9046, 9050, ["selective", "search", "fast", "mode"], "Method"], [9069, 9071, ["fc", "layers"], "Method"], [9075, 9076, ["SVMs"], "Method"], [9087, 9090, ["bounding", "box", "regression"], "Method"], [9104, 9105, ["mAP"], "Metric"], [9126, 9127, ["mAP"], "Metric"], [9149, 9151, ["ILSVRC", "2014"], "Task"], [9165, 9166, ["NUS"], "Method"], [9179, 9180, ["speed"], "Metric"], [9194, 9195, ["conv"], "Method"], [9198, 9199, ["fc"], "Method"], [9232, 9233, ["RCNN"], "Method"], [9252, 9253, ["RCNN"], "Method"], [9262, 9263, ["SPP"], "Method"], [9283, 9285, ["visual", "recognition"], "Task"], [9294, 9296, ["deep", "networks"], "Method"], [9305, 9307, ["deep", "network"], "Method"], [9316, 9319, ["SPP", "-", "net"], "Method"], [9321, 9322, ["accuracy"], "Metric"], [9323, 9327, ["classification", "/", "detection", "tasks"], "Task"], [9330, 9334, ["DNN", "-", "based", "detection"], "Method"], [9348, 9350, ["computer", "vision"], "Task"], [9356, 9362, ["deep", "-", "networks", "-", "based", "recognition"], "Task"], [9375, 9377, ["Mean", "Subtraction"], "Method"], [9408, 9413, ["fixed", "-", "size", "mean", "image"], "Method"], [9420, 9422, ["ImageNet", "dataset"], "Material"], [9440, 9443, ["Pascal", "VOC", "2007"], "Material"], [9444, 9445, ["Caltech101"], "Material"], [9554, 9556, ["detection", "algorithm"], "Method"], [9558, 9565, ["multi", "-", "view", "testing", "on", "feature", "maps"], "Task"], [9656, 9657, ["mapping"], "Task"], [9661, 9668, ["padding", "of", "all", "convolutional", "and", "pooling", "layers"], "Method"], [9722, 9725, ["ZF", "-", "5"], "Task"], [9726, 9727, ["conv"], "Task"], [9808, 9809, ["SPP"], "Method"], [9816, 9818, ["ILSVRC", "2014"], "Task"], [9825, 9826, ["TPAMI"], "Task"], [9830, 9833, ["R", "-", "CNN"], "Method"], [9839, 9840, ["detection"], "Task"], [716, 717, ["SPP"], "Method"], [727, 728, ["SPP"], "Method"], [813, 814, ["SPP"], "Method"], [822, 823, ["SPP"], "Method"], [1234, 1235, ["SPP"], "Method"], [1324, 1326, ["object", "detection"], "Task"], [1507, 1508, ["SPP"], "Method"], [1642, 1643, ["SPP"], "Method"], [1669, 1670, ["SPP"], "Method"], [2197, 2200, ["Spatial", "Pyramid", "Pooling"], "Method"], [2356, 2359, ["spatial", "pyramid", "pooling"], "Method"], [2732, 2735, ["spatial", "pyramid", "pooling"], "Method"], [3079, 3082, ["spatial", "pyramid", "pooling"], "Method"], [3134, 3135, ["SPP"], "Method"], [3872, 3873, ["SPP"], "Method"], [3952, 3953, ["SPP"], "Method"], [4101, 4102, ["SPP"], "Method"], [4162, 4163, ["SPP"], "Method"], [4245, 4248, ["Overfeat", "-", "7"], "Method"], [4263, 4264, ["SPP"], "Method"], [4312, 4315, ["Overfeat", "-", "7"], "Method"], [4558, 4561, ["Overfeat", "-", "7"], "Method"], [5207, 5209, ["ILSVRC", "2014"], "Task"], [5293, 5295, ["ILSVRC", "2014"], "Task"], [5347, 5349, ["ILSVRC", "2014"], "Task"], [5389, 5390, ["VOC"], "Material"], [5602, 5603, ["SPP"], "Method"], [5770, 5771, ["VOC"], "Material"], [5832, 5835, ["Overfeat", "-", "7"], "Method"], [5957, 5958, ["Caltech101"], "Material"], [6032, 6033, ["SPP"], "Method"], [6077, 6078, ["VOC"], "Material"], [6090, 6091, ["SPP"], "Method"], [6195, 6196, ["SPP"], "Method"], [6210, 6211, ["SPP"], "Method"], [6657, 6658, ["SPP"], "Method"], [7347, 7348, ["VOC"], "Material"], [7662, 7663, ["SPP"], "Method"], [7906, 7907, ["SPP"], "Method"], [7954, 7955, ["VOC"], "Material"], [8589, 8591, ["ILSVRC", "2014"], "Task"], [8593, 8595, ["ILSVRC", "2014"], "Task"], [9022, 9023, ["VOC"], "Material"], [9107, 9109, ["ILSVRC", "2014"], "Task"], [9309, 9312, ["spatial", "pyramid", "pooling"], "Method"]]}